1 | //===- OpenMPIRBuilder.cpp - Builder for LLVM-IR for OpenMP directives ----===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// \file |
9 | /// |
10 | /// This file implements the OpenMPIRBuilder class, which is used as a |
11 | /// convenient way to create LLVM instructions for OpenMP directives. |
12 | /// |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" |
16 | #include "llvm/ADT/SmallSet.h" |
17 | #include "llvm/ADT/StringExtras.h" |
18 | #include "llvm/ADT/StringRef.h" |
19 | #include "llvm/Analysis/AssumptionCache.h" |
20 | #include "llvm/Analysis/CodeMetrics.h" |
21 | #include "llvm/Analysis/LoopInfo.h" |
22 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
23 | #include "llvm/Analysis/ScalarEvolution.h" |
24 | #include "llvm/Analysis/TargetLibraryInfo.h" |
25 | #include "llvm/Bitcode/BitcodeReader.h" |
26 | #include "llvm/Frontend/Offloading/Utility.h" |
27 | #include "llvm/Frontend/OpenMP/OMPGridValues.h" |
28 | #include "llvm/IR/Attributes.h" |
29 | #include "llvm/IR/BasicBlock.h" |
30 | #include "llvm/IR/CFG.h" |
31 | #include "llvm/IR/CallingConv.h" |
32 | #include "llvm/IR/Constant.h" |
33 | #include "llvm/IR/Constants.h" |
34 | #include "llvm/IR/DebugInfoMetadata.h" |
35 | #include "llvm/IR/DerivedTypes.h" |
36 | #include "llvm/IR/Function.h" |
37 | #include "llvm/IR/GlobalVariable.h" |
38 | #include "llvm/IR/IRBuilder.h" |
39 | #include "llvm/IR/LLVMContext.h" |
40 | #include "llvm/IR/MDBuilder.h" |
41 | #include "llvm/IR/Metadata.h" |
42 | #include "llvm/IR/PassManager.h" |
43 | #include "llvm/IR/PassInstrumentation.h" |
44 | #include "llvm/IR/ReplaceConstant.h" |
45 | #include "llvm/IR/Value.h" |
46 | #include "llvm/MC/TargetRegistry.h" |
47 | #include "llvm/Support/CommandLine.h" |
48 | #include "llvm/Support/ErrorHandling.h" |
49 | #include "llvm/Support/FileSystem.h" |
50 | #include "llvm/Target/TargetMachine.h" |
51 | #include "llvm/Target/TargetOptions.h" |
52 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
53 | #include "llvm/Transforms/Utils/Cloning.h" |
54 | #include "llvm/Transforms/Utils/CodeExtractor.h" |
55 | #include "llvm/Transforms/Utils/LoopPeel.h" |
56 | #include "llvm/Transforms/Utils/UnrollLoop.h" |
57 | |
58 | #include <cstdint> |
59 | #include <optional> |
60 | #include <stack> |
61 | |
62 | #define DEBUG_TYPE "openmp-ir-builder" |
63 | |
64 | using namespace llvm; |
65 | using namespace omp; |
66 | |
67 | static cl::opt<bool> |
68 | OptimisticAttributes("openmp-ir-builder-optimistic-attributes" , cl::Hidden, |
69 | cl::desc("Use optimistic attributes describing " |
70 | "'as-if' properties of runtime calls." ), |
71 | cl::init(Val: false)); |
72 | |
73 | static cl::opt<double> UnrollThresholdFactor( |
74 | "openmp-ir-builder-unroll-threshold-factor" , cl::Hidden, |
75 | cl::desc("Factor for the unroll threshold to account for code " |
76 | "simplifications still taking place" ), |
77 | cl::init(Val: 1.5)); |
78 | |
79 | #ifndef NDEBUG |
80 | /// Return whether IP1 and IP2 are ambiguous, i.e. that inserting instructions |
81 | /// at position IP1 may change the meaning of IP2 or vice-versa. This is because |
82 | /// an InsertPoint stores the instruction before something is inserted. For |
83 | /// instance, if both point to the same instruction, two IRBuilders alternating |
84 | /// creating instruction will cause the instructions to be interleaved. |
85 | static bool isConflictIP(IRBuilder<>::InsertPoint IP1, |
86 | IRBuilder<>::InsertPoint IP2) { |
87 | if (!IP1.isSet() || !IP2.isSet()) |
88 | return false; |
89 | return IP1.getBlock() == IP2.getBlock() && IP1.getPoint() == IP2.getPoint(); |
90 | } |
91 | |
92 | static bool isValidWorkshareLoopScheduleType(OMPScheduleType SchedType) { |
93 | // Valid ordered/unordered and base algorithm combinations. |
94 | switch (SchedType & ~OMPScheduleType::MonotonicityMask) { |
95 | case OMPScheduleType::UnorderedStaticChunked: |
96 | case OMPScheduleType::UnorderedStatic: |
97 | case OMPScheduleType::UnorderedDynamicChunked: |
98 | case OMPScheduleType::UnorderedGuidedChunked: |
99 | case OMPScheduleType::UnorderedRuntime: |
100 | case OMPScheduleType::UnorderedAuto: |
101 | case OMPScheduleType::UnorderedTrapezoidal: |
102 | case OMPScheduleType::UnorderedGreedy: |
103 | case OMPScheduleType::UnorderedBalanced: |
104 | case OMPScheduleType::UnorderedGuidedIterativeChunked: |
105 | case OMPScheduleType::UnorderedGuidedAnalyticalChunked: |
106 | case OMPScheduleType::UnorderedSteal: |
107 | case OMPScheduleType::UnorderedStaticBalancedChunked: |
108 | case OMPScheduleType::UnorderedGuidedSimd: |
109 | case OMPScheduleType::UnorderedRuntimeSimd: |
110 | case OMPScheduleType::OrderedStaticChunked: |
111 | case OMPScheduleType::OrderedStatic: |
112 | case OMPScheduleType::OrderedDynamicChunked: |
113 | case OMPScheduleType::OrderedGuidedChunked: |
114 | case OMPScheduleType::OrderedRuntime: |
115 | case OMPScheduleType::OrderedAuto: |
116 | case OMPScheduleType::OrderdTrapezoidal: |
117 | case OMPScheduleType::NomergeUnorderedStaticChunked: |
118 | case OMPScheduleType::NomergeUnorderedStatic: |
119 | case OMPScheduleType::NomergeUnorderedDynamicChunked: |
120 | case OMPScheduleType::NomergeUnorderedGuidedChunked: |
121 | case OMPScheduleType::NomergeUnorderedRuntime: |
122 | case OMPScheduleType::NomergeUnorderedAuto: |
123 | case OMPScheduleType::NomergeUnorderedTrapezoidal: |
124 | case OMPScheduleType::NomergeUnorderedGreedy: |
125 | case OMPScheduleType::NomergeUnorderedBalanced: |
126 | case OMPScheduleType::NomergeUnorderedGuidedIterativeChunked: |
127 | case OMPScheduleType::NomergeUnorderedGuidedAnalyticalChunked: |
128 | case OMPScheduleType::NomergeUnorderedSteal: |
129 | case OMPScheduleType::NomergeOrderedStaticChunked: |
130 | case OMPScheduleType::NomergeOrderedStatic: |
131 | case OMPScheduleType::NomergeOrderedDynamicChunked: |
132 | case OMPScheduleType::NomergeOrderedGuidedChunked: |
133 | case OMPScheduleType::NomergeOrderedRuntime: |
134 | case OMPScheduleType::NomergeOrderedAuto: |
135 | case OMPScheduleType::NomergeOrderedTrapezoidal: |
136 | break; |
137 | default: |
138 | return false; |
139 | } |
140 | |
141 | // Must not set both monotonicity modifiers at the same time. |
142 | OMPScheduleType MonotonicityFlags = |
143 | SchedType & OMPScheduleType::MonotonicityMask; |
144 | if (MonotonicityFlags == OMPScheduleType::MonotonicityMask) |
145 | return false; |
146 | |
147 | return true; |
148 | } |
149 | #endif |
150 | |
151 | static const omp::GV &getGridValue(const Triple &T, Function *Kernel) { |
152 | if (T.isAMDGPU()) { |
153 | StringRef Features = |
154 | Kernel->getFnAttribute(Kind: "target-features" ).getValueAsString(); |
155 | if (Features.count(Str: "+wavefrontsize64" )) |
156 | return omp::getAMDGPUGridValues<64>(); |
157 | return omp::getAMDGPUGridValues<32>(); |
158 | } |
159 | if (T.isNVPTX()) |
160 | return omp::NVPTXGridValues; |
161 | llvm_unreachable("No grid value available for this architecture!" ); |
162 | } |
163 | |
164 | /// Determine which scheduling algorithm to use, determined from schedule clause |
165 | /// arguments. |
166 | static OMPScheduleType |
167 | getOpenMPBaseScheduleType(llvm::omp::ScheduleKind ClauseKind, bool HasChunks, |
168 | bool HasSimdModifier) { |
169 | // Currently, the default schedule it static. |
170 | switch (ClauseKind) { |
171 | case OMP_SCHEDULE_Default: |
172 | case OMP_SCHEDULE_Static: |
173 | return HasChunks ? OMPScheduleType::BaseStaticChunked |
174 | : OMPScheduleType::BaseStatic; |
175 | case OMP_SCHEDULE_Dynamic: |
176 | return OMPScheduleType::BaseDynamicChunked; |
177 | case OMP_SCHEDULE_Guided: |
178 | return HasSimdModifier ? OMPScheduleType::BaseGuidedSimd |
179 | : OMPScheduleType::BaseGuidedChunked; |
180 | case OMP_SCHEDULE_Auto: |
181 | return llvm::omp::OMPScheduleType::BaseAuto; |
182 | case OMP_SCHEDULE_Runtime: |
183 | return HasSimdModifier ? OMPScheduleType::BaseRuntimeSimd |
184 | : OMPScheduleType::BaseRuntime; |
185 | } |
186 | llvm_unreachable("unhandled schedule clause argument" ); |
187 | } |
188 | |
189 | /// Adds ordering modifier flags to schedule type. |
190 | static OMPScheduleType |
191 | getOpenMPOrderingScheduleType(OMPScheduleType BaseScheduleType, |
192 | bool HasOrderedClause) { |
193 | assert((BaseScheduleType & OMPScheduleType::ModifierMask) == |
194 | OMPScheduleType::None && |
195 | "Must not have ordering nor monotonicity flags already set" ); |
196 | |
197 | OMPScheduleType OrderingModifier = HasOrderedClause |
198 | ? OMPScheduleType::ModifierOrdered |
199 | : OMPScheduleType::ModifierUnordered; |
200 | OMPScheduleType OrderingScheduleType = BaseScheduleType | OrderingModifier; |
201 | |
202 | // Unsupported combinations |
203 | if (OrderingScheduleType == |
204 | (OMPScheduleType::BaseGuidedSimd | OMPScheduleType::ModifierOrdered)) |
205 | return OMPScheduleType::OrderedGuidedChunked; |
206 | else if (OrderingScheduleType == (OMPScheduleType::BaseRuntimeSimd | |
207 | OMPScheduleType::ModifierOrdered)) |
208 | return OMPScheduleType::OrderedRuntime; |
209 | |
210 | return OrderingScheduleType; |
211 | } |
212 | |
213 | /// Adds monotonicity modifier flags to schedule type. |
214 | static OMPScheduleType |
215 | getOpenMPMonotonicityScheduleType(OMPScheduleType ScheduleType, |
216 | bool HasSimdModifier, bool HasMonotonic, |
217 | bool HasNonmonotonic, bool HasOrderedClause) { |
218 | assert((ScheduleType & OMPScheduleType::MonotonicityMask) == |
219 | OMPScheduleType::None && |
220 | "Must not have monotonicity flags already set" ); |
221 | assert((!HasMonotonic || !HasNonmonotonic) && |
222 | "Monotonic and Nonmonotonic are contradicting each other" ); |
223 | |
224 | if (HasMonotonic) { |
225 | return ScheduleType | OMPScheduleType::ModifierMonotonic; |
226 | } else if (HasNonmonotonic) { |
227 | return ScheduleType | OMPScheduleType::ModifierNonmonotonic; |
228 | } else { |
229 | // OpenMP 5.1, 2.11.4 Worksharing-Loop Construct, Description. |
230 | // If the static schedule kind is specified or if the ordered clause is |
231 | // specified, and if the nonmonotonic modifier is not specified, the |
232 | // effect is as if the monotonic modifier is specified. Otherwise, unless |
233 | // the monotonic modifier is specified, the effect is as if the |
234 | // nonmonotonic modifier is specified. |
235 | OMPScheduleType BaseScheduleType = |
236 | ScheduleType & ~OMPScheduleType::ModifierMask; |
237 | if ((BaseScheduleType == OMPScheduleType::BaseStatic) || |
238 | (BaseScheduleType == OMPScheduleType::BaseStaticChunked) || |
239 | HasOrderedClause) { |
240 | // The monotonic is used by default in openmp runtime library, so no need |
241 | // to set it. |
242 | return ScheduleType; |
243 | } else { |
244 | return ScheduleType | OMPScheduleType::ModifierNonmonotonic; |
245 | } |
246 | } |
247 | } |
248 | |
249 | /// Determine the schedule type using schedule and ordering clause arguments. |
250 | static OMPScheduleType |
251 | computeOpenMPScheduleType(ScheduleKind ClauseKind, bool HasChunks, |
252 | bool HasSimdModifier, bool HasMonotonicModifier, |
253 | bool HasNonmonotonicModifier, bool HasOrderedClause) { |
254 | OMPScheduleType BaseSchedule = |
255 | getOpenMPBaseScheduleType(ClauseKind, HasChunks, HasSimdModifier); |
256 | OMPScheduleType OrderedSchedule = |
257 | getOpenMPOrderingScheduleType(BaseScheduleType: BaseSchedule, HasOrderedClause); |
258 | OMPScheduleType Result = getOpenMPMonotonicityScheduleType( |
259 | ScheduleType: OrderedSchedule, HasSimdModifier, HasMonotonic: HasMonotonicModifier, |
260 | HasNonmonotonic: HasNonmonotonicModifier, HasOrderedClause); |
261 | |
262 | assert(isValidWorkshareLoopScheduleType(Result)); |
263 | return Result; |
264 | } |
265 | |
266 | /// Make \p Source branch to \p Target. |
267 | /// |
268 | /// Handles two situations: |
269 | /// * \p Source already has an unconditional branch. |
270 | /// * \p Source is a degenerate block (no terminator because the BB is |
271 | /// the current head of the IR construction). |
272 | static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL) { |
273 | if (Instruction *Term = Source->getTerminator()) { |
274 | auto *Br = cast<BranchInst>(Val: Term); |
275 | assert(!Br->isConditional() && |
276 | "BB's terminator must be an unconditional branch (or degenerate)" ); |
277 | BasicBlock *Succ = Br->getSuccessor(i: 0); |
278 | Succ->removePredecessor(Pred: Source, /*KeepOneInputPHIs=*/true); |
279 | Br->setSuccessor(idx: 0, NewSucc: Target); |
280 | return; |
281 | } |
282 | |
283 | auto *NewBr = BranchInst::Create(IfTrue: Target, InsertBefore: Source); |
284 | NewBr->setDebugLoc(DL); |
285 | } |
286 | |
287 | void llvm::spliceBB(IRBuilderBase::InsertPoint IP, BasicBlock *New, |
288 | bool CreateBranch) { |
289 | assert(New->getFirstInsertionPt() == New->begin() && |
290 | "Target BB must not have PHI nodes" ); |
291 | |
292 | // Move instructions to new block. |
293 | BasicBlock *Old = IP.getBlock(); |
294 | New->splice(ToIt: New->begin(), FromBB: Old, FromBeginIt: IP.getPoint(), FromEndIt: Old->end()); |
295 | |
296 | if (CreateBranch) |
297 | BranchInst::Create(IfTrue: New, InsertBefore: Old); |
298 | } |
299 | |
300 | void llvm::spliceBB(IRBuilder<> &Builder, BasicBlock *New, bool CreateBranch) { |
301 | DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); |
302 | BasicBlock *Old = Builder.GetInsertBlock(); |
303 | |
304 | spliceBB(IP: Builder.saveIP(), New, CreateBranch); |
305 | if (CreateBranch) |
306 | Builder.SetInsertPoint(Old->getTerminator()); |
307 | else |
308 | Builder.SetInsertPoint(Old); |
309 | |
310 | // SetInsertPoint also updates the Builder's debug location, but we want to |
311 | // keep the one the Builder was configured to use. |
312 | Builder.SetCurrentDebugLocation(DebugLoc); |
313 | } |
314 | |
315 | BasicBlock *llvm::splitBB(IRBuilderBase::InsertPoint IP, bool CreateBranch, |
316 | llvm::Twine Name) { |
317 | BasicBlock *Old = IP.getBlock(); |
318 | BasicBlock *New = BasicBlock::Create( |
319 | Context&: Old->getContext(), Name: Name.isTriviallyEmpty() ? Old->getName() : Name, |
320 | Parent: Old->getParent(), InsertBefore: Old->getNextNode()); |
321 | spliceBB(IP, New, CreateBranch); |
322 | New->replaceSuccessorsPhiUsesWith(Old, New); |
323 | return New; |
324 | } |
325 | |
326 | BasicBlock *llvm::splitBB(IRBuilderBase &Builder, bool CreateBranch, |
327 | llvm::Twine Name) { |
328 | DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); |
329 | BasicBlock *New = splitBB(IP: Builder.saveIP(), CreateBranch, Name); |
330 | if (CreateBranch) |
331 | Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator()); |
332 | else |
333 | Builder.SetInsertPoint(Builder.GetInsertBlock()); |
334 | // SetInsertPoint also updates the Builder's debug location, but we want to |
335 | // keep the one the Builder was configured to use. |
336 | Builder.SetCurrentDebugLocation(DebugLoc); |
337 | return New; |
338 | } |
339 | |
340 | BasicBlock *llvm::splitBB(IRBuilder<> &Builder, bool CreateBranch, |
341 | llvm::Twine Name) { |
342 | DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); |
343 | BasicBlock *New = splitBB(IP: Builder.saveIP(), CreateBranch, Name); |
344 | if (CreateBranch) |
345 | Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator()); |
346 | else |
347 | Builder.SetInsertPoint(Builder.GetInsertBlock()); |
348 | // SetInsertPoint also updates the Builder's debug location, but we want to |
349 | // keep the one the Builder was configured to use. |
350 | Builder.SetCurrentDebugLocation(DebugLoc); |
351 | return New; |
352 | } |
353 | |
354 | BasicBlock *llvm::splitBBWithSuffix(IRBuilderBase &Builder, bool CreateBranch, |
355 | llvm::Twine Suffix) { |
356 | BasicBlock *Old = Builder.GetInsertBlock(); |
357 | return splitBB(Builder, CreateBranch, Name: Old->getName() + Suffix); |
358 | } |
359 | |
360 | // This function creates a fake integer value and a fake use for the integer |
361 | // value. It returns the fake value created. This is useful in modeling the |
362 | // extra arguments to the outlined functions. |
363 | Value *createFakeIntVal(IRBuilderBase &Builder, |
364 | OpenMPIRBuilder::InsertPointTy OuterAllocaIP, |
365 | llvm::SmallVectorImpl<Instruction *> &ToBeDeleted, |
366 | OpenMPIRBuilder::InsertPointTy InnerAllocaIP, |
367 | const Twine &Name = "" , bool AsPtr = true) { |
368 | Builder.restoreIP(IP: OuterAllocaIP); |
369 | Instruction *FakeVal; |
370 | AllocaInst *FakeValAddr = |
371 | Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, Name: Name + ".addr" ); |
372 | ToBeDeleted.push_back(Elt: FakeValAddr); |
373 | |
374 | if (AsPtr) { |
375 | FakeVal = FakeValAddr; |
376 | } else { |
377 | FakeVal = |
378 | Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: FakeValAddr, Name: Name + ".val" ); |
379 | ToBeDeleted.push_back(Elt: FakeVal); |
380 | } |
381 | |
382 | // Generate a fake use of this value |
383 | Builder.restoreIP(IP: InnerAllocaIP); |
384 | Instruction *UseFakeVal; |
385 | if (AsPtr) { |
386 | UseFakeVal = |
387 | Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: FakeVal, Name: Name + ".use" ); |
388 | } else { |
389 | UseFakeVal = |
390 | cast<BinaryOperator>(Val: Builder.CreateAdd(LHS: FakeVal, RHS: Builder.getInt32(C: 10))); |
391 | } |
392 | ToBeDeleted.push_back(Elt: UseFakeVal); |
393 | return FakeVal; |
394 | } |
395 | |
396 | //===----------------------------------------------------------------------===// |
397 | // OpenMPIRBuilderConfig |
398 | //===----------------------------------------------------------------------===// |
399 | |
400 | namespace { |
401 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE(); |
402 | /// Values for bit flags for marking which requires clauses have been used. |
403 | enum OpenMPOffloadingRequiresDirFlags { |
404 | /// flag undefined. |
405 | OMP_REQ_UNDEFINED = 0x000, |
406 | /// no requires directive present. |
407 | OMP_REQ_NONE = 0x001, |
408 | /// reverse_offload clause. |
409 | OMP_REQ_REVERSE_OFFLOAD = 0x002, |
410 | /// unified_address clause. |
411 | OMP_REQ_UNIFIED_ADDRESS = 0x004, |
412 | /// unified_shared_memory clause. |
413 | OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008, |
414 | /// dynamic_allocators clause. |
415 | OMP_REQ_DYNAMIC_ALLOCATORS = 0x010, |
416 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS) |
417 | }; |
418 | |
419 | } // anonymous namespace |
420 | |
421 | OpenMPIRBuilderConfig::OpenMPIRBuilderConfig() |
422 | : RequiresFlags(OMP_REQ_UNDEFINED) {} |
423 | |
424 | OpenMPIRBuilderConfig::OpenMPIRBuilderConfig( |
425 | bool IsTargetDevice, bool IsGPU, bool OpenMPOffloadMandatory, |
426 | bool HasRequiresReverseOffload, bool HasRequiresUnifiedAddress, |
427 | bool HasRequiresUnifiedSharedMemory, bool HasRequiresDynamicAllocators) |
428 | : IsTargetDevice(IsTargetDevice), IsGPU(IsGPU), |
429 | OpenMPOffloadMandatory(OpenMPOffloadMandatory), |
430 | RequiresFlags(OMP_REQ_UNDEFINED) { |
431 | if (HasRequiresReverseOffload) |
432 | RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD; |
433 | if (HasRequiresUnifiedAddress) |
434 | RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS; |
435 | if (HasRequiresUnifiedSharedMemory) |
436 | RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY; |
437 | if (HasRequiresDynamicAllocators) |
438 | RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS; |
439 | } |
440 | |
441 | bool OpenMPIRBuilderConfig::hasRequiresReverseOffload() const { |
442 | return RequiresFlags & OMP_REQ_REVERSE_OFFLOAD; |
443 | } |
444 | |
445 | bool OpenMPIRBuilderConfig::hasRequiresUnifiedAddress() const { |
446 | return RequiresFlags & OMP_REQ_UNIFIED_ADDRESS; |
447 | } |
448 | |
449 | bool OpenMPIRBuilderConfig::hasRequiresUnifiedSharedMemory() const { |
450 | return RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY; |
451 | } |
452 | |
453 | bool OpenMPIRBuilderConfig::hasRequiresDynamicAllocators() const { |
454 | return RequiresFlags & OMP_REQ_DYNAMIC_ALLOCATORS; |
455 | } |
456 | |
457 | int64_t OpenMPIRBuilderConfig::getRequiresFlags() const { |
458 | return hasRequiresFlags() ? RequiresFlags |
459 | : static_cast<int64_t>(OMP_REQ_NONE); |
460 | } |
461 | |
462 | void OpenMPIRBuilderConfig::setHasRequiresReverseOffload(bool Value) { |
463 | if (Value) |
464 | RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD; |
465 | else |
466 | RequiresFlags &= ~OMP_REQ_REVERSE_OFFLOAD; |
467 | } |
468 | |
469 | void OpenMPIRBuilderConfig::setHasRequiresUnifiedAddress(bool Value) { |
470 | if (Value) |
471 | RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS; |
472 | else |
473 | RequiresFlags &= ~OMP_REQ_UNIFIED_ADDRESS; |
474 | } |
475 | |
476 | void OpenMPIRBuilderConfig::setHasRequiresUnifiedSharedMemory(bool Value) { |
477 | if (Value) |
478 | RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY; |
479 | else |
480 | RequiresFlags &= ~OMP_REQ_UNIFIED_SHARED_MEMORY; |
481 | } |
482 | |
483 | void OpenMPIRBuilderConfig::setHasRequiresDynamicAllocators(bool Value) { |
484 | if (Value) |
485 | RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS; |
486 | else |
487 | RequiresFlags &= ~OMP_REQ_DYNAMIC_ALLOCATORS; |
488 | } |
489 | |
490 | //===----------------------------------------------------------------------===// |
491 | // OpenMPIRBuilder |
492 | //===----------------------------------------------------------------------===// |
493 | |
494 | void OpenMPIRBuilder::getKernelArgsVector(TargetKernelArgs &KernelArgs, |
495 | IRBuilderBase &Builder, |
496 | SmallVector<Value *> &ArgsVector) { |
497 | Value *Version = Builder.getInt32(OMP_KERNEL_ARG_VERSION); |
498 | Value *PointerNum = Builder.getInt32(C: KernelArgs.NumTargetItems); |
499 | auto Int32Ty = Type::getInt32Ty(C&: Builder.getContext()); |
500 | Value *ZeroArray = Constant::getNullValue(Ty: ArrayType::get(ElementType: Int32Ty, NumElements: 3)); |
501 | Value *Flags = Builder.getInt64(C: KernelArgs.HasNoWait); |
502 | |
503 | Value *NumTeams3D = |
504 | Builder.CreateInsertValue(Agg: ZeroArray, Val: KernelArgs.NumTeams, Idxs: {0}); |
505 | Value *NumThreads3D = |
506 | Builder.CreateInsertValue(Agg: ZeroArray, Val: KernelArgs.NumThreads, Idxs: {0}); |
507 | |
508 | ArgsVector = {Version, |
509 | PointerNum, |
510 | KernelArgs.RTArgs.BasePointersArray, |
511 | KernelArgs.RTArgs.PointersArray, |
512 | KernelArgs.RTArgs.SizesArray, |
513 | KernelArgs.RTArgs.MapTypesArray, |
514 | KernelArgs.RTArgs.MapNamesArray, |
515 | KernelArgs.RTArgs.MappersArray, |
516 | KernelArgs.NumIterations, |
517 | Flags, |
518 | NumTeams3D, |
519 | NumThreads3D, |
520 | KernelArgs.DynCGGroupMem}; |
521 | } |
522 | |
523 | void OpenMPIRBuilder::addAttributes(omp::RuntimeFunction FnID, Function &Fn) { |
524 | LLVMContext &Ctx = Fn.getContext(); |
525 | |
526 | // Get the function's current attributes. |
527 | auto Attrs = Fn.getAttributes(); |
528 | auto FnAttrs = Attrs.getFnAttrs(); |
529 | auto RetAttrs = Attrs.getRetAttrs(); |
530 | SmallVector<AttributeSet, 4> ArgAttrs; |
531 | for (size_t ArgNo = 0; ArgNo < Fn.arg_size(); ++ArgNo) |
532 | ArgAttrs.emplace_back(Args: Attrs.getParamAttrs(ArgNo)); |
533 | |
534 | // Add AS to FnAS while taking special care with integer extensions. |
535 | auto addAttrSet = [&](AttributeSet &FnAS, const AttributeSet &AS, |
536 | bool Param = true) -> void { |
537 | bool HasSignExt = AS.hasAttribute(Kind: Attribute::SExt); |
538 | bool HasZeroExt = AS.hasAttribute(Kind: Attribute::ZExt); |
539 | if (HasSignExt || HasZeroExt) { |
540 | assert(AS.getNumAttributes() == 1 && |
541 | "Currently not handling extension attr combined with others." ); |
542 | if (Param) { |
543 | if (auto AK = TargetLibraryInfo::getExtAttrForI32Param(T, Signed: HasSignExt)) |
544 | FnAS = FnAS.addAttribute(C&: Ctx, Kind: AK); |
545 | } else if (auto AK = |
546 | TargetLibraryInfo::getExtAttrForI32Return(T, Signed: HasSignExt)) |
547 | FnAS = FnAS.addAttribute(C&: Ctx, Kind: AK); |
548 | } else { |
549 | FnAS = FnAS.addAttributes(C&: Ctx, AS); |
550 | } |
551 | }; |
552 | |
553 | #define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet; |
554 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
555 | |
556 | // Add attributes to the function declaration. |
557 | switch (FnID) { |
558 | #define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets) \ |
559 | case Enum: \ |
560 | FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet); \ |
561 | addAttrSet(RetAttrs, RetAttrSet, /*Param*/ false); \ |
562 | for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo) \ |
563 | addAttrSet(ArgAttrs[ArgNo], ArgAttrSets[ArgNo]); \ |
564 | Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs)); \ |
565 | break; |
566 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
567 | default: |
568 | // Attributes are optional. |
569 | break; |
570 | } |
571 | } |
572 | |
573 | FunctionCallee |
574 | OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) { |
575 | FunctionType *FnTy = nullptr; |
576 | Function *Fn = nullptr; |
577 | |
578 | // Try to find the declation in the module first. |
579 | switch (FnID) { |
580 | #define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...) \ |
581 | case Enum: \ |
582 | FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__}, \ |
583 | IsVarArg); \ |
584 | Fn = M.getFunction(Str); \ |
585 | break; |
586 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
587 | } |
588 | |
589 | if (!Fn) { |
590 | // Create a new declaration if we need one. |
591 | switch (FnID) { |
592 | #define OMP_RTL(Enum, Str, ...) \ |
593 | case Enum: \ |
594 | Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M); \ |
595 | break; |
596 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
597 | } |
598 | |
599 | // Add information if the runtime function takes a callback function |
600 | if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) { |
601 | if (!Fn->hasMetadata(KindID: LLVMContext::MD_callback)) { |
602 | LLVMContext &Ctx = Fn->getContext(); |
603 | MDBuilder MDB(Ctx); |
604 | // Annotate the callback behavior of the runtime function: |
605 | // - The callback callee is argument number 2 (microtask). |
606 | // - The first two arguments of the callback callee are unknown (-1). |
607 | // - All variadic arguments to the runtime function are passed to the |
608 | // callback callee. |
609 | Fn->addMetadata( |
610 | KindID: LLVMContext::MD_callback, |
611 | MD&: *MDNode::get(Context&: Ctx, MDs: {MDB.createCallbackEncoding( |
612 | CalleeArgNo: 2, Arguments: {-1, -1}, /* VarArgsArePassed */ true)})); |
613 | } |
614 | } |
615 | |
616 | LLVM_DEBUG(dbgs() << "Created OpenMP runtime function " << Fn->getName() |
617 | << " with type " << *Fn->getFunctionType() << "\n" ); |
618 | addAttributes(FnID, Fn&: *Fn); |
619 | |
620 | } else { |
621 | LLVM_DEBUG(dbgs() << "Found OpenMP runtime function " << Fn->getName() |
622 | << " with type " << *Fn->getFunctionType() << "\n" ); |
623 | } |
624 | |
625 | assert(Fn && "Failed to create OpenMP runtime function" ); |
626 | |
627 | return {FnTy, Fn}; |
628 | } |
629 | |
630 | Function *OpenMPIRBuilder::getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID) { |
631 | FunctionCallee RTLFn = getOrCreateRuntimeFunction(M, FnID); |
632 | auto *Fn = dyn_cast<llvm::Function>(Val: RTLFn.getCallee()); |
633 | assert(Fn && "Failed to create OpenMP runtime function pointer" ); |
634 | return Fn; |
635 | } |
636 | |
637 | void OpenMPIRBuilder::initialize() { initializeTypes(M); } |
638 | |
639 | static void raiseUserConstantDataAllocasToEntryBlock(IRBuilderBase &Builder, |
640 | Function *Function) { |
641 | BasicBlock &EntryBlock = Function->getEntryBlock(); |
642 | Instruction *MoveLocInst = EntryBlock.getFirstNonPHI(); |
643 | |
644 | // Loop over blocks looking for constant allocas, skipping the entry block |
645 | // as any allocas there are already in the desired location. |
646 | for (auto Block = std::next(x: Function->begin(), n: 1); Block != Function->end(); |
647 | Block++) { |
648 | for (auto Inst = Block->getReverseIterator()->begin(); |
649 | Inst != Block->getReverseIterator()->end();) { |
650 | if (auto *AllocaInst = dyn_cast_if_present<llvm::AllocaInst>(Val&: Inst)) { |
651 | Inst++; |
652 | if (!isa<ConstantData>(Val: AllocaInst->getArraySize())) |
653 | continue; |
654 | AllocaInst->moveBeforePreserving(MovePos: MoveLocInst); |
655 | } else { |
656 | Inst++; |
657 | } |
658 | } |
659 | } |
660 | } |
661 | |
662 | void OpenMPIRBuilder::finalize(Function *Fn) { |
663 | SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet; |
664 | SmallVector<BasicBlock *, 32> Blocks; |
665 | SmallVector<OutlineInfo, 16> DeferredOutlines; |
666 | for (OutlineInfo &OI : OutlineInfos) { |
667 | // Skip functions that have not finalized yet; may happen with nested |
668 | // function generation. |
669 | if (Fn && OI.getFunction() != Fn) { |
670 | DeferredOutlines.push_back(Elt: OI); |
671 | continue; |
672 | } |
673 | |
674 | ParallelRegionBlockSet.clear(); |
675 | Blocks.clear(); |
676 | OI.collectBlocks(BlockSet&: ParallelRegionBlockSet, BlockVector&: Blocks); |
677 | |
678 | Function *OuterFn = OI.getFunction(); |
679 | CodeExtractorAnalysisCache CEAC(*OuterFn); |
680 | // If we generate code for the target device, we need to allocate |
681 | // struct for aggregate params in the device default alloca address space. |
682 | // OpenMP runtime requires that the params of the extracted functions are |
683 | // passed as zero address space pointers. This flag ensures that |
684 | // CodeExtractor generates correct code for extracted functions |
685 | // which are used by OpenMP runtime. |
686 | bool ArgsInZeroAddressSpace = Config.isTargetDevice(); |
687 | CodeExtractor (Blocks, /* DominatorTree */ nullptr, |
688 | /* AggregateArgs */ true, |
689 | /* BlockFrequencyInfo */ nullptr, |
690 | /* BranchProbabilityInfo */ nullptr, |
691 | /* AssumptionCache */ nullptr, |
692 | /* AllowVarArgs */ true, |
693 | /* AllowAlloca */ true, |
694 | /* AllocaBlock*/ OI.OuterAllocaBB, |
695 | /* Suffix */ ".omp_par" , ArgsInZeroAddressSpace); |
696 | |
697 | LLVM_DEBUG(dbgs() << "Before outlining: " << *OuterFn << "\n" ); |
698 | LLVM_DEBUG(dbgs() << "Entry " << OI.EntryBB->getName() |
699 | << " Exit: " << OI.ExitBB->getName() << "\n" ); |
700 | assert(Extractor.isEligible() && |
701 | "Expected OpenMP outlining to be possible!" ); |
702 | |
703 | for (auto *V : OI.ExcludeArgsFromAggregate) |
704 | Extractor.excludeArgFromAggregate(Arg: V); |
705 | |
706 | Function *OutlinedFn = Extractor.extractCodeRegion(CEAC); |
707 | |
708 | // Forward target-cpu, target-features attributes to the outlined function. |
709 | auto TargetCpuAttr = OuterFn->getFnAttribute(Kind: "target-cpu" ); |
710 | if (TargetCpuAttr.isStringAttribute()) |
711 | OutlinedFn->addFnAttr(Attr: TargetCpuAttr); |
712 | |
713 | auto TargetFeaturesAttr = OuterFn->getFnAttribute(Kind: "target-features" ); |
714 | if (TargetFeaturesAttr.isStringAttribute()) |
715 | OutlinedFn->addFnAttr(Attr: TargetFeaturesAttr); |
716 | |
717 | LLVM_DEBUG(dbgs() << "After outlining: " << *OuterFn << "\n" ); |
718 | LLVM_DEBUG(dbgs() << " Outlined function: " << *OutlinedFn << "\n" ); |
719 | assert(OutlinedFn->getReturnType()->isVoidTy() && |
720 | "OpenMP outlined functions should not return a value!" ); |
721 | |
722 | // For compability with the clang CG we move the outlined function after the |
723 | // one with the parallel region. |
724 | OutlinedFn->removeFromParent(); |
725 | M.getFunctionList().insertAfter(where: OuterFn->getIterator(), New: OutlinedFn); |
726 | |
727 | // Remove the artificial entry introduced by the extractor right away, we |
728 | // made our own entry block after all. |
729 | { |
730 | BasicBlock &ArtificialEntry = OutlinedFn->getEntryBlock(); |
731 | assert(ArtificialEntry.getUniqueSuccessor() == OI.EntryBB); |
732 | assert(OI.EntryBB->getUniquePredecessor() == &ArtificialEntry); |
733 | // Move instructions from the to-be-deleted ArtificialEntry to the entry |
734 | // basic block of the parallel region. CodeExtractor generates |
735 | // instructions to unwrap the aggregate argument and may sink |
736 | // allocas/bitcasts for values that are solely used in the outlined region |
737 | // and do not escape. |
738 | assert(!ArtificialEntry.empty() && |
739 | "Expected instructions to add in the outlined region entry" ); |
740 | for (BasicBlock::reverse_iterator It = ArtificialEntry.rbegin(), |
741 | End = ArtificialEntry.rend(); |
742 | It != End;) { |
743 | Instruction &I = *It; |
744 | It++; |
745 | |
746 | if (I.isTerminator()) |
747 | continue; |
748 | |
749 | I.moveBeforePreserving(BB&: *OI.EntryBB, I: OI.EntryBB->getFirstInsertionPt()); |
750 | } |
751 | |
752 | OI.EntryBB->moveBefore(MovePos: &ArtificialEntry); |
753 | ArtificialEntry.eraseFromParent(); |
754 | } |
755 | assert(&OutlinedFn->getEntryBlock() == OI.EntryBB); |
756 | assert(OutlinedFn && OutlinedFn->getNumUses() == 1); |
757 | |
758 | // Run a user callback, e.g. to add attributes. |
759 | if (OI.PostOutlineCB) |
760 | OI.PostOutlineCB(*OutlinedFn); |
761 | } |
762 | |
763 | // Remove work items that have been completed. |
764 | OutlineInfos = std::move(DeferredOutlines); |
765 | |
766 | // The createTarget functions embeds user written code into |
767 | // the target region which may inject allocas which need to |
768 | // be moved to the entry block of our target or risk malformed |
769 | // optimisations by later passes, this is only relevant for |
770 | // the device pass which appears to be a little more delicate |
771 | // when it comes to optimisations (however, we do not block on |
772 | // that here, it's up to the inserter to the list to do so). |
773 | // This notbaly has to occur after the OutlinedInfo candidates |
774 | // have been extracted so we have an end product that will not |
775 | // be implicitly adversely affected by any raises unless |
776 | // intentionally appended to the list. |
777 | // NOTE: This only does so for ConstantData, it could be extended |
778 | // to ConstantExpr's with further effort, however, they should |
779 | // largely be folded when they get here. Extending it to runtime |
780 | // defined/read+writeable allocation sizes would be non-trivial |
781 | // (need to factor in movement of any stores to variables the |
782 | // allocation size depends on, as well as the usual loads, |
783 | // otherwise it'll yield the wrong result after movement) and |
784 | // likely be more suitable as an LLVM optimisation pass. |
785 | for (Function *F : ConstantAllocaRaiseCandidates) |
786 | raiseUserConstantDataAllocasToEntryBlock(Builder, Function: F); |
787 | |
788 | EmitMetadataErrorReportFunctionTy &&ErrorReportFn = |
789 | [](EmitMetadataErrorKind Kind, |
790 | const TargetRegionEntryInfo &EntryInfo) -> void { |
791 | errs() << "Error of kind: " << Kind |
792 | << " when emitting offload entries and metadata during " |
793 | "OMPIRBuilder finalization \n" ; |
794 | }; |
795 | |
796 | if (!OffloadInfoManager.empty()) |
797 | createOffloadEntriesAndInfoMetadata(ErrorReportFunction&: ErrorReportFn); |
798 | |
799 | if (Config.EmitLLVMUsedMetaInfo.value_or(u: false)) { |
800 | std::vector<WeakTrackingVH> LLVMCompilerUsed = { |
801 | M.getGlobalVariable(Name: "__openmp_nvptx_data_transfer_temporary_storage" )}; |
802 | emitUsed(Name: "llvm.compiler.used" , List&: LLVMCompilerUsed); |
803 | } |
804 | } |
805 | |
806 | OpenMPIRBuilder::~OpenMPIRBuilder() { |
807 | assert(OutlineInfos.empty() && "There must be no outstanding outlinings" ); |
808 | } |
809 | |
810 | GlobalValue *OpenMPIRBuilder::createGlobalFlag(unsigned Value, StringRef Name) { |
811 | IntegerType *I32Ty = Type::getInt32Ty(C&: M.getContext()); |
812 | auto *GV = |
813 | new GlobalVariable(M, I32Ty, |
814 | /* isConstant = */ true, GlobalValue::WeakODRLinkage, |
815 | ConstantInt::get(Ty: I32Ty, V: Value), Name); |
816 | GV->setVisibility(GlobalValue::HiddenVisibility); |
817 | |
818 | return GV; |
819 | } |
820 | |
821 | Constant *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr, |
822 | uint32_t SrcLocStrSize, |
823 | IdentFlag LocFlags, |
824 | unsigned Reserve2Flags) { |
825 | // Enable "C-mode". |
826 | LocFlags |= OMP_IDENT_FLAG_KMPC; |
827 | |
828 | Constant *&Ident = |
829 | IdentMap[{SrcLocStr, uint64_t(LocFlags) << 31 | Reserve2Flags}]; |
830 | if (!Ident) { |
831 | Constant *I32Null = ConstantInt::getNullValue(Ty: Int32); |
832 | Constant *IdentData[] = {I32Null, |
833 | ConstantInt::get(Ty: Int32, V: uint32_t(LocFlags)), |
834 | ConstantInt::get(Ty: Int32, V: Reserve2Flags), |
835 | ConstantInt::get(Ty: Int32, V: SrcLocStrSize), SrcLocStr}; |
836 | Constant *Initializer = |
837 | ConstantStruct::get(T: OpenMPIRBuilder::Ident, V: IdentData); |
838 | |
839 | // Look for existing encoding of the location + flags, not needed but |
840 | // minimizes the difference to the existing solution while we transition. |
841 | for (GlobalVariable &GV : M.globals()) |
842 | if (GV.getValueType() == OpenMPIRBuilder::Ident && GV.hasInitializer()) |
843 | if (GV.getInitializer() == Initializer) |
844 | Ident = &GV; |
845 | |
846 | if (!Ident) { |
847 | auto *GV = new GlobalVariable( |
848 | M, OpenMPIRBuilder::Ident, |
849 | /* isConstant = */ true, GlobalValue::PrivateLinkage, Initializer, "" , |
850 | nullptr, GlobalValue::NotThreadLocal, |
851 | M.getDataLayout().getDefaultGlobalsAddressSpace()); |
852 | GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); |
853 | GV->setAlignment(Align(8)); |
854 | Ident = GV; |
855 | } |
856 | } |
857 | |
858 | return ConstantExpr::getPointerBitCastOrAddrSpaceCast(C: Ident, Ty: IdentPtr); |
859 | } |
860 | |
861 | Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr, |
862 | uint32_t &SrcLocStrSize) { |
863 | SrcLocStrSize = LocStr.size(); |
864 | Constant *&SrcLocStr = SrcLocStrMap[LocStr]; |
865 | if (!SrcLocStr) { |
866 | Constant *Initializer = |
867 | ConstantDataArray::getString(Context&: M.getContext(), Initializer: LocStr); |
868 | |
869 | // Look for existing encoding of the location, not needed but minimizes the |
870 | // difference to the existing solution while we transition. |
871 | for (GlobalVariable &GV : M.globals()) |
872 | if (GV.isConstant() && GV.hasInitializer() && |
873 | GV.getInitializer() == Initializer) |
874 | return SrcLocStr = ConstantExpr::getPointerCast(C: &GV, Ty: Int8Ptr); |
875 | |
876 | SrcLocStr = Builder.CreateGlobalStringPtr(Str: LocStr, /* Name */ "" , |
877 | /* AddressSpace */ 0, M: &M); |
878 | } |
879 | return SrcLocStr; |
880 | } |
881 | |
882 | Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef FunctionName, |
883 | StringRef FileName, |
884 | unsigned Line, unsigned Column, |
885 | uint32_t &SrcLocStrSize) { |
886 | SmallString<128> Buffer; |
887 | Buffer.push_back(Elt: ';'); |
888 | Buffer.append(RHS: FileName); |
889 | Buffer.push_back(Elt: ';'); |
890 | Buffer.append(RHS: FunctionName); |
891 | Buffer.push_back(Elt: ';'); |
892 | Buffer.append(RHS: std::to_string(val: Line)); |
893 | Buffer.push_back(Elt: ';'); |
894 | Buffer.append(RHS: std::to_string(val: Column)); |
895 | Buffer.push_back(Elt: ';'); |
896 | Buffer.push_back(Elt: ';'); |
897 | return getOrCreateSrcLocStr(LocStr: Buffer.str(), SrcLocStrSize); |
898 | } |
899 | |
900 | Constant * |
901 | OpenMPIRBuilder::getOrCreateDefaultSrcLocStr(uint32_t &SrcLocStrSize) { |
902 | StringRef UnknownLoc = ";unknown;unknown;0;0;;" ; |
903 | return getOrCreateSrcLocStr(LocStr: UnknownLoc, SrcLocStrSize); |
904 | } |
905 | |
906 | Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(DebugLoc DL, |
907 | uint32_t &SrcLocStrSize, |
908 | Function *F) { |
909 | DILocation *DIL = DL.get(); |
910 | if (!DIL) |
911 | return getOrCreateDefaultSrcLocStr(SrcLocStrSize); |
912 | StringRef FileName = M.getName(); |
913 | if (DIFile *DIF = DIL->getFile()) |
914 | if (std::optional<StringRef> Source = DIF->getSource()) |
915 | FileName = *Source; |
916 | StringRef Function = DIL->getScope()->getSubprogram()->getName(); |
917 | if (Function.empty() && F) |
918 | Function = F->getName(); |
919 | return getOrCreateSrcLocStr(FunctionName: Function, FileName, Line: DIL->getLine(), |
920 | Column: DIL->getColumn(), SrcLocStrSize); |
921 | } |
922 | |
923 | Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(const LocationDescription &Loc, |
924 | uint32_t &SrcLocStrSize) { |
925 | return getOrCreateSrcLocStr(DL: Loc.DL, SrcLocStrSize, |
926 | F: Loc.IP.getBlock()->getParent()); |
927 | } |
928 | |
929 | Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) { |
930 | return Builder.CreateCall( |
931 | Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_global_thread_num), Args: Ident, |
932 | Name: "omp_global_thread_num" ); |
933 | } |
934 | |
935 | OpenMPIRBuilder::InsertPointTy |
936 | OpenMPIRBuilder::createBarrier(const LocationDescription &Loc, Directive Kind, |
937 | bool ForceSimpleCall, bool CheckCancelFlag) { |
938 | if (!updateToLocation(Loc)) |
939 | return Loc.IP; |
940 | |
941 | // Build call __kmpc_cancel_barrier(loc, thread_id) or |
942 | // __kmpc_barrier(loc, thread_id); |
943 | |
944 | IdentFlag BarrierLocFlags; |
945 | switch (Kind) { |
946 | case OMPD_for: |
947 | BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR; |
948 | break; |
949 | case OMPD_sections: |
950 | BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS; |
951 | break; |
952 | case OMPD_single: |
953 | BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE; |
954 | break; |
955 | case OMPD_barrier: |
956 | BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL; |
957 | break; |
958 | default: |
959 | BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL; |
960 | break; |
961 | } |
962 | |
963 | uint32_t SrcLocStrSize; |
964 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
965 | Value *Args[] = { |
966 | getOrCreateIdent(SrcLocStr, SrcLocStrSize, LocFlags: BarrierLocFlags), |
967 | getOrCreateThreadID(Ident: getOrCreateIdent(SrcLocStr, SrcLocStrSize))}; |
968 | |
969 | // If we are in a cancellable parallel region, barriers are cancellation |
970 | // points. |
971 | // TODO: Check why we would force simple calls or to ignore the cancel flag. |
972 | bool UseCancelBarrier = |
973 | !ForceSimpleCall && isLastFinalizationInfoCancellable(DK: OMPD_parallel); |
974 | |
975 | Value *Result = |
976 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr( |
977 | FnID: UseCancelBarrier ? OMPRTL___kmpc_cancel_barrier |
978 | : OMPRTL___kmpc_barrier), |
979 | Args); |
980 | |
981 | if (UseCancelBarrier && CheckCancelFlag) |
982 | emitCancelationCheckImpl(CancelFlag: Result, CanceledDirective: OMPD_parallel); |
983 | |
984 | return Builder.saveIP(); |
985 | } |
986 | |
987 | OpenMPIRBuilder::InsertPointTy |
988 | OpenMPIRBuilder::createCancel(const LocationDescription &Loc, |
989 | Value *IfCondition, |
990 | omp::Directive CanceledDirective) { |
991 | if (!updateToLocation(Loc)) |
992 | return Loc.IP; |
993 | |
994 | // LLVM utilities like blocks with terminators. |
995 | auto *UI = Builder.CreateUnreachable(); |
996 | |
997 | Instruction *ThenTI = UI, *ElseTI = nullptr; |
998 | if (IfCondition) |
999 | SplitBlockAndInsertIfThenElse(Cond: IfCondition, SplitBefore: UI, ThenTerm: &ThenTI, ElseTerm: &ElseTI); |
1000 | Builder.SetInsertPoint(ThenTI); |
1001 | |
1002 | Value *CancelKind = nullptr; |
1003 | switch (CanceledDirective) { |
1004 | #define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value) \ |
1005 | case DirectiveEnum: \ |
1006 | CancelKind = Builder.getInt32(Value); \ |
1007 | break; |
1008 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
1009 | default: |
1010 | llvm_unreachable("Unknown cancel kind!" ); |
1011 | } |
1012 | |
1013 | uint32_t SrcLocStrSize; |
1014 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1015 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
1016 | Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind}; |
1017 | Value *Result = Builder.CreateCall( |
1018 | Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_cancel), Args); |
1019 | auto ExitCB = [this, CanceledDirective, Loc](InsertPointTy IP) { |
1020 | if (CanceledDirective == OMPD_parallel) { |
1021 | IRBuilder<>::InsertPointGuard IPG(Builder); |
1022 | Builder.restoreIP(IP); |
1023 | createBarrier(Loc: LocationDescription(Builder.saveIP(), Loc.DL), |
1024 | Kind: omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false, |
1025 | /* CheckCancelFlag */ false); |
1026 | } |
1027 | }; |
1028 | |
1029 | // The actual cancel logic is shared with others, e.g., cancel_barriers. |
1030 | emitCancelationCheckImpl(CancelFlag: Result, CanceledDirective, ExitCB); |
1031 | |
1032 | // Update the insertion point and remove the terminator we introduced. |
1033 | Builder.SetInsertPoint(UI->getParent()); |
1034 | UI->eraseFromParent(); |
1035 | |
1036 | return Builder.saveIP(); |
1037 | } |
1038 | |
1039 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetKernel( |
1040 | const LocationDescription &Loc, InsertPointTy AllocaIP, Value *&Return, |
1041 | Value *Ident, Value *DeviceID, Value *NumTeams, Value *NumThreads, |
1042 | Value *HostPtr, ArrayRef<Value *> KernelArgs) { |
1043 | if (!updateToLocation(Loc)) |
1044 | return Loc.IP; |
1045 | |
1046 | Builder.restoreIP(IP: AllocaIP); |
1047 | auto *KernelArgsPtr = |
1048 | Builder.CreateAlloca(Ty: OpenMPIRBuilder::KernelArgs, ArraySize: nullptr, Name: "kernel_args" ); |
1049 | Builder.restoreIP(IP: Loc.IP); |
1050 | |
1051 | for (unsigned I = 0, Size = KernelArgs.size(); I != Size; ++I) { |
1052 | llvm::Value *Arg = |
1053 | Builder.CreateStructGEP(Ty: OpenMPIRBuilder::KernelArgs, Ptr: KernelArgsPtr, Idx: I); |
1054 | Builder.CreateAlignedStore( |
1055 | Val: KernelArgs[I], Ptr: Arg, |
1056 | Align: M.getDataLayout().getPrefTypeAlign(Ty: KernelArgs[I]->getType())); |
1057 | } |
1058 | |
1059 | SmallVector<Value *> OffloadingArgs{Ident, DeviceID, NumTeams, |
1060 | NumThreads, HostPtr, KernelArgsPtr}; |
1061 | |
1062 | Return = Builder.CreateCall( |
1063 | Callee: getOrCreateRuntimeFunction(M, FnID: OMPRTL___tgt_target_kernel), |
1064 | Args: OffloadingArgs); |
1065 | |
1066 | return Builder.saveIP(); |
1067 | } |
1068 | |
1069 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitKernelLaunch( |
1070 | const LocationDescription &Loc, Function *OutlinedFn, Value *OutlinedFnID, |
1071 | EmitFallbackCallbackTy emitTargetCallFallbackCB, TargetKernelArgs &Args, |
1072 | Value *DeviceID, Value *RTLoc, InsertPointTy AllocaIP) { |
1073 | |
1074 | if (!updateToLocation(Loc)) |
1075 | return Loc.IP; |
1076 | |
1077 | Builder.restoreIP(IP: Loc.IP); |
1078 | // On top of the arrays that were filled up, the target offloading call |
1079 | // takes as arguments the device id as well as the host pointer. The host |
1080 | // pointer is used by the runtime library to identify the current target |
1081 | // region, so it only has to be unique and not necessarily point to |
1082 | // anything. It could be the pointer to the outlined function that |
1083 | // implements the target region, but we aren't using that so that the |
1084 | // compiler doesn't need to keep that, and could therefore inline the host |
1085 | // function if proven worthwhile during optimization. |
1086 | |
1087 | // From this point on, we need to have an ID of the target region defined. |
1088 | assert(OutlinedFnID && "Invalid outlined function ID!" ); |
1089 | (void)OutlinedFnID; |
1090 | |
1091 | // Return value of the runtime offloading call. |
1092 | Value *Return = nullptr; |
1093 | |
1094 | // Arguments for the target kernel. |
1095 | SmallVector<Value *> ArgsVector; |
1096 | getKernelArgsVector(KernelArgs&: Args, Builder, ArgsVector); |
1097 | |
1098 | // The target region is an outlined function launched by the runtime |
1099 | // via calls to __tgt_target_kernel(). |
1100 | // |
1101 | // Note that on the host and CPU targets, the runtime implementation of |
1102 | // these calls simply call the outlined function without forking threads. |
1103 | // The outlined functions themselves have runtime calls to |
1104 | // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by |
1105 | // the compiler in emitTeamsCall() and emitParallelCall(). |
1106 | // |
1107 | // In contrast, on the NVPTX target, the implementation of |
1108 | // __tgt_target_teams() launches a GPU kernel with the requested number |
1109 | // of teams and threads so no additional calls to the runtime are required. |
1110 | // Check the error code and execute the host version if required. |
1111 | Builder.restoreIP(IP: emitTargetKernel(Loc: Builder, AllocaIP, Return, Ident: RTLoc, DeviceID, |
1112 | NumTeams: Args.NumTeams, NumThreads: Args.NumThreads, |
1113 | HostPtr: OutlinedFnID, KernelArgs: ArgsVector)); |
1114 | |
1115 | BasicBlock *OffloadFailedBlock = |
1116 | BasicBlock::Create(Context&: Builder.getContext(), Name: "omp_offload.failed" ); |
1117 | BasicBlock *OffloadContBlock = |
1118 | BasicBlock::Create(Context&: Builder.getContext(), Name: "omp_offload.cont" ); |
1119 | Value *Failed = Builder.CreateIsNotNull(Arg: Return); |
1120 | Builder.CreateCondBr(Cond: Failed, True: OffloadFailedBlock, False: OffloadContBlock); |
1121 | |
1122 | auto CurFn = Builder.GetInsertBlock()->getParent(); |
1123 | emitBlock(BB: OffloadFailedBlock, CurFn); |
1124 | Builder.restoreIP(IP: emitTargetCallFallbackCB(Builder.saveIP())); |
1125 | emitBranch(Target: OffloadContBlock); |
1126 | emitBlock(BB: OffloadContBlock, CurFn, /*IsFinished=*/true); |
1127 | return Builder.saveIP(); |
1128 | } |
1129 | |
1130 | void OpenMPIRBuilder::emitCancelationCheckImpl(Value *CancelFlag, |
1131 | omp::Directive CanceledDirective, |
1132 | FinalizeCallbackTy ExitCB) { |
1133 | assert(isLastFinalizationInfoCancellable(CanceledDirective) && |
1134 | "Unexpected cancellation!" ); |
1135 | |
1136 | // For a cancel barrier we create two new blocks. |
1137 | BasicBlock *BB = Builder.GetInsertBlock(); |
1138 | BasicBlock *NonCancellationBlock; |
1139 | if (Builder.GetInsertPoint() == BB->end()) { |
1140 | // TODO: This branch will not be needed once we moved to the |
1141 | // OpenMPIRBuilder codegen completely. |
1142 | NonCancellationBlock = BasicBlock::Create( |
1143 | Context&: BB->getContext(), Name: BB->getName() + ".cont" , Parent: BB->getParent()); |
1144 | } else { |
1145 | NonCancellationBlock = SplitBlock(Old: BB, SplitPt: &*Builder.GetInsertPoint()); |
1146 | BB->getTerminator()->eraseFromParent(); |
1147 | Builder.SetInsertPoint(BB); |
1148 | } |
1149 | BasicBlock *CancellationBlock = BasicBlock::Create( |
1150 | Context&: BB->getContext(), Name: BB->getName() + ".cncl" , Parent: BB->getParent()); |
1151 | |
1152 | // Jump to them based on the return value. |
1153 | Value *Cmp = Builder.CreateIsNull(Arg: CancelFlag); |
1154 | Builder.CreateCondBr(Cond: Cmp, True: NonCancellationBlock, False: CancellationBlock, |
1155 | /* TODO weight */ BranchWeights: nullptr, Unpredictable: nullptr); |
1156 | |
1157 | // From the cancellation block we finalize all variables and go to the |
1158 | // post finalization block that is known to the FiniCB callback. |
1159 | Builder.SetInsertPoint(CancellationBlock); |
1160 | if (ExitCB) |
1161 | ExitCB(Builder.saveIP()); |
1162 | auto &FI = FinalizationStack.back(); |
1163 | FI.FiniCB(Builder.saveIP()); |
1164 | |
1165 | // The continuation block is where code generation continues. |
1166 | Builder.SetInsertPoint(TheBB: NonCancellationBlock, IP: NonCancellationBlock->begin()); |
1167 | } |
1168 | |
1169 | // Callback used to create OpenMP runtime calls to support |
1170 | // omp parallel clause for the device. |
1171 | // We need to use this callback to replace call to the OutlinedFn in OuterFn |
1172 | // by the call to the OpenMP DeviceRTL runtime function (kmpc_parallel_51) |
1173 | static void targetParallelCallback( |
1174 | OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn, Function *OuterFn, |
1175 | BasicBlock *OuterAllocaBB, Value *Ident, Value *IfCondition, |
1176 | Value *NumThreads, Instruction *PrivTID, AllocaInst *PrivTIDAddr, |
1177 | Value *ThreadID, const SmallVector<Instruction *, 4> &ToBeDeleted) { |
1178 | // Add some known attributes. |
1179 | IRBuilder<> &Builder = OMPIRBuilder->Builder; |
1180 | OutlinedFn.addParamAttr(ArgNo: 0, Kind: Attribute::NoAlias); |
1181 | OutlinedFn.addParamAttr(ArgNo: 1, Kind: Attribute::NoAlias); |
1182 | OutlinedFn.addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
1183 | OutlinedFn.addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
1184 | OutlinedFn.addFnAttr(Kind: Attribute::NoUnwind); |
1185 | |
1186 | assert(OutlinedFn.arg_size() >= 2 && |
1187 | "Expected at least tid and bounded tid as arguments" ); |
1188 | unsigned NumCapturedVars = OutlinedFn.arg_size() - /* tid & bounded tid */ 2; |
1189 | |
1190 | CallInst *CI = cast<CallInst>(Val: OutlinedFn.user_back()); |
1191 | assert(CI && "Expected call instruction to outlined function" ); |
1192 | CI->getParent()->setName("omp_parallel" ); |
1193 | |
1194 | Builder.SetInsertPoint(CI); |
1195 | Type *PtrTy = OMPIRBuilder->VoidPtr; |
1196 | Value *NullPtrValue = Constant::getNullValue(Ty: PtrTy); |
1197 | |
1198 | // Add alloca for kernel args |
1199 | OpenMPIRBuilder ::InsertPointTy CurrentIP = Builder.saveIP(); |
1200 | Builder.SetInsertPoint(TheBB: OuterAllocaBB, IP: OuterAllocaBB->getFirstInsertionPt()); |
1201 | AllocaInst *ArgsAlloca = |
1202 | Builder.CreateAlloca(Ty: ArrayType::get(ElementType: PtrTy, NumElements: NumCapturedVars)); |
1203 | Value *Args = ArgsAlloca; |
1204 | // Add address space cast if array for storing arguments is not allocated |
1205 | // in address space 0 |
1206 | if (ArgsAlloca->getAddressSpace()) |
1207 | Args = Builder.CreatePointerCast(V: ArgsAlloca, DestTy: PtrTy); |
1208 | Builder.restoreIP(IP: CurrentIP); |
1209 | |
1210 | // Store captured vars which are used by kmpc_parallel_51 |
1211 | for (unsigned Idx = 0; Idx < NumCapturedVars; Idx++) { |
1212 | Value *V = *(CI->arg_begin() + 2 + Idx); |
1213 | Value *StoreAddress = Builder.CreateConstInBoundsGEP2_64( |
1214 | Ty: ArrayType::get(ElementType: PtrTy, NumElements: NumCapturedVars), Ptr: Args, Idx0: 0, Idx1: Idx); |
1215 | Builder.CreateStore(Val: V, Ptr: StoreAddress); |
1216 | } |
1217 | |
1218 | Value *Cond = |
1219 | IfCondition ? Builder.CreateSExtOrTrunc(V: IfCondition, DestTy: OMPIRBuilder->Int32) |
1220 | : Builder.getInt32(C: 1); |
1221 | |
1222 | // Build kmpc_parallel_51 call |
1223 | Value *Parallel51CallArgs[] = { |
1224 | /* identifier*/ Ident, |
1225 | /* global thread num*/ ThreadID, |
1226 | /* if expression */ Cond, |
1227 | /* number of threads */ NumThreads ? NumThreads : Builder.getInt32(C: -1), |
1228 | /* Proc bind */ Builder.getInt32(C: -1), |
1229 | /* outlined function */ |
1230 | Builder.CreateBitCast(V: &OutlinedFn, DestTy: OMPIRBuilder->ParallelTaskPtr), |
1231 | /* wrapper function */ NullPtrValue, |
1232 | /* arguments of the outlined funciton*/ Args, |
1233 | /* number of arguments */ Builder.getInt64(C: NumCapturedVars)}; |
1234 | |
1235 | FunctionCallee RTLFn = |
1236 | OMPIRBuilder->getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_parallel_51); |
1237 | |
1238 | Builder.CreateCall(Callee: RTLFn, Args: Parallel51CallArgs); |
1239 | |
1240 | LLVM_DEBUG(dbgs() << "With kmpc_parallel_51 placed: " |
1241 | << *Builder.GetInsertBlock()->getParent() << "\n" ); |
1242 | |
1243 | // Initialize the local TID stack location with the argument value. |
1244 | Builder.SetInsertPoint(PrivTID); |
1245 | Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin(); |
1246 | Builder.CreateStore(Val: Builder.CreateLoad(Ty: OMPIRBuilder->Int32, Ptr: OutlinedAI), |
1247 | Ptr: PrivTIDAddr); |
1248 | |
1249 | // Remove redundant call to the outlined function. |
1250 | CI->eraseFromParent(); |
1251 | |
1252 | for (Instruction *I : ToBeDeleted) { |
1253 | I->eraseFromParent(); |
1254 | } |
1255 | } |
1256 | |
1257 | // Callback used to create OpenMP runtime calls to support |
1258 | // omp parallel clause for the host. |
1259 | // We need to use this callback to replace call to the OutlinedFn in OuterFn |
1260 | // by the call to the OpenMP host runtime function ( __kmpc_fork_call[_if]) |
1261 | static void |
1262 | hostParallelCallback(OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn, |
1263 | Function *OuterFn, Value *Ident, Value *IfCondition, |
1264 | Instruction *PrivTID, AllocaInst *PrivTIDAddr, |
1265 | const SmallVector<Instruction *, 4> &ToBeDeleted) { |
1266 | IRBuilder<> &Builder = OMPIRBuilder->Builder; |
1267 | FunctionCallee RTLFn; |
1268 | if (IfCondition) { |
1269 | RTLFn = |
1270 | OMPIRBuilder->getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_fork_call_if); |
1271 | } else { |
1272 | RTLFn = |
1273 | OMPIRBuilder->getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_fork_call); |
1274 | } |
1275 | if (auto *F = dyn_cast<Function>(Val: RTLFn.getCallee())) { |
1276 | if (!F->hasMetadata(KindID: LLVMContext::MD_callback)) { |
1277 | LLVMContext &Ctx = F->getContext(); |
1278 | MDBuilder MDB(Ctx); |
1279 | // Annotate the callback behavior of the __kmpc_fork_call: |
1280 | // - The callback callee is argument number 2 (microtask). |
1281 | // - The first two arguments of the callback callee are unknown (-1). |
1282 | // - All variadic arguments to the __kmpc_fork_call are passed to the |
1283 | // callback callee. |
1284 | F->addMetadata(KindID: LLVMContext::MD_callback, |
1285 | MD&: *MDNode::get(Context&: Ctx, MDs: {MDB.createCallbackEncoding( |
1286 | CalleeArgNo: 2, Arguments: {-1, -1}, |
1287 | /* VarArgsArePassed */ true)})); |
1288 | } |
1289 | } |
1290 | // Add some known attributes. |
1291 | OutlinedFn.addParamAttr(ArgNo: 0, Kind: Attribute::NoAlias); |
1292 | OutlinedFn.addParamAttr(ArgNo: 1, Kind: Attribute::NoAlias); |
1293 | OutlinedFn.addFnAttr(Kind: Attribute::NoUnwind); |
1294 | |
1295 | assert(OutlinedFn.arg_size() >= 2 && |
1296 | "Expected at least tid and bounded tid as arguments" ); |
1297 | unsigned NumCapturedVars = OutlinedFn.arg_size() - /* tid & bounded tid */ 2; |
1298 | |
1299 | CallInst *CI = cast<CallInst>(Val: OutlinedFn.user_back()); |
1300 | CI->getParent()->setName("omp_parallel" ); |
1301 | Builder.SetInsertPoint(CI); |
1302 | |
1303 | // Build call __kmpc_fork_call[_if](Ident, n, microtask, var1, .., varn); |
1304 | Value *ForkCallArgs[] = { |
1305 | Ident, Builder.getInt32(C: NumCapturedVars), |
1306 | Builder.CreateBitCast(V: &OutlinedFn, DestTy: OMPIRBuilder->ParallelTaskPtr)}; |
1307 | |
1308 | SmallVector<Value *, 16> RealArgs; |
1309 | RealArgs.append(in_start: std::begin(arr&: ForkCallArgs), in_end: std::end(arr&: ForkCallArgs)); |
1310 | if (IfCondition) { |
1311 | Value *Cond = Builder.CreateSExtOrTrunc(V: IfCondition, DestTy: OMPIRBuilder->Int32); |
1312 | RealArgs.push_back(Elt: Cond); |
1313 | } |
1314 | RealArgs.append(in_start: CI->arg_begin() + /* tid & bound tid */ 2, in_end: CI->arg_end()); |
1315 | |
1316 | // __kmpc_fork_call_if always expects a void ptr as the last argument |
1317 | // If there are no arguments, pass a null pointer. |
1318 | auto PtrTy = OMPIRBuilder->VoidPtr; |
1319 | if (IfCondition && NumCapturedVars == 0) { |
1320 | Value *NullPtrValue = Constant::getNullValue(Ty: PtrTy); |
1321 | RealArgs.push_back(Elt: NullPtrValue); |
1322 | } |
1323 | if (IfCondition && RealArgs.back()->getType() != PtrTy) |
1324 | RealArgs.back() = Builder.CreateBitCast(V: RealArgs.back(), DestTy: PtrTy); |
1325 | |
1326 | Builder.CreateCall(Callee: RTLFn, Args: RealArgs); |
1327 | |
1328 | LLVM_DEBUG(dbgs() << "With fork_call placed: " |
1329 | << *Builder.GetInsertBlock()->getParent() << "\n" ); |
1330 | |
1331 | // Initialize the local TID stack location with the argument value. |
1332 | Builder.SetInsertPoint(PrivTID); |
1333 | Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin(); |
1334 | Builder.CreateStore(Val: Builder.CreateLoad(Ty: OMPIRBuilder->Int32, Ptr: OutlinedAI), |
1335 | Ptr: PrivTIDAddr); |
1336 | |
1337 | // Remove redundant call to the outlined function. |
1338 | CI->eraseFromParent(); |
1339 | |
1340 | for (Instruction *I : ToBeDeleted) { |
1341 | I->eraseFromParent(); |
1342 | } |
1343 | } |
1344 | |
1345 | IRBuilder<>::InsertPoint OpenMPIRBuilder::createParallel( |
1346 | const LocationDescription &Loc, InsertPointTy OuterAllocaIP, |
1347 | BodyGenCallbackTy BodyGenCB, PrivatizeCallbackTy PrivCB, |
1348 | FinalizeCallbackTy FiniCB, Value *IfCondition, Value *NumThreads, |
1349 | omp::ProcBindKind ProcBind, bool IsCancellable) { |
1350 | assert(!isConflictIP(Loc.IP, OuterAllocaIP) && "IPs must not be ambiguous" ); |
1351 | |
1352 | if (!updateToLocation(Loc)) |
1353 | return Loc.IP; |
1354 | |
1355 | uint32_t SrcLocStrSize; |
1356 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1357 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
1358 | Value *ThreadID = getOrCreateThreadID(Ident); |
1359 | // If we generate code for the target device, we need to allocate |
1360 | // struct for aggregate params in the device default alloca address space. |
1361 | // OpenMP runtime requires that the params of the extracted functions are |
1362 | // passed as zero address space pointers. This flag ensures that extracted |
1363 | // function arguments are declared in zero address space |
1364 | bool ArgsInZeroAddressSpace = Config.isTargetDevice(); |
1365 | |
1366 | // Build call __kmpc_push_num_threads(&Ident, global_tid, num_threads) |
1367 | // only if we compile for host side. |
1368 | if (NumThreads && !Config.isTargetDevice()) { |
1369 | Value *Args[] = { |
1370 | Ident, ThreadID, |
1371 | Builder.CreateIntCast(V: NumThreads, DestTy: Int32, /*isSigned*/ false)}; |
1372 | Builder.CreateCall( |
1373 | Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_push_num_threads), Args); |
1374 | } |
1375 | |
1376 | if (ProcBind != OMP_PROC_BIND_default) { |
1377 | // Build call __kmpc_push_proc_bind(&Ident, global_tid, proc_bind) |
1378 | Value *Args[] = { |
1379 | Ident, ThreadID, |
1380 | ConstantInt::get(Ty: Int32, V: unsigned(ProcBind), /*isSigned=*/IsSigned: true)}; |
1381 | Builder.CreateCall( |
1382 | Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_push_proc_bind), Args); |
1383 | } |
1384 | |
1385 | BasicBlock *InsertBB = Builder.GetInsertBlock(); |
1386 | Function *OuterFn = InsertBB->getParent(); |
1387 | |
1388 | // Save the outer alloca block because the insertion iterator may get |
1389 | // invalidated and we still need this later. |
1390 | BasicBlock *OuterAllocaBlock = OuterAllocaIP.getBlock(); |
1391 | |
1392 | // Vector to remember instructions we used only during the modeling but which |
1393 | // we want to delete at the end. |
1394 | SmallVector<Instruction *, 4> ToBeDeleted; |
1395 | |
1396 | // Change the location to the outer alloca insertion point to create and |
1397 | // initialize the allocas we pass into the parallel region. |
1398 | InsertPointTy NewOuter(OuterAllocaBlock, OuterAllocaBlock->begin()); |
1399 | Builder.restoreIP(IP: NewOuter); |
1400 | AllocaInst *TIDAddrAlloca = Builder.CreateAlloca(Ty: Int32, ArraySize: nullptr, Name: "tid.addr" ); |
1401 | AllocaInst *ZeroAddrAlloca = |
1402 | Builder.CreateAlloca(Ty: Int32, ArraySize: nullptr, Name: "zero.addr" ); |
1403 | Instruction *TIDAddr = TIDAddrAlloca; |
1404 | Instruction *ZeroAddr = ZeroAddrAlloca; |
1405 | if (ArgsInZeroAddressSpace && M.getDataLayout().getAllocaAddrSpace() != 0) { |
1406 | // Add additional casts to enforce pointers in zero address space |
1407 | TIDAddr = new AddrSpaceCastInst( |
1408 | TIDAddrAlloca, PointerType ::get(C&: M.getContext(), AddressSpace: 0), "tid.addr.ascast" ); |
1409 | TIDAddr->insertAfter(InsertPos: TIDAddrAlloca); |
1410 | ToBeDeleted.push_back(Elt: TIDAddr); |
1411 | ZeroAddr = new AddrSpaceCastInst(ZeroAddrAlloca, |
1412 | PointerType ::get(C&: M.getContext(), AddressSpace: 0), |
1413 | "zero.addr.ascast" ); |
1414 | ZeroAddr->insertAfter(InsertPos: ZeroAddrAlloca); |
1415 | ToBeDeleted.push_back(Elt: ZeroAddr); |
1416 | } |
1417 | |
1418 | // We only need TIDAddr and ZeroAddr for modeling purposes to get the |
1419 | // associated arguments in the outlined function, so we delete them later. |
1420 | ToBeDeleted.push_back(Elt: TIDAddrAlloca); |
1421 | ToBeDeleted.push_back(Elt: ZeroAddrAlloca); |
1422 | |
1423 | // Create an artificial insertion point that will also ensure the blocks we |
1424 | // are about to split are not degenerated. |
1425 | auto *UI = new UnreachableInst(Builder.getContext(), InsertBB); |
1426 | |
1427 | BasicBlock *EntryBB = UI->getParent(); |
1428 | BasicBlock *PRegEntryBB = EntryBB->splitBasicBlock(I: UI, BBName: "omp.par.entry" ); |
1429 | BasicBlock *PRegBodyBB = PRegEntryBB->splitBasicBlock(I: UI, BBName: "omp.par.region" ); |
1430 | BasicBlock *PRegPreFiniBB = |
1431 | PRegBodyBB->splitBasicBlock(I: UI, BBName: "omp.par.pre_finalize" ); |
1432 | BasicBlock *PRegExitBB = PRegPreFiniBB->splitBasicBlock(I: UI, BBName: "omp.par.exit" ); |
1433 | |
1434 | auto FiniCBWrapper = [&](InsertPointTy IP) { |
1435 | // Hide "open-ended" blocks from the given FiniCB by setting the right jump |
1436 | // target to the region exit block. |
1437 | if (IP.getBlock()->end() == IP.getPoint()) { |
1438 | IRBuilder<>::InsertPointGuard IPG(Builder); |
1439 | Builder.restoreIP(IP); |
1440 | Instruction *I = Builder.CreateBr(Dest: PRegExitBB); |
1441 | IP = InsertPointTy(I->getParent(), I->getIterator()); |
1442 | } |
1443 | assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 && |
1444 | IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB && |
1445 | "Unexpected insertion point for finalization call!" ); |
1446 | return FiniCB(IP); |
1447 | }; |
1448 | |
1449 | FinalizationStack.push_back(Elt: {.FiniCB: FiniCBWrapper, .DK: OMPD_parallel, .IsCancellable: IsCancellable}); |
1450 | |
1451 | // Generate the privatization allocas in the block that will become the entry |
1452 | // of the outlined function. |
1453 | Builder.SetInsertPoint(PRegEntryBB->getTerminator()); |
1454 | InsertPointTy InnerAllocaIP = Builder.saveIP(); |
1455 | |
1456 | AllocaInst *PrivTIDAddr = |
1457 | Builder.CreateAlloca(Ty: Int32, ArraySize: nullptr, Name: "tid.addr.local" ); |
1458 | Instruction *PrivTID = Builder.CreateLoad(Ty: Int32, Ptr: PrivTIDAddr, Name: "tid" ); |
1459 | |
1460 | // Add some fake uses for OpenMP provided arguments. |
1461 | ToBeDeleted.push_back(Elt: Builder.CreateLoad(Ty: Int32, Ptr: TIDAddr, Name: "tid.addr.use" )); |
1462 | Instruction *ZeroAddrUse = |
1463 | Builder.CreateLoad(Ty: Int32, Ptr: ZeroAddr, Name: "zero.addr.use" ); |
1464 | ToBeDeleted.push_back(Elt: ZeroAddrUse); |
1465 | |
1466 | // EntryBB |
1467 | // | |
1468 | // V |
1469 | // PRegionEntryBB <- Privatization allocas are placed here. |
1470 | // | |
1471 | // V |
1472 | // PRegionBodyBB <- BodeGen is invoked here. |
1473 | // | |
1474 | // V |
1475 | // PRegPreFiniBB <- The block we will start finalization from. |
1476 | // | |
1477 | // V |
1478 | // PRegionExitBB <- A common exit to simplify block collection. |
1479 | // |
1480 | |
1481 | LLVM_DEBUG(dbgs() << "Before body codegen: " << *OuterFn << "\n" ); |
1482 | |
1483 | // Let the caller create the body. |
1484 | assert(BodyGenCB && "Expected body generation callback!" ); |
1485 | InsertPointTy CodeGenIP(PRegBodyBB, PRegBodyBB->begin()); |
1486 | BodyGenCB(InnerAllocaIP, CodeGenIP); |
1487 | |
1488 | LLVM_DEBUG(dbgs() << "After body codegen: " << *OuterFn << "\n" ); |
1489 | |
1490 | OutlineInfo OI; |
1491 | if (Config.isTargetDevice()) { |
1492 | // Generate OpenMP target specific runtime call |
1493 | OI.PostOutlineCB = [=, ToBeDeletedVec = |
1494 | std::move(ToBeDeleted)](Function &OutlinedFn) { |
1495 | targetParallelCallback(OMPIRBuilder: this, OutlinedFn, OuterFn, OuterAllocaBB: OuterAllocaBlock, Ident, |
1496 | IfCondition, NumThreads, PrivTID, PrivTIDAddr, |
1497 | ThreadID, ToBeDeleted: ToBeDeletedVec); |
1498 | }; |
1499 | } else { |
1500 | // Generate OpenMP host runtime call |
1501 | OI.PostOutlineCB = [=, ToBeDeletedVec = |
1502 | std::move(ToBeDeleted)](Function &OutlinedFn) { |
1503 | hostParallelCallback(OMPIRBuilder: this, OutlinedFn, OuterFn, Ident, IfCondition, |
1504 | PrivTID, PrivTIDAddr, ToBeDeleted: ToBeDeletedVec); |
1505 | }; |
1506 | } |
1507 | |
1508 | OI.OuterAllocaBB = OuterAllocaBlock; |
1509 | OI.EntryBB = PRegEntryBB; |
1510 | OI.ExitBB = PRegExitBB; |
1511 | |
1512 | SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet; |
1513 | SmallVector<BasicBlock *, 32> Blocks; |
1514 | OI.collectBlocks(BlockSet&: ParallelRegionBlockSet, BlockVector&: Blocks); |
1515 | |
1516 | // Ensure a single exit node for the outlined region by creating one. |
1517 | // We might have multiple incoming edges to the exit now due to finalizations, |
1518 | // e.g., cancel calls that cause the control flow to leave the region. |
1519 | BasicBlock *PRegOutlinedExitBB = PRegExitBB; |
1520 | PRegExitBB = SplitBlock(Old: PRegExitBB, SplitPt: &*PRegExitBB->getFirstInsertionPt()); |
1521 | PRegOutlinedExitBB->setName("omp.par.outlined.exit" ); |
1522 | Blocks.push_back(Elt: PRegOutlinedExitBB); |
1523 | |
1524 | CodeExtractorAnalysisCache CEAC(*OuterFn); |
1525 | CodeExtractor (Blocks, /* DominatorTree */ nullptr, |
1526 | /* AggregateArgs */ false, |
1527 | /* BlockFrequencyInfo */ nullptr, |
1528 | /* BranchProbabilityInfo */ nullptr, |
1529 | /* AssumptionCache */ nullptr, |
1530 | /* AllowVarArgs */ true, |
1531 | /* AllowAlloca */ true, |
1532 | /* AllocationBlock */ OuterAllocaBlock, |
1533 | /* Suffix */ ".omp_par" , ArgsInZeroAddressSpace); |
1534 | |
1535 | // Find inputs to, outputs from the code region. |
1536 | BasicBlock *CommonExit = nullptr; |
1537 | SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands; |
1538 | Extractor.findAllocas(CEAC, SinkCands&: SinkingCands, HoistCands&: HoistingCands, ExitBlock&: CommonExit); |
1539 | Extractor.findInputsOutputs(Inputs, Outputs, Allocas: SinkingCands); |
1540 | |
1541 | LLVM_DEBUG(dbgs() << "Before privatization: " << *OuterFn << "\n" ); |
1542 | |
1543 | FunctionCallee TIDRTLFn = |
1544 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_global_thread_num); |
1545 | |
1546 | auto PrivHelper = [&](Value &V) { |
1547 | if (&V == TIDAddr || &V == ZeroAddr) { |
1548 | OI.ExcludeArgsFromAggregate.push_back(Elt: &V); |
1549 | return; |
1550 | } |
1551 | |
1552 | SetVector<Use *> Uses; |
1553 | for (Use &U : V.uses()) |
1554 | if (auto *UserI = dyn_cast<Instruction>(Val: U.getUser())) |
1555 | if (ParallelRegionBlockSet.count(Ptr: UserI->getParent())) |
1556 | Uses.insert(X: &U); |
1557 | |
1558 | // __kmpc_fork_call expects extra arguments as pointers. If the input |
1559 | // already has a pointer type, everything is fine. Otherwise, store the |
1560 | // value onto stack and load it back inside the to-be-outlined region. This |
1561 | // will ensure only the pointer will be passed to the function. |
1562 | // FIXME: if there are more than 15 trailing arguments, they must be |
1563 | // additionally packed in a struct. |
1564 | Value *Inner = &V; |
1565 | if (!V.getType()->isPointerTy()) { |
1566 | IRBuilder<>::InsertPointGuard Guard(Builder); |
1567 | LLVM_DEBUG(llvm::dbgs() << "Forwarding input as pointer: " << V << "\n" ); |
1568 | |
1569 | Builder.restoreIP(IP: OuterAllocaIP); |
1570 | Value *Ptr = |
1571 | Builder.CreateAlloca(Ty: V.getType(), ArraySize: nullptr, Name: V.getName() + ".reloaded" ); |
1572 | |
1573 | // Store to stack at end of the block that currently branches to the entry |
1574 | // block of the to-be-outlined region. |
1575 | Builder.SetInsertPoint(TheBB: InsertBB, |
1576 | IP: InsertBB->getTerminator()->getIterator()); |
1577 | Builder.CreateStore(Val: &V, Ptr); |
1578 | |
1579 | // Load back next to allocations in the to-be-outlined region. |
1580 | Builder.restoreIP(IP: InnerAllocaIP); |
1581 | Inner = Builder.CreateLoad(Ty: V.getType(), Ptr); |
1582 | } |
1583 | |
1584 | Value *ReplacementValue = nullptr; |
1585 | CallInst *CI = dyn_cast<CallInst>(Val: &V); |
1586 | if (CI && CI->getCalledFunction() == TIDRTLFn.getCallee()) { |
1587 | ReplacementValue = PrivTID; |
1588 | } else { |
1589 | Builder.restoreIP( |
1590 | IP: PrivCB(InnerAllocaIP, Builder.saveIP(), V, *Inner, ReplacementValue)); |
1591 | InnerAllocaIP = { |
1592 | InnerAllocaIP.getBlock(), |
1593 | InnerAllocaIP.getBlock()->getTerminator()->getIterator()}; |
1594 | |
1595 | assert(ReplacementValue && |
1596 | "Expected copy/create callback to set replacement value!" ); |
1597 | if (ReplacementValue == &V) |
1598 | return; |
1599 | } |
1600 | |
1601 | for (Use *UPtr : Uses) |
1602 | UPtr->set(ReplacementValue); |
1603 | }; |
1604 | |
1605 | // Reset the inner alloca insertion as it will be used for loading the values |
1606 | // wrapped into pointers before passing them into the to-be-outlined region. |
1607 | // Configure it to insert immediately after the fake use of zero address so |
1608 | // that they are available in the generated body and so that the |
1609 | // OpenMP-related values (thread ID and zero address pointers) remain leading |
1610 | // in the argument list. |
1611 | InnerAllocaIP = IRBuilder<>::InsertPoint( |
1612 | ZeroAddrUse->getParent(), ZeroAddrUse->getNextNode()->getIterator()); |
1613 | |
1614 | // Reset the outer alloca insertion point to the entry of the relevant block |
1615 | // in case it was invalidated. |
1616 | OuterAllocaIP = IRBuilder<>::InsertPoint( |
1617 | OuterAllocaBlock, OuterAllocaBlock->getFirstInsertionPt()); |
1618 | |
1619 | for (Value *Input : Inputs) { |
1620 | LLVM_DEBUG(dbgs() << "Captured input: " << *Input << "\n" ); |
1621 | PrivHelper(*Input); |
1622 | } |
1623 | LLVM_DEBUG({ |
1624 | for (Value *Output : Outputs) |
1625 | LLVM_DEBUG(dbgs() << "Captured output: " << *Output << "\n" ); |
1626 | }); |
1627 | assert(Outputs.empty() && |
1628 | "OpenMP outlining should not produce live-out values!" ); |
1629 | |
1630 | LLVM_DEBUG(dbgs() << "After privatization: " << *OuterFn << "\n" ); |
1631 | LLVM_DEBUG({ |
1632 | for (auto *BB : Blocks) |
1633 | dbgs() << " PBR: " << BB->getName() << "\n" ; |
1634 | }); |
1635 | |
1636 | // Adjust the finalization stack, verify the adjustment, and call the |
1637 | // finalize function a last time to finalize values between the pre-fini |
1638 | // block and the exit block if we left the parallel "the normal way". |
1639 | auto FiniInfo = FinalizationStack.pop_back_val(); |
1640 | (void)FiniInfo; |
1641 | assert(FiniInfo.DK == OMPD_parallel && |
1642 | "Unexpected finalization stack state!" ); |
1643 | |
1644 | Instruction *PRegPreFiniTI = PRegPreFiniBB->getTerminator(); |
1645 | |
1646 | InsertPointTy PreFiniIP(PRegPreFiniBB, PRegPreFiniTI->getIterator()); |
1647 | FiniCB(PreFiniIP); |
1648 | |
1649 | // Register the outlined info. |
1650 | addOutlineInfo(OI: std::move(OI)); |
1651 | |
1652 | InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end()); |
1653 | UI->eraseFromParent(); |
1654 | |
1655 | return AfterIP; |
1656 | } |
1657 | |
1658 | void OpenMPIRBuilder::emitFlush(const LocationDescription &Loc) { |
1659 | // Build call void __kmpc_flush(ident_t *loc) |
1660 | uint32_t SrcLocStrSize; |
1661 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1662 | Value *Args[] = {getOrCreateIdent(SrcLocStr, SrcLocStrSize)}; |
1663 | |
1664 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_flush), Args); |
1665 | } |
1666 | |
1667 | void OpenMPIRBuilder::createFlush(const LocationDescription &Loc) { |
1668 | if (!updateToLocation(Loc)) |
1669 | return; |
1670 | emitFlush(Loc); |
1671 | } |
1672 | |
1673 | void OpenMPIRBuilder::emitTaskwaitImpl(const LocationDescription &Loc) { |
1674 | // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 |
1675 | // global_tid); |
1676 | uint32_t SrcLocStrSize; |
1677 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1678 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
1679 | Value *Args[] = {Ident, getOrCreateThreadID(Ident)}; |
1680 | |
1681 | // Ignore return result until untied tasks are supported. |
1682 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_taskwait), |
1683 | Args); |
1684 | } |
1685 | |
1686 | void OpenMPIRBuilder::createTaskwait(const LocationDescription &Loc) { |
1687 | if (!updateToLocation(Loc)) |
1688 | return; |
1689 | emitTaskwaitImpl(Loc); |
1690 | } |
1691 | |
1692 | void OpenMPIRBuilder::emitTaskyieldImpl(const LocationDescription &Loc) { |
1693 | // Build call __kmpc_omp_taskyield(loc, thread_id, 0); |
1694 | uint32_t SrcLocStrSize; |
1695 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1696 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
1697 | Constant *I32Null = ConstantInt::getNullValue(Ty: Int32); |
1698 | Value *Args[] = {Ident, getOrCreateThreadID(Ident), I32Null}; |
1699 | |
1700 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_taskyield), |
1701 | Args); |
1702 | } |
1703 | |
1704 | void OpenMPIRBuilder::createTaskyield(const LocationDescription &Loc) { |
1705 | if (!updateToLocation(Loc)) |
1706 | return; |
1707 | emitTaskyieldImpl(Loc); |
1708 | } |
1709 | |
1710 | // Processes the dependencies in Dependencies and does the following |
1711 | // - Allocates space on the stack of an array of DependInfo objects |
1712 | // - Populates each DependInfo object with relevant information of |
1713 | // the corresponding dependence. |
1714 | // - All code is inserted in the entry block of the current function. |
1715 | static Value *emitTaskDependencies( |
1716 | OpenMPIRBuilder &OMPBuilder, |
1717 | SmallVectorImpl<OpenMPIRBuilder::DependData> &Dependencies) { |
1718 | // Early return if we have no dependencies to process |
1719 | if (Dependencies.empty()) |
1720 | return nullptr; |
1721 | |
1722 | // Given a vector of DependData objects, in this function we create an |
1723 | // array on the stack that holds kmp_dep_info objects corresponding |
1724 | // to each dependency. This is then passed to the OpenMP runtime. |
1725 | // For example, if there are 'n' dependencies then the following psedo |
1726 | // code is generated. Assume the first dependence is on a variable 'a' |
1727 | // |
1728 | // \code{c} |
1729 | // DepArray = alloc(n x sizeof(kmp_depend_info); |
1730 | // idx = 0; |
1731 | // DepArray[idx].base_addr = ptrtoint(&a); |
1732 | // DepArray[idx].len = 8; |
1733 | // DepArray[idx].flags = Dep.DepKind; /*(See OMPContants.h for DepKind)*/ |
1734 | // ++idx; |
1735 | // DepArray[idx].base_addr = ...; |
1736 | // \endcode |
1737 | |
1738 | IRBuilderBase &Builder = OMPBuilder.Builder; |
1739 | Type *DependInfo = OMPBuilder.DependInfo; |
1740 | Module &M = OMPBuilder.M; |
1741 | |
1742 | Value *DepArray = nullptr; |
1743 | OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP(); |
1744 | Builder.SetInsertPoint( |
1745 | OldIP.getBlock()->getParent()->getEntryBlock().getTerminator()); |
1746 | |
1747 | Type *DepArrayTy = ArrayType::get(ElementType: DependInfo, NumElements: Dependencies.size()); |
1748 | DepArray = Builder.CreateAlloca(Ty: DepArrayTy, ArraySize: nullptr, Name: ".dep.arr.addr" ); |
1749 | |
1750 | for (const auto &[DepIdx, Dep] : enumerate(First&: Dependencies)) { |
1751 | Value *Base = |
1752 | Builder.CreateConstInBoundsGEP2_64(Ty: DepArrayTy, Ptr: DepArray, Idx0: 0, Idx1: DepIdx); |
1753 | // Store the pointer to the variable |
1754 | Value *Addr = Builder.CreateStructGEP( |
1755 | Ty: DependInfo, Ptr: Base, |
1756 | Idx: static_cast<unsigned int>(RTLDependInfoFields::BaseAddr)); |
1757 | Value *DepValPtr = Builder.CreatePtrToInt(V: Dep.DepVal, DestTy: Builder.getInt64Ty()); |
1758 | Builder.CreateStore(Val: DepValPtr, Ptr: Addr); |
1759 | // Store the size of the variable |
1760 | Value *Size = Builder.CreateStructGEP( |
1761 | Ty: DependInfo, Ptr: Base, Idx: static_cast<unsigned int>(RTLDependInfoFields::Len)); |
1762 | Builder.CreateStore( |
1763 | Val: Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: Dep.DepValueType)), |
1764 | Ptr: Size); |
1765 | // Store the dependency kind |
1766 | Value *Flags = Builder.CreateStructGEP( |
1767 | Ty: DependInfo, Ptr: Base, |
1768 | Idx: static_cast<unsigned int>(RTLDependInfoFields::Flags)); |
1769 | Builder.CreateStore( |
1770 | Val: ConstantInt::get(Ty: Builder.getInt8Ty(), |
1771 | V: static_cast<unsigned int>(Dep.DepKind)), |
1772 | Ptr: Flags); |
1773 | } |
1774 | Builder.restoreIP(IP: OldIP); |
1775 | return DepArray; |
1776 | } |
1777 | |
1778 | OpenMPIRBuilder::InsertPointTy |
1779 | OpenMPIRBuilder::createTask(const LocationDescription &Loc, |
1780 | InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB, |
1781 | bool Tied, Value *Final, Value *IfCondition, |
1782 | SmallVector<DependData> Dependencies) { |
1783 | |
1784 | if (!updateToLocation(Loc)) |
1785 | return InsertPointTy(); |
1786 | |
1787 | uint32_t SrcLocStrSize; |
1788 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
1789 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
1790 | // The current basic block is split into four basic blocks. After outlining, |
1791 | // they will be mapped as follows: |
1792 | // ``` |
1793 | // def current_fn() { |
1794 | // current_basic_block: |
1795 | // br label %task.exit |
1796 | // task.exit: |
1797 | // ; instructions after task |
1798 | // } |
1799 | // def outlined_fn() { |
1800 | // task.alloca: |
1801 | // br label %task.body |
1802 | // task.body: |
1803 | // ret void |
1804 | // } |
1805 | // ``` |
1806 | BasicBlock *TaskExitBB = splitBB(Builder, /*CreateBranch=*/true, Name: "task.exit" ); |
1807 | BasicBlock *TaskBodyBB = splitBB(Builder, /*CreateBranch=*/true, Name: "task.body" ); |
1808 | BasicBlock *TaskAllocaBB = |
1809 | splitBB(Builder, /*CreateBranch=*/true, Name: "task.alloca" ); |
1810 | |
1811 | InsertPointTy TaskAllocaIP = |
1812 | InsertPointTy(TaskAllocaBB, TaskAllocaBB->begin()); |
1813 | InsertPointTy TaskBodyIP = InsertPointTy(TaskBodyBB, TaskBodyBB->begin()); |
1814 | BodyGenCB(TaskAllocaIP, TaskBodyIP); |
1815 | |
1816 | OutlineInfo OI; |
1817 | OI.EntryBB = TaskAllocaBB; |
1818 | OI.OuterAllocaBB = AllocaIP.getBlock(); |
1819 | OI.ExitBB = TaskExitBB; |
1820 | |
1821 | // Add the thread ID argument. |
1822 | SmallVector<Instruction *, 4> ToBeDeleted; |
1823 | OI.ExcludeArgsFromAggregate.push_back(Elt: createFakeIntVal( |
1824 | Builder, OuterAllocaIP: AllocaIP, ToBeDeleted, InnerAllocaIP: TaskAllocaIP, Name: "global.tid" , AsPtr: false)); |
1825 | |
1826 | OI.PostOutlineCB = [this, Ident, Tied, Final, IfCondition, Dependencies, |
1827 | TaskAllocaBB, ToBeDeleted](Function &OutlinedFn) mutable { |
1828 | // Replace the Stale CI by appropriate RTL function call. |
1829 | assert(OutlinedFn.getNumUses() == 1 && |
1830 | "there must be a single user for the outlined function" ); |
1831 | CallInst *StaleCI = cast<CallInst>(Val: OutlinedFn.user_back()); |
1832 | |
1833 | // HasShareds is true if any variables are captured in the outlined region, |
1834 | // false otherwise. |
1835 | bool HasShareds = StaleCI->arg_size() > 1; |
1836 | Builder.SetInsertPoint(StaleCI); |
1837 | |
1838 | // Gather the arguments for emitting the runtime call for |
1839 | // @__kmpc_omp_task_alloc |
1840 | Function *TaskAllocFn = |
1841 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_alloc); |
1842 | |
1843 | // Arguments - `loc_ref` (Ident) and `gtid` (ThreadID) |
1844 | // call. |
1845 | Value *ThreadID = getOrCreateThreadID(Ident); |
1846 | |
1847 | // Argument - `flags` |
1848 | // Task is tied iff (Flags & 1) == 1. |
1849 | // Task is untied iff (Flags & 1) == 0. |
1850 | // Task is final iff (Flags & 2) == 2. |
1851 | // Task is not final iff (Flags & 2) == 0. |
1852 | // TODO: Handle the other flags. |
1853 | Value *Flags = Builder.getInt32(C: Tied); |
1854 | if (Final) { |
1855 | Value *FinalFlag = |
1856 | Builder.CreateSelect(C: Final, True: Builder.getInt32(C: 2), False: Builder.getInt32(C: 0)); |
1857 | Flags = Builder.CreateOr(LHS: FinalFlag, RHS: Flags); |
1858 | } |
1859 | |
1860 | // Argument - `sizeof_kmp_task_t` (TaskSize) |
1861 | // Tasksize refers to the size in bytes of kmp_task_t data structure |
1862 | // including private vars accessed in task. |
1863 | // TODO: add kmp_task_t_with_privates (privates) |
1864 | Value *TaskSize = Builder.getInt64( |
1865 | C: divideCeil(Numerator: M.getDataLayout().getTypeSizeInBits(Ty: Task), Denominator: 8)); |
1866 | |
1867 | // Argument - `sizeof_shareds` (SharedsSize) |
1868 | // SharedsSize refers to the shareds array size in the kmp_task_t data |
1869 | // structure. |
1870 | Value *SharedsSize = Builder.getInt64(C: 0); |
1871 | if (HasShareds) { |
1872 | AllocaInst *ArgStructAlloca = |
1873 | dyn_cast<AllocaInst>(Val: StaleCI->getArgOperand(i: 1)); |
1874 | assert(ArgStructAlloca && |
1875 | "Unable to find the alloca instruction corresponding to arguments " |
1876 | "for extracted function" ); |
1877 | StructType *ArgStructType = |
1878 | dyn_cast<StructType>(Val: ArgStructAlloca->getAllocatedType()); |
1879 | assert(ArgStructType && "Unable to find struct type corresponding to " |
1880 | "arguments for extracted function" ); |
1881 | SharedsSize = |
1882 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: ArgStructType)); |
1883 | } |
1884 | // Emit the @__kmpc_omp_task_alloc runtime call |
1885 | // The runtime call returns a pointer to an area where the task captured |
1886 | // variables must be copied before the task is run (TaskData) |
1887 | CallInst *TaskData = Builder.CreateCall( |
1888 | Callee: TaskAllocFn, Args: {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags, |
1889 | /*sizeof_task=*/TaskSize, /*sizeof_shared=*/SharedsSize, |
1890 | /*task_func=*/&OutlinedFn}); |
1891 | |
1892 | // Copy the arguments for outlined function |
1893 | if (HasShareds) { |
1894 | Value *Shareds = StaleCI->getArgOperand(i: 1); |
1895 | Align Alignment = TaskData->getPointerAlignment(DL: M.getDataLayout()); |
1896 | Value *TaskShareds = Builder.CreateLoad(Ty: VoidPtr, Ptr: TaskData); |
1897 | Builder.CreateMemCpy(Dst: TaskShareds, DstAlign: Alignment, Src: Shareds, SrcAlign: Alignment, |
1898 | Size: SharedsSize); |
1899 | } |
1900 | |
1901 | Value *DepArray = nullptr; |
1902 | if (Dependencies.size()) { |
1903 | InsertPointTy OldIP = Builder.saveIP(); |
1904 | Builder.SetInsertPoint( |
1905 | &OldIP.getBlock()->getParent()->getEntryBlock().back()); |
1906 | |
1907 | Type *DepArrayTy = ArrayType::get(ElementType: DependInfo, NumElements: Dependencies.size()); |
1908 | DepArray = Builder.CreateAlloca(Ty: DepArrayTy, ArraySize: nullptr, Name: ".dep.arr.addr" ); |
1909 | |
1910 | unsigned P = 0; |
1911 | for (const DependData &Dep : Dependencies) { |
1912 | Value *Base = |
1913 | Builder.CreateConstInBoundsGEP2_64(Ty: DepArrayTy, Ptr: DepArray, Idx0: 0, Idx1: P); |
1914 | // Store the pointer to the variable |
1915 | Value *Addr = Builder.CreateStructGEP( |
1916 | Ty: DependInfo, Ptr: Base, |
1917 | Idx: static_cast<unsigned int>(RTLDependInfoFields::BaseAddr)); |
1918 | Value *DepValPtr = |
1919 | Builder.CreatePtrToInt(V: Dep.DepVal, DestTy: Builder.getInt64Ty()); |
1920 | Builder.CreateStore(Val: DepValPtr, Ptr: Addr); |
1921 | // Store the size of the variable |
1922 | Value *Size = Builder.CreateStructGEP( |
1923 | Ty: DependInfo, Ptr: Base, |
1924 | Idx: static_cast<unsigned int>(RTLDependInfoFields::Len)); |
1925 | Builder.CreateStore(Val: Builder.getInt64(C: M.getDataLayout().getTypeStoreSize( |
1926 | Ty: Dep.DepValueType)), |
1927 | Ptr: Size); |
1928 | // Store the dependency kind |
1929 | Value *Flags = Builder.CreateStructGEP( |
1930 | Ty: DependInfo, Ptr: Base, |
1931 | Idx: static_cast<unsigned int>(RTLDependInfoFields::Flags)); |
1932 | Builder.CreateStore( |
1933 | Val: ConstantInt::get(Ty: Builder.getInt8Ty(), |
1934 | V: static_cast<unsigned int>(Dep.DepKind)), |
1935 | Ptr: Flags); |
1936 | ++P; |
1937 | } |
1938 | |
1939 | Builder.restoreIP(IP: OldIP); |
1940 | } |
1941 | |
1942 | // In the presence of the `if` clause, the following IR is generated: |
1943 | // ... |
1944 | // %data = call @__kmpc_omp_task_alloc(...) |
1945 | // br i1 %if_condition, label %then, label %else |
1946 | // then: |
1947 | // call @__kmpc_omp_task(...) |
1948 | // br label %exit |
1949 | // else: |
1950 | // ;; Wait for resolution of dependencies, if any, before |
1951 | // ;; beginning the task |
1952 | // call @__kmpc_omp_wait_deps(...) |
1953 | // call @__kmpc_omp_task_begin_if0(...) |
1954 | // call @outlined_fn(...) |
1955 | // call @__kmpc_omp_task_complete_if0(...) |
1956 | // br label %exit |
1957 | // exit: |
1958 | // ... |
1959 | if (IfCondition) { |
1960 | // `SplitBlockAndInsertIfThenElse` requires the block to have a |
1961 | // terminator. |
1962 | splitBB(Builder, /*CreateBranch=*/true, Name: "if.end" ); |
1963 | Instruction *IfTerminator = |
1964 | Builder.GetInsertPoint()->getParent()->getTerminator(); |
1965 | Instruction *ThenTI = IfTerminator, *ElseTI = nullptr; |
1966 | Builder.SetInsertPoint(IfTerminator); |
1967 | SplitBlockAndInsertIfThenElse(Cond: IfCondition, SplitBefore: IfTerminator, ThenTerm: &ThenTI, |
1968 | ElseTerm: &ElseTI); |
1969 | Builder.SetInsertPoint(ElseTI); |
1970 | |
1971 | if (Dependencies.size()) { |
1972 | Function *TaskWaitFn = |
1973 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_wait_deps); |
1974 | Builder.CreateCall( |
1975 | Callee: TaskWaitFn, |
1976 | Args: {Ident, ThreadID, Builder.getInt32(C: Dependencies.size()), DepArray, |
1977 | ConstantInt::get(Ty: Builder.getInt32Ty(), V: 0), |
1978 | ConstantPointerNull::get(T: PointerType::getUnqual(C&: M.getContext()))}); |
1979 | } |
1980 | Function *TaskBeginFn = |
1981 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_begin_if0); |
1982 | Function *TaskCompleteFn = |
1983 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_complete_if0); |
1984 | Builder.CreateCall(Callee: TaskBeginFn, Args: {Ident, ThreadID, TaskData}); |
1985 | CallInst *CI = nullptr; |
1986 | if (HasShareds) |
1987 | CI = Builder.CreateCall(Callee: &OutlinedFn, Args: {ThreadID, TaskData}); |
1988 | else |
1989 | CI = Builder.CreateCall(Callee: &OutlinedFn, Args: {ThreadID}); |
1990 | CI->setDebugLoc(StaleCI->getDebugLoc()); |
1991 | Builder.CreateCall(Callee: TaskCompleteFn, Args: {Ident, ThreadID, TaskData}); |
1992 | Builder.SetInsertPoint(ThenTI); |
1993 | } |
1994 | |
1995 | if (Dependencies.size()) { |
1996 | Function *TaskFn = |
1997 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_with_deps); |
1998 | Builder.CreateCall( |
1999 | Callee: TaskFn, |
2000 | Args: {Ident, ThreadID, TaskData, Builder.getInt32(C: Dependencies.size()), |
2001 | DepArray, ConstantInt::get(Ty: Builder.getInt32Ty(), V: 0), |
2002 | ConstantPointerNull::get(T: PointerType::getUnqual(C&: M.getContext()))}); |
2003 | |
2004 | } else { |
2005 | // Emit the @__kmpc_omp_task runtime call to spawn the task |
2006 | Function *TaskFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task); |
2007 | Builder.CreateCall(Callee: TaskFn, Args: {Ident, ThreadID, TaskData}); |
2008 | } |
2009 | |
2010 | StaleCI->eraseFromParent(); |
2011 | |
2012 | Builder.SetInsertPoint(TheBB: TaskAllocaBB, IP: TaskAllocaBB->begin()); |
2013 | if (HasShareds) { |
2014 | LoadInst *Shareds = Builder.CreateLoad(Ty: VoidPtr, Ptr: OutlinedFn.getArg(i: 1)); |
2015 | OutlinedFn.getArg(i: 1)->replaceUsesWithIf( |
2016 | New: Shareds, ShouldReplace: [Shareds](Use &U) { return U.getUser() != Shareds; }); |
2017 | } |
2018 | |
2019 | llvm::for_each(Range: llvm::reverse(C&: ToBeDeleted), |
2020 | F: [](Instruction *I) { I->eraseFromParent(); }); |
2021 | }; |
2022 | |
2023 | addOutlineInfo(OI: std::move(OI)); |
2024 | Builder.SetInsertPoint(TheBB: TaskExitBB, IP: TaskExitBB->begin()); |
2025 | |
2026 | return Builder.saveIP(); |
2027 | } |
2028 | |
2029 | OpenMPIRBuilder::InsertPointTy |
2030 | OpenMPIRBuilder::createTaskgroup(const LocationDescription &Loc, |
2031 | InsertPointTy AllocaIP, |
2032 | BodyGenCallbackTy BodyGenCB) { |
2033 | if (!updateToLocation(Loc)) |
2034 | return InsertPointTy(); |
2035 | |
2036 | uint32_t SrcLocStrSize; |
2037 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
2038 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
2039 | Value *ThreadID = getOrCreateThreadID(Ident); |
2040 | |
2041 | // Emit the @__kmpc_taskgroup runtime call to start the taskgroup |
2042 | Function *TaskgroupFn = |
2043 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_taskgroup); |
2044 | Builder.CreateCall(Callee: TaskgroupFn, Args: {Ident, ThreadID}); |
2045 | |
2046 | BasicBlock *TaskgroupExitBB = splitBB(Builder, CreateBranch: true, Name: "taskgroup.exit" ); |
2047 | BodyGenCB(AllocaIP, Builder.saveIP()); |
2048 | |
2049 | Builder.SetInsertPoint(TaskgroupExitBB); |
2050 | // Emit the @__kmpc_end_taskgroup runtime call to end the taskgroup |
2051 | Function *EndTaskgroupFn = |
2052 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_taskgroup); |
2053 | Builder.CreateCall(Callee: EndTaskgroupFn, Args: {Ident, ThreadID}); |
2054 | |
2055 | return Builder.saveIP(); |
2056 | } |
2057 | |
2058 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSections( |
2059 | const LocationDescription &Loc, InsertPointTy AllocaIP, |
2060 | ArrayRef<StorableBodyGenCallbackTy> SectionCBs, PrivatizeCallbackTy PrivCB, |
2061 | FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait) { |
2062 | assert(!isConflictIP(AllocaIP, Loc.IP) && "Dedicated IP allocas required" ); |
2063 | |
2064 | if (!updateToLocation(Loc)) |
2065 | return Loc.IP; |
2066 | |
2067 | auto FiniCBWrapper = [&](InsertPointTy IP) { |
2068 | if (IP.getBlock()->end() != IP.getPoint()) |
2069 | return FiniCB(IP); |
2070 | // This must be done otherwise any nested constructs using FinalizeOMPRegion |
2071 | // will fail because that function requires the Finalization Basic Block to |
2072 | // have a terminator, which is already removed by EmitOMPRegionBody. |
2073 | // IP is currently at cancelation block. |
2074 | // We need to backtrack to the condition block to fetch |
2075 | // the exit block and create a branch from cancelation |
2076 | // to exit block. |
2077 | IRBuilder<>::InsertPointGuard IPG(Builder); |
2078 | Builder.restoreIP(IP); |
2079 | auto *CaseBB = IP.getBlock()->getSinglePredecessor(); |
2080 | auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor(); |
2081 | auto *ExitBB = CondBB->getTerminator()->getSuccessor(Idx: 1); |
2082 | Instruction *I = Builder.CreateBr(Dest: ExitBB); |
2083 | IP = InsertPointTy(I->getParent(), I->getIterator()); |
2084 | return FiniCB(IP); |
2085 | }; |
2086 | |
2087 | FinalizationStack.push_back(Elt: {.FiniCB: FiniCBWrapper, .DK: OMPD_sections, .IsCancellable: IsCancellable}); |
2088 | |
2089 | // Each section is emitted as a switch case |
2090 | // Each finalization callback is handled from clang.EmitOMPSectionDirective() |
2091 | // -> OMP.createSection() which generates the IR for each section |
2092 | // Iterate through all sections and emit a switch construct: |
2093 | // switch (IV) { |
2094 | // case 0: |
2095 | // <SectionStmt[0]>; |
2096 | // break; |
2097 | // ... |
2098 | // case <NumSection> - 1: |
2099 | // <SectionStmt[<NumSection> - 1]>; |
2100 | // break; |
2101 | // } |
2102 | // ... |
2103 | // section_loop.after: |
2104 | // <FiniCB>; |
2105 | auto LoopBodyGenCB = [&](InsertPointTy CodeGenIP, Value *IndVar) { |
2106 | Builder.restoreIP(IP: CodeGenIP); |
2107 | BasicBlock *Continue = |
2108 | splitBBWithSuffix(Builder, /*CreateBranch=*/false, Suffix: ".sections.after" ); |
2109 | Function *CurFn = Continue->getParent(); |
2110 | SwitchInst *SwitchStmt = Builder.CreateSwitch(V: IndVar, Dest: Continue); |
2111 | |
2112 | unsigned CaseNumber = 0; |
2113 | for (auto SectionCB : SectionCBs) { |
2114 | BasicBlock *CaseBB = BasicBlock::Create( |
2115 | Context&: M.getContext(), Name: "omp_section_loop.body.case" , Parent: CurFn, InsertBefore: Continue); |
2116 | SwitchStmt->addCase(OnVal: Builder.getInt32(C: CaseNumber), Dest: CaseBB); |
2117 | Builder.SetInsertPoint(CaseBB); |
2118 | BranchInst *CaseEndBr = Builder.CreateBr(Dest: Continue); |
2119 | SectionCB(InsertPointTy(), |
2120 | {CaseEndBr->getParent(), CaseEndBr->getIterator()}); |
2121 | CaseNumber++; |
2122 | } |
2123 | // remove the existing terminator from body BB since there can be no |
2124 | // terminators after switch/case |
2125 | }; |
2126 | // Loop body ends here |
2127 | // LowerBound, UpperBound, and STride for createCanonicalLoop |
2128 | Type *I32Ty = Type::getInt32Ty(C&: M.getContext()); |
2129 | Value *LB = ConstantInt::get(Ty: I32Ty, V: 0); |
2130 | Value *UB = ConstantInt::get(Ty: I32Ty, V: SectionCBs.size()); |
2131 | Value *ST = ConstantInt::get(Ty: I32Ty, V: 1); |
2132 | llvm::CanonicalLoopInfo *LoopInfo = createCanonicalLoop( |
2133 | Loc, BodyGenCB: LoopBodyGenCB, Start: LB, Stop: UB, Step: ST, IsSigned: true, InclusiveStop: false, ComputeIP: AllocaIP, Name: "section_loop" ); |
2134 | InsertPointTy AfterIP = |
2135 | applyStaticWorkshareLoop(DL: Loc.DL, CLI: LoopInfo, AllocaIP, NeedsBarrier: !IsNowait); |
2136 | |
2137 | // Apply the finalization callback in LoopAfterBB |
2138 | auto FiniInfo = FinalizationStack.pop_back_val(); |
2139 | assert(FiniInfo.DK == OMPD_sections && |
2140 | "Unexpected finalization stack state!" ); |
2141 | if (FinalizeCallbackTy &CB = FiniInfo.FiniCB) { |
2142 | Builder.restoreIP(IP: AfterIP); |
2143 | BasicBlock *FiniBB = |
2144 | splitBBWithSuffix(Builder, /*CreateBranch=*/true, Suffix: "sections.fini" ); |
2145 | CB(Builder.saveIP()); |
2146 | AfterIP = {FiniBB, FiniBB->begin()}; |
2147 | } |
2148 | |
2149 | return AfterIP; |
2150 | } |
2151 | |
2152 | OpenMPIRBuilder::InsertPointTy |
2153 | OpenMPIRBuilder::createSection(const LocationDescription &Loc, |
2154 | BodyGenCallbackTy BodyGenCB, |
2155 | FinalizeCallbackTy FiniCB) { |
2156 | if (!updateToLocation(Loc)) |
2157 | return Loc.IP; |
2158 | |
2159 | auto FiniCBWrapper = [&](InsertPointTy IP) { |
2160 | if (IP.getBlock()->end() != IP.getPoint()) |
2161 | return FiniCB(IP); |
2162 | // This must be done otherwise any nested constructs using FinalizeOMPRegion |
2163 | // will fail because that function requires the Finalization Basic Block to |
2164 | // have a terminator, which is already removed by EmitOMPRegionBody. |
2165 | // IP is currently at cancelation block. |
2166 | // We need to backtrack to the condition block to fetch |
2167 | // the exit block and create a branch from cancelation |
2168 | // to exit block. |
2169 | IRBuilder<>::InsertPointGuard IPG(Builder); |
2170 | Builder.restoreIP(IP); |
2171 | auto *CaseBB = Loc.IP.getBlock(); |
2172 | auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor(); |
2173 | auto *ExitBB = CondBB->getTerminator()->getSuccessor(Idx: 1); |
2174 | Instruction *I = Builder.CreateBr(Dest: ExitBB); |
2175 | IP = InsertPointTy(I->getParent(), I->getIterator()); |
2176 | return FiniCB(IP); |
2177 | }; |
2178 | |
2179 | Directive OMPD = Directive::OMPD_sections; |
2180 | // Since we are using Finalization Callback here, HasFinalize |
2181 | // and IsCancellable have to be true |
2182 | return EmitOMPInlinedRegion(OMPD, EntryCall: nullptr, ExitCall: nullptr, BodyGenCB, FiniCB: FiniCBWrapper, |
2183 | /*Conditional*/ false, /*hasFinalize*/ HasFinalize: true, |
2184 | /*IsCancellable*/ true); |
2185 | } |
2186 | |
2187 | static OpenMPIRBuilder::InsertPointTy getInsertPointAfterInstr(Instruction *I) { |
2188 | BasicBlock::iterator IT(I); |
2189 | IT++; |
2190 | return OpenMPIRBuilder::InsertPointTy(I->getParent(), IT); |
2191 | } |
2192 | |
2193 | void OpenMPIRBuilder::emitUsed(StringRef Name, |
2194 | std::vector<WeakTrackingVH> &List) { |
2195 | if (List.empty()) |
2196 | return; |
2197 | |
2198 | // Convert List to what ConstantArray needs. |
2199 | SmallVector<Constant *, 8> UsedArray; |
2200 | UsedArray.resize(N: List.size()); |
2201 | for (unsigned I = 0, E = List.size(); I != E; ++I) |
2202 | UsedArray[I] = ConstantExpr::getPointerBitCastOrAddrSpaceCast( |
2203 | C: cast<Constant>(Val: &*List[I]), Ty: Builder.getPtrTy()); |
2204 | |
2205 | if (UsedArray.empty()) |
2206 | return; |
2207 | ArrayType *ATy = ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: UsedArray.size()); |
2208 | |
2209 | auto *GV = new GlobalVariable(M, ATy, false, GlobalValue::AppendingLinkage, |
2210 | ConstantArray::get(T: ATy, V: UsedArray), Name); |
2211 | |
2212 | GV->setSection("llvm.metadata" ); |
2213 | } |
2214 | |
2215 | Value *OpenMPIRBuilder::getGPUThreadID() { |
2216 | return Builder.CreateCall( |
2217 | Callee: getOrCreateRuntimeFunction(M, |
2218 | FnID: OMPRTL___kmpc_get_hardware_thread_id_in_block), |
2219 | Args: {}); |
2220 | } |
2221 | |
2222 | Value *OpenMPIRBuilder::getGPUWarpSize() { |
2223 | return Builder.CreateCall( |
2224 | Callee: getOrCreateRuntimeFunction(M, FnID: OMPRTL___kmpc_get_warp_size), Args: {}); |
2225 | } |
2226 | |
2227 | Value *OpenMPIRBuilder::getNVPTXWarpID() { |
2228 | unsigned LaneIDBits = Log2_32(Value: Config.getGridValue().GV_Warp_Size); |
2229 | return Builder.CreateAShr(LHS: getGPUThreadID(), RHS: LaneIDBits, Name: "nvptx_warp_id" ); |
2230 | } |
2231 | |
2232 | Value *OpenMPIRBuilder::getNVPTXLaneID() { |
2233 | unsigned LaneIDBits = Log2_32(Value: Config.getGridValue().GV_Warp_Size); |
2234 | assert(LaneIDBits < 32 && "Invalid LaneIDBits size in NVPTX device." ); |
2235 | unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); |
2236 | return Builder.CreateAnd(LHS: getGPUThreadID(), RHS: Builder.getInt32(C: LaneIDMask), |
2237 | Name: "nvptx_lane_id" ); |
2238 | } |
2239 | |
2240 | Value *OpenMPIRBuilder::castValueToType(InsertPointTy AllocaIP, Value *From, |
2241 | Type *ToType) { |
2242 | Type *FromType = From->getType(); |
2243 | uint64_t FromSize = M.getDataLayout().getTypeStoreSize(Ty: FromType); |
2244 | uint64_t ToSize = M.getDataLayout().getTypeStoreSize(Ty: ToType); |
2245 | assert(FromSize > 0 && "From size must be greater than zero" ); |
2246 | assert(ToSize > 0 && "To size must be greater than zero" ); |
2247 | if (FromType == ToType) |
2248 | return From; |
2249 | if (FromSize == ToSize) |
2250 | return Builder.CreateBitCast(V: From, DestTy: ToType); |
2251 | if (ToType->isIntegerTy() && FromType->isIntegerTy()) |
2252 | return Builder.CreateIntCast(V: From, DestTy: ToType, /*isSigned*/ true); |
2253 | InsertPointTy SaveIP = Builder.saveIP(); |
2254 | Builder.restoreIP(IP: AllocaIP); |
2255 | Value *CastItem = Builder.CreateAlloca(Ty: ToType); |
2256 | Builder.restoreIP(IP: SaveIP); |
2257 | |
2258 | Value *ValCastItem = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2259 | V: CastItem, DestTy: FromType->getPointerTo()); |
2260 | Builder.CreateStore(Val: From, Ptr: ValCastItem); |
2261 | return Builder.CreateLoad(Ty: ToType, Ptr: CastItem); |
2262 | } |
2263 | |
2264 | Value *OpenMPIRBuilder::createRuntimeShuffleFunction(InsertPointTy AllocaIP, |
2265 | Value *Element, |
2266 | Type *ElementType, |
2267 | Value *Offset) { |
2268 | uint64_t Size = M.getDataLayout().getTypeStoreSize(Ty: ElementType); |
2269 | assert(Size <= 8 && "Unsupported bitwidth in shuffle instruction" ); |
2270 | |
2271 | // Cast all types to 32- or 64-bit values before calling shuffle routines. |
2272 | Type *CastTy = Builder.getIntNTy(N: Size <= 4 ? 32 : 64); |
2273 | Value *ElemCast = castValueToType(AllocaIP, From: Element, ToType: CastTy); |
2274 | Value *WarpSize = |
2275 | Builder.CreateIntCast(V: getGPUWarpSize(), DestTy: Builder.getInt16Ty(), isSigned: true); |
2276 | Function *ShuffleFunc = getOrCreateRuntimeFunctionPtr( |
2277 | FnID: Size <= 4 ? RuntimeFunction::OMPRTL___kmpc_shuffle_int32 |
2278 | : RuntimeFunction::OMPRTL___kmpc_shuffle_int64); |
2279 | Value *WarpSizeCast = |
2280 | Builder.CreateIntCast(V: WarpSize, DestTy: Builder.getInt16Ty(), /*isSigned=*/true); |
2281 | Value *ShuffleCall = |
2282 | Builder.CreateCall(Callee: ShuffleFunc, Args: {ElemCast, Offset, WarpSizeCast}); |
2283 | return castValueToType(AllocaIP, From: ShuffleCall, ToType: CastTy); |
2284 | } |
2285 | |
2286 | void OpenMPIRBuilder::shuffleAndStore(InsertPointTy AllocaIP, Value *SrcAddr, |
2287 | Value *DstAddr, Type *ElemType, |
2288 | Value *Offset, Type *ReductionArrayTy) { |
2289 | uint64_t Size = M.getDataLayout().getTypeStoreSize(Ty: ElemType); |
2290 | // Create the loop over the big sized data. |
2291 | // ptr = (void*)Elem; |
2292 | // ptrEnd = (void*) Elem + 1; |
2293 | // Step = 8; |
2294 | // while (ptr + Step < ptrEnd) |
2295 | // shuffle((int64_t)*ptr); |
2296 | // Step = 4; |
2297 | // while (ptr + Step < ptrEnd) |
2298 | // shuffle((int32_t)*ptr); |
2299 | // ... |
2300 | Type *IndexTy = Builder.getIndexTy( |
2301 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
2302 | Value *ElemPtr = DstAddr; |
2303 | Value *Ptr = SrcAddr; |
2304 | for (unsigned IntSize = 8; IntSize >= 1; IntSize /= 2) { |
2305 | if (Size < IntSize) |
2306 | continue; |
2307 | Type *IntType = Builder.getIntNTy(N: IntSize * 8); |
2308 | Ptr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2309 | V: Ptr, DestTy: IntType->getPointerTo(), Name: Ptr->getName() + ".ascast" ); |
2310 | Value *SrcAddrGEP = |
2311 | Builder.CreateGEP(Ty: ElemType, Ptr: SrcAddr, IdxList: {ConstantInt::get(Ty: IndexTy, V: 1)}); |
2312 | ElemPtr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2313 | V: ElemPtr, DestTy: IntType->getPointerTo(), Name: ElemPtr->getName() + ".ascast" ); |
2314 | |
2315 | Function *CurFunc = Builder.GetInsertBlock()->getParent(); |
2316 | if ((Size / IntSize) > 1) { |
2317 | Value *PtrEnd = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2318 | V: SrcAddrGEP, DestTy: Builder.getPtrTy()); |
2319 | BasicBlock *PreCondBB = |
2320 | BasicBlock::Create(Context&: M.getContext(), Name: ".shuffle.pre_cond" ); |
2321 | BasicBlock *ThenBB = BasicBlock::Create(Context&: M.getContext(), Name: ".shuffle.then" ); |
2322 | BasicBlock *ExitBB = BasicBlock::Create(Context&: M.getContext(), Name: ".shuffle.exit" ); |
2323 | BasicBlock *CurrentBB = Builder.GetInsertBlock(); |
2324 | emitBlock(BB: PreCondBB, CurFn: CurFunc); |
2325 | PHINode *PhiSrc = |
2326 | Builder.CreatePHI(Ty: Ptr->getType(), /*NumReservedValues=*/2); |
2327 | PhiSrc->addIncoming(V: Ptr, BB: CurrentBB); |
2328 | PHINode *PhiDest = |
2329 | Builder.CreatePHI(Ty: ElemPtr->getType(), /*NumReservedValues=*/2); |
2330 | PhiDest->addIncoming(V: ElemPtr, BB: CurrentBB); |
2331 | Ptr = PhiSrc; |
2332 | ElemPtr = PhiDest; |
2333 | Value *PtrDiff = Builder.CreatePtrDiff( |
2334 | ElemTy: Builder.getInt8Ty(), LHS: PtrEnd, |
2335 | RHS: Builder.CreatePointerBitCastOrAddrSpaceCast(V: Ptr, DestTy: Builder.getPtrTy())); |
2336 | Builder.CreateCondBr( |
2337 | Cond: Builder.CreateICmpSGT(LHS: PtrDiff, RHS: Builder.getInt64(C: IntSize - 1)), True: ThenBB, |
2338 | False: ExitBB); |
2339 | emitBlock(BB: ThenBB, CurFn: CurFunc); |
2340 | Value *Res = createRuntimeShuffleFunction( |
2341 | AllocaIP, |
2342 | Element: Builder.CreateAlignedLoad( |
2343 | Ty: IntType, Ptr, Align: M.getDataLayout().getPrefTypeAlign(Ty: ElemType)), |
2344 | ElementType: IntType, Offset); |
2345 | Builder.CreateAlignedStore(Val: Res, Ptr: ElemPtr, |
2346 | Align: M.getDataLayout().getPrefTypeAlign(Ty: ElemType)); |
2347 | Value *LocalPtr = |
2348 | Builder.CreateGEP(Ty: IntType, Ptr, IdxList: {ConstantInt::get(Ty: IndexTy, V: 1)}); |
2349 | Value *LocalElemPtr = |
2350 | Builder.CreateGEP(Ty: IntType, Ptr: ElemPtr, IdxList: {ConstantInt::get(Ty: IndexTy, V: 1)}); |
2351 | PhiSrc->addIncoming(V: LocalPtr, BB: ThenBB); |
2352 | PhiDest->addIncoming(V: LocalElemPtr, BB: ThenBB); |
2353 | emitBranch(Target: PreCondBB); |
2354 | emitBlock(BB: ExitBB, CurFn: CurFunc); |
2355 | } else { |
2356 | Value *Res = createRuntimeShuffleFunction( |
2357 | AllocaIP, Element: Builder.CreateLoad(Ty: IntType, Ptr), ElementType: IntType, Offset); |
2358 | if (ElemType->isIntegerTy() && ElemType->getScalarSizeInBits() < |
2359 | Res->getType()->getScalarSizeInBits()) |
2360 | Res = Builder.CreateTrunc(V: Res, DestTy: ElemType); |
2361 | Builder.CreateStore(Val: Res, Ptr: ElemPtr); |
2362 | Ptr = Builder.CreateGEP(Ty: IntType, Ptr, IdxList: {ConstantInt::get(Ty: IndexTy, V: 1)}); |
2363 | ElemPtr = |
2364 | Builder.CreateGEP(Ty: IntType, Ptr: ElemPtr, IdxList: {ConstantInt::get(Ty: IndexTy, V: 1)}); |
2365 | } |
2366 | Size = Size % IntSize; |
2367 | } |
2368 | } |
2369 | |
2370 | void OpenMPIRBuilder::emitReductionListCopy( |
2371 | InsertPointTy AllocaIP, CopyAction Action, Type *ReductionArrayTy, |
2372 | ArrayRef<ReductionInfo> ReductionInfos, Value *SrcBase, Value *DestBase, |
2373 | CopyOptionsTy CopyOptions) { |
2374 | Type *IndexTy = Builder.getIndexTy( |
2375 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
2376 | Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; |
2377 | |
2378 | // Iterates, element-by-element, through the source Reduce list and |
2379 | // make a copy. |
2380 | for (auto En : enumerate(First&: ReductionInfos)) { |
2381 | const ReductionInfo &RI = En.value(); |
2382 | Value *SrcElementAddr = nullptr; |
2383 | Value *DestElementAddr = nullptr; |
2384 | Value *DestElementPtrAddr = nullptr; |
2385 | // Should we shuffle in an element from a remote lane? |
2386 | bool ShuffleInElement = false; |
2387 | // Set to true to update the pointer in the dest Reduce list to a |
2388 | // newly created element. |
2389 | bool UpdateDestListPtr = false; |
2390 | |
2391 | // Step 1.1: Get the address for the src element in the Reduce list. |
2392 | Value *SrcElementPtrAddr = Builder.CreateInBoundsGEP( |
2393 | Ty: ReductionArrayTy, Ptr: SrcBase, |
2394 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
2395 | SrcElementAddr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: SrcElementPtrAddr); |
2396 | |
2397 | // Step 1.2: Create a temporary to store the element in the destination |
2398 | // Reduce list. |
2399 | DestElementPtrAddr = Builder.CreateInBoundsGEP( |
2400 | Ty: ReductionArrayTy, Ptr: DestBase, |
2401 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
2402 | switch (Action) { |
2403 | case CopyAction::RemoteLaneToThread: { |
2404 | InsertPointTy CurIP = Builder.saveIP(); |
2405 | Builder.restoreIP(IP: AllocaIP); |
2406 | AllocaInst *DestAlloca = Builder.CreateAlloca(Ty: RI.ElementType, ArraySize: nullptr, |
2407 | Name: ".omp.reduction.element" ); |
2408 | DestAlloca->setAlignment( |
2409 | M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType)); |
2410 | DestElementAddr = DestAlloca; |
2411 | DestElementAddr = |
2412 | Builder.CreateAddrSpaceCast(V: DestElementAddr, DestTy: Builder.getPtrTy(), |
2413 | Name: DestElementAddr->getName() + ".ascast" ); |
2414 | Builder.restoreIP(IP: CurIP); |
2415 | ShuffleInElement = true; |
2416 | UpdateDestListPtr = true; |
2417 | break; |
2418 | } |
2419 | case CopyAction::ThreadCopy: { |
2420 | DestElementAddr = |
2421 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: DestElementPtrAddr); |
2422 | break; |
2423 | } |
2424 | } |
2425 | |
2426 | // Now that all active lanes have read the element in the |
2427 | // Reduce list, shuffle over the value from the remote lane. |
2428 | if (ShuffleInElement) { |
2429 | shuffleAndStore(AllocaIP, SrcAddr: SrcElementAddr, DstAddr: DestElementAddr, ElemType: RI.ElementType, |
2430 | Offset: RemoteLaneOffset, ReductionArrayTy); |
2431 | } else { |
2432 | switch (RI.EvaluationKind) { |
2433 | case EvalKind::Scalar: { |
2434 | Value *Elem = Builder.CreateLoad(Ty: RI.ElementType, Ptr: SrcElementAddr); |
2435 | // Store the source element value to the dest element address. |
2436 | Builder.CreateStore(Val: Elem, Ptr: DestElementAddr); |
2437 | break; |
2438 | } |
2439 | case EvalKind::Complex: { |
2440 | Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32( |
2441 | Ty: RI.ElementType, Ptr: SrcElementAddr, Idx0: 0, Idx1: 0, Name: ".realp" ); |
2442 | Value *SrcReal = Builder.CreateLoad( |
2443 | Ty: RI.ElementType->getStructElementType(N: 0), Ptr: SrcRealPtr, Name: ".real" ); |
2444 | Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32( |
2445 | Ty: RI.ElementType, Ptr: SrcElementAddr, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
2446 | Value *SrcImg = Builder.CreateLoad( |
2447 | Ty: RI.ElementType->getStructElementType(N: 1), Ptr: SrcImgPtr, Name: ".imag" ); |
2448 | |
2449 | Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32( |
2450 | Ty: RI.ElementType, Ptr: DestElementAddr, Idx0: 0, Idx1: 0, Name: ".realp" ); |
2451 | Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32( |
2452 | Ty: RI.ElementType, Ptr: DestElementAddr, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
2453 | Builder.CreateStore(Val: SrcReal, Ptr: DestRealPtr); |
2454 | Builder.CreateStore(Val: SrcImg, Ptr: DestImgPtr); |
2455 | break; |
2456 | } |
2457 | case EvalKind::Aggregate: { |
2458 | Value *SizeVal = Builder.getInt64( |
2459 | C: M.getDataLayout().getTypeStoreSize(Ty: RI.ElementType)); |
2460 | Builder.CreateMemCpy( |
2461 | Dst: DestElementAddr, DstAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), |
2462 | Src: SrcElementAddr, SrcAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), |
2463 | Size: SizeVal, isVolatile: false); |
2464 | break; |
2465 | } |
2466 | }; |
2467 | } |
2468 | |
2469 | // Step 3.1: Modify reference in dest Reduce list as needed. |
2470 | // Modifying the reference in Reduce list to point to the newly |
2471 | // created element. The element is live in the current function |
2472 | // scope and that of functions it invokes (i.e., reduce_function). |
2473 | // RemoteReduceData[i] = (void*)&RemoteElem |
2474 | if (UpdateDestListPtr) { |
2475 | Value *CastDestAddr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2476 | V: DestElementAddr, DestTy: Builder.getPtrTy(), |
2477 | Name: DestElementAddr->getName() + ".ascast" ); |
2478 | Builder.CreateStore(Val: CastDestAddr, Ptr: DestElementPtrAddr); |
2479 | } |
2480 | } |
2481 | } |
2482 | |
2483 | Function *OpenMPIRBuilder::emitInterWarpCopyFunction( |
2484 | const LocationDescription &Loc, ArrayRef<ReductionInfo> ReductionInfos, |
2485 | AttributeList FuncAttrs) { |
2486 | InsertPointTy SavedIP = Builder.saveIP(); |
2487 | LLVMContext &Ctx = M.getContext(); |
2488 | FunctionType *FuncTy = FunctionType::get( |
2489 | Result: Builder.getVoidTy(), Params: {Builder.getPtrTy(), Builder.getInt32Ty()}, |
2490 | /* IsVarArg */ isVarArg: false); |
2491 | Function *WcFunc = |
2492 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
2493 | N: "_omp_reduction_inter_warp_copy_func" , M: &M); |
2494 | WcFunc->setAttributes(FuncAttrs); |
2495 | WcFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
2496 | WcFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
2497 | BasicBlock *EntryBB = BasicBlock::Create(Context&: M.getContext(), Name: "entry" , Parent: WcFunc); |
2498 | Builder.SetInsertPoint(EntryBB); |
2499 | |
2500 | // ReduceList: thread local Reduce list. |
2501 | // At the stage of the computation when this function is called, partially |
2502 | // aggregated values reside in the first lane of every active warp. |
2503 | Argument *ReduceListArg = WcFunc->getArg(i: 0); |
2504 | // NumWarps: number of warps active in the parallel region. This could |
2505 | // be smaller than 32 (max warps in a CTA) for partial block reduction. |
2506 | Argument *NumWarpsArg = WcFunc->getArg(i: 1); |
2507 | |
2508 | // This array is used as a medium to transfer, one reduce element at a time, |
2509 | // the data from the first lane of every warp to lanes in the first warp |
2510 | // in order to perform the final step of a reduction in a parallel region |
2511 | // (reduction across warps). The array is placed in NVPTX __shared__ memory |
2512 | // for reduced latency, as well as to have a distinct copy for concurrently |
2513 | // executing target regions. The array is declared with common linkage so |
2514 | // as to be shared across compilation units. |
2515 | StringRef TransferMediumName = |
2516 | "__openmp_nvptx_data_transfer_temporary_storage" ; |
2517 | GlobalVariable *TransferMedium = M.getGlobalVariable(Name: TransferMediumName); |
2518 | unsigned WarpSize = Config.getGridValue().GV_Warp_Size; |
2519 | ArrayType *ArrayTy = ArrayType::get(ElementType: Builder.getInt32Ty(), NumElements: WarpSize); |
2520 | if (!TransferMedium) { |
2521 | TransferMedium = new GlobalVariable( |
2522 | M, ArrayTy, /*isConstant=*/false, GlobalVariable::WeakAnyLinkage, |
2523 | UndefValue::get(T: ArrayTy), TransferMediumName, |
2524 | /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal, |
2525 | /*AddressSpace=*/3); |
2526 | } |
2527 | |
2528 | // Get the CUDA thread id of the current OpenMP thread on the GPU. |
2529 | Value *GPUThreadID = getGPUThreadID(); |
2530 | // nvptx_lane_id = nvptx_id % warpsize |
2531 | Value *LaneID = getNVPTXLaneID(); |
2532 | // nvptx_warp_id = nvptx_id / warpsize |
2533 | Value *WarpID = getNVPTXWarpID(); |
2534 | |
2535 | InsertPointTy AllocaIP = |
2536 | InsertPointTy(Builder.GetInsertBlock(), |
2537 | Builder.GetInsertBlock()->getFirstInsertionPt()); |
2538 | Type *Arg0Type = ReduceListArg->getType(); |
2539 | Type *Arg1Type = NumWarpsArg->getType(); |
2540 | Builder.restoreIP(IP: AllocaIP); |
2541 | AllocaInst *ReduceListAlloca = Builder.CreateAlloca( |
2542 | Ty: Arg0Type, ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
2543 | AllocaInst *NumWarpsAlloca = |
2544 | Builder.CreateAlloca(Ty: Arg1Type, ArraySize: nullptr, Name: NumWarpsArg->getName() + ".addr" ); |
2545 | Value *ReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2546 | V: ReduceListAlloca, DestTy: Arg0Type, Name: ReduceListAlloca->getName() + ".ascast" ); |
2547 | Value *NumWarpsAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2548 | V: NumWarpsAlloca, DestTy: Arg1Type->getPointerTo(), |
2549 | Name: NumWarpsAlloca->getName() + ".ascast" ); |
2550 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListAddrCast); |
2551 | Builder.CreateStore(Val: NumWarpsArg, Ptr: NumWarpsAddrCast); |
2552 | AllocaIP = getInsertPointAfterInstr(I: NumWarpsAlloca); |
2553 | InsertPointTy CodeGenIP = |
2554 | getInsertPointAfterInstr(I: &Builder.GetInsertBlock()->back()); |
2555 | Builder.restoreIP(IP: CodeGenIP); |
2556 | |
2557 | Value *ReduceList = |
2558 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ReduceListAddrCast); |
2559 | |
2560 | for (auto En : enumerate(First&: ReductionInfos)) { |
2561 | // |
2562 | // Warp master copies reduce element to transfer medium in __shared__ |
2563 | // memory. |
2564 | // |
2565 | const ReductionInfo &RI = En.value(); |
2566 | unsigned RealTySize = M.getDataLayout().getTypeAllocSize(Ty: RI.ElementType); |
2567 | for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /= 2) { |
2568 | Type *CType = Builder.getIntNTy(N: TySize * 8); |
2569 | |
2570 | unsigned NumIters = RealTySize / TySize; |
2571 | if (NumIters == 0) |
2572 | continue; |
2573 | Value *Cnt = nullptr; |
2574 | Value *CntAddr = nullptr; |
2575 | BasicBlock *PrecondBB = nullptr; |
2576 | BasicBlock *ExitBB = nullptr; |
2577 | if (NumIters > 1) { |
2578 | CodeGenIP = Builder.saveIP(); |
2579 | Builder.restoreIP(IP: AllocaIP); |
2580 | CntAddr = |
2581 | Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, Name: ".cnt.addr" ); |
2582 | |
2583 | CntAddr = Builder.CreateAddrSpaceCast(V: CntAddr, DestTy: Builder.getPtrTy(), |
2584 | Name: CntAddr->getName() + ".ascast" ); |
2585 | Builder.restoreIP(IP: CodeGenIP); |
2586 | Builder.CreateStore(Val: Constant::getNullValue(Ty: Builder.getInt32Ty()), |
2587 | Ptr: CntAddr, |
2588 | /*Volatile=*/isVolatile: false); |
2589 | PrecondBB = BasicBlock::Create(Context&: Ctx, Name: "precond" ); |
2590 | ExitBB = BasicBlock::Create(Context&: Ctx, Name: "exit" ); |
2591 | BasicBlock *BodyBB = BasicBlock::Create(Context&: Ctx, Name: "body" ); |
2592 | emitBlock(BB: PrecondBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2593 | Cnt = Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: CntAddr, |
2594 | /*Volatile=*/isVolatile: false); |
2595 | Value *Cmp = Builder.CreateICmpULT( |
2596 | LHS: Cnt, RHS: ConstantInt::get(Ty: Builder.getInt32Ty(), V: NumIters)); |
2597 | Builder.CreateCondBr(Cond: Cmp, True: BodyBB, False: ExitBB); |
2598 | emitBlock(BB: BodyBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2599 | } |
2600 | |
2601 | // kmpc_barrier. |
2602 | createBarrier(Loc: LocationDescription(Builder.saveIP(), Loc.DL), |
2603 | Kind: omp::Directive::OMPD_unknown, |
2604 | /* ForceSimpleCall */ false, |
2605 | /* CheckCancelFlag */ true); |
2606 | BasicBlock *ThenBB = BasicBlock::Create(Context&: Ctx, Name: "then" ); |
2607 | BasicBlock *ElseBB = BasicBlock::Create(Context&: Ctx, Name: "else" ); |
2608 | BasicBlock *MergeBB = BasicBlock::Create(Context&: Ctx, Name: "ifcont" ); |
2609 | |
2610 | // if (lane_id == 0) |
2611 | Value *IsWarpMaster = Builder.CreateIsNull(Arg: LaneID, Name: "warp_master" ); |
2612 | Builder.CreateCondBr(Cond: IsWarpMaster, True: ThenBB, False: ElseBB); |
2613 | emitBlock(BB: ThenBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2614 | |
2615 | // Reduce element = LocalReduceList[i] |
2616 | auto *RedListArrayTy = |
2617 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
2618 | Type *IndexTy = Builder.getIndexTy( |
2619 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
2620 | Value *ElemPtrPtr = |
2621 | Builder.CreateInBoundsGEP(Ty: RedListArrayTy, Ptr: ReduceList, |
2622 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), |
2623 | ConstantInt::get(Ty: IndexTy, V: En.index())}); |
2624 | // elemptr = ((CopyType*)(elemptrptr)) + I |
2625 | Value *ElemPtr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ElemPtrPtr); |
2626 | if (NumIters > 1) |
2627 | ElemPtr = Builder.CreateGEP(Ty: Builder.getInt32Ty(), Ptr: ElemPtr, IdxList: Cnt); |
2628 | |
2629 | // Get pointer to location in transfer medium. |
2630 | // MediumPtr = &medium[warp_id] |
2631 | Value *MediumPtr = Builder.CreateInBoundsGEP( |
2632 | Ty: ArrayTy, Ptr: TransferMedium, IdxList: {Builder.getInt64(C: 0), WarpID}); |
2633 | // elem = *elemptr |
2634 | //*MediumPtr = elem |
2635 | Value *Elem = Builder.CreateLoad(Ty: CType, Ptr: ElemPtr); |
2636 | // Store the source element value to the dest element address. |
2637 | Builder.CreateStore(Val: Elem, Ptr: MediumPtr, |
2638 | /*IsVolatile*/ isVolatile: true); |
2639 | Builder.CreateBr(Dest: MergeBB); |
2640 | |
2641 | // else |
2642 | emitBlock(BB: ElseBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2643 | Builder.CreateBr(Dest: MergeBB); |
2644 | |
2645 | // endif |
2646 | emitBlock(BB: MergeBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2647 | createBarrier(Loc: LocationDescription(Builder.saveIP(), Loc.DL), |
2648 | Kind: omp::Directive::OMPD_unknown, |
2649 | /* ForceSimpleCall */ false, |
2650 | /* CheckCancelFlag */ true); |
2651 | |
2652 | // Warp 0 copies reduce element from transfer medium |
2653 | BasicBlock *W0ThenBB = BasicBlock::Create(Context&: Ctx, Name: "then" ); |
2654 | BasicBlock *W0ElseBB = BasicBlock::Create(Context&: Ctx, Name: "else" ); |
2655 | BasicBlock *W0MergeBB = BasicBlock::Create(Context&: Ctx, Name: "ifcont" ); |
2656 | |
2657 | Value *NumWarpsVal = |
2658 | Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: NumWarpsAddrCast); |
2659 | // Up to 32 threads in warp 0 are active. |
2660 | Value *IsActiveThread = |
2661 | Builder.CreateICmpULT(LHS: GPUThreadID, RHS: NumWarpsVal, Name: "is_active_thread" ); |
2662 | Builder.CreateCondBr(Cond: IsActiveThread, True: W0ThenBB, False: W0ElseBB); |
2663 | |
2664 | emitBlock(BB: W0ThenBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2665 | |
2666 | // SecMediumPtr = &medium[tid] |
2667 | // SrcMediumVal = *SrcMediumPtr |
2668 | Value *SrcMediumPtrVal = Builder.CreateInBoundsGEP( |
2669 | Ty: ArrayTy, Ptr: TransferMedium, IdxList: {Builder.getInt64(C: 0), GPUThreadID}); |
2670 | // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I |
2671 | Value *TargetElemPtrPtr = |
2672 | Builder.CreateInBoundsGEP(Ty: RedListArrayTy, Ptr: ReduceList, |
2673 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), |
2674 | ConstantInt::get(Ty: IndexTy, V: En.index())}); |
2675 | Value *TargetElemPtrVal = |
2676 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: TargetElemPtrPtr); |
2677 | Value *TargetElemPtr = TargetElemPtrVal; |
2678 | if (NumIters > 1) |
2679 | TargetElemPtr = |
2680 | Builder.CreateGEP(Ty: Builder.getInt32Ty(), Ptr: TargetElemPtr, IdxList: Cnt); |
2681 | |
2682 | // *TargetElemPtr = SrcMediumVal; |
2683 | Value *SrcMediumValue = |
2684 | Builder.CreateLoad(Ty: CType, Ptr: SrcMediumPtrVal, /*IsVolatile*/ isVolatile: true); |
2685 | Builder.CreateStore(Val: SrcMediumValue, Ptr: TargetElemPtr); |
2686 | Builder.CreateBr(Dest: W0MergeBB); |
2687 | |
2688 | emitBlock(BB: W0ElseBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2689 | Builder.CreateBr(Dest: W0MergeBB); |
2690 | |
2691 | emitBlock(BB: W0MergeBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2692 | |
2693 | if (NumIters > 1) { |
2694 | Cnt = Builder.CreateNSWAdd( |
2695 | LHS: Cnt, RHS: ConstantInt::get(Ty: Builder.getInt32Ty(), /*V=*/1)); |
2696 | Builder.CreateStore(Val: Cnt, Ptr: CntAddr, /*Volatile=*/isVolatile: false); |
2697 | |
2698 | auto *CurFn = Builder.GetInsertBlock()->getParent(); |
2699 | emitBranch(Target: PrecondBB); |
2700 | emitBlock(BB: ExitBB, CurFn); |
2701 | } |
2702 | RealTySize %= TySize; |
2703 | } |
2704 | } |
2705 | |
2706 | Builder.CreateRetVoid(); |
2707 | Builder.restoreIP(IP: SavedIP); |
2708 | |
2709 | return WcFunc; |
2710 | } |
2711 | |
2712 | Function *OpenMPIRBuilder::emitShuffleAndReduceFunction( |
2713 | ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn, |
2714 | AttributeList FuncAttrs) { |
2715 | LLVMContext &Ctx = M.getContext(); |
2716 | FunctionType *FuncTy = |
2717 | FunctionType::get(Result: Builder.getVoidTy(), |
2718 | Params: {Builder.getPtrTy(), Builder.getInt16Ty(), |
2719 | Builder.getInt16Ty(), Builder.getInt16Ty()}, |
2720 | /* IsVarArg */ isVarArg: false); |
2721 | Function *SarFunc = |
2722 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
2723 | N: "_omp_reduction_shuffle_and_reduce_func" , M: &M); |
2724 | SarFunc->setAttributes(FuncAttrs); |
2725 | SarFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
2726 | SarFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
2727 | SarFunc->addParamAttr(ArgNo: 2, Kind: Attribute::NoUndef); |
2728 | SarFunc->addParamAttr(ArgNo: 3, Kind: Attribute::NoUndef); |
2729 | SarFunc->addParamAttr(ArgNo: 1, Kind: Attribute::SExt); |
2730 | SarFunc->addParamAttr(ArgNo: 2, Kind: Attribute::SExt); |
2731 | SarFunc->addParamAttr(ArgNo: 3, Kind: Attribute::SExt); |
2732 | BasicBlock *EntryBB = BasicBlock::Create(Context&: M.getContext(), Name: "entry" , Parent: SarFunc); |
2733 | Builder.SetInsertPoint(EntryBB); |
2734 | |
2735 | // Thread local Reduce list used to host the values of data to be reduced. |
2736 | Argument *ReduceListArg = SarFunc->getArg(i: 0); |
2737 | // Current lane id; could be logical. |
2738 | Argument *LaneIDArg = SarFunc->getArg(i: 1); |
2739 | // Offset of the remote source lane relative to the current lane. |
2740 | Argument *RemoteLaneOffsetArg = SarFunc->getArg(i: 2); |
2741 | // Algorithm version. This is expected to be known at compile time. |
2742 | Argument *AlgoVerArg = SarFunc->getArg(i: 3); |
2743 | |
2744 | Type *ReduceListArgType = ReduceListArg->getType(); |
2745 | Type *LaneIDArgType = LaneIDArg->getType(); |
2746 | Type *LaneIDArgPtrType = LaneIDArg->getType()->getPointerTo(); |
2747 | Value *ReduceListAlloca = Builder.CreateAlloca( |
2748 | Ty: ReduceListArgType, ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
2749 | Value *LaneIdAlloca = Builder.CreateAlloca(Ty: LaneIDArgType, ArraySize: nullptr, |
2750 | Name: LaneIDArg->getName() + ".addr" ); |
2751 | Value *RemoteLaneOffsetAlloca = Builder.CreateAlloca( |
2752 | Ty: LaneIDArgType, ArraySize: nullptr, Name: RemoteLaneOffsetArg->getName() + ".addr" ); |
2753 | Value *AlgoVerAlloca = Builder.CreateAlloca(Ty: LaneIDArgType, ArraySize: nullptr, |
2754 | Name: AlgoVerArg->getName() + ".addr" ); |
2755 | ArrayType *RedListArrayTy = |
2756 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
2757 | |
2758 | // Create a local thread-private variable to host the Reduce list |
2759 | // from a remote lane. |
2760 | Instruction *RemoteReductionListAlloca = Builder.CreateAlloca( |
2761 | Ty: RedListArrayTy, ArraySize: nullptr, Name: ".omp.reduction.remote_reduce_list" ); |
2762 | |
2763 | Value *ReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2764 | V: ReduceListAlloca, DestTy: ReduceListArgType, |
2765 | Name: ReduceListAlloca->getName() + ".ascast" ); |
2766 | Value *LaneIdAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2767 | V: LaneIdAlloca, DestTy: LaneIDArgPtrType, Name: LaneIdAlloca->getName() + ".ascast" ); |
2768 | Value *RemoteLaneOffsetAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2769 | V: RemoteLaneOffsetAlloca, DestTy: LaneIDArgPtrType, |
2770 | Name: RemoteLaneOffsetAlloca->getName() + ".ascast" ); |
2771 | Value *AlgoVerAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2772 | V: AlgoVerAlloca, DestTy: LaneIDArgPtrType, Name: AlgoVerAlloca->getName() + ".ascast" ); |
2773 | Value *RemoteListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2774 | V: RemoteReductionListAlloca, DestTy: Builder.getPtrTy(), |
2775 | Name: RemoteReductionListAlloca->getName() + ".ascast" ); |
2776 | |
2777 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListAddrCast); |
2778 | Builder.CreateStore(Val: LaneIDArg, Ptr: LaneIdAddrCast); |
2779 | Builder.CreateStore(Val: RemoteLaneOffsetArg, Ptr: RemoteLaneOffsetAddrCast); |
2780 | Builder.CreateStore(Val: AlgoVerArg, Ptr: AlgoVerAddrCast); |
2781 | |
2782 | Value *ReduceList = Builder.CreateLoad(Ty: ReduceListArgType, Ptr: ReduceListAddrCast); |
2783 | Value *LaneId = Builder.CreateLoad(Ty: LaneIDArgType, Ptr: LaneIdAddrCast); |
2784 | Value *RemoteLaneOffset = |
2785 | Builder.CreateLoad(Ty: LaneIDArgType, Ptr: RemoteLaneOffsetAddrCast); |
2786 | Value *AlgoVer = Builder.CreateLoad(Ty: LaneIDArgType, Ptr: AlgoVerAddrCast); |
2787 | |
2788 | InsertPointTy AllocaIP = getInsertPointAfterInstr(I: RemoteReductionListAlloca); |
2789 | |
2790 | // This loop iterates through the list of reduce elements and copies, |
2791 | // element by element, from a remote lane in the warp to RemoteReduceList, |
2792 | // hosted on the thread's stack. |
2793 | emitReductionListCopy( |
2794 | AllocaIP, Action: CopyAction::RemoteLaneToThread, ReductionArrayTy: RedListArrayTy, ReductionInfos, |
2795 | SrcBase: ReduceList, DestBase: RemoteListAddrCast, CopyOptions: {.RemoteLaneOffset: RemoteLaneOffset, .ScratchpadIndex: nullptr, .ScratchpadWidth: nullptr}); |
2796 | |
2797 | // The actions to be performed on the Remote Reduce list is dependent |
2798 | // on the algorithm version. |
2799 | // |
2800 | // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && |
2801 | // LaneId % 2 == 0 && Offset > 0): |
2802 | // do the reduction value aggregation |
2803 | // |
2804 | // The thread local variable Reduce list is mutated in place to host the |
2805 | // reduced data, which is the aggregated value produced from local and |
2806 | // remote lanes. |
2807 | // |
2808 | // Note that AlgoVer is expected to be a constant integer known at compile |
2809 | // time. |
2810 | // When AlgoVer==0, the first conjunction evaluates to true, making |
2811 | // the entire predicate true during compile time. |
2812 | // When AlgoVer==1, the second conjunction has only the second part to be |
2813 | // evaluated during runtime. Other conjunctions evaluates to false |
2814 | // during compile time. |
2815 | // When AlgoVer==2, the third conjunction has only the second part to be |
2816 | // evaluated during runtime. Other conjunctions evaluates to false |
2817 | // during compile time. |
2818 | Value *CondAlgo0 = Builder.CreateIsNull(Arg: AlgoVer); |
2819 | Value *Algo1 = Builder.CreateICmpEQ(LHS: AlgoVer, RHS: Builder.getInt16(C: 1)); |
2820 | Value *LaneComp = Builder.CreateICmpULT(LHS: LaneId, RHS: RemoteLaneOffset); |
2821 | Value *CondAlgo1 = Builder.CreateAnd(LHS: Algo1, RHS: LaneComp); |
2822 | Value *Algo2 = Builder.CreateICmpEQ(LHS: AlgoVer, RHS: Builder.getInt16(C: 2)); |
2823 | Value *LaneIdAnd1 = Builder.CreateAnd(LHS: LaneId, RHS: Builder.getInt16(C: 1)); |
2824 | Value *LaneIdComp = Builder.CreateIsNull(Arg: LaneIdAnd1); |
2825 | Value *Algo2AndLaneIdComp = Builder.CreateAnd(LHS: Algo2, RHS: LaneIdComp); |
2826 | Value *RemoteOffsetComp = |
2827 | Builder.CreateICmpSGT(LHS: RemoteLaneOffset, RHS: Builder.getInt16(C: 0)); |
2828 | Value *CondAlgo2 = Builder.CreateAnd(LHS: Algo2AndLaneIdComp, RHS: RemoteOffsetComp); |
2829 | Value *CA0OrCA1 = Builder.CreateOr(LHS: CondAlgo0, RHS: CondAlgo1); |
2830 | Value *CondReduce = Builder.CreateOr(LHS: CA0OrCA1, RHS: CondAlgo2); |
2831 | |
2832 | BasicBlock *ThenBB = BasicBlock::Create(Context&: Ctx, Name: "then" ); |
2833 | BasicBlock *ElseBB = BasicBlock::Create(Context&: Ctx, Name: "else" ); |
2834 | BasicBlock *MergeBB = BasicBlock::Create(Context&: Ctx, Name: "ifcont" ); |
2835 | |
2836 | Builder.CreateCondBr(Cond: CondReduce, True: ThenBB, False: ElseBB); |
2837 | emitBlock(BB: ThenBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2838 | Value *LocalReduceListPtr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2839 | V: ReduceList, DestTy: Builder.getPtrTy()); |
2840 | Value *RemoteReduceListPtr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2841 | V: RemoteListAddrCast, DestTy: Builder.getPtrTy()); |
2842 | Builder.CreateCall(Callee: ReduceFn, Args: {LocalReduceListPtr, RemoteReduceListPtr}) |
2843 | ->addFnAttr(Kind: Attribute::NoUnwind); |
2844 | Builder.CreateBr(Dest: MergeBB); |
2845 | |
2846 | emitBlock(BB: ElseBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2847 | Builder.CreateBr(Dest: MergeBB); |
2848 | |
2849 | emitBlock(BB: MergeBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2850 | |
2851 | // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local |
2852 | // Reduce list. |
2853 | Algo1 = Builder.CreateICmpEQ(LHS: AlgoVer, RHS: Builder.getInt16(C: 1)); |
2854 | Value *LaneIdGtOffset = Builder.CreateICmpUGE(LHS: LaneId, RHS: RemoteLaneOffset); |
2855 | Value *CondCopy = Builder.CreateAnd(LHS: Algo1, RHS: LaneIdGtOffset); |
2856 | |
2857 | BasicBlock *CpyThenBB = BasicBlock::Create(Context&: Ctx, Name: "then" ); |
2858 | BasicBlock *CpyElseBB = BasicBlock::Create(Context&: Ctx, Name: "else" ); |
2859 | BasicBlock *CpyMergeBB = BasicBlock::Create(Context&: Ctx, Name: "ifcont" ); |
2860 | Builder.CreateCondBr(Cond: CondCopy, True: CpyThenBB, False: CpyElseBB); |
2861 | |
2862 | emitBlock(BB: CpyThenBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2863 | emitReductionListCopy(AllocaIP, Action: CopyAction::ThreadCopy, ReductionArrayTy: RedListArrayTy, |
2864 | ReductionInfos, SrcBase: RemoteListAddrCast, DestBase: ReduceList); |
2865 | Builder.CreateBr(Dest: CpyMergeBB); |
2866 | |
2867 | emitBlock(BB: CpyElseBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2868 | Builder.CreateBr(Dest: CpyMergeBB); |
2869 | |
2870 | emitBlock(BB: CpyMergeBB, CurFn: Builder.GetInsertBlock()->getParent()); |
2871 | |
2872 | Builder.CreateRetVoid(); |
2873 | |
2874 | return SarFunc; |
2875 | } |
2876 | |
2877 | Function *OpenMPIRBuilder::emitListToGlobalCopyFunction( |
2878 | ArrayRef<ReductionInfo> ReductionInfos, Type *ReductionsBufferTy, |
2879 | AttributeList FuncAttrs) { |
2880 | OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP(); |
2881 | LLVMContext &Ctx = M.getContext(); |
2882 | FunctionType *FuncTy = FunctionType::get( |
2883 | Result: Builder.getVoidTy(), |
2884 | Params: {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()}, |
2885 | /* IsVarArg */ isVarArg: false); |
2886 | Function *LtGCFunc = |
2887 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
2888 | N: "_omp_reduction_list_to_global_copy_func" , M: &M); |
2889 | LtGCFunc->setAttributes(FuncAttrs); |
2890 | LtGCFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
2891 | LtGCFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
2892 | LtGCFunc->addParamAttr(ArgNo: 2, Kind: Attribute::NoUndef); |
2893 | |
2894 | BasicBlock *EntryBlock = BasicBlock::Create(Context&: Ctx, Name: "entry" , Parent: LtGCFunc); |
2895 | Builder.SetInsertPoint(EntryBlock); |
2896 | |
2897 | // Buffer: global reduction buffer. |
2898 | Argument *BufferArg = LtGCFunc->getArg(i: 0); |
2899 | // Idx: index of the buffer. |
2900 | Argument *IdxArg = LtGCFunc->getArg(i: 1); |
2901 | // ReduceList: thread local Reduce list. |
2902 | Argument *ReduceListArg = LtGCFunc->getArg(i: 2); |
2903 | |
2904 | Value *BufferArgAlloca = Builder.CreateAlloca(Ty: Builder.getPtrTy(), ArraySize: nullptr, |
2905 | Name: BufferArg->getName() + ".addr" ); |
2906 | Value *IdxArgAlloca = Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, |
2907 | Name: IdxArg->getName() + ".addr" ); |
2908 | Value *ReduceListArgAlloca = Builder.CreateAlloca( |
2909 | Ty: Builder.getPtrTy(), ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
2910 | Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2911 | V: BufferArgAlloca, DestTy: Builder.getPtrTy(), |
2912 | Name: BufferArgAlloca->getName() + ".ascast" ); |
2913 | Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2914 | V: IdxArgAlloca, DestTy: Builder.getPtrTy(), Name: IdxArgAlloca->getName() + ".ascast" ); |
2915 | Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
2916 | V: ReduceListArgAlloca, DestTy: Builder.getPtrTy(), |
2917 | Name: ReduceListArgAlloca->getName() + ".ascast" ); |
2918 | |
2919 | Builder.CreateStore(Val: BufferArg, Ptr: BufferArgAddrCast); |
2920 | Builder.CreateStore(Val: IdxArg, Ptr: IdxArgAddrCast); |
2921 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListArgAddrCast); |
2922 | |
2923 | Value *LocalReduceList = |
2924 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ReduceListArgAddrCast); |
2925 | Value *BufferArgVal = |
2926 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: BufferArgAddrCast); |
2927 | Value *Idxs[] = {Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: IdxArgAddrCast)}; |
2928 | Type *IndexTy = Builder.getIndexTy( |
2929 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
2930 | for (auto En : enumerate(First&: ReductionInfos)) { |
2931 | const ReductionInfo &RI = En.value(); |
2932 | auto *RedListArrayTy = |
2933 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
2934 | // Reduce element = LocalReduceList[i] |
2935 | Value *ElemPtrPtr = Builder.CreateInBoundsGEP( |
2936 | Ty: RedListArrayTy, Ptr: LocalReduceList, |
2937 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
2938 | // elemptr = ((CopyType*)(elemptrptr)) + I |
2939 | Value *ElemPtr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ElemPtrPtr); |
2940 | |
2941 | // Global = Buffer.VD[Idx]; |
2942 | Value *BufferVD = |
2943 | Builder.CreateInBoundsGEP(Ty: ReductionsBufferTy, Ptr: BufferArgVal, IdxList: Idxs); |
2944 | Value *GlobVal = Builder.CreateConstInBoundsGEP2_32( |
2945 | Ty: ReductionsBufferTy, Ptr: BufferVD, Idx0: 0, Idx1: En.index()); |
2946 | |
2947 | switch (RI.EvaluationKind) { |
2948 | case EvalKind::Scalar: { |
2949 | Value *TargetElement = Builder.CreateLoad(Ty: RI.ElementType, Ptr: ElemPtr); |
2950 | Builder.CreateStore(Val: TargetElement, Ptr: GlobVal); |
2951 | break; |
2952 | } |
2953 | case EvalKind::Complex: { |
2954 | Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32( |
2955 | Ty: RI.ElementType, Ptr: ElemPtr, Idx0: 0, Idx1: 0, Name: ".realp" ); |
2956 | Value *SrcReal = Builder.CreateLoad( |
2957 | Ty: RI.ElementType->getStructElementType(N: 0), Ptr: SrcRealPtr, Name: ".real" ); |
2958 | Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32( |
2959 | Ty: RI.ElementType, Ptr: ElemPtr, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
2960 | Value *SrcImg = Builder.CreateLoad( |
2961 | Ty: RI.ElementType->getStructElementType(N: 1), Ptr: SrcImgPtr, Name: ".imag" ); |
2962 | |
2963 | Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32( |
2964 | Ty: RI.ElementType, Ptr: GlobVal, Idx0: 0, Idx1: 0, Name: ".realp" ); |
2965 | Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32( |
2966 | Ty: RI.ElementType, Ptr: GlobVal, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
2967 | Builder.CreateStore(Val: SrcReal, Ptr: DestRealPtr); |
2968 | Builder.CreateStore(Val: SrcImg, Ptr: DestImgPtr); |
2969 | break; |
2970 | } |
2971 | case EvalKind::Aggregate: { |
2972 | Value *SizeVal = |
2973 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: RI.ElementType)); |
2974 | Builder.CreateMemCpy( |
2975 | Dst: GlobVal, DstAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), Src: ElemPtr, |
2976 | SrcAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), Size: SizeVal, isVolatile: false); |
2977 | break; |
2978 | } |
2979 | } |
2980 | } |
2981 | |
2982 | Builder.CreateRetVoid(); |
2983 | Builder.restoreIP(IP: OldIP); |
2984 | return LtGCFunc; |
2985 | } |
2986 | |
2987 | Function *OpenMPIRBuilder::emitListToGlobalReduceFunction( |
2988 | ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn, |
2989 | Type *ReductionsBufferTy, AttributeList FuncAttrs) { |
2990 | OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP(); |
2991 | LLVMContext &Ctx = M.getContext(); |
2992 | FunctionType *FuncTy = FunctionType::get( |
2993 | Result: Builder.getVoidTy(), |
2994 | Params: {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()}, |
2995 | /* IsVarArg */ isVarArg: false); |
2996 | Function *LtGRFunc = |
2997 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
2998 | N: "_omp_reduction_list_to_global_reduce_func" , M: &M); |
2999 | LtGRFunc->setAttributes(FuncAttrs); |
3000 | LtGRFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
3001 | LtGRFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
3002 | LtGRFunc->addParamAttr(ArgNo: 2, Kind: Attribute::NoUndef); |
3003 | |
3004 | BasicBlock *EntryBlock = BasicBlock::Create(Context&: Ctx, Name: "entry" , Parent: LtGRFunc); |
3005 | Builder.SetInsertPoint(EntryBlock); |
3006 | |
3007 | // Buffer: global reduction buffer. |
3008 | Argument *BufferArg = LtGRFunc->getArg(i: 0); |
3009 | // Idx: index of the buffer. |
3010 | Argument *IdxArg = LtGRFunc->getArg(i: 1); |
3011 | // ReduceList: thread local Reduce list. |
3012 | Argument *ReduceListArg = LtGRFunc->getArg(i: 2); |
3013 | |
3014 | Value *BufferArgAlloca = Builder.CreateAlloca(Ty: Builder.getPtrTy(), ArraySize: nullptr, |
3015 | Name: BufferArg->getName() + ".addr" ); |
3016 | Value *IdxArgAlloca = Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, |
3017 | Name: IdxArg->getName() + ".addr" ); |
3018 | Value *ReduceListArgAlloca = Builder.CreateAlloca( |
3019 | Ty: Builder.getPtrTy(), ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
3020 | auto *RedListArrayTy = |
3021 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
3022 | |
3023 | // 1. Build a list of reduction variables. |
3024 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
3025 | Value *LocalReduceList = |
3026 | Builder.CreateAlloca(Ty: RedListArrayTy, ArraySize: nullptr, Name: ".omp.reduction.red_list" ); |
3027 | |
3028 | Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3029 | V: BufferArgAlloca, DestTy: Builder.getPtrTy(), |
3030 | Name: BufferArgAlloca->getName() + ".ascast" ); |
3031 | Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3032 | V: IdxArgAlloca, DestTy: Builder.getPtrTy(), Name: IdxArgAlloca->getName() + ".ascast" ); |
3033 | Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3034 | V: ReduceListArgAlloca, DestTy: Builder.getPtrTy(), |
3035 | Name: ReduceListArgAlloca->getName() + ".ascast" ); |
3036 | Value *LocalReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3037 | V: LocalReduceList, DestTy: Builder.getPtrTy(), |
3038 | Name: LocalReduceList->getName() + ".ascast" ); |
3039 | |
3040 | Builder.CreateStore(Val: BufferArg, Ptr: BufferArgAddrCast); |
3041 | Builder.CreateStore(Val: IdxArg, Ptr: IdxArgAddrCast); |
3042 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListArgAddrCast); |
3043 | |
3044 | Value *BufferVal = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: BufferArgAddrCast); |
3045 | Value *Idxs[] = {Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: IdxArgAddrCast)}; |
3046 | Type *IndexTy = Builder.getIndexTy( |
3047 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
3048 | for (auto En : enumerate(First&: ReductionInfos)) { |
3049 | Value *TargetElementPtrPtr = Builder.CreateInBoundsGEP( |
3050 | Ty: RedListArrayTy, Ptr: LocalReduceListAddrCast, |
3051 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3052 | Value *BufferVD = |
3053 | Builder.CreateInBoundsGEP(Ty: ReductionsBufferTy, Ptr: BufferVal, IdxList: Idxs); |
3054 | // Global = Buffer.VD[Idx]; |
3055 | Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32( |
3056 | Ty: ReductionsBufferTy, Ptr: BufferVD, Idx0: 0, Idx1: En.index()); |
3057 | Builder.CreateStore(Val: GlobValPtr, Ptr: TargetElementPtrPtr); |
3058 | } |
3059 | |
3060 | // Call reduce_function(GlobalReduceList, ReduceList) |
3061 | Value *ReduceList = |
3062 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ReduceListArgAddrCast); |
3063 | Builder.CreateCall(Callee: ReduceFn, Args: {LocalReduceListAddrCast, ReduceList}) |
3064 | ->addFnAttr(Kind: Attribute::NoUnwind); |
3065 | Builder.CreateRetVoid(); |
3066 | Builder.restoreIP(IP: OldIP); |
3067 | return LtGRFunc; |
3068 | } |
3069 | |
3070 | Function *OpenMPIRBuilder::emitGlobalToListCopyFunction( |
3071 | ArrayRef<ReductionInfo> ReductionInfos, Type *ReductionsBufferTy, |
3072 | AttributeList FuncAttrs) { |
3073 | OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP(); |
3074 | LLVMContext &Ctx = M.getContext(); |
3075 | FunctionType *FuncTy = FunctionType::get( |
3076 | Result: Builder.getVoidTy(), |
3077 | Params: {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()}, |
3078 | /* IsVarArg */ isVarArg: false); |
3079 | Function *LtGCFunc = |
3080 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
3081 | N: "_omp_reduction_global_to_list_copy_func" , M: &M); |
3082 | LtGCFunc->setAttributes(FuncAttrs); |
3083 | LtGCFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
3084 | LtGCFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
3085 | LtGCFunc->addParamAttr(ArgNo: 2, Kind: Attribute::NoUndef); |
3086 | |
3087 | BasicBlock *EntryBlock = BasicBlock::Create(Context&: Ctx, Name: "entry" , Parent: LtGCFunc); |
3088 | Builder.SetInsertPoint(EntryBlock); |
3089 | |
3090 | // Buffer: global reduction buffer. |
3091 | Argument *BufferArg = LtGCFunc->getArg(i: 0); |
3092 | // Idx: index of the buffer. |
3093 | Argument *IdxArg = LtGCFunc->getArg(i: 1); |
3094 | // ReduceList: thread local Reduce list. |
3095 | Argument *ReduceListArg = LtGCFunc->getArg(i: 2); |
3096 | |
3097 | Value *BufferArgAlloca = Builder.CreateAlloca(Ty: Builder.getPtrTy(), ArraySize: nullptr, |
3098 | Name: BufferArg->getName() + ".addr" ); |
3099 | Value *IdxArgAlloca = Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, |
3100 | Name: IdxArg->getName() + ".addr" ); |
3101 | Value *ReduceListArgAlloca = Builder.CreateAlloca( |
3102 | Ty: Builder.getPtrTy(), ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
3103 | Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3104 | V: BufferArgAlloca, DestTy: Builder.getPtrTy(), |
3105 | Name: BufferArgAlloca->getName() + ".ascast" ); |
3106 | Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3107 | V: IdxArgAlloca, DestTy: Builder.getPtrTy(), Name: IdxArgAlloca->getName() + ".ascast" ); |
3108 | Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3109 | V: ReduceListArgAlloca, DestTy: Builder.getPtrTy(), |
3110 | Name: ReduceListArgAlloca->getName() + ".ascast" ); |
3111 | Builder.CreateStore(Val: BufferArg, Ptr: BufferArgAddrCast); |
3112 | Builder.CreateStore(Val: IdxArg, Ptr: IdxArgAddrCast); |
3113 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListArgAddrCast); |
3114 | |
3115 | Value *LocalReduceList = |
3116 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ReduceListArgAddrCast); |
3117 | Value *BufferVal = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: BufferArgAddrCast); |
3118 | Value *Idxs[] = {Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: IdxArgAddrCast)}; |
3119 | Type *IndexTy = Builder.getIndexTy( |
3120 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
3121 | for (auto En : enumerate(First&: ReductionInfos)) { |
3122 | const OpenMPIRBuilder::ReductionInfo &RI = En.value(); |
3123 | auto *RedListArrayTy = |
3124 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
3125 | // Reduce element = LocalReduceList[i] |
3126 | Value *ElemPtrPtr = Builder.CreateInBoundsGEP( |
3127 | Ty: RedListArrayTy, Ptr: LocalReduceList, |
3128 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3129 | // elemptr = ((CopyType*)(elemptrptr)) + I |
3130 | Value *ElemPtr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ElemPtrPtr); |
3131 | // Global = Buffer.VD[Idx]; |
3132 | Value *BufferVD = |
3133 | Builder.CreateInBoundsGEP(Ty: ReductionsBufferTy, Ptr: BufferVal, IdxList: Idxs); |
3134 | Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32( |
3135 | Ty: ReductionsBufferTy, Ptr: BufferVD, Idx0: 0, Idx1: En.index()); |
3136 | |
3137 | switch (RI.EvaluationKind) { |
3138 | case EvalKind::Scalar: { |
3139 | Value *TargetElement = Builder.CreateLoad(Ty: RI.ElementType, Ptr: GlobValPtr); |
3140 | Builder.CreateStore(Val: TargetElement, Ptr: ElemPtr); |
3141 | break; |
3142 | } |
3143 | case EvalKind::Complex: { |
3144 | Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32( |
3145 | Ty: RI.ElementType, Ptr: GlobValPtr, Idx0: 0, Idx1: 0, Name: ".realp" ); |
3146 | Value *SrcReal = Builder.CreateLoad( |
3147 | Ty: RI.ElementType->getStructElementType(N: 0), Ptr: SrcRealPtr, Name: ".real" ); |
3148 | Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32( |
3149 | Ty: RI.ElementType, Ptr: GlobValPtr, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
3150 | Value *SrcImg = Builder.CreateLoad( |
3151 | Ty: RI.ElementType->getStructElementType(N: 1), Ptr: SrcImgPtr, Name: ".imag" ); |
3152 | |
3153 | Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32( |
3154 | Ty: RI.ElementType, Ptr: ElemPtr, Idx0: 0, Idx1: 0, Name: ".realp" ); |
3155 | Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32( |
3156 | Ty: RI.ElementType, Ptr: ElemPtr, Idx0: 0, Idx1: 1, Name: ".imagp" ); |
3157 | Builder.CreateStore(Val: SrcReal, Ptr: DestRealPtr); |
3158 | Builder.CreateStore(Val: SrcImg, Ptr: DestImgPtr); |
3159 | break; |
3160 | } |
3161 | case EvalKind::Aggregate: { |
3162 | Value *SizeVal = |
3163 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: RI.ElementType)); |
3164 | Builder.CreateMemCpy( |
3165 | Dst: ElemPtr, DstAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), |
3166 | Src: GlobValPtr, SrcAlign: M.getDataLayout().getPrefTypeAlign(Ty: RI.ElementType), |
3167 | Size: SizeVal, isVolatile: false); |
3168 | break; |
3169 | } |
3170 | } |
3171 | } |
3172 | |
3173 | Builder.CreateRetVoid(); |
3174 | Builder.restoreIP(IP: OldIP); |
3175 | return LtGCFunc; |
3176 | } |
3177 | |
3178 | Function *OpenMPIRBuilder::emitGlobalToListReduceFunction( |
3179 | ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn, |
3180 | Type *ReductionsBufferTy, AttributeList FuncAttrs) { |
3181 | OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP(); |
3182 | LLVMContext &Ctx = M.getContext(); |
3183 | auto *FuncTy = FunctionType::get( |
3184 | Result: Builder.getVoidTy(), |
3185 | Params: {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()}, |
3186 | /* IsVarArg */ isVarArg: false); |
3187 | Function *LtGRFunc = |
3188 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
3189 | N: "_omp_reduction_global_to_list_reduce_func" , M: &M); |
3190 | LtGRFunc->setAttributes(FuncAttrs); |
3191 | LtGRFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
3192 | LtGRFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
3193 | LtGRFunc->addParamAttr(ArgNo: 2, Kind: Attribute::NoUndef); |
3194 | |
3195 | BasicBlock *EntryBlock = BasicBlock::Create(Context&: Ctx, Name: "entry" , Parent: LtGRFunc); |
3196 | Builder.SetInsertPoint(EntryBlock); |
3197 | |
3198 | // Buffer: global reduction buffer. |
3199 | Argument *BufferArg = LtGRFunc->getArg(i: 0); |
3200 | // Idx: index of the buffer. |
3201 | Argument *IdxArg = LtGRFunc->getArg(i: 1); |
3202 | // ReduceList: thread local Reduce list. |
3203 | Argument *ReduceListArg = LtGRFunc->getArg(i: 2); |
3204 | |
3205 | Value *BufferArgAlloca = Builder.CreateAlloca(Ty: Builder.getPtrTy(), ArraySize: nullptr, |
3206 | Name: BufferArg->getName() + ".addr" ); |
3207 | Value *IdxArgAlloca = Builder.CreateAlloca(Ty: Builder.getInt32Ty(), ArraySize: nullptr, |
3208 | Name: IdxArg->getName() + ".addr" ); |
3209 | Value *ReduceListArgAlloca = Builder.CreateAlloca( |
3210 | Ty: Builder.getPtrTy(), ArraySize: nullptr, Name: ReduceListArg->getName() + ".addr" ); |
3211 | ArrayType *RedListArrayTy = |
3212 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
3213 | |
3214 | // 1. Build a list of reduction variables. |
3215 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
3216 | Value *LocalReduceList = |
3217 | Builder.CreateAlloca(Ty: RedListArrayTy, ArraySize: nullptr, Name: ".omp.reduction.red_list" ); |
3218 | |
3219 | Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3220 | V: BufferArgAlloca, DestTy: Builder.getPtrTy(), |
3221 | Name: BufferArgAlloca->getName() + ".ascast" ); |
3222 | Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3223 | V: IdxArgAlloca, DestTy: Builder.getPtrTy(), Name: IdxArgAlloca->getName() + ".ascast" ); |
3224 | Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3225 | V: ReduceListArgAlloca, DestTy: Builder.getPtrTy(), |
3226 | Name: ReduceListArgAlloca->getName() + ".ascast" ); |
3227 | Value *ReductionList = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3228 | V: LocalReduceList, DestTy: Builder.getPtrTy(), |
3229 | Name: LocalReduceList->getName() + ".ascast" ); |
3230 | |
3231 | Builder.CreateStore(Val: BufferArg, Ptr: BufferArgAddrCast); |
3232 | Builder.CreateStore(Val: IdxArg, Ptr: IdxArgAddrCast); |
3233 | Builder.CreateStore(Val: ReduceListArg, Ptr: ReduceListArgAddrCast); |
3234 | |
3235 | Value *BufferVal = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: BufferArgAddrCast); |
3236 | Value *Idxs[] = {Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: IdxArgAddrCast)}; |
3237 | Type *IndexTy = Builder.getIndexTy( |
3238 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
3239 | for (auto En : enumerate(First&: ReductionInfos)) { |
3240 | Value *TargetElementPtrPtr = Builder.CreateInBoundsGEP( |
3241 | Ty: RedListArrayTy, Ptr: ReductionList, |
3242 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3243 | // Global = Buffer.VD[Idx]; |
3244 | Value *BufferVD = |
3245 | Builder.CreateInBoundsGEP(Ty: ReductionsBufferTy, Ptr: BufferVal, IdxList: Idxs); |
3246 | Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32( |
3247 | Ty: ReductionsBufferTy, Ptr: BufferVD, Idx0: 0, Idx1: En.index()); |
3248 | Builder.CreateStore(Val: GlobValPtr, Ptr: TargetElementPtrPtr); |
3249 | } |
3250 | |
3251 | // Call reduce_function(ReduceList, GlobalReduceList) |
3252 | Value *ReduceList = |
3253 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: ReduceListArgAddrCast); |
3254 | Builder.CreateCall(Callee: ReduceFn, Args: {ReduceList, ReductionList}) |
3255 | ->addFnAttr(Kind: Attribute::NoUnwind); |
3256 | Builder.CreateRetVoid(); |
3257 | Builder.restoreIP(IP: OldIP); |
3258 | return LtGRFunc; |
3259 | } |
3260 | |
3261 | std::string OpenMPIRBuilder::getReductionFuncName(StringRef Name) const { |
3262 | std::string Suffix = |
3263 | createPlatformSpecificName(Parts: {"omp" , "reduction" , "reduction_func" }); |
3264 | return (Name + Suffix).str(); |
3265 | } |
3266 | |
3267 | Function *OpenMPIRBuilder::createReductionFunction( |
3268 | StringRef ReducerName, ArrayRef<ReductionInfo> ReductionInfos, |
3269 | ReductionGenCBKind ReductionGenCBKind, AttributeList FuncAttrs) { |
3270 | auto *FuncTy = FunctionType::get(Result: Builder.getVoidTy(), |
3271 | Params: {Builder.getPtrTy(), Builder.getPtrTy()}, |
3272 | /* IsVarArg */ isVarArg: false); |
3273 | std::string Name = getReductionFuncName(Name: ReducerName); |
3274 | Function *ReductionFunc = |
3275 | Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, N: Name, M: &M); |
3276 | ReductionFunc->setAttributes(FuncAttrs); |
3277 | ReductionFunc->addParamAttr(ArgNo: 0, Kind: Attribute::NoUndef); |
3278 | ReductionFunc->addParamAttr(ArgNo: 1, Kind: Attribute::NoUndef); |
3279 | BasicBlock *EntryBB = |
3280 | BasicBlock::Create(Context&: M.getContext(), Name: "entry" , Parent: ReductionFunc); |
3281 | Builder.SetInsertPoint(EntryBB); |
3282 | |
3283 | // Need to alloca memory here and deal with the pointers before getting |
3284 | // LHS/RHS pointers out |
3285 | Value *LHSArrayPtr = nullptr; |
3286 | Value *RHSArrayPtr = nullptr; |
3287 | Argument *Arg0 = ReductionFunc->getArg(i: 0); |
3288 | Argument *Arg1 = ReductionFunc->getArg(i: 1); |
3289 | Type *Arg0Type = Arg0->getType(); |
3290 | Type *Arg1Type = Arg1->getType(); |
3291 | |
3292 | Value *LHSAlloca = |
3293 | Builder.CreateAlloca(Ty: Arg0Type, ArraySize: nullptr, Name: Arg0->getName() + ".addr" ); |
3294 | Value *RHSAlloca = |
3295 | Builder.CreateAlloca(Ty: Arg1Type, ArraySize: nullptr, Name: Arg1->getName() + ".addr" ); |
3296 | Value *LHSAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3297 | V: LHSAlloca, DestTy: Arg0Type, Name: LHSAlloca->getName() + ".ascast" ); |
3298 | Value *RHSAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3299 | V: RHSAlloca, DestTy: Arg1Type, Name: RHSAlloca->getName() + ".ascast" ); |
3300 | Builder.CreateStore(Val: Arg0, Ptr: LHSAddrCast); |
3301 | Builder.CreateStore(Val: Arg1, Ptr: RHSAddrCast); |
3302 | LHSArrayPtr = Builder.CreateLoad(Ty: Arg0Type, Ptr: LHSAddrCast); |
3303 | RHSArrayPtr = Builder.CreateLoad(Ty: Arg1Type, Ptr: RHSAddrCast); |
3304 | |
3305 | Type *RedArrayTy = ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: ReductionInfos.size()); |
3306 | Type *IndexTy = Builder.getIndexTy( |
3307 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
3308 | SmallVector<Value *> LHSPtrs, RHSPtrs; |
3309 | for (auto En : enumerate(First&: ReductionInfos)) { |
3310 | const ReductionInfo &RI = En.value(); |
3311 | Value *RHSI8PtrPtr = Builder.CreateInBoundsGEP( |
3312 | Ty: RedArrayTy, Ptr: RHSArrayPtr, |
3313 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3314 | Value *RHSI8Ptr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: RHSI8PtrPtr); |
3315 | Value *RHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3316 | V: RHSI8Ptr, DestTy: RI.PrivateVariable->getType(), |
3317 | Name: RHSI8Ptr->getName() + ".ascast" ); |
3318 | |
3319 | Value *LHSI8PtrPtr = Builder.CreateInBoundsGEP( |
3320 | Ty: RedArrayTy, Ptr: LHSArrayPtr, |
3321 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3322 | Value *LHSI8Ptr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: LHSI8PtrPtr); |
3323 | Value *LHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3324 | V: LHSI8Ptr, DestTy: RI.Variable->getType(), Name: LHSI8Ptr->getName() + ".ascast" ); |
3325 | |
3326 | if (ReductionGenCBKind == ReductionGenCBKind::Clang) { |
3327 | LHSPtrs.emplace_back(Args&: LHSPtr); |
3328 | RHSPtrs.emplace_back(Args&: RHSPtr); |
3329 | } else { |
3330 | Value *LHS = Builder.CreateLoad(Ty: RI.ElementType, Ptr: LHSPtr); |
3331 | Value *RHS = Builder.CreateLoad(Ty: RI.ElementType, Ptr: RHSPtr); |
3332 | Value *Reduced; |
3333 | RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced); |
3334 | if (!Builder.GetInsertBlock()) |
3335 | return ReductionFunc; |
3336 | Builder.CreateStore(Val: Reduced, Ptr: LHSPtr); |
3337 | } |
3338 | } |
3339 | |
3340 | if (ReductionGenCBKind == ReductionGenCBKind::Clang) |
3341 | for (auto En : enumerate(First&: ReductionInfos)) { |
3342 | unsigned Index = En.index(); |
3343 | const ReductionInfo &RI = En.value(); |
3344 | Value *LHSFixupPtr, *RHSFixupPtr; |
3345 | Builder.restoreIP(IP: RI.ReductionGenClang( |
3346 | Builder.saveIP(), Index, &LHSFixupPtr, &RHSFixupPtr, ReductionFunc)); |
3347 | |
3348 | // Fix the CallBack code genereated to use the correct Values for the LHS |
3349 | // and RHS |
3350 | LHSFixupPtr->replaceUsesWithIf( |
3351 | New: LHSPtrs[Index], ShouldReplace: [ReductionFunc](const Use &U) { |
3352 | return cast<Instruction>(Val: U.getUser())->getParent()->getParent() == |
3353 | ReductionFunc; |
3354 | }); |
3355 | RHSFixupPtr->replaceUsesWithIf( |
3356 | New: RHSPtrs[Index], ShouldReplace: [ReductionFunc](const Use &U) { |
3357 | return cast<Instruction>(Val: U.getUser())->getParent()->getParent() == |
3358 | ReductionFunc; |
3359 | }); |
3360 | } |
3361 | |
3362 | Builder.CreateRetVoid(); |
3363 | return ReductionFunc; |
3364 | } |
3365 | |
3366 | static void |
3367 | checkReductionInfos(ArrayRef<OpenMPIRBuilder::ReductionInfo> ReductionInfos, |
3368 | bool IsGPU) { |
3369 | for (const OpenMPIRBuilder::ReductionInfo &RI : ReductionInfos) { |
3370 | (void)RI; |
3371 | assert(RI.Variable && "expected non-null variable" ); |
3372 | assert(RI.PrivateVariable && "expected non-null private variable" ); |
3373 | assert((RI.ReductionGen || RI.ReductionGenClang) && |
3374 | "expected non-null reduction generator callback" ); |
3375 | if (!IsGPU) { |
3376 | assert( |
3377 | RI.Variable->getType() == RI.PrivateVariable->getType() && |
3378 | "expected variables and their private equivalents to have the same " |
3379 | "type" ); |
3380 | } |
3381 | assert(RI.Variable->getType()->isPointerTy() && |
3382 | "expected variables to be pointers" ); |
3383 | } |
3384 | } |
3385 | |
3386 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createReductionsGPU( |
3387 | const LocationDescription &Loc, InsertPointTy AllocaIP, |
3388 | InsertPointTy CodeGenIP, ArrayRef<ReductionInfo> ReductionInfos, |
3389 | bool IsNoWait, bool IsTeamsReduction, bool HasDistribute, |
3390 | ReductionGenCBKind ReductionGenCBKind, std::optional<omp::GV> GridValue, |
3391 | unsigned ReductionBufNum, Value *SrcLocInfo) { |
3392 | if (!updateToLocation(Loc)) |
3393 | return InsertPointTy(); |
3394 | Builder.restoreIP(IP: CodeGenIP); |
3395 | checkReductionInfos(ReductionInfos, /*IsGPU*/ true); |
3396 | LLVMContext &Ctx = M.getContext(); |
3397 | |
3398 | // Source location for the ident struct |
3399 | if (!SrcLocInfo) { |
3400 | uint32_t SrcLocStrSize; |
3401 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
3402 | SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
3403 | } |
3404 | |
3405 | if (ReductionInfos.size() == 0) |
3406 | return Builder.saveIP(); |
3407 | |
3408 | Function *CurFunc = Builder.GetInsertBlock()->getParent(); |
3409 | AttributeList FuncAttrs; |
3410 | AttrBuilder AttrBldr(Ctx); |
3411 | for (auto Attr : CurFunc->getAttributes().getFnAttrs()) |
3412 | AttrBldr.addAttribute(A: Attr); |
3413 | AttrBldr.removeAttribute(Val: Attribute::OptimizeNone); |
3414 | FuncAttrs = FuncAttrs.addFnAttributes(C&: Ctx, B: AttrBldr); |
3415 | |
3416 | Function *ReductionFunc = nullptr; |
3417 | CodeGenIP = Builder.saveIP(); |
3418 | ReductionFunc = |
3419 | createReductionFunction(ReducerName: Builder.GetInsertBlock()->getParent()->getName(), |
3420 | ReductionInfos, ReductionGenCBKind, FuncAttrs); |
3421 | Builder.restoreIP(IP: CodeGenIP); |
3422 | |
3423 | // Set the grid value in the config needed for lowering later on |
3424 | if (GridValue.has_value()) |
3425 | Config.setGridValue(GridValue.value()); |
3426 | else |
3427 | Config.setGridValue(getGridValue(T, Kernel: ReductionFunc)); |
3428 | |
3429 | uint32_t SrcLocStrSize; |
3430 | Constant *SrcLocStr = getOrCreateDefaultSrcLocStr(SrcLocStrSize); |
3431 | Value *RTLoc = |
3432 | getOrCreateIdent(SrcLocStr, SrcLocStrSize, LocFlags: omp::IdentFlag(0), Reserve2Flags: 0); |
3433 | |
3434 | // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), |
3435 | // RedList, shuffle_reduce_func, interwarp_copy_func); |
3436 | // or |
3437 | // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); |
3438 | Value *Res; |
3439 | |
3440 | // 1. Build a list of reduction variables. |
3441 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
3442 | auto Size = ReductionInfos.size(); |
3443 | Type *PtrTy = PointerType::getUnqual(C&: Ctx); |
3444 | Type *RedArrayTy = ArrayType::get(ElementType: PtrTy, NumElements: Size); |
3445 | CodeGenIP = Builder.saveIP(); |
3446 | Builder.restoreIP(IP: AllocaIP); |
3447 | Value *ReductionListAlloca = |
3448 | Builder.CreateAlloca(Ty: RedArrayTy, ArraySize: nullptr, Name: ".omp.reduction.red_list" ); |
3449 | Value *ReductionList = Builder.CreatePointerBitCastOrAddrSpaceCast( |
3450 | V: ReductionListAlloca, DestTy: PtrTy, Name: ReductionListAlloca->getName() + ".ascast" ); |
3451 | Builder.restoreIP(IP: CodeGenIP); |
3452 | Type *IndexTy = Builder.getIndexTy( |
3453 | DL: M.getDataLayout(), AddrSpace: M.getDataLayout().getDefaultGlobalsAddressSpace()); |
3454 | for (auto En : enumerate(First&: ReductionInfos)) { |
3455 | const ReductionInfo &RI = En.value(); |
3456 | Value *ElemPtr = Builder.CreateInBoundsGEP( |
3457 | Ty: RedArrayTy, Ptr: ReductionList, |
3458 | IdxList: {ConstantInt::get(Ty: IndexTy, V: 0), ConstantInt::get(Ty: IndexTy, V: En.index())}); |
3459 | Value *CastElem = |
3460 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: RI.PrivateVariable, DestTy: PtrTy); |
3461 | Builder.CreateStore(Val: CastElem, Ptr: ElemPtr); |
3462 | } |
3463 | CodeGenIP = Builder.saveIP(); |
3464 | Function *SarFunc = |
3465 | emitShuffleAndReduceFunction(ReductionInfos, ReduceFn: ReductionFunc, FuncAttrs); |
3466 | Function *WcFunc = emitInterWarpCopyFunction(Loc, ReductionInfos, FuncAttrs); |
3467 | Builder.restoreIP(IP: CodeGenIP); |
3468 | |
3469 | Value *RL = Builder.CreatePointerBitCastOrAddrSpaceCast(V: ReductionList, DestTy: PtrTy); |
3470 | |
3471 | unsigned MaxDataSize = 0; |
3472 | SmallVector<Type *> ReductionTypeArgs; |
3473 | for (auto En : enumerate(First&: ReductionInfos)) { |
3474 | auto Size = M.getDataLayout().getTypeStoreSize(Ty: En.value().ElementType); |
3475 | if (Size > MaxDataSize) |
3476 | MaxDataSize = Size; |
3477 | ReductionTypeArgs.emplace_back(Args: En.value().ElementType); |
3478 | } |
3479 | Value *ReductionDataSize = |
3480 | Builder.getInt64(C: MaxDataSize * ReductionInfos.size()); |
3481 | if (!IsTeamsReduction) { |
3482 | Value *SarFuncCast = |
3483 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: SarFunc, DestTy: PtrTy); |
3484 | Value *WcFuncCast = |
3485 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: WcFunc, DestTy: PtrTy); |
3486 | Value *Args[] = {RTLoc, ReductionDataSize, RL, SarFuncCast, WcFuncCast}; |
3487 | Function *Pv2Ptr = getOrCreateRuntimeFunctionPtr( |
3488 | FnID: RuntimeFunction::OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2); |
3489 | Res = Builder.CreateCall(Callee: Pv2Ptr, Args); |
3490 | } else { |
3491 | CodeGenIP = Builder.saveIP(); |
3492 | StructType *ReductionsBufferTy = StructType::create( |
3493 | Context&: Ctx, Elements: ReductionTypeArgs, Name: "struct._globalized_locals_ty" ); |
3494 | Function *RedFixedBuferFn = getOrCreateRuntimeFunctionPtr( |
3495 | FnID: RuntimeFunction::OMPRTL___kmpc_reduction_get_fixed_buffer); |
3496 | Function *LtGCFunc = emitListToGlobalCopyFunction( |
3497 | ReductionInfos, ReductionsBufferTy, FuncAttrs); |
3498 | Function *LtGRFunc = emitListToGlobalReduceFunction( |
3499 | ReductionInfos, ReduceFn: ReductionFunc, ReductionsBufferTy, FuncAttrs); |
3500 | Function *GtLCFunc = emitGlobalToListCopyFunction( |
3501 | ReductionInfos, ReductionsBufferTy, FuncAttrs); |
3502 | Function *GtLRFunc = emitGlobalToListReduceFunction( |
3503 | ReductionInfos, ReduceFn: ReductionFunc, ReductionsBufferTy, FuncAttrs); |
3504 | Builder.restoreIP(IP: CodeGenIP); |
3505 | |
3506 | Value *KernelTeamsReductionPtr = Builder.CreateCall( |
3507 | Callee: RedFixedBuferFn, Args: {}, Name: "_openmp_teams_reductions_buffer_$_$ptr" ); |
3508 | |
3509 | Value *Args3[] = {RTLoc, |
3510 | KernelTeamsReductionPtr, |
3511 | Builder.getInt32(C: ReductionBufNum), |
3512 | ReductionDataSize, |
3513 | RL, |
3514 | SarFunc, |
3515 | WcFunc, |
3516 | LtGCFunc, |
3517 | LtGRFunc, |
3518 | GtLCFunc, |
3519 | GtLRFunc}; |
3520 | |
3521 | Function *TeamsReduceFn = getOrCreateRuntimeFunctionPtr( |
3522 | FnID: RuntimeFunction::OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2); |
3523 | Res = Builder.CreateCall(Callee: TeamsReduceFn, Args: Args3); |
3524 | } |
3525 | |
3526 | // 5. Build if (res == 1) |
3527 | BasicBlock *ExitBB = BasicBlock::Create(Context&: Ctx, Name: ".omp.reduction.done" ); |
3528 | BasicBlock *ThenBB = BasicBlock::Create(Context&: Ctx, Name: ".omp.reduction.then" ); |
3529 | Value *Cond = Builder.CreateICmpEQ(LHS: Res, RHS: Builder.getInt32(C: 1)); |
3530 | Builder.CreateCondBr(Cond, True: ThenBB, False: ExitBB); |
3531 | |
3532 | // 6. Build then branch: where we have reduced values in the master |
3533 | // thread in each team. |
3534 | // __kmpc_end_reduce{_nowait}(<gtid>); |
3535 | // break; |
3536 | emitBlock(BB: ThenBB, CurFn: CurFunc); |
3537 | |
3538 | // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); |
3539 | for (auto En : enumerate(First&: ReductionInfos)) { |
3540 | const ReductionInfo &RI = En.value(); |
3541 | Value *LHS = RI.Variable; |
3542 | Value *RHS = |
3543 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: RI.PrivateVariable, DestTy: PtrTy); |
3544 | |
3545 | if (ReductionGenCBKind == ReductionGenCBKind::Clang) { |
3546 | Value *LHSPtr, *RHSPtr; |
3547 | Builder.restoreIP(IP: RI.ReductionGenClang(Builder.saveIP(), En.index(), |
3548 | &LHSPtr, &RHSPtr, CurFunc)); |
3549 | |
3550 | // Fix the CallBack code genereated to use the correct Values for the LHS |
3551 | // and RHS |
3552 | LHSPtr->replaceUsesWithIf(New: LHS, ShouldReplace: [ReductionFunc](const Use &U) { |
3553 | return cast<Instruction>(Val: U.getUser())->getParent()->getParent() == |
3554 | ReductionFunc; |
3555 | }); |
3556 | RHSPtr->replaceUsesWithIf(New: RHS, ShouldReplace: [ReductionFunc](const Use &U) { |
3557 | return cast<Instruction>(Val: U.getUser())->getParent()->getParent() == |
3558 | ReductionFunc; |
3559 | }); |
3560 | } else { |
3561 | assert(false && "Unhandled ReductionGenCBKind" ); |
3562 | } |
3563 | } |
3564 | emitBlock(BB: ExitBB, CurFn: CurFunc); |
3565 | |
3566 | Config.setEmitLLVMUsed(); |
3567 | |
3568 | return Builder.saveIP(); |
3569 | } |
3570 | |
3571 | static Function *getFreshReductionFunc(Module &M) { |
3572 | Type *VoidTy = Type::getVoidTy(C&: M.getContext()); |
3573 | Type *Int8PtrTy = PointerType::getUnqual(C&: M.getContext()); |
3574 | auto *FuncTy = |
3575 | FunctionType::get(Result: VoidTy, Params: {Int8PtrTy, Int8PtrTy}, /* IsVarArg */ isVarArg: false); |
3576 | return Function::Create(Ty: FuncTy, Linkage: GlobalVariable::InternalLinkage, |
3577 | N: ".omp.reduction.func" , M: &M); |
3578 | } |
3579 | |
3580 | OpenMPIRBuilder::InsertPointTy |
3581 | OpenMPIRBuilder::createReductions(const LocationDescription &Loc, |
3582 | InsertPointTy AllocaIP, |
3583 | ArrayRef<ReductionInfo> ReductionInfos, |
3584 | ArrayRef<bool> IsByRef, bool IsNoWait) { |
3585 | assert(ReductionInfos.size() == IsByRef.size()); |
3586 | for (const ReductionInfo &RI : ReductionInfos) { |
3587 | (void)RI; |
3588 | assert(RI.Variable && "expected non-null variable" ); |
3589 | assert(RI.PrivateVariable && "expected non-null private variable" ); |
3590 | assert(RI.ReductionGen && "expected non-null reduction generator callback" ); |
3591 | assert(RI.Variable->getType() == RI.PrivateVariable->getType() && |
3592 | "expected variables and their private equivalents to have the same " |
3593 | "type" ); |
3594 | assert(RI.Variable->getType()->isPointerTy() && |
3595 | "expected variables to be pointers" ); |
3596 | } |
3597 | |
3598 | if (!updateToLocation(Loc)) |
3599 | return InsertPointTy(); |
3600 | |
3601 | BasicBlock *InsertBlock = Loc.IP.getBlock(); |
3602 | BasicBlock *ContinuationBlock = |
3603 | InsertBlock->splitBasicBlock(I: Loc.IP.getPoint(), BBName: "reduce.finalize" ); |
3604 | InsertBlock->getTerminator()->eraseFromParent(); |
3605 | |
3606 | // Create and populate array of type-erased pointers to private reduction |
3607 | // values. |
3608 | unsigned NumReductions = ReductionInfos.size(); |
3609 | Type *RedArrayTy = ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: NumReductions); |
3610 | Builder.SetInsertPoint(AllocaIP.getBlock()->getTerminator()); |
3611 | Value *RedArray = Builder.CreateAlloca(Ty: RedArrayTy, ArraySize: nullptr, Name: "red.array" ); |
3612 | |
3613 | Builder.SetInsertPoint(TheBB: InsertBlock, IP: InsertBlock->end()); |
3614 | |
3615 | for (auto En : enumerate(First&: ReductionInfos)) { |
3616 | unsigned Index = En.index(); |
3617 | const ReductionInfo &RI = En.value(); |
3618 | Value *RedArrayElemPtr = Builder.CreateConstInBoundsGEP2_64( |
3619 | Ty: RedArrayTy, Ptr: RedArray, Idx0: 0, Idx1: Index, Name: "red.array.elem." + Twine(Index)); |
3620 | Builder.CreateStore(Val: RI.PrivateVariable, Ptr: RedArrayElemPtr); |
3621 | } |
3622 | |
3623 | // Emit a call to the runtime function that orchestrates the reduction. |
3624 | // Declare the reduction function in the process. |
3625 | Function *Func = Builder.GetInsertBlock()->getParent(); |
3626 | Module *Module = Func->getParent(); |
3627 | uint32_t SrcLocStrSize; |
3628 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
3629 | bool CanGenerateAtomic = all_of(Range&: ReductionInfos, P: [](const ReductionInfo &RI) { |
3630 | return RI.AtomicReductionGen; |
3631 | }); |
3632 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize, |
3633 | LocFlags: CanGenerateAtomic |
3634 | ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE |
3635 | : IdentFlag(0)); |
3636 | Value *ThreadId = getOrCreateThreadID(Ident); |
3637 | Constant *NumVariables = Builder.getInt32(C: NumReductions); |
3638 | const DataLayout &DL = Module->getDataLayout(); |
3639 | unsigned RedArrayByteSize = DL.getTypeStoreSize(Ty: RedArrayTy); |
3640 | Constant *RedArraySize = Builder.getInt64(C: RedArrayByteSize); |
3641 | Function *ReductionFunc = getFreshReductionFunc(M&: *Module); |
3642 | Value *Lock = getOMPCriticalRegionLock(CriticalName: ".reduction" ); |
3643 | Function *ReduceFunc = getOrCreateRuntimeFunctionPtr( |
3644 | FnID: IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait |
3645 | : RuntimeFunction::OMPRTL___kmpc_reduce); |
3646 | CallInst *ReduceCall = |
3647 | Builder.CreateCall(Callee: ReduceFunc, |
3648 | Args: {Ident, ThreadId, NumVariables, RedArraySize, RedArray, |
3649 | ReductionFunc, Lock}, |
3650 | Name: "reduce" ); |
3651 | |
3652 | // Create final reduction entry blocks for the atomic and non-atomic case. |
3653 | // Emit IR that dispatches control flow to one of the blocks based on the |
3654 | // reduction supporting the atomic mode. |
3655 | BasicBlock *NonAtomicRedBlock = |
3656 | BasicBlock::Create(Context&: Module->getContext(), Name: "reduce.switch.nonatomic" , Parent: Func); |
3657 | BasicBlock *AtomicRedBlock = |
3658 | BasicBlock::Create(Context&: Module->getContext(), Name: "reduce.switch.atomic" , Parent: Func); |
3659 | SwitchInst *Switch = |
3660 | Builder.CreateSwitch(V: ReduceCall, Dest: ContinuationBlock, /* NumCases */ 2); |
3661 | Switch->addCase(OnVal: Builder.getInt32(C: 1), Dest: NonAtomicRedBlock); |
3662 | Switch->addCase(OnVal: Builder.getInt32(C: 2), Dest: AtomicRedBlock); |
3663 | |
3664 | // Populate the non-atomic reduction using the elementwise reduction function. |
3665 | // This loads the elements from the global and private variables and reduces |
3666 | // them before storing back the result to the global variable. |
3667 | Builder.SetInsertPoint(NonAtomicRedBlock); |
3668 | for (auto En : enumerate(First&: ReductionInfos)) { |
3669 | const ReductionInfo &RI = En.value(); |
3670 | Type *ValueType = RI.ElementType; |
3671 | // We have one less load for by-ref case because that load is now inside of |
3672 | // the reduction region |
3673 | Value *RedValue = nullptr; |
3674 | if (!IsByRef[En.index()]) { |
3675 | RedValue = Builder.CreateLoad(Ty: ValueType, Ptr: RI.Variable, |
3676 | Name: "red.value." + Twine(En.index())); |
3677 | } |
3678 | Value *PrivateRedValue = |
3679 | Builder.CreateLoad(Ty: ValueType, Ptr: RI.PrivateVariable, |
3680 | Name: "red.private.value." + Twine(En.index())); |
3681 | Value *Reduced; |
3682 | if (IsByRef[En.index()]) { |
3683 | Builder.restoreIP(IP: RI.ReductionGen(Builder.saveIP(), RI.Variable, |
3684 | PrivateRedValue, Reduced)); |
3685 | } else { |
3686 | Builder.restoreIP(IP: RI.ReductionGen(Builder.saveIP(), RedValue, |
3687 | PrivateRedValue, Reduced)); |
3688 | } |
3689 | if (!Builder.GetInsertBlock()) |
3690 | return InsertPointTy(); |
3691 | // for by-ref case, the load is inside of the reduction region |
3692 | if (!IsByRef[En.index()]) |
3693 | Builder.CreateStore(Val: Reduced, Ptr: RI.Variable); |
3694 | } |
3695 | Function *EndReduceFunc = getOrCreateRuntimeFunctionPtr( |
3696 | FnID: IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait |
3697 | : RuntimeFunction::OMPRTL___kmpc_end_reduce); |
3698 | Builder.CreateCall(Callee: EndReduceFunc, Args: {Ident, ThreadId, Lock}); |
3699 | Builder.CreateBr(Dest: ContinuationBlock); |
3700 | |
3701 | // Populate the atomic reduction using the atomic elementwise reduction |
3702 | // function. There are no loads/stores here because they will be happening |
3703 | // inside the atomic elementwise reduction. |
3704 | Builder.SetInsertPoint(AtomicRedBlock); |
3705 | if (CanGenerateAtomic && llvm::none_of(Range&: IsByRef, P: [](bool P) { return P; })) { |
3706 | for (const ReductionInfo &RI : ReductionInfos) { |
3707 | Builder.restoreIP(IP: RI.AtomicReductionGen(Builder.saveIP(), RI.ElementType, |
3708 | RI.Variable, RI.PrivateVariable)); |
3709 | if (!Builder.GetInsertBlock()) |
3710 | return InsertPointTy(); |
3711 | } |
3712 | Builder.CreateBr(Dest: ContinuationBlock); |
3713 | } else { |
3714 | Builder.CreateUnreachable(); |
3715 | } |
3716 | |
3717 | // Populate the outlined reduction function using the elementwise reduction |
3718 | // function. Partial values are extracted from the type-erased array of |
3719 | // pointers to private variables. |
3720 | BasicBlock *ReductionFuncBlock = |
3721 | BasicBlock::Create(Context&: Module->getContext(), Name: "" , Parent: ReductionFunc); |
3722 | Builder.SetInsertPoint(ReductionFuncBlock); |
3723 | Value *LHSArrayPtr = ReductionFunc->getArg(i: 0); |
3724 | Value *RHSArrayPtr = ReductionFunc->getArg(i: 1); |
3725 | |
3726 | for (auto En : enumerate(First&: ReductionInfos)) { |
3727 | const ReductionInfo &RI = En.value(); |
3728 | Value *LHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64( |
3729 | Ty: RedArrayTy, Ptr: LHSArrayPtr, Idx0: 0, Idx1: En.index()); |
3730 | Value *LHSI8Ptr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: LHSI8PtrPtr); |
3731 | Value *LHSPtr = Builder.CreateBitCast(V: LHSI8Ptr, DestTy: RI.Variable->getType()); |
3732 | Value *LHS = Builder.CreateLoad(Ty: RI.ElementType, Ptr: LHSPtr); |
3733 | Value *RHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64( |
3734 | Ty: RedArrayTy, Ptr: RHSArrayPtr, Idx0: 0, Idx1: En.index()); |
3735 | Value *RHSI8Ptr = Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: RHSI8PtrPtr); |
3736 | Value *RHSPtr = |
3737 | Builder.CreateBitCast(V: RHSI8Ptr, DestTy: RI.PrivateVariable->getType()); |
3738 | Value *RHS = Builder.CreateLoad(Ty: RI.ElementType, Ptr: RHSPtr); |
3739 | Value *Reduced; |
3740 | Builder.restoreIP(IP: RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced)); |
3741 | if (!Builder.GetInsertBlock()) |
3742 | return InsertPointTy(); |
3743 | // store is inside of the reduction region when using by-ref |
3744 | if (!IsByRef[En.index()]) |
3745 | Builder.CreateStore(Val: Reduced, Ptr: LHSPtr); |
3746 | } |
3747 | Builder.CreateRetVoid(); |
3748 | |
3749 | Builder.SetInsertPoint(ContinuationBlock); |
3750 | return Builder.saveIP(); |
3751 | } |
3752 | |
3753 | OpenMPIRBuilder::InsertPointTy |
3754 | OpenMPIRBuilder::createMaster(const LocationDescription &Loc, |
3755 | BodyGenCallbackTy BodyGenCB, |
3756 | FinalizeCallbackTy FiniCB) { |
3757 | |
3758 | if (!updateToLocation(Loc)) |
3759 | return Loc.IP; |
3760 | |
3761 | Directive OMPD = Directive::OMPD_master; |
3762 | uint32_t SrcLocStrSize; |
3763 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
3764 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
3765 | Value *ThreadId = getOrCreateThreadID(Ident); |
3766 | Value *Args[] = {Ident, ThreadId}; |
3767 | |
3768 | Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_master); |
3769 | Instruction *EntryCall = Builder.CreateCall(Callee: EntryRTLFn, Args); |
3770 | |
3771 | Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_master); |
3772 | Instruction *ExitCall = Builder.CreateCall(Callee: ExitRTLFn, Args); |
3773 | |
3774 | return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, |
3775 | /*Conditional*/ true, /*hasFinalize*/ HasFinalize: true); |
3776 | } |
3777 | |
3778 | OpenMPIRBuilder::InsertPointTy |
3779 | OpenMPIRBuilder::createMasked(const LocationDescription &Loc, |
3780 | BodyGenCallbackTy BodyGenCB, |
3781 | FinalizeCallbackTy FiniCB, Value *Filter) { |
3782 | if (!updateToLocation(Loc)) |
3783 | return Loc.IP; |
3784 | |
3785 | Directive OMPD = Directive::OMPD_masked; |
3786 | uint32_t SrcLocStrSize; |
3787 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
3788 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
3789 | Value *ThreadId = getOrCreateThreadID(Ident); |
3790 | Value *Args[] = {Ident, ThreadId, Filter}; |
3791 | Value *ArgsEnd[] = {Ident, ThreadId}; |
3792 | |
3793 | Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_masked); |
3794 | Instruction *EntryCall = Builder.CreateCall(Callee: EntryRTLFn, Args); |
3795 | |
3796 | Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_masked); |
3797 | Instruction *ExitCall = Builder.CreateCall(Callee: ExitRTLFn, Args: ArgsEnd); |
3798 | |
3799 | return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, |
3800 | /*Conditional*/ true, /*hasFinalize*/ HasFinalize: true); |
3801 | } |
3802 | |
3803 | CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton( |
3804 | DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore, |
3805 | BasicBlock *PostInsertBefore, const Twine &Name) { |
3806 | Module *M = F->getParent(); |
3807 | LLVMContext &Ctx = M->getContext(); |
3808 | Type *IndVarTy = TripCount->getType(); |
3809 | |
3810 | // Create the basic block structure. |
3811 | BasicBlock * = |
3812 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".preheader" , Parent: F, InsertBefore: PreInsertBefore); |
3813 | BasicBlock * = |
3814 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".header" , Parent: F, InsertBefore: PreInsertBefore); |
3815 | BasicBlock *Cond = |
3816 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".cond" , Parent: F, InsertBefore: PreInsertBefore); |
3817 | BasicBlock *Body = |
3818 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".body" , Parent: F, InsertBefore: PreInsertBefore); |
3819 | BasicBlock *Latch = |
3820 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".inc" , Parent: F, InsertBefore: PostInsertBefore); |
3821 | BasicBlock *Exit = |
3822 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".exit" , Parent: F, InsertBefore: PostInsertBefore); |
3823 | BasicBlock *After = |
3824 | BasicBlock::Create(Context&: Ctx, Name: "omp_" + Name + ".after" , Parent: F, InsertBefore: PostInsertBefore); |
3825 | |
3826 | // Use specified DebugLoc for new instructions. |
3827 | Builder.SetCurrentDebugLocation(DL); |
3828 | |
3829 | Builder.SetInsertPoint(Preheader); |
3830 | Builder.CreateBr(Dest: Header); |
3831 | |
3832 | Builder.SetInsertPoint(Header); |
3833 | PHINode *IndVarPHI = Builder.CreatePHI(Ty: IndVarTy, NumReservedValues: 2, Name: "omp_" + Name + ".iv" ); |
3834 | IndVarPHI->addIncoming(V: ConstantInt::get(Ty: IndVarTy, V: 0), BB: Preheader); |
3835 | Builder.CreateBr(Dest: Cond); |
3836 | |
3837 | Builder.SetInsertPoint(Cond); |
3838 | Value *Cmp = |
3839 | Builder.CreateICmpULT(LHS: IndVarPHI, RHS: TripCount, Name: "omp_" + Name + ".cmp" ); |
3840 | Builder.CreateCondBr(Cond: Cmp, True: Body, False: Exit); |
3841 | |
3842 | Builder.SetInsertPoint(Body); |
3843 | Builder.CreateBr(Dest: Latch); |
3844 | |
3845 | Builder.SetInsertPoint(Latch); |
3846 | Value *Next = Builder.CreateAdd(LHS: IndVarPHI, RHS: ConstantInt::get(Ty: IndVarTy, V: 1), |
3847 | Name: "omp_" + Name + ".next" , /*HasNUW=*/true); |
3848 | Builder.CreateBr(Dest: Header); |
3849 | IndVarPHI->addIncoming(V: Next, BB: Latch); |
3850 | |
3851 | Builder.SetInsertPoint(Exit); |
3852 | Builder.CreateBr(Dest: After); |
3853 | |
3854 | // Remember and return the canonical control flow. |
3855 | LoopInfos.emplace_front(); |
3856 | CanonicalLoopInfo *CL = &LoopInfos.front(); |
3857 | |
3858 | CL->Header = Header; |
3859 | CL->Cond = Cond; |
3860 | CL->Latch = Latch; |
3861 | CL->Exit = Exit; |
3862 | |
3863 | #ifndef NDEBUG |
3864 | CL->assertOK(); |
3865 | #endif |
3866 | return CL; |
3867 | } |
3868 | |
3869 | CanonicalLoopInfo * |
3870 | OpenMPIRBuilder::createCanonicalLoop(const LocationDescription &Loc, |
3871 | LoopBodyGenCallbackTy BodyGenCB, |
3872 | Value *TripCount, const Twine &Name) { |
3873 | BasicBlock *BB = Loc.IP.getBlock(); |
3874 | BasicBlock *NextBB = BB->getNextNode(); |
3875 | |
3876 | CanonicalLoopInfo *CL = createLoopSkeleton(DL: Loc.DL, TripCount, F: BB->getParent(), |
3877 | PreInsertBefore: NextBB, PostInsertBefore: NextBB, Name); |
3878 | BasicBlock *After = CL->getAfter(); |
3879 | |
3880 | // If location is not set, don't connect the loop. |
3881 | if (updateToLocation(Loc)) { |
3882 | // Split the loop at the insertion point: Branch to the preheader and move |
3883 | // every following instruction to after the loop (the After BB). Also, the |
3884 | // new successor is the loop's after block. |
3885 | spliceBB(Builder, New: After, /*CreateBranch=*/false); |
3886 | Builder.CreateBr(Dest: CL->getPreheader()); |
3887 | } |
3888 | |
3889 | // Emit the body content. We do it after connecting the loop to the CFG to |
3890 | // avoid that the callback encounters degenerate BBs. |
3891 | BodyGenCB(CL->getBodyIP(), CL->getIndVar()); |
3892 | |
3893 | #ifndef NDEBUG |
3894 | CL->assertOK(); |
3895 | #endif |
3896 | return CL; |
3897 | } |
3898 | |
3899 | CanonicalLoopInfo *OpenMPIRBuilder::createCanonicalLoop( |
3900 | const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB, |
3901 | Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop, |
3902 | InsertPointTy ComputeIP, const Twine &Name) { |
3903 | |
3904 | // Consider the following difficulties (assuming 8-bit signed integers): |
3905 | // * Adding \p Step to the loop counter which passes \p Stop may overflow: |
3906 | // DO I = 1, 100, 50 |
3907 | /// * A \p Step of INT_MIN cannot not be normalized to a positive direction: |
3908 | // DO I = 100, 0, -128 |
3909 | |
3910 | // Start, Stop and Step must be of the same integer type. |
3911 | auto *IndVarTy = cast<IntegerType>(Val: Start->getType()); |
3912 | assert(IndVarTy == Stop->getType() && "Stop type mismatch" ); |
3913 | assert(IndVarTy == Step->getType() && "Step type mismatch" ); |
3914 | |
3915 | LocationDescription ComputeLoc = |
3916 | ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc; |
3917 | updateToLocation(Loc: ComputeLoc); |
3918 | |
3919 | ConstantInt *Zero = ConstantInt::get(Ty: IndVarTy, V: 0); |
3920 | ConstantInt *One = ConstantInt::get(Ty: IndVarTy, V: 1); |
3921 | |
3922 | // Like Step, but always positive. |
3923 | Value *Incr = Step; |
3924 | |
3925 | // Distance between Start and Stop; always positive. |
3926 | Value *Span; |
3927 | |
3928 | // Condition whether there are no iterations are executed at all, e.g. because |
3929 | // UB < LB. |
3930 | Value *ZeroCmp; |
3931 | |
3932 | if (IsSigned) { |
3933 | // Ensure that increment is positive. If not, negate and invert LB and UB. |
3934 | Value *IsNeg = Builder.CreateICmpSLT(LHS: Step, RHS: Zero); |
3935 | Incr = Builder.CreateSelect(C: IsNeg, True: Builder.CreateNeg(V: Step), False: Step); |
3936 | Value *LB = Builder.CreateSelect(C: IsNeg, True: Stop, False: Start); |
3937 | Value *UB = Builder.CreateSelect(C: IsNeg, True: Start, False: Stop); |
3938 | Span = Builder.CreateSub(LHS: UB, RHS: LB, Name: "" , HasNUW: false, HasNSW: true); |
3939 | ZeroCmp = Builder.CreateICmp( |
3940 | P: InclusiveStop ? CmpInst::ICMP_SLT : CmpInst::ICMP_SLE, LHS: UB, RHS: LB); |
3941 | } else { |
3942 | Span = Builder.CreateSub(LHS: Stop, RHS: Start, Name: "" , HasNUW: true); |
3943 | ZeroCmp = Builder.CreateICmp( |
3944 | P: InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, LHS: Stop, RHS: Start); |
3945 | } |
3946 | |
3947 | Value *CountIfLooping; |
3948 | if (InclusiveStop) { |
3949 | CountIfLooping = Builder.CreateAdd(LHS: Builder.CreateUDiv(LHS: Span, RHS: Incr), RHS: One); |
3950 | } else { |
3951 | // Avoid incrementing past stop since it could overflow. |
3952 | Value *CountIfTwo = Builder.CreateAdd( |
3953 | LHS: Builder.CreateUDiv(LHS: Builder.CreateSub(LHS: Span, RHS: One), RHS: Incr), RHS: One); |
3954 | Value *OneCmp = Builder.CreateICmp(P: CmpInst::ICMP_ULE, LHS: Span, RHS: Incr); |
3955 | CountIfLooping = Builder.CreateSelect(C: OneCmp, True: One, False: CountIfTwo); |
3956 | } |
3957 | Value *TripCount = Builder.CreateSelect(C: ZeroCmp, True: Zero, False: CountIfLooping, |
3958 | Name: "omp_" + Name + ".tripcount" ); |
3959 | |
3960 | auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) { |
3961 | Builder.restoreIP(IP: CodeGenIP); |
3962 | Value *Span = Builder.CreateMul(LHS: IV, RHS: Step); |
3963 | Value *IndVar = Builder.CreateAdd(LHS: Span, RHS: Start); |
3964 | BodyGenCB(Builder.saveIP(), IndVar); |
3965 | }; |
3966 | LocationDescription LoopLoc = ComputeIP.isSet() ? Loc.IP : Builder.saveIP(); |
3967 | return createCanonicalLoop(Loc: LoopLoc, BodyGenCB: BodyGen, TripCount, Name); |
3968 | } |
3969 | |
3970 | // Returns an LLVM function to call for initializing loop bounds using OpenMP |
3971 | // static scheduling depending on `type`. Only i32 and i64 are supported by the |
3972 | // runtime. Always interpret integers as unsigned similarly to |
3973 | // CanonicalLoopInfo. |
3974 | static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M, |
3975 | OpenMPIRBuilder &OMPBuilder) { |
3976 | unsigned Bitwidth = Ty->getIntegerBitWidth(); |
3977 | if (Bitwidth == 32) |
3978 | return OMPBuilder.getOrCreateRuntimeFunction( |
3979 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u); |
3980 | if (Bitwidth == 64) |
3981 | return OMPBuilder.getOrCreateRuntimeFunction( |
3982 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u); |
3983 | llvm_unreachable("unknown OpenMP loop iterator bitwidth" ); |
3984 | } |
3985 | |
3986 | OpenMPIRBuilder::InsertPointTy |
3987 | OpenMPIRBuilder::applyStaticWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI, |
3988 | InsertPointTy AllocaIP, |
3989 | bool NeedsBarrier) { |
3990 | assert(CLI->isValid() && "Requires a valid canonical loop" ); |
3991 | assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) && |
3992 | "Require dedicated allocate IP" ); |
3993 | |
3994 | // Set up the source location value for OpenMP runtime. |
3995 | Builder.restoreIP(IP: CLI->getPreheaderIP()); |
3996 | Builder.SetCurrentDebugLocation(DL); |
3997 | |
3998 | uint32_t SrcLocStrSize; |
3999 | Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); |
4000 | Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
4001 | |
4002 | // Declare useful OpenMP runtime functions. |
4003 | Value *IV = CLI->getIndVar(); |
4004 | Type *IVTy = IV->getType(); |
4005 | FunctionCallee StaticInit = getKmpcForStaticInitForType(Ty: IVTy, M, OMPBuilder&: *this); |
4006 | FunctionCallee StaticFini = |
4007 | getOrCreateRuntimeFunction(M, FnID: omp::OMPRTL___kmpc_for_static_fini); |
4008 | |
4009 | // Allocate space for computed loop bounds as expected by the "init" function. |
4010 | Builder.SetInsertPoint(AllocaIP.getBlock()->getFirstNonPHIOrDbgOrAlloca()); |
4011 | |
4012 | Type *I32Type = Type::getInt32Ty(C&: M.getContext()); |
4013 | Value *PLastIter = Builder.CreateAlloca(Ty: I32Type, ArraySize: nullptr, Name: "p.lastiter" ); |
4014 | Value *PLowerBound = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.lowerbound" ); |
4015 | Value *PUpperBound = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.upperbound" ); |
4016 | Value *PStride = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.stride" ); |
4017 | |
4018 | // At the end of the preheader, prepare for calling the "init" function by |
4019 | // storing the current loop bounds into the allocated space. A canonical loop |
4020 | // always iterates from 0 to trip-count with step 1. Note that "init" expects |
4021 | // and produces an inclusive upper bound. |
4022 | Builder.SetInsertPoint(CLI->getPreheader()->getTerminator()); |
4023 | Constant *Zero = ConstantInt::get(Ty: IVTy, V: 0); |
4024 | Constant *One = ConstantInt::get(Ty: IVTy, V: 1); |
4025 | Builder.CreateStore(Val: Zero, Ptr: PLowerBound); |
4026 | Value *UpperBound = Builder.CreateSub(LHS: CLI->getTripCount(), RHS: One); |
4027 | Builder.CreateStore(Val: UpperBound, Ptr: PUpperBound); |
4028 | Builder.CreateStore(Val: One, Ptr: PStride); |
4029 | |
4030 | Value *ThreadNum = getOrCreateThreadID(Ident: SrcLoc); |
4031 | |
4032 | Constant *SchedulingType = ConstantInt::get( |
4033 | Ty: I32Type, V: static_cast<int>(OMPScheduleType::UnorderedStatic)); |
4034 | |
4035 | // Call the "init" function and update the trip count of the loop with the |
4036 | // value it produced. |
4037 | Builder.CreateCall(Callee: StaticInit, |
4038 | Args: {SrcLoc, ThreadNum, SchedulingType, PLastIter, PLowerBound, |
4039 | PUpperBound, PStride, One, Zero}); |
4040 | Value *LowerBound = Builder.CreateLoad(Ty: IVTy, Ptr: PLowerBound); |
4041 | Value *InclusiveUpperBound = Builder.CreateLoad(Ty: IVTy, Ptr: PUpperBound); |
4042 | Value *TripCountMinusOne = Builder.CreateSub(LHS: InclusiveUpperBound, RHS: LowerBound); |
4043 | Value *TripCount = Builder.CreateAdd(LHS: TripCountMinusOne, RHS: One); |
4044 | CLI->setTripCount(TripCount); |
4045 | |
4046 | // Update all uses of the induction variable except the one in the condition |
4047 | // block that compares it with the actual upper bound, and the increment in |
4048 | // the latch block. |
4049 | |
4050 | CLI->mapIndVar(Updater: [&](Instruction *OldIV) -> Value * { |
4051 | Builder.SetInsertPoint(TheBB: CLI->getBody(), |
4052 | IP: CLI->getBody()->getFirstInsertionPt()); |
4053 | Builder.SetCurrentDebugLocation(DL); |
4054 | return Builder.CreateAdd(LHS: OldIV, RHS: LowerBound); |
4055 | }); |
4056 | |
4057 | // In the "exit" block, call the "fini" function. |
4058 | Builder.SetInsertPoint(TheBB: CLI->getExit(), |
4059 | IP: CLI->getExit()->getTerminator()->getIterator()); |
4060 | Builder.CreateCall(Callee: StaticFini, Args: {SrcLoc, ThreadNum}); |
4061 | |
4062 | // Add the barrier if requested. |
4063 | if (NeedsBarrier) |
4064 | createBarrier(Loc: LocationDescription(Builder.saveIP(), DL), |
4065 | Kind: omp::Directive::OMPD_for, /* ForceSimpleCall */ false, |
4066 | /* CheckCancelFlag */ false); |
4067 | |
4068 | InsertPointTy AfterIP = CLI->getAfterIP(); |
4069 | CLI->invalidate(); |
4070 | |
4071 | return AfterIP; |
4072 | } |
4073 | |
4074 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyStaticChunkedWorkshareLoop( |
4075 | DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, |
4076 | bool NeedsBarrier, Value *ChunkSize) { |
4077 | assert(CLI->isValid() && "Requires a valid canonical loop" ); |
4078 | assert(ChunkSize && "Chunk size is required" ); |
4079 | |
4080 | LLVMContext &Ctx = CLI->getFunction()->getContext(); |
4081 | Value *IV = CLI->getIndVar(); |
4082 | Value *OrigTripCount = CLI->getTripCount(); |
4083 | Type *IVTy = IV->getType(); |
4084 | assert(IVTy->getIntegerBitWidth() <= 64 && |
4085 | "Max supported tripcount bitwidth is 64 bits" ); |
4086 | Type *InternalIVTy = IVTy->getIntegerBitWidth() <= 32 ? Type::getInt32Ty(C&: Ctx) |
4087 | : Type::getInt64Ty(C&: Ctx); |
4088 | Type *I32Type = Type::getInt32Ty(C&: M.getContext()); |
4089 | Constant *Zero = ConstantInt::get(Ty: InternalIVTy, V: 0); |
4090 | Constant *One = ConstantInt::get(Ty: InternalIVTy, V: 1); |
4091 | |
4092 | // Declare useful OpenMP runtime functions. |
4093 | FunctionCallee StaticInit = |
4094 | getKmpcForStaticInitForType(Ty: InternalIVTy, M, OMPBuilder&: *this); |
4095 | FunctionCallee StaticFini = |
4096 | getOrCreateRuntimeFunction(M, FnID: omp::OMPRTL___kmpc_for_static_fini); |
4097 | |
4098 | // Allocate space for computed loop bounds as expected by the "init" function. |
4099 | Builder.restoreIP(IP: AllocaIP); |
4100 | Builder.SetCurrentDebugLocation(DL); |
4101 | Value *PLastIter = Builder.CreateAlloca(Ty: I32Type, ArraySize: nullptr, Name: "p.lastiter" ); |
4102 | Value *PLowerBound = |
4103 | Builder.CreateAlloca(Ty: InternalIVTy, ArraySize: nullptr, Name: "p.lowerbound" ); |
4104 | Value *PUpperBound = |
4105 | Builder.CreateAlloca(Ty: InternalIVTy, ArraySize: nullptr, Name: "p.upperbound" ); |
4106 | Value *PStride = Builder.CreateAlloca(Ty: InternalIVTy, ArraySize: nullptr, Name: "p.stride" ); |
4107 | |
4108 | // Set up the source location value for the OpenMP runtime. |
4109 | Builder.restoreIP(IP: CLI->getPreheaderIP()); |
4110 | Builder.SetCurrentDebugLocation(DL); |
4111 | |
4112 | // TODO: Detect overflow in ubsan or max-out with current tripcount. |
4113 | Value *CastedChunkSize = |
4114 | Builder.CreateZExtOrTrunc(V: ChunkSize, DestTy: InternalIVTy, Name: "chunksize" ); |
4115 | Value *CastedTripCount = |
4116 | Builder.CreateZExt(V: OrigTripCount, DestTy: InternalIVTy, Name: "tripcount" ); |
4117 | |
4118 | Constant *SchedulingType = ConstantInt::get( |
4119 | Ty: I32Type, V: static_cast<int>(OMPScheduleType::UnorderedStaticChunked)); |
4120 | Builder.CreateStore(Val: Zero, Ptr: PLowerBound); |
4121 | Value *OrigUpperBound = Builder.CreateSub(LHS: CastedTripCount, RHS: One); |
4122 | Builder.CreateStore(Val: OrigUpperBound, Ptr: PUpperBound); |
4123 | Builder.CreateStore(Val: One, Ptr: PStride); |
4124 | |
4125 | // Call the "init" function and update the trip count of the loop with the |
4126 | // value it produced. |
4127 | uint32_t SrcLocStrSize; |
4128 | Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); |
4129 | Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
4130 | Value *ThreadNum = getOrCreateThreadID(Ident: SrcLoc); |
4131 | Builder.CreateCall(Callee: StaticInit, |
4132 | Args: {/*loc=*/SrcLoc, /*global_tid=*/ThreadNum, |
4133 | /*schedtype=*/SchedulingType, /*plastiter=*/PLastIter, |
4134 | /*plower=*/PLowerBound, /*pupper=*/PUpperBound, |
4135 | /*pstride=*/PStride, /*incr=*/One, |
4136 | /*chunk=*/CastedChunkSize}); |
4137 | |
4138 | // Load values written by the "init" function. |
4139 | Value *FirstChunkStart = |
4140 | Builder.CreateLoad(Ty: InternalIVTy, Ptr: PLowerBound, Name: "omp_firstchunk.lb" ); |
4141 | Value *FirstChunkStop = |
4142 | Builder.CreateLoad(Ty: InternalIVTy, Ptr: PUpperBound, Name: "omp_firstchunk.ub" ); |
4143 | Value *FirstChunkEnd = Builder.CreateAdd(LHS: FirstChunkStop, RHS: One); |
4144 | Value *ChunkRange = |
4145 | Builder.CreateSub(LHS: FirstChunkEnd, RHS: FirstChunkStart, Name: "omp_chunk.range" ); |
4146 | Value *NextChunkStride = |
4147 | Builder.CreateLoad(Ty: InternalIVTy, Ptr: PStride, Name: "omp_dispatch.stride" ); |
4148 | |
4149 | // Create outer "dispatch" loop for enumerating the chunks. |
4150 | BasicBlock *DispatchEnter = splitBB(Builder, CreateBranch: true); |
4151 | Value *DispatchCounter; |
4152 | CanonicalLoopInfo *DispatchCLI = createCanonicalLoop( |
4153 | Loc: {Builder.saveIP(), DL}, |
4154 | BodyGenCB: [&](InsertPointTy BodyIP, Value *Counter) { DispatchCounter = Counter; }, |
4155 | Start: FirstChunkStart, Stop: CastedTripCount, Step: NextChunkStride, |
4156 | /*IsSigned=*/false, /*InclusiveStop=*/false, /*ComputeIP=*/{}, |
4157 | Name: "dispatch" ); |
4158 | |
4159 | // Remember the BasicBlocks of the dispatch loop we need, then invalidate to |
4160 | // not have to preserve the canonical invariant. |
4161 | BasicBlock *DispatchBody = DispatchCLI->getBody(); |
4162 | BasicBlock *DispatchLatch = DispatchCLI->getLatch(); |
4163 | BasicBlock *DispatchExit = DispatchCLI->getExit(); |
4164 | BasicBlock *DispatchAfter = DispatchCLI->getAfter(); |
4165 | DispatchCLI->invalidate(); |
4166 | |
4167 | // Rewire the original loop to become the chunk loop inside the dispatch loop. |
4168 | redirectTo(Source: DispatchAfter, Target: CLI->getAfter(), DL); |
4169 | redirectTo(Source: CLI->getExit(), Target: DispatchLatch, DL); |
4170 | redirectTo(Source: DispatchBody, Target: DispatchEnter, DL); |
4171 | |
4172 | // Prepare the prolog of the chunk loop. |
4173 | Builder.restoreIP(IP: CLI->getPreheaderIP()); |
4174 | Builder.SetCurrentDebugLocation(DL); |
4175 | |
4176 | // Compute the number of iterations of the chunk loop. |
4177 | Builder.SetInsertPoint(CLI->getPreheader()->getTerminator()); |
4178 | Value *ChunkEnd = Builder.CreateAdd(LHS: DispatchCounter, RHS: ChunkRange); |
4179 | Value *IsLastChunk = |
4180 | Builder.CreateICmpUGE(LHS: ChunkEnd, RHS: CastedTripCount, Name: "omp_chunk.is_last" ); |
4181 | Value *CountUntilOrigTripCount = |
4182 | Builder.CreateSub(LHS: CastedTripCount, RHS: DispatchCounter); |
4183 | Value *ChunkTripCount = Builder.CreateSelect( |
4184 | C: IsLastChunk, True: CountUntilOrigTripCount, False: ChunkRange, Name: "omp_chunk.tripcount" ); |
4185 | Value *BackcastedChunkTC = |
4186 | Builder.CreateTrunc(V: ChunkTripCount, DestTy: IVTy, Name: "omp_chunk.tripcount.trunc" ); |
4187 | CLI->setTripCount(BackcastedChunkTC); |
4188 | |
4189 | // Update all uses of the induction variable except the one in the condition |
4190 | // block that compares it with the actual upper bound, and the increment in |
4191 | // the latch block. |
4192 | Value *BackcastedDispatchCounter = |
4193 | Builder.CreateTrunc(V: DispatchCounter, DestTy: IVTy, Name: "omp_dispatch.iv.trunc" ); |
4194 | CLI->mapIndVar(Updater: [&](Instruction *) -> Value * { |
4195 | Builder.restoreIP(IP: CLI->getBodyIP()); |
4196 | return Builder.CreateAdd(LHS: IV, RHS: BackcastedDispatchCounter); |
4197 | }); |
4198 | |
4199 | // In the "exit" block, call the "fini" function. |
4200 | Builder.SetInsertPoint(TheBB: DispatchExit, IP: DispatchExit->getFirstInsertionPt()); |
4201 | Builder.CreateCall(Callee: StaticFini, Args: {SrcLoc, ThreadNum}); |
4202 | |
4203 | // Add the barrier if requested. |
4204 | if (NeedsBarrier) |
4205 | createBarrier(Loc: LocationDescription(Builder.saveIP(), DL), Kind: OMPD_for, |
4206 | /*ForceSimpleCall=*/false, /*CheckCancelFlag=*/false); |
4207 | |
4208 | #ifndef NDEBUG |
4209 | // Even though we currently do not support applying additional methods to it, |
4210 | // the chunk loop should remain a canonical loop. |
4211 | CLI->assertOK(); |
4212 | #endif |
4213 | |
4214 | return {DispatchAfter, DispatchAfter->getFirstInsertionPt()}; |
4215 | } |
4216 | |
4217 | // Returns an LLVM function to call for executing an OpenMP static worksharing |
4218 | // for loop depending on `type`. Only i32 and i64 are supported by the runtime. |
4219 | // Always interpret integers as unsigned similarly to CanonicalLoopInfo. |
4220 | static FunctionCallee |
4221 | getKmpcForStaticLoopForType(Type *Ty, OpenMPIRBuilder *OMPBuilder, |
4222 | WorksharingLoopType LoopType) { |
4223 | unsigned Bitwidth = Ty->getIntegerBitWidth(); |
4224 | Module &M = OMPBuilder->M; |
4225 | switch (LoopType) { |
4226 | case WorksharingLoopType::ForStaticLoop: |
4227 | if (Bitwidth == 32) |
4228 | return OMPBuilder->getOrCreateRuntimeFunction( |
4229 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_4u); |
4230 | if (Bitwidth == 64) |
4231 | return OMPBuilder->getOrCreateRuntimeFunction( |
4232 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_8u); |
4233 | break; |
4234 | case WorksharingLoopType::DistributeStaticLoop: |
4235 | if (Bitwidth == 32) |
4236 | return OMPBuilder->getOrCreateRuntimeFunction( |
4237 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_4u); |
4238 | if (Bitwidth == 64) |
4239 | return OMPBuilder->getOrCreateRuntimeFunction( |
4240 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_8u); |
4241 | break; |
4242 | case WorksharingLoopType::DistributeForStaticLoop: |
4243 | if (Bitwidth == 32) |
4244 | return OMPBuilder->getOrCreateRuntimeFunction( |
4245 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_4u); |
4246 | if (Bitwidth == 64) |
4247 | return OMPBuilder->getOrCreateRuntimeFunction( |
4248 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_8u); |
4249 | break; |
4250 | } |
4251 | if (Bitwidth != 32 && Bitwidth != 64) { |
4252 | llvm_unreachable("Unknown OpenMP loop iterator bitwidth" ); |
4253 | } |
4254 | llvm_unreachable("Unknown type of OpenMP worksharing loop" ); |
4255 | } |
4256 | |
4257 | // Inserts a call to proper OpenMP Device RTL function which handles |
4258 | // loop worksharing. |
4259 | static void createTargetLoopWorkshareCall( |
4260 | OpenMPIRBuilder *OMPBuilder, WorksharingLoopType LoopType, |
4261 | BasicBlock *InsertBlock, Value *Ident, Value *LoopBodyArg, |
4262 | Type *ParallelTaskPtr, Value *TripCount, Function &LoopBodyFn) { |
4263 | Type *TripCountTy = TripCount->getType(); |
4264 | Module &M = OMPBuilder->M; |
4265 | IRBuilder<> &Builder = OMPBuilder->Builder; |
4266 | FunctionCallee RTLFn = |
4267 | getKmpcForStaticLoopForType(Ty: TripCountTy, OMPBuilder, LoopType); |
4268 | SmallVector<Value *, 8> RealArgs; |
4269 | RealArgs.push_back(Elt: Ident); |
4270 | RealArgs.push_back(Elt: Builder.CreateBitCast(V: &LoopBodyFn, DestTy: ParallelTaskPtr)); |
4271 | RealArgs.push_back(Elt: LoopBodyArg); |
4272 | RealArgs.push_back(Elt: TripCount); |
4273 | if (LoopType == WorksharingLoopType::DistributeStaticLoop) { |
4274 | RealArgs.push_back(Elt: ConstantInt::get(Ty: TripCountTy, V: 0)); |
4275 | Builder.CreateCall(Callee: RTLFn, Args: RealArgs); |
4276 | return; |
4277 | } |
4278 | FunctionCallee RTLNumThreads = OMPBuilder->getOrCreateRuntimeFunction( |
4279 | M, FnID: omp::RuntimeFunction::OMPRTL_omp_get_num_threads); |
4280 | Builder.restoreIP(IP: {InsertBlock, std::prev(x: InsertBlock->end())}); |
4281 | Value *NumThreads = Builder.CreateCall(Callee: RTLNumThreads, Args: {}); |
4282 | |
4283 | RealArgs.push_back( |
4284 | Elt: Builder.CreateZExtOrTrunc(V: NumThreads, DestTy: TripCountTy, Name: "num.threads.cast" )); |
4285 | RealArgs.push_back(Elt: ConstantInt::get(Ty: TripCountTy, V: 0)); |
4286 | if (LoopType == WorksharingLoopType::DistributeForStaticLoop) { |
4287 | RealArgs.push_back(Elt: ConstantInt::get(Ty: TripCountTy, V: 0)); |
4288 | } |
4289 | |
4290 | Builder.CreateCall(Callee: RTLFn, Args: RealArgs); |
4291 | } |
4292 | |
4293 | static void |
4294 | workshareLoopTargetCallback(OpenMPIRBuilder *OMPIRBuilder, |
4295 | CanonicalLoopInfo *CLI, Value *Ident, |
4296 | Function &OutlinedFn, Type *ParallelTaskPtr, |
4297 | const SmallVector<Instruction *, 4> &ToBeDeleted, |
4298 | WorksharingLoopType LoopType) { |
4299 | IRBuilder<> &Builder = OMPIRBuilder->Builder; |
4300 | BasicBlock * = CLI->getPreheader(); |
4301 | Value *TripCount = CLI->getTripCount(); |
4302 | |
4303 | // After loop body outling, the loop body contains only set up |
4304 | // of loop body argument structure and the call to the outlined |
4305 | // loop body function. Firstly, we need to move setup of loop body args |
4306 | // into loop preheader. |
4307 | Preheader->splice(ToIt: std::prev(x: Preheader->end()), FromBB: CLI->getBody(), |
4308 | FromBeginIt: CLI->getBody()->begin(), FromEndIt: std::prev(x: CLI->getBody()->end())); |
4309 | |
4310 | // The next step is to remove the whole loop. We do not it need anymore. |
4311 | // That's why make an unconditional branch from loop preheader to loop |
4312 | // exit block |
4313 | Builder.restoreIP(IP: {Preheader, Preheader->end()}); |
4314 | Preheader->getTerminator()->eraseFromParent(); |
4315 | Builder.CreateBr(Dest: CLI->getExit()); |
4316 | |
4317 | // Delete dead loop blocks |
4318 | OpenMPIRBuilder::OutlineInfo CleanUpInfo; |
4319 | SmallPtrSet<BasicBlock *, 32> RegionBlockSet; |
4320 | SmallVector<BasicBlock *, 32> BlocksToBeRemoved; |
4321 | CleanUpInfo.EntryBB = CLI->getHeader(); |
4322 | CleanUpInfo.ExitBB = CLI->getExit(); |
4323 | CleanUpInfo.collectBlocks(BlockSet&: RegionBlockSet, BlockVector&: BlocksToBeRemoved); |
4324 | DeleteDeadBlocks(BBs: BlocksToBeRemoved); |
4325 | |
4326 | // Find the instruction which corresponds to loop body argument structure |
4327 | // and remove the call to loop body function instruction. |
4328 | Value *LoopBodyArg; |
4329 | User *OutlinedFnUser = OutlinedFn.getUniqueUndroppableUser(); |
4330 | assert(OutlinedFnUser && |
4331 | "Expected unique undroppable user of outlined function" ); |
4332 | CallInst *OutlinedFnCallInstruction = dyn_cast<CallInst>(Val: OutlinedFnUser); |
4333 | assert(OutlinedFnCallInstruction && "Expected outlined function call" ); |
4334 | assert((OutlinedFnCallInstruction->getParent() == Preheader) && |
4335 | "Expected outlined function call to be located in loop preheader" ); |
4336 | // Check in case no argument structure has been passed. |
4337 | if (OutlinedFnCallInstruction->arg_size() > 1) |
4338 | LoopBodyArg = OutlinedFnCallInstruction->getArgOperand(i: 1); |
4339 | else |
4340 | LoopBodyArg = Constant::getNullValue(Ty: Builder.getPtrTy()); |
4341 | OutlinedFnCallInstruction->eraseFromParent(); |
4342 | |
4343 | createTargetLoopWorkshareCall(OMPBuilder: OMPIRBuilder, LoopType, InsertBlock: Preheader, Ident, |
4344 | LoopBodyArg, ParallelTaskPtr, TripCount, |
4345 | LoopBodyFn&: OutlinedFn); |
4346 | |
4347 | for (auto &ToBeDeletedItem : ToBeDeleted) |
4348 | ToBeDeletedItem->eraseFromParent(); |
4349 | CLI->invalidate(); |
4350 | } |
4351 | |
4352 | OpenMPIRBuilder::InsertPointTy |
4353 | OpenMPIRBuilder::applyWorkshareLoopTarget(DebugLoc DL, CanonicalLoopInfo *CLI, |
4354 | InsertPointTy AllocaIP, |
4355 | WorksharingLoopType LoopType) { |
4356 | uint32_t SrcLocStrSize; |
4357 | Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); |
4358 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
4359 | |
4360 | OutlineInfo OI; |
4361 | OI.OuterAllocaBB = CLI->getPreheader(); |
4362 | Function *OuterFn = CLI->getPreheader()->getParent(); |
4363 | |
4364 | // Instructions which need to be deleted at the end of code generation |
4365 | SmallVector<Instruction *, 4> ToBeDeleted; |
4366 | |
4367 | OI.OuterAllocaBB = AllocaIP.getBlock(); |
4368 | |
4369 | // Mark the body loop as region which needs to be extracted |
4370 | OI.EntryBB = CLI->getBody(); |
4371 | OI.ExitBB = CLI->getLatch()->splitBasicBlock(I: CLI->getLatch()->begin(), |
4372 | BBName: "omp.prelatch" , Before: true); |
4373 | |
4374 | // Prepare loop body for extraction |
4375 | Builder.restoreIP(IP: {CLI->getPreheader(), CLI->getPreheader()->begin()}); |
4376 | |
4377 | // Insert new loop counter variable which will be used only in loop |
4378 | // body. |
4379 | AllocaInst *NewLoopCnt = Builder.CreateAlloca(Ty: CLI->getIndVarType(), ArraySize: 0, Name: "" ); |
4380 | Instruction *NewLoopCntLoad = |
4381 | Builder.CreateLoad(Ty: CLI->getIndVarType(), Ptr: NewLoopCnt); |
4382 | // New loop counter instructions are redundant in the loop preheader when |
4383 | // code generation for workshare loop is finshed. That's why mark them as |
4384 | // ready for deletion. |
4385 | ToBeDeleted.push_back(Elt: NewLoopCntLoad); |
4386 | ToBeDeleted.push_back(Elt: NewLoopCnt); |
4387 | |
4388 | // Analyse loop body region. Find all input variables which are used inside |
4389 | // loop body region. |
4390 | SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet; |
4391 | SmallVector<BasicBlock *, 32> Blocks; |
4392 | OI.collectBlocks(BlockSet&: ParallelRegionBlockSet, BlockVector&: Blocks); |
4393 | SmallVector<BasicBlock *, 32> BlocksT(ParallelRegionBlockSet.begin(), |
4394 | ParallelRegionBlockSet.end()); |
4395 | |
4396 | CodeExtractorAnalysisCache CEAC(*OuterFn); |
4397 | CodeExtractor (Blocks, |
4398 | /* DominatorTree */ nullptr, |
4399 | /* AggregateArgs */ true, |
4400 | /* BlockFrequencyInfo */ nullptr, |
4401 | /* BranchProbabilityInfo */ nullptr, |
4402 | /* AssumptionCache */ nullptr, |
4403 | /* AllowVarArgs */ true, |
4404 | /* AllowAlloca */ true, |
4405 | /* AllocationBlock */ CLI->getPreheader(), |
4406 | /* Suffix */ ".omp_wsloop" , |
4407 | /* AggrArgsIn0AddrSpace */ true); |
4408 | |
4409 | BasicBlock *CommonExit = nullptr; |
4410 | SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands; |
4411 | |
4412 | // Find allocas outside the loop body region which are used inside loop |
4413 | // body |
4414 | Extractor.findAllocas(CEAC, SinkCands&: SinkingCands, HoistCands&: HoistingCands, ExitBlock&: CommonExit); |
4415 | |
4416 | // We need to model loop body region as the function f(cnt, loop_arg). |
4417 | // That's why we replace loop induction variable by the new counter |
4418 | // which will be one of loop body function argument |
4419 | SmallVector<User *> Users(CLI->getIndVar()->user_begin(), |
4420 | CLI->getIndVar()->user_end()); |
4421 | for (auto Use : Users) { |
4422 | if (Instruction *Inst = dyn_cast<Instruction>(Val: Use)) { |
4423 | if (ParallelRegionBlockSet.count(Ptr: Inst->getParent())) { |
4424 | Inst->replaceUsesOfWith(From: CLI->getIndVar(), To: NewLoopCntLoad); |
4425 | } |
4426 | } |
4427 | } |
4428 | // Make sure that loop counter variable is not merged into loop body |
4429 | // function argument structure and it is passed as separate variable |
4430 | OI.ExcludeArgsFromAggregate.push_back(Elt: NewLoopCntLoad); |
4431 | |
4432 | // PostOutline CB is invoked when loop body function is outlined and |
4433 | // loop body is replaced by call to outlined function. We need to add |
4434 | // call to OpenMP device rtl inside loop preheader. OpenMP device rtl |
4435 | // function will handle loop control logic. |
4436 | // |
4437 | OI.PostOutlineCB = [=, ToBeDeletedVec = |
4438 | std::move(ToBeDeleted)](Function &OutlinedFn) { |
4439 | workshareLoopTargetCallback(OMPIRBuilder: this, CLI, Ident, OutlinedFn, ParallelTaskPtr, |
4440 | ToBeDeleted: ToBeDeletedVec, LoopType); |
4441 | }; |
4442 | addOutlineInfo(OI: std::move(OI)); |
4443 | return CLI->getAfterIP(); |
4444 | } |
4445 | |
4446 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyWorkshareLoop( |
4447 | DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, |
4448 | bool NeedsBarrier, omp::ScheduleKind SchedKind, Value *ChunkSize, |
4449 | bool HasSimdModifier, bool HasMonotonicModifier, |
4450 | bool HasNonmonotonicModifier, bool HasOrderedClause, |
4451 | WorksharingLoopType LoopType) { |
4452 | if (Config.isTargetDevice()) |
4453 | return applyWorkshareLoopTarget(DL, CLI, AllocaIP, LoopType); |
4454 | OMPScheduleType EffectiveScheduleType = computeOpenMPScheduleType( |
4455 | ClauseKind: SchedKind, HasChunks: ChunkSize, HasSimdModifier, HasMonotonicModifier, |
4456 | HasNonmonotonicModifier, HasOrderedClause); |
4457 | |
4458 | bool IsOrdered = (EffectiveScheduleType & OMPScheduleType::ModifierOrdered) == |
4459 | OMPScheduleType::ModifierOrdered; |
4460 | switch (EffectiveScheduleType & ~OMPScheduleType::ModifierMask) { |
4461 | case OMPScheduleType::BaseStatic: |
4462 | assert(!ChunkSize && "No chunk size with static-chunked schedule" ); |
4463 | if (IsOrdered) |
4464 | return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, SchedType: EffectiveScheduleType, |
4465 | NeedsBarrier, Chunk: ChunkSize); |
4466 | // FIXME: Monotonicity ignored? |
4467 | return applyStaticWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier); |
4468 | |
4469 | case OMPScheduleType::BaseStaticChunked: |
4470 | if (IsOrdered) |
4471 | return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, SchedType: EffectiveScheduleType, |
4472 | NeedsBarrier, Chunk: ChunkSize); |
4473 | // FIXME: Monotonicity ignored? |
4474 | return applyStaticChunkedWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier, |
4475 | ChunkSize); |
4476 | |
4477 | case OMPScheduleType::BaseRuntime: |
4478 | case OMPScheduleType::BaseAuto: |
4479 | case OMPScheduleType::BaseGreedy: |
4480 | case OMPScheduleType::BaseBalanced: |
4481 | case OMPScheduleType::BaseSteal: |
4482 | case OMPScheduleType::BaseGuidedSimd: |
4483 | case OMPScheduleType::BaseRuntimeSimd: |
4484 | assert(!ChunkSize && |
4485 | "schedule type does not support user-defined chunk sizes" ); |
4486 | [[fallthrough]]; |
4487 | case OMPScheduleType::BaseDynamicChunked: |
4488 | case OMPScheduleType::BaseGuidedChunked: |
4489 | case OMPScheduleType::BaseGuidedIterativeChunked: |
4490 | case OMPScheduleType::BaseGuidedAnalyticalChunked: |
4491 | case OMPScheduleType::BaseStaticBalancedChunked: |
4492 | return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, SchedType: EffectiveScheduleType, |
4493 | NeedsBarrier, Chunk: ChunkSize); |
4494 | |
4495 | default: |
4496 | llvm_unreachable("Unknown/unimplemented schedule kind" ); |
4497 | } |
4498 | } |
4499 | |
4500 | /// Returns an LLVM function to call for initializing loop bounds using OpenMP |
4501 | /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by |
4502 | /// the runtime. Always interpret integers as unsigned similarly to |
4503 | /// CanonicalLoopInfo. |
4504 | static FunctionCallee |
4505 | getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { |
4506 | unsigned Bitwidth = Ty->getIntegerBitWidth(); |
4507 | if (Bitwidth == 32) |
4508 | return OMPBuilder.getOrCreateRuntimeFunction( |
4509 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u); |
4510 | if (Bitwidth == 64) |
4511 | return OMPBuilder.getOrCreateRuntimeFunction( |
4512 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u); |
4513 | llvm_unreachable("unknown OpenMP loop iterator bitwidth" ); |
4514 | } |
4515 | |
4516 | /// Returns an LLVM function to call for updating the next loop using OpenMP |
4517 | /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by |
4518 | /// the runtime. Always interpret integers as unsigned similarly to |
4519 | /// CanonicalLoopInfo. |
4520 | static FunctionCallee |
4521 | getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { |
4522 | unsigned Bitwidth = Ty->getIntegerBitWidth(); |
4523 | if (Bitwidth == 32) |
4524 | return OMPBuilder.getOrCreateRuntimeFunction( |
4525 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u); |
4526 | if (Bitwidth == 64) |
4527 | return OMPBuilder.getOrCreateRuntimeFunction( |
4528 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u); |
4529 | llvm_unreachable("unknown OpenMP loop iterator bitwidth" ); |
4530 | } |
4531 | |
4532 | /// Returns an LLVM function to call for finalizing the dynamic loop using |
4533 | /// depending on `type`. Only i32 and i64 are supported by the runtime. Always |
4534 | /// interpret integers as unsigned similarly to CanonicalLoopInfo. |
4535 | static FunctionCallee |
4536 | getKmpcForDynamicFiniForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { |
4537 | unsigned Bitwidth = Ty->getIntegerBitWidth(); |
4538 | if (Bitwidth == 32) |
4539 | return OMPBuilder.getOrCreateRuntimeFunction( |
4540 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_4u); |
4541 | if (Bitwidth == 64) |
4542 | return OMPBuilder.getOrCreateRuntimeFunction( |
4543 | M, FnID: omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_8u); |
4544 | llvm_unreachable("unknown OpenMP loop iterator bitwidth" ); |
4545 | } |
4546 | |
4547 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyDynamicWorkshareLoop( |
4548 | DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, |
4549 | OMPScheduleType SchedType, bool NeedsBarrier, Value *Chunk) { |
4550 | assert(CLI->isValid() && "Requires a valid canonical loop" ); |
4551 | assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) && |
4552 | "Require dedicated allocate IP" ); |
4553 | assert(isValidWorkshareLoopScheduleType(SchedType) && |
4554 | "Require valid schedule type" ); |
4555 | |
4556 | bool Ordered = (SchedType & OMPScheduleType::ModifierOrdered) == |
4557 | OMPScheduleType::ModifierOrdered; |
4558 | |
4559 | // Set up the source location value for OpenMP runtime. |
4560 | Builder.SetCurrentDebugLocation(DL); |
4561 | |
4562 | uint32_t SrcLocStrSize; |
4563 | Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); |
4564 | Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
4565 | |
4566 | // Declare useful OpenMP runtime functions. |
4567 | Value *IV = CLI->getIndVar(); |
4568 | Type *IVTy = IV->getType(); |
4569 | FunctionCallee DynamicInit = getKmpcForDynamicInitForType(Ty: IVTy, M, OMPBuilder&: *this); |
4570 | FunctionCallee DynamicNext = getKmpcForDynamicNextForType(Ty: IVTy, M, OMPBuilder&: *this); |
4571 | |
4572 | // Allocate space for computed loop bounds as expected by the "init" function. |
4573 | Builder.SetInsertPoint(AllocaIP.getBlock()->getFirstNonPHIOrDbgOrAlloca()); |
4574 | Type *I32Type = Type::getInt32Ty(C&: M.getContext()); |
4575 | Value *PLastIter = Builder.CreateAlloca(Ty: I32Type, ArraySize: nullptr, Name: "p.lastiter" ); |
4576 | Value *PLowerBound = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.lowerbound" ); |
4577 | Value *PUpperBound = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.upperbound" ); |
4578 | Value *PStride = Builder.CreateAlloca(Ty: IVTy, ArraySize: nullptr, Name: "p.stride" ); |
4579 | |
4580 | // At the end of the preheader, prepare for calling the "init" function by |
4581 | // storing the current loop bounds into the allocated space. A canonical loop |
4582 | // always iterates from 0 to trip-count with step 1. Note that "init" expects |
4583 | // and produces an inclusive upper bound. |
4584 | BasicBlock * = CLI->getPreheader(); |
4585 | Builder.SetInsertPoint(PreHeader->getTerminator()); |
4586 | Constant *One = ConstantInt::get(Ty: IVTy, V: 1); |
4587 | Builder.CreateStore(Val: One, Ptr: PLowerBound); |
4588 | Value *UpperBound = CLI->getTripCount(); |
4589 | Builder.CreateStore(Val: UpperBound, Ptr: PUpperBound); |
4590 | Builder.CreateStore(Val: One, Ptr: PStride); |
4591 | |
4592 | BasicBlock * = CLI->getHeader(); |
4593 | BasicBlock *Exit = CLI->getExit(); |
4594 | BasicBlock *Cond = CLI->getCond(); |
4595 | BasicBlock *Latch = CLI->getLatch(); |
4596 | InsertPointTy AfterIP = CLI->getAfterIP(); |
4597 | |
4598 | // The CLI will be "broken" in the code below, as the loop is no longer |
4599 | // a valid canonical loop. |
4600 | |
4601 | if (!Chunk) |
4602 | Chunk = One; |
4603 | |
4604 | Value *ThreadNum = getOrCreateThreadID(Ident: SrcLoc); |
4605 | |
4606 | Constant *SchedulingType = |
4607 | ConstantInt::get(Ty: I32Type, V: static_cast<int>(SchedType)); |
4608 | |
4609 | // Call the "init" function. |
4610 | Builder.CreateCall(Callee: DynamicInit, |
4611 | Args: {SrcLoc, ThreadNum, SchedulingType, /* LowerBound */ One, |
4612 | UpperBound, /* step */ One, Chunk}); |
4613 | |
4614 | // An outer loop around the existing one. |
4615 | BasicBlock *OuterCond = BasicBlock::Create( |
4616 | Context&: PreHeader->getContext(), Name: Twine(PreHeader->getName()) + ".outer.cond" , |
4617 | Parent: PreHeader->getParent()); |
4618 | // This needs to be 32-bit always, so can't use the IVTy Zero above. |
4619 | Builder.SetInsertPoint(TheBB: OuterCond, IP: OuterCond->getFirstInsertionPt()); |
4620 | Value *Res = |
4621 | Builder.CreateCall(Callee: DynamicNext, Args: {SrcLoc, ThreadNum, PLastIter, |
4622 | PLowerBound, PUpperBound, PStride}); |
4623 | Constant *Zero32 = ConstantInt::get(Ty: I32Type, V: 0); |
4624 | Value *MoreWork = Builder.CreateCmp(Pred: CmpInst::ICMP_NE, LHS: Res, RHS: Zero32); |
4625 | Value *LowerBound = |
4626 | Builder.CreateSub(LHS: Builder.CreateLoad(Ty: IVTy, Ptr: PLowerBound), RHS: One, Name: "lb" ); |
4627 | Builder.CreateCondBr(Cond: MoreWork, True: Header, False: Exit); |
4628 | |
4629 | // Change PHI-node in loop header to use outer cond rather than preheader, |
4630 | // and set IV to the LowerBound. |
4631 | Instruction *Phi = &Header->front(); |
4632 | auto *PI = cast<PHINode>(Val: Phi); |
4633 | PI->setIncomingBlock(i: 0, BB: OuterCond); |
4634 | PI->setIncomingValue(i: 0, V: LowerBound); |
4635 | |
4636 | // Then set the pre-header to jump to the OuterCond |
4637 | Instruction *Term = PreHeader->getTerminator(); |
4638 | auto *Br = cast<BranchInst>(Val: Term); |
4639 | Br->setSuccessor(idx: 0, NewSucc: OuterCond); |
4640 | |
4641 | // Modify the inner condition: |
4642 | // * Use the UpperBound returned from the DynamicNext call. |
4643 | // * jump to the loop outer loop when done with one of the inner loops. |
4644 | Builder.SetInsertPoint(TheBB: Cond, IP: Cond->getFirstInsertionPt()); |
4645 | UpperBound = Builder.CreateLoad(Ty: IVTy, Ptr: PUpperBound, Name: "ub" ); |
4646 | Instruction *Comp = &*Builder.GetInsertPoint(); |
4647 | auto *CI = cast<CmpInst>(Val: Comp); |
4648 | CI->setOperand(i_nocapture: 1, Val_nocapture: UpperBound); |
4649 | // Redirect the inner exit to branch to outer condition. |
4650 | Instruction *Branch = &Cond->back(); |
4651 | auto *BI = cast<BranchInst>(Val: Branch); |
4652 | assert(BI->getSuccessor(1) == Exit); |
4653 | BI->setSuccessor(idx: 1, NewSucc: OuterCond); |
4654 | |
4655 | // Call the "fini" function if "ordered" is present in wsloop directive. |
4656 | if (Ordered) { |
4657 | Builder.SetInsertPoint(&Latch->back()); |
4658 | FunctionCallee DynamicFini = getKmpcForDynamicFiniForType(Ty: IVTy, M, OMPBuilder&: *this); |
4659 | Builder.CreateCall(Callee: DynamicFini, Args: {SrcLoc, ThreadNum}); |
4660 | } |
4661 | |
4662 | // Add the barrier if requested. |
4663 | if (NeedsBarrier) { |
4664 | Builder.SetInsertPoint(&Exit->back()); |
4665 | createBarrier(Loc: LocationDescription(Builder.saveIP(), DL), |
4666 | Kind: omp::Directive::OMPD_for, /* ForceSimpleCall */ false, |
4667 | /* CheckCancelFlag */ false); |
4668 | } |
4669 | |
4670 | CLI->invalidate(); |
4671 | return AfterIP; |
4672 | } |
4673 | |
4674 | /// Redirect all edges that branch to \p OldTarget to \p NewTarget. That is, |
4675 | /// after this \p OldTarget will be orphaned. |
4676 | static void redirectAllPredecessorsTo(BasicBlock *OldTarget, |
4677 | BasicBlock *NewTarget, DebugLoc DL) { |
4678 | for (BasicBlock *Pred : make_early_inc_range(Range: predecessors(BB: OldTarget))) |
4679 | redirectTo(Source: Pred, Target: NewTarget, DL); |
4680 | } |
4681 | |
4682 | /// Determine which blocks in \p BBs are reachable from outside and remove the |
4683 | /// ones that are not reachable from the function. |
4684 | static void removeUnusedBlocksFromParent(ArrayRef<BasicBlock *> BBs) { |
4685 | SmallPtrSet<BasicBlock *, 6> BBsToErase{BBs.begin(), BBs.end()}; |
4686 | auto HasRemainingUses = [&BBsToErase](BasicBlock *BB) { |
4687 | for (Use &U : BB->uses()) { |
4688 | auto *UseInst = dyn_cast<Instruction>(Val: U.getUser()); |
4689 | if (!UseInst) |
4690 | continue; |
4691 | if (BBsToErase.count(Ptr: UseInst->getParent())) |
4692 | continue; |
4693 | return true; |
4694 | } |
4695 | return false; |
4696 | }; |
4697 | |
4698 | while (BBsToErase.remove_if(P: HasRemainingUses)) { |
4699 | // Try again if anything was removed. |
4700 | } |
4701 | |
4702 | SmallVector<BasicBlock *, 7> BBVec(BBsToErase.begin(), BBsToErase.end()); |
4703 | DeleteDeadBlocks(BBs: BBVec); |
4704 | } |
4705 | |
4706 | CanonicalLoopInfo * |
4707 | OpenMPIRBuilder::collapseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops, |
4708 | InsertPointTy ComputeIP) { |
4709 | assert(Loops.size() >= 1 && "At least one loop required" ); |
4710 | size_t NumLoops = Loops.size(); |
4711 | |
4712 | // Nothing to do if there is already just one loop. |
4713 | if (NumLoops == 1) |
4714 | return Loops.front(); |
4715 | |
4716 | CanonicalLoopInfo *Outermost = Loops.front(); |
4717 | CanonicalLoopInfo *Innermost = Loops.back(); |
4718 | BasicBlock * = Outermost->getPreheader(); |
4719 | BasicBlock *OrigAfter = Outermost->getAfter(); |
4720 | Function *F = OrigPreheader->getParent(); |
4721 | |
4722 | // Loop control blocks that may become orphaned later. |
4723 | SmallVector<BasicBlock *, 12> OldControlBBs; |
4724 | OldControlBBs.reserve(N: 6 * Loops.size()); |
4725 | for (CanonicalLoopInfo *Loop : Loops) |
4726 | Loop->collectControlBlocks(BBs&: OldControlBBs); |
4727 | |
4728 | // Setup the IRBuilder for inserting the trip count computation. |
4729 | Builder.SetCurrentDebugLocation(DL); |
4730 | if (ComputeIP.isSet()) |
4731 | Builder.restoreIP(IP: ComputeIP); |
4732 | else |
4733 | Builder.restoreIP(IP: Outermost->getPreheaderIP()); |
4734 | |
4735 | // Derive the collapsed' loop trip count. |
4736 | // TODO: Find common/largest indvar type. |
4737 | Value *CollapsedTripCount = nullptr; |
4738 | for (CanonicalLoopInfo *L : Loops) { |
4739 | assert(L->isValid() && |
4740 | "All loops to collapse must be valid canonical loops" ); |
4741 | Value *OrigTripCount = L->getTripCount(); |
4742 | if (!CollapsedTripCount) { |
4743 | CollapsedTripCount = OrigTripCount; |
4744 | continue; |
4745 | } |
4746 | |
4747 | // TODO: Enable UndefinedSanitizer to diagnose an overflow here. |
4748 | CollapsedTripCount = Builder.CreateMul(LHS: CollapsedTripCount, RHS: OrigTripCount, |
4749 | Name: {}, /*HasNUW=*/true); |
4750 | } |
4751 | |
4752 | // Create the collapsed loop control flow. |
4753 | CanonicalLoopInfo *Result = |
4754 | createLoopSkeleton(DL, TripCount: CollapsedTripCount, F, |
4755 | PreInsertBefore: OrigPreheader->getNextNode(), PostInsertBefore: OrigAfter, Name: "collapsed" ); |
4756 | |
4757 | // Build the collapsed loop body code. |
4758 | // Start with deriving the input loop induction variables from the collapsed |
4759 | // one, using a divmod scheme. To preserve the original loops' order, the |
4760 | // innermost loop use the least significant bits. |
4761 | Builder.restoreIP(IP: Result->getBodyIP()); |
4762 | |
4763 | Value *Leftover = Result->getIndVar(); |
4764 | SmallVector<Value *> NewIndVars; |
4765 | NewIndVars.resize(N: NumLoops); |
4766 | for (int i = NumLoops - 1; i >= 1; --i) { |
4767 | Value *OrigTripCount = Loops[i]->getTripCount(); |
4768 | |
4769 | Value *NewIndVar = Builder.CreateURem(LHS: Leftover, RHS: OrigTripCount); |
4770 | NewIndVars[i] = NewIndVar; |
4771 | |
4772 | Leftover = Builder.CreateUDiv(LHS: Leftover, RHS: OrigTripCount); |
4773 | } |
4774 | // Outermost loop gets all the remaining bits. |
4775 | NewIndVars[0] = Leftover; |
4776 | |
4777 | // Construct the loop body control flow. |
4778 | // We progressively construct the branch structure following in direction of |
4779 | // the control flow, from the leading in-between code, the loop nest body, the |
4780 | // trailing in-between code, and rejoining the collapsed loop's latch. |
4781 | // ContinueBlock and ContinuePred keep track of the source(s) of next edge. If |
4782 | // the ContinueBlock is set, continue with that block. If ContinuePred, use |
4783 | // its predecessors as sources. |
4784 | BasicBlock *ContinueBlock = Result->getBody(); |
4785 | BasicBlock *ContinuePred = nullptr; |
4786 | auto ContinueWith = [&ContinueBlock, &ContinuePred, DL](BasicBlock *Dest, |
4787 | BasicBlock *NextSrc) { |
4788 | if (ContinueBlock) |
4789 | redirectTo(Source: ContinueBlock, Target: Dest, DL); |
4790 | else |
4791 | redirectAllPredecessorsTo(OldTarget: ContinuePred, NewTarget: Dest, DL); |
4792 | |
4793 | ContinueBlock = nullptr; |
4794 | ContinuePred = NextSrc; |
4795 | }; |
4796 | |
4797 | // The code before the nested loop of each level. |
4798 | // Because we are sinking it into the nest, it will be executed more often |
4799 | // that the original loop. More sophisticated schemes could keep track of what |
4800 | // the in-between code is and instantiate it only once per thread. |
4801 | for (size_t i = 0; i < NumLoops - 1; ++i) |
4802 | ContinueWith(Loops[i]->getBody(), Loops[i + 1]->getHeader()); |
4803 | |
4804 | // Connect the loop nest body. |
4805 | ContinueWith(Innermost->getBody(), Innermost->getLatch()); |
4806 | |
4807 | // The code after the nested loop at each level. |
4808 | for (size_t i = NumLoops - 1; i > 0; --i) |
4809 | ContinueWith(Loops[i]->getAfter(), Loops[i - 1]->getLatch()); |
4810 | |
4811 | // Connect the finished loop to the collapsed loop latch. |
4812 | ContinueWith(Result->getLatch(), nullptr); |
4813 | |
4814 | // Replace the input loops with the new collapsed loop. |
4815 | redirectTo(Source: Outermost->getPreheader(), Target: Result->getPreheader(), DL); |
4816 | redirectTo(Source: Result->getAfter(), Target: Outermost->getAfter(), DL); |
4817 | |
4818 | // Replace the input loop indvars with the derived ones. |
4819 | for (size_t i = 0; i < NumLoops; ++i) |
4820 | Loops[i]->getIndVar()->replaceAllUsesWith(V: NewIndVars[i]); |
4821 | |
4822 | // Remove unused parts of the input loops. |
4823 | removeUnusedBlocksFromParent(BBs: OldControlBBs); |
4824 | |
4825 | for (CanonicalLoopInfo *L : Loops) |
4826 | L->invalidate(); |
4827 | |
4828 | #ifndef NDEBUG |
4829 | Result->assertOK(); |
4830 | #endif |
4831 | return Result; |
4832 | } |
4833 | |
4834 | std::vector<CanonicalLoopInfo *> |
4835 | OpenMPIRBuilder::tileLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops, |
4836 | ArrayRef<Value *> TileSizes) { |
4837 | assert(TileSizes.size() == Loops.size() && |
4838 | "Must pass as many tile sizes as there are loops" ); |
4839 | int NumLoops = Loops.size(); |
4840 | assert(NumLoops >= 1 && "At least one loop to tile required" ); |
4841 | |
4842 | CanonicalLoopInfo *OutermostLoop = Loops.front(); |
4843 | CanonicalLoopInfo *InnermostLoop = Loops.back(); |
4844 | Function *F = OutermostLoop->getBody()->getParent(); |
4845 | BasicBlock *InnerEnter = InnermostLoop->getBody(); |
4846 | BasicBlock *InnerLatch = InnermostLoop->getLatch(); |
4847 | |
4848 | // Loop control blocks that may become orphaned later. |
4849 | SmallVector<BasicBlock *, 12> OldControlBBs; |
4850 | OldControlBBs.reserve(N: 6 * Loops.size()); |
4851 | for (CanonicalLoopInfo *Loop : Loops) |
4852 | Loop->collectControlBlocks(BBs&: OldControlBBs); |
4853 | |
4854 | // Collect original trip counts and induction variable to be accessible by |
4855 | // index. Also, the structure of the original loops is not preserved during |
4856 | // the construction of the tiled loops, so do it before we scavenge the BBs of |
4857 | // any original CanonicalLoopInfo. |
4858 | SmallVector<Value *, 4> OrigTripCounts, OrigIndVars; |
4859 | for (CanonicalLoopInfo *L : Loops) { |
4860 | assert(L->isValid() && "All input loops must be valid canonical loops" ); |
4861 | OrigTripCounts.push_back(Elt: L->getTripCount()); |
4862 | OrigIndVars.push_back(Elt: L->getIndVar()); |
4863 | } |
4864 | |
4865 | // Collect the code between loop headers. These may contain SSA definitions |
4866 | // that are used in the loop nest body. To be usable with in the innermost |
4867 | // body, these BasicBlocks will be sunk into the loop nest body. That is, |
4868 | // these instructions may be executed more often than before the tiling. |
4869 | // TODO: It would be sufficient to only sink them into body of the |
4870 | // corresponding tile loop. |
4871 | SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> InbetweenCode; |
4872 | for (int i = 0; i < NumLoops - 1; ++i) { |
4873 | CanonicalLoopInfo *Surrounding = Loops[i]; |
4874 | CanonicalLoopInfo *Nested = Loops[i + 1]; |
4875 | |
4876 | BasicBlock *EnterBB = Surrounding->getBody(); |
4877 | BasicBlock *ExitBB = Nested->getHeader(); |
4878 | InbetweenCode.emplace_back(Args&: EnterBB, Args&: ExitBB); |
4879 | } |
4880 | |
4881 | // Compute the trip counts of the floor loops. |
4882 | Builder.SetCurrentDebugLocation(DL); |
4883 | Builder.restoreIP(IP: OutermostLoop->getPreheaderIP()); |
4884 | SmallVector<Value *, 4> FloorCount, FloorRems; |
4885 | for (int i = 0; i < NumLoops; ++i) { |
4886 | Value *TileSize = TileSizes[i]; |
4887 | Value *OrigTripCount = OrigTripCounts[i]; |
4888 | Type *IVType = OrigTripCount->getType(); |
4889 | |
4890 | Value *FloorTripCount = Builder.CreateUDiv(LHS: OrigTripCount, RHS: TileSize); |
4891 | Value *FloorTripRem = Builder.CreateURem(LHS: OrigTripCount, RHS: TileSize); |
4892 | |
4893 | // 0 if tripcount divides the tilesize, 1 otherwise. |
4894 | // 1 means we need an additional iteration for a partial tile. |
4895 | // |
4896 | // Unfortunately we cannot just use the roundup-formula |
4897 | // (tripcount + tilesize - 1)/tilesize |
4898 | // because the summation might overflow. We do not want introduce undefined |
4899 | // behavior when the untiled loop nest did not. |
4900 | Value *FloorTripOverflow = |
4901 | Builder.CreateICmpNE(LHS: FloorTripRem, RHS: ConstantInt::get(Ty: IVType, V: 0)); |
4902 | |
4903 | FloorTripOverflow = Builder.CreateZExt(V: FloorTripOverflow, DestTy: IVType); |
4904 | FloorTripCount = |
4905 | Builder.CreateAdd(LHS: FloorTripCount, RHS: FloorTripOverflow, |
4906 | Name: "omp_floor" + Twine(i) + ".tripcount" , HasNUW: true); |
4907 | |
4908 | // Remember some values for later use. |
4909 | FloorCount.push_back(Elt: FloorTripCount); |
4910 | FloorRems.push_back(Elt: FloorTripRem); |
4911 | } |
4912 | |
4913 | // Generate the new loop nest, from the outermost to the innermost. |
4914 | std::vector<CanonicalLoopInfo *> Result; |
4915 | Result.reserve(n: NumLoops * 2); |
4916 | |
4917 | // The basic block of the surrounding loop that enters the nest generated |
4918 | // loop. |
4919 | BasicBlock *Enter = OutermostLoop->getPreheader(); |
4920 | |
4921 | // The basic block of the surrounding loop where the inner code should |
4922 | // continue. |
4923 | BasicBlock *Continue = OutermostLoop->getAfter(); |
4924 | |
4925 | // Where the next loop basic block should be inserted. |
4926 | BasicBlock *OutroInsertBefore = InnermostLoop->getExit(); |
4927 | |
4928 | auto EmbeddNewLoop = |
4929 | [this, DL, F, InnerEnter, &Enter, &Continue, &OutroInsertBefore]( |
4930 | Value *TripCount, const Twine &Name) -> CanonicalLoopInfo * { |
4931 | CanonicalLoopInfo *EmbeddedLoop = createLoopSkeleton( |
4932 | DL, TripCount, F, PreInsertBefore: InnerEnter, PostInsertBefore: OutroInsertBefore, Name); |
4933 | redirectTo(Source: Enter, Target: EmbeddedLoop->getPreheader(), DL); |
4934 | redirectTo(Source: EmbeddedLoop->getAfter(), Target: Continue, DL); |
4935 | |
4936 | // Setup the position where the next embedded loop connects to this loop. |
4937 | Enter = EmbeddedLoop->getBody(); |
4938 | Continue = EmbeddedLoop->getLatch(); |
4939 | OutroInsertBefore = EmbeddedLoop->getLatch(); |
4940 | return EmbeddedLoop; |
4941 | }; |
4942 | |
4943 | auto EmbeddNewLoops = [&Result, &EmbeddNewLoop](ArrayRef<Value *> TripCounts, |
4944 | const Twine &NameBase) { |
4945 | for (auto P : enumerate(First&: TripCounts)) { |
4946 | CanonicalLoopInfo *EmbeddedLoop = |
4947 | EmbeddNewLoop(P.value(), NameBase + Twine(P.index())); |
4948 | Result.push_back(x: EmbeddedLoop); |
4949 | } |
4950 | }; |
4951 | |
4952 | EmbeddNewLoops(FloorCount, "floor" ); |
4953 | |
4954 | // Within the innermost floor loop, emit the code that computes the tile |
4955 | // sizes. |
4956 | Builder.SetInsertPoint(Enter->getTerminator()); |
4957 | SmallVector<Value *, 4> TileCounts; |
4958 | for (int i = 0; i < NumLoops; ++i) { |
4959 | CanonicalLoopInfo *FloorLoop = Result[i]; |
4960 | Value *TileSize = TileSizes[i]; |
4961 | |
4962 | Value *FloorIsEpilogue = |
4963 | Builder.CreateICmpEQ(LHS: FloorLoop->getIndVar(), RHS: FloorCount[i]); |
4964 | Value *TileTripCount = |
4965 | Builder.CreateSelect(C: FloorIsEpilogue, True: FloorRems[i], False: TileSize); |
4966 | |
4967 | TileCounts.push_back(Elt: TileTripCount); |
4968 | } |
4969 | |
4970 | // Create the tile loops. |
4971 | EmbeddNewLoops(TileCounts, "tile" ); |
4972 | |
4973 | // Insert the inbetween code into the body. |
4974 | BasicBlock *BodyEnter = Enter; |
4975 | BasicBlock *BodyEntered = nullptr; |
4976 | for (std::pair<BasicBlock *, BasicBlock *> P : InbetweenCode) { |
4977 | BasicBlock *EnterBB = P.first; |
4978 | BasicBlock *ExitBB = P.second; |
4979 | |
4980 | if (BodyEnter) |
4981 | redirectTo(Source: BodyEnter, Target: EnterBB, DL); |
4982 | else |
4983 | redirectAllPredecessorsTo(OldTarget: BodyEntered, NewTarget: EnterBB, DL); |
4984 | |
4985 | BodyEnter = nullptr; |
4986 | BodyEntered = ExitBB; |
4987 | } |
4988 | |
4989 | // Append the original loop nest body into the generated loop nest body. |
4990 | if (BodyEnter) |
4991 | redirectTo(Source: BodyEnter, Target: InnerEnter, DL); |
4992 | else |
4993 | redirectAllPredecessorsTo(OldTarget: BodyEntered, NewTarget: InnerEnter, DL); |
4994 | redirectAllPredecessorsTo(OldTarget: InnerLatch, NewTarget: Continue, DL); |
4995 | |
4996 | // Replace the original induction variable with an induction variable computed |
4997 | // from the tile and floor induction variables. |
4998 | Builder.restoreIP(IP: Result.back()->getBodyIP()); |
4999 | for (int i = 0; i < NumLoops; ++i) { |
5000 | CanonicalLoopInfo *FloorLoop = Result[i]; |
5001 | CanonicalLoopInfo *TileLoop = Result[NumLoops + i]; |
5002 | Value *OrigIndVar = OrigIndVars[i]; |
5003 | Value *Size = TileSizes[i]; |
5004 | |
5005 | Value *Scale = |
5006 | Builder.CreateMul(LHS: Size, RHS: FloorLoop->getIndVar(), Name: {}, /*HasNUW=*/true); |
5007 | Value *Shift = |
5008 | Builder.CreateAdd(LHS: Scale, RHS: TileLoop->getIndVar(), Name: {}, /*HasNUW=*/true); |
5009 | OrigIndVar->replaceAllUsesWith(V: Shift); |
5010 | } |
5011 | |
5012 | // Remove unused parts of the original loops. |
5013 | removeUnusedBlocksFromParent(BBs: OldControlBBs); |
5014 | |
5015 | for (CanonicalLoopInfo *L : Loops) |
5016 | L->invalidate(); |
5017 | |
5018 | #ifndef NDEBUG |
5019 | for (CanonicalLoopInfo *GenL : Result) |
5020 | GenL->assertOK(); |
5021 | #endif |
5022 | return Result; |
5023 | } |
5024 | |
5025 | /// Attach metadata \p Properties to the basic block described by \p BB. If the |
5026 | /// basic block already has metadata, the basic block properties are appended. |
5027 | static void addBasicBlockMetadata(BasicBlock *BB, |
5028 | ArrayRef<Metadata *> Properties) { |
5029 | // Nothing to do if no property to attach. |
5030 | if (Properties.empty()) |
5031 | return; |
5032 | |
5033 | LLVMContext &Ctx = BB->getContext(); |
5034 | SmallVector<Metadata *> NewProperties; |
5035 | NewProperties.push_back(Elt: nullptr); |
5036 | |
5037 | // If the basic block already has metadata, prepend it to the new metadata. |
5038 | MDNode *Existing = BB->getTerminator()->getMetadata(KindID: LLVMContext::MD_loop); |
5039 | if (Existing) |
5040 | append_range(C&: NewProperties, R: drop_begin(RangeOrContainer: Existing->operands(), N: 1)); |
5041 | |
5042 | append_range(C&: NewProperties, R&: Properties); |
5043 | MDNode *BasicBlockID = MDNode::getDistinct(Context&: Ctx, MDs: NewProperties); |
5044 | BasicBlockID->replaceOperandWith(I: 0, New: BasicBlockID); |
5045 | |
5046 | BB->getTerminator()->setMetadata(KindID: LLVMContext::MD_loop, Node: BasicBlockID); |
5047 | } |
5048 | |
5049 | /// Attach loop metadata \p Properties to the loop described by \p Loop. If the |
5050 | /// loop already has metadata, the loop properties are appended. |
5051 | static void addLoopMetadata(CanonicalLoopInfo *Loop, |
5052 | ArrayRef<Metadata *> Properties) { |
5053 | assert(Loop->isValid() && "Expecting a valid CanonicalLoopInfo" ); |
5054 | |
5055 | // Attach metadata to the loop's latch |
5056 | BasicBlock *Latch = Loop->getLatch(); |
5057 | assert(Latch && "A valid CanonicalLoopInfo must have a unique latch" ); |
5058 | addBasicBlockMetadata(BB: Latch, Properties); |
5059 | } |
5060 | |
5061 | /// Attach llvm.access.group metadata to the memref instructions of \p Block |
5062 | static void addSimdMetadata(BasicBlock *Block, MDNode *AccessGroup, |
5063 | LoopInfo &LI) { |
5064 | for (Instruction &I : *Block) { |
5065 | if (I.mayReadOrWriteMemory()) { |
5066 | // TODO: This instruction may already have access group from |
5067 | // other pragmas e.g. #pragma clang loop vectorize. Append |
5068 | // so that the existing metadata is not overwritten. |
5069 | I.setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessGroup); |
5070 | } |
5071 | } |
5072 | } |
5073 | |
5074 | void OpenMPIRBuilder::unrollLoopFull(DebugLoc, CanonicalLoopInfo *Loop) { |
5075 | LLVMContext &Ctx = Builder.getContext(); |
5076 | addLoopMetadata( |
5077 | Loop, Properties: {MDNode::get(Context&: Ctx, MDs: MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.enable" )), |
5078 | MDNode::get(Context&: Ctx, MDs: MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.full" ))}); |
5079 | } |
5080 | |
5081 | void OpenMPIRBuilder::unrollLoopHeuristic(DebugLoc, CanonicalLoopInfo *Loop) { |
5082 | LLVMContext &Ctx = Builder.getContext(); |
5083 | addLoopMetadata( |
5084 | Loop, Properties: { |
5085 | MDNode::get(Context&: Ctx, MDs: MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.enable" )), |
5086 | }); |
5087 | } |
5088 | |
5089 | void OpenMPIRBuilder::createIfVersion(CanonicalLoopInfo *CanonicalLoop, |
5090 | Value *IfCond, ValueToValueMapTy &VMap, |
5091 | const Twine &NamePrefix) { |
5092 | Function *F = CanonicalLoop->getFunction(); |
5093 | |
5094 | // Define where if branch should be inserted |
5095 | Instruction *SplitBefore; |
5096 | if (Instruction::classof(V: IfCond)) { |
5097 | SplitBefore = dyn_cast<Instruction>(Val: IfCond); |
5098 | } else { |
5099 | SplitBefore = CanonicalLoop->getPreheader()->getTerminator(); |
5100 | } |
5101 | |
5102 | // TODO: We should not rely on pass manager. Currently we use pass manager |
5103 | // only for getting llvm::Loop which corresponds to given CanonicalLoopInfo |
5104 | // object. We should have a method which returns all blocks between |
5105 | // CanonicalLoopInfo::getHeader() and CanonicalLoopInfo::getAfter() |
5106 | FunctionAnalysisManager FAM; |
5107 | FAM.registerPass(PassBuilder: []() { return DominatorTreeAnalysis(); }); |
5108 | FAM.registerPass(PassBuilder: []() { return LoopAnalysis(); }); |
5109 | FAM.registerPass(PassBuilder: []() { return PassInstrumentationAnalysis(); }); |
5110 | |
5111 | // Get the loop which needs to be cloned |
5112 | LoopAnalysis LIA; |
5113 | LoopInfo &&LI = LIA.run(F&: *F, AM&: FAM); |
5114 | Loop *L = LI.getLoopFor(BB: CanonicalLoop->getHeader()); |
5115 | |
5116 | // Create additional blocks for the if statement |
5117 | BasicBlock *Head = SplitBefore->getParent(); |
5118 | Instruction *HeadOldTerm = Head->getTerminator(); |
5119 | llvm::LLVMContext &C = Head->getContext(); |
5120 | llvm::BasicBlock *ThenBlock = llvm::BasicBlock::Create( |
5121 | Context&: C, Name: NamePrefix + ".if.then" , Parent: Head->getParent(), InsertBefore: Head->getNextNode()); |
5122 | llvm::BasicBlock *ElseBlock = llvm::BasicBlock::Create( |
5123 | Context&: C, Name: NamePrefix + ".if.else" , Parent: Head->getParent(), InsertBefore: CanonicalLoop->getExit()); |
5124 | |
5125 | // Create if condition branch. |
5126 | Builder.SetInsertPoint(HeadOldTerm); |
5127 | Instruction *BrInstr = |
5128 | Builder.CreateCondBr(Cond: IfCond, True: ThenBlock, /*ifFalse*/ False: ElseBlock); |
5129 | InsertPointTy IP{BrInstr->getParent(), ++BrInstr->getIterator()}; |
5130 | // Then block contains branch to omp loop which needs to be vectorized |
5131 | spliceBB(IP, New: ThenBlock, CreateBranch: false); |
5132 | ThenBlock->replaceSuccessorsPhiUsesWith(Old: Head, New: ThenBlock); |
5133 | |
5134 | Builder.SetInsertPoint(ElseBlock); |
5135 | |
5136 | // Clone loop for the else branch |
5137 | SmallVector<BasicBlock *, 8> NewBlocks; |
5138 | |
5139 | VMap[CanonicalLoop->getPreheader()] = ElseBlock; |
5140 | for (BasicBlock *Block : L->getBlocks()) { |
5141 | BasicBlock *NewBB = CloneBasicBlock(BB: Block, VMap, NameSuffix: "" , F); |
5142 | NewBB->moveBefore(MovePos: CanonicalLoop->getExit()); |
5143 | VMap[Block] = NewBB; |
5144 | NewBlocks.push_back(Elt: NewBB); |
5145 | } |
5146 | remapInstructionsInBlocks(Blocks: NewBlocks, VMap); |
5147 | Builder.CreateBr(Dest: NewBlocks.front()); |
5148 | } |
5149 | |
5150 | unsigned |
5151 | OpenMPIRBuilder::getOpenMPDefaultSimdAlign(const Triple &TargetTriple, |
5152 | const StringMap<bool> &Features) { |
5153 | if (TargetTriple.isX86()) { |
5154 | if (Features.lookup(Key: "avx512f" )) |
5155 | return 512; |
5156 | else if (Features.lookup(Key: "avx" )) |
5157 | return 256; |
5158 | return 128; |
5159 | } |
5160 | if (TargetTriple.isPPC()) |
5161 | return 128; |
5162 | if (TargetTriple.isWasm()) |
5163 | return 128; |
5164 | return 0; |
5165 | } |
5166 | |
5167 | void OpenMPIRBuilder::applySimd(CanonicalLoopInfo *CanonicalLoop, |
5168 | MapVector<Value *, Value *> AlignedVars, |
5169 | Value *IfCond, OrderKind Order, |
5170 | ConstantInt *Simdlen, ConstantInt *Safelen) { |
5171 | LLVMContext &Ctx = Builder.getContext(); |
5172 | |
5173 | Function *F = CanonicalLoop->getFunction(); |
5174 | |
5175 | // TODO: We should not rely on pass manager. Currently we use pass manager |
5176 | // only for getting llvm::Loop which corresponds to given CanonicalLoopInfo |
5177 | // object. We should have a method which returns all blocks between |
5178 | // CanonicalLoopInfo::getHeader() and CanonicalLoopInfo::getAfter() |
5179 | FunctionAnalysisManager FAM; |
5180 | FAM.registerPass(PassBuilder: []() { return DominatorTreeAnalysis(); }); |
5181 | FAM.registerPass(PassBuilder: []() { return LoopAnalysis(); }); |
5182 | FAM.registerPass(PassBuilder: []() { return PassInstrumentationAnalysis(); }); |
5183 | |
5184 | LoopAnalysis LIA; |
5185 | LoopInfo &&LI = LIA.run(F&: *F, AM&: FAM); |
5186 | |
5187 | Loop *L = LI.getLoopFor(BB: CanonicalLoop->getHeader()); |
5188 | if (AlignedVars.size()) { |
5189 | InsertPointTy IP = Builder.saveIP(); |
5190 | Builder.SetInsertPoint(CanonicalLoop->getPreheader()->getTerminator()); |
5191 | for (auto &AlignedItem : AlignedVars) { |
5192 | Value *AlignedPtr = AlignedItem.first; |
5193 | Value *Alignment = AlignedItem.second; |
5194 | Builder.CreateAlignmentAssumption(DL: F->getDataLayout(), |
5195 | PtrValue: AlignedPtr, Alignment); |
5196 | } |
5197 | Builder.restoreIP(IP); |
5198 | } |
5199 | |
5200 | if (IfCond) { |
5201 | ValueToValueMapTy VMap; |
5202 | createIfVersion(CanonicalLoop, IfCond, VMap, NamePrefix: "simd" ); |
5203 | // Add metadata to the cloned loop which disables vectorization |
5204 | Value *MappedLatch = VMap.lookup(Val: CanonicalLoop->getLatch()); |
5205 | assert(MappedLatch && |
5206 | "Cannot find value which corresponds to original loop latch" ); |
5207 | assert(isa<BasicBlock>(MappedLatch) && |
5208 | "Cannot cast mapped latch block value to BasicBlock" ); |
5209 | BasicBlock *NewLatchBlock = dyn_cast<BasicBlock>(Val: MappedLatch); |
5210 | ConstantAsMetadata *BoolConst = |
5211 | ConstantAsMetadata::get(C: ConstantInt::getFalse(Ty: Type::getInt1Ty(C&: Ctx))); |
5212 | addBasicBlockMetadata( |
5213 | BB: NewLatchBlock, |
5214 | Properties: {MDNode::get(Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.vectorize.enable" ), |
5215 | BoolConst})}); |
5216 | } |
5217 | |
5218 | SmallSet<BasicBlock *, 8> Reachable; |
5219 | |
5220 | // Get the basic blocks from the loop in which memref instructions |
5221 | // can be found. |
5222 | // TODO: Generalize getting all blocks inside a CanonicalizeLoopInfo, |
5223 | // preferably without running any passes. |
5224 | for (BasicBlock *Block : L->getBlocks()) { |
5225 | if (Block == CanonicalLoop->getCond() || |
5226 | Block == CanonicalLoop->getHeader()) |
5227 | continue; |
5228 | Reachable.insert(Ptr: Block); |
5229 | } |
5230 | |
5231 | SmallVector<Metadata *> LoopMDList; |
5232 | |
5233 | // In presence of finite 'safelen', it may be unsafe to mark all |
5234 | // the memory instructions parallel, because loop-carried |
5235 | // dependences of 'safelen' iterations are possible. |
5236 | // If clause order(concurrent) is specified then the memory instructions |
5237 | // are marked parallel even if 'safelen' is finite. |
5238 | if ((Safelen == nullptr) || (Order == OrderKind::OMP_ORDER_concurrent)) { |
5239 | // Add access group metadata to memory-access instructions. |
5240 | MDNode *AccessGroup = MDNode::getDistinct(Context&: Ctx, MDs: {}); |
5241 | for (BasicBlock *BB : Reachable) |
5242 | addSimdMetadata(Block: BB, AccessGroup, LI); |
5243 | // TODO: If the loop has existing parallel access metadata, have |
5244 | // to combine two lists. |
5245 | LoopMDList.push_back(Elt: MDNode::get( |
5246 | Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.parallel_accesses" ), AccessGroup})); |
5247 | } |
5248 | |
5249 | // Use the above access group metadata to create loop level |
5250 | // metadata, which should be distinct for each loop. |
5251 | ConstantAsMetadata *BoolConst = |
5252 | ConstantAsMetadata::get(C: ConstantInt::getTrue(Ty: Type::getInt1Ty(C&: Ctx))); |
5253 | LoopMDList.push_back(Elt: MDNode::get( |
5254 | Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.vectorize.enable" ), BoolConst})); |
5255 | |
5256 | if (Simdlen || Safelen) { |
5257 | // If both simdlen and safelen clauses are specified, the value of the |
5258 | // simdlen parameter must be less than or equal to the value of the safelen |
5259 | // parameter. Therefore, use safelen only in the absence of simdlen. |
5260 | ConstantInt *VectorizeWidth = Simdlen == nullptr ? Safelen : Simdlen; |
5261 | LoopMDList.push_back( |
5262 | Elt: MDNode::get(Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.vectorize.width" ), |
5263 | ConstantAsMetadata::get(C: VectorizeWidth)})); |
5264 | } |
5265 | |
5266 | addLoopMetadata(Loop: CanonicalLoop, Properties: LoopMDList); |
5267 | } |
5268 | |
5269 | /// Create the TargetMachine object to query the backend for optimization |
5270 | /// preferences. |
5271 | /// |
5272 | /// Ideally, this would be passed from the front-end to the OpenMPBuilder, but |
5273 | /// e.g. Clang does not pass it to its CodeGen layer and creates it only when |
5274 | /// needed for the LLVM pass pipline. We use some default options to avoid |
5275 | /// having to pass too many settings from the frontend that probably do not |
5276 | /// matter. |
5277 | /// |
5278 | /// Currently, TargetMachine is only used sometimes by the unrollLoopPartial |
5279 | /// method. If we are going to use TargetMachine for more purposes, especially |
5280 | /// those that are sensitive to TargetOptions, RelocModel and CodeModel, it |
5281 | /// might become be worth requiring front-ends to pass on their TargetMachine, |
5282 | /// or at least cache it between methods. Note that while fontends such as Clang |
5283 | /// have just a single main TargetMachine per translation unit, "target-cpu" and |
5284 | /// "target-features" that determine the TargetMachine are per-function and can |
5285 | /// be overrided using __attribute__((target("OPTIONS"))). |
5286 | static std::unique_ptr<TargetMachine> |
5287 | createTargetMachine(Function *F, CodeGenOptLevel OptLevel) { |
5288 | Module *M = F->getParent(); |
5289 | |
5290 | StringRef CPU = F->getFnAttribute(Kind: "target-cpu" ).getValueAsString(); |
5291 | StringRef Features = F->getFnAttribute(Kind: "target-features" ).getValueAsString(); |
5292 | const std::string &Triple = M->getTargetTriple(); |
5293 | |
5294 | std::string Error; |
5295 | const llvm::Target *TheTarget = TargetRegistry::lookupTarget(Triple, Error); |
5296 | if (!TheTarget) |
5297 | return {}; |
5298 | |
5299 | llvm::TargetOptions Options; |
5300 | return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine( |
5301 | TT: Triple, CPU, Features, Options, /*RelocModel=*/RM: std::nullopt, |
5302 | /*CodeModel=*/CM: std::nullopt, OL: OptLevel)); |
5303 | } |
5304 | |
5305 | /// Heuristically determine the best-performant unroll factor for \p CLI. This |
5306 | /// depends on the target processor. We are re-using the same heuristics as the |
5307 | /// LoopUnrollPass. |
5308 | static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI) { |
5309 | Function *F = CLI->getFunction(); |
5310 | |
5311 | // Assume the user requests the most aggressive unrolling, even if the rest of |
5312 | // the code is optimized using a lower setting. |
5313 | CodeGenOptLevel OptLevel = CodeGenOptLevel::Aggressive; |
5314 | std::unique_ptr<TargetMachine> TM = createTargetMachine(F, OptLevel); |
5315 | |
5316 | FunctionAnalysisManager FAM; |
5317 | FAM.registerPass(PassBuilder: []() { return TargetLibraryAnalysis(); }); |
5318 | FAM.registerPass(PassBuilder: []() { return AssumptionAnalysis(); }); |
5319 | FAM.registerPass(PassBuilder: []() { return DominatorTreeAnalysis(); }); |
5320 | FAM.registerPass(PassBuilder: []() { return LoopAnalysis(); }); |
5321 | FAM.registerPass(PassBuilder: []() { return ScalarEvolutionAnalysis(); }); |
5322 | FAM.registerPass(PassBuilder: []() { return PassInstrumentationAnalysis(); }); |
5323 | TargetIRAnalysis TIRA; |
5324 | if (TM) |
5325 | TIRA = TargetIRAnalysis( |
5326 | [&](const Function &F) { return TM->getTargetTransformInfo(F); }); |
5327 | FAM.registerPass(PassBuilder: [&]() { return TIRA; }); |
5328 | |
5329 | TargetIRAnalysis::Result &&TTI = TIRA.run(F: *F, FAM); |
5330 | ScalarEvolutionAnalysis SEA; |
5331 | ScalarEvolution &&SE = SEA.run(F&: *F, AM&: FAM); |
5332 | DominatorTreeAnalysis DTA; |
5333 | DominatorTree &&DT = DTA.run(F&: *F, FAM); |
5334 | LoopAnalysis LIA; |
5335 | LoopInfo &&LI = LIA.run(F&: *F, AM&: FAM); |
5336 | AssumptionAnalysis ACT; |
5337 | AssumptionCache &&AC = ACT.run(F&: *F, FAM); |
5338 | OptimizationRemarkEmitter ORE{F}; |
5339 | |
5340 | Loop *L = LI.getLoopFor(BB: CLI->getHeader()); |
5341 | assert(L && "Expecting CanonicalLoopInfo to be recognized as a loop" ); |
5342 | |
5343 | TargetTransformInfo::UnrollingPreferences UP = |
5344 | gatherUnrollingPreferences(L, SE, TTI, |
5345 | /*BlockFrequencyInfo=*/BFI: nullptr, |
5346 | /*ProfileSummaryInfo=*/PSI: nullptr, ORE, OptLevel: static_cast<int>(OptLevel), |
5347 | /*UserThreshold=*/std::nullopt, |
5348 | /*UserCount=*/std::nullopt, |
5349 | /*UserAllowPartial=*/true, |
5350 | /*UserAllowRuntime=*/UserRuntime: true, |
5351 | /*UserUpperBound=*/std::nullopt, |
5352 | /*UserFullUnrollMaxCount=*/std::nullopt); |
5353 | |
5354 | UP.Force = true; |
5355 | |
5356 | // Account for additional optimizations taking place before the LoopUnrollPass |
5357 | // would unroll the loop. |
5358 | UP.Threshold *= UnrollThresholdFactor; |
5359 | UP.PartialThreshold *= UnrollThresholdFactor; |
5360 | |
5361 | // Use normal unroll factors even if the rest of the code is optimized for |
5362 | // size. |
5363 | UP.OptSizeThreshold = UP.Threshold; |
5364 | UP.PartialOptSizeThreshold = UP.PartialThreshold; |
5365 | |
5366 | LLVM_DEBUG(dbgs() << "Unroll heuristic thresholds:\n" |
5367 | << " Threshold=" << UP.Threshold << "\n" |
5368 | << " PartialThreshold=" << UP.PartialThreshold << "\n" |
5369 | << " OptSizeThreshold=" << UP.OptSizeThreshold << "\n" |
5370 | << " PartialOptSizeThreshold=" |
5371 | << UP.PartialOptSizeThreshold << "\n" ); |
5372 | |
5373 | // Disable peeling. |
5374 | TargetTransformInfo::PeelingPreferences PP = |
5375 | gatherPeelingPreferences(L, SE, TTI, |
5376 | /*UserAllowPeeling=*/false, |
5377 | /*UserAllowProfileBasedPeeling=*/false, |
5378 | /*UnrollingSpecficValues=*/false); |
5379 | |
5380 | SmallPtrSet<const Value *, 32> EphValues; |
5381 | CodeMetrics::collectEphemeralValues(L, AC: &AC, EphValues); |
5382 | |
5383 | // Assume that reads and writes to stack variables can be eliminated by |
5384 | // Mem2Reg, SROA or LICM. That is, don't count them towards the loop body's |
5385 | // size. |
5386 | for (BasicBlock *BB : L->blocks()) { |
5387 | for (Instruction &I : *BB) { |
5388 | Value *Ptr; |
5389 | if (auto *Load = dyn_cast<LoadInst>(Val: &I)) { |
5390 | Ptr = Load->getPointerOperand(); |
5391 | } else if (auto *Store = dyn_cast<StoreInst>(Val: &I)) { |
5392 | Ptr = Store->getPointerOperand(); |
5393 | } else |
5394 | continue; |
5395 | |
5396 | Ptr = Ptr->stripPointerCasts(); |
5397 | |
5398 | if (auto *Alloca = dyn_cast<AllocaInst>(Val: Ptr)) { |
5399 | if (Alloca->getParent() == &F->getEntryBlock()) |
5400 | EphValues.insert(Ptr: &I); |
5401 | } |
5402 | } |
5403 | } |
5404 | |
5405 | UnrollCostEstimator UCE(L, TTI, EphValues, UP.BEInsns); |
5406 | |
5407 | // Loop is not unrollable if the loop contains certain instructions. |
5408 | if (!UCE.canUnroll()) { |
5409 | LLVM_DEBUG(dbgs() << "Loop not considered unrollable\n" ); |
5410 | return 1; |
5411 | } |
5412 | |
5413 | LLVM_DEBUG(dbgs() << "Estimated loop size is " << UCE.getRolledLoopSize() |
5414 | << "\n" ); |
5415 | |
5416 | // TODO: Determine trip count of \p CLI if constant, computeUnrollCount might |
5417 | // be able to use it. |
5418 | int TripCount = 0; |
5419 | int MaxTripCount = 0; |
5420 | bool MaxOrZero = false; |
5421 | unsigned TripMultiple = 0; |
5422 | |
5423 | bool UseUpperBound = false; |
5424 | computeUnrollCount(L, TTI, DT, LI: &LI, AC: &AC, SE, EphValues, ORE: &ORE, TripCount, |
5425 | MaxTripCount, MaxOrZero, TripMultiple, UCE, UP, PP, |
5426 | UseUpperBound); |
5427 | unsigned Factor = UP.Count; |
5428 | LLVM_DEBUG(dbgs() << "Suggesting unroll factor of " << Factor << "\n" ); |
5429 | |
5430 | // This function returns 1 to signal to not unroll a loop. |
5431 | if (Factor == 0) |
5432 | return 1; |
5433 | return Factor; |
5434 | } |
5435 | |
5436 | void OpenMPIRBuilder::unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop, |
5437 | int32_t Factor, |
5438 | CanonicalLoopInfo **UnrolledCLI) { |
5439 | assert(Factor >= 0 && "Unroll factor must not be negative" ); |
5440 | |
5441 | Function *F = Loop->getFunction(); |
5442 | LLVMContext &Ctx = F->getContext(); |
5443 | |
5444 | // If the unrolled loop is not used for another loop-associated directive, it |
5445 | // is sufficient to add metadata for the LoopUnrollPass. |
5446 | if (!UnrolledCLI) { |
5447 | SmallVector<Metadata *, 2> LoopMetadata; |
5448 | LoopMetadata.push_back( |
5449 | Elt: MDNode::get(Context&: Ctx, MDs: MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.enable" ))); |
5450 | |
5451 | if (Factor >= 1) { |
5452 | ConstantAsMetadata *FactorConst = ConstantAsMetadata::get( |
5453 | C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: APInt(32, Factor))); |
5454 | LoopMetadata.push_back(Elt: MDNode::get( |
5455 | Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.count" ), FactorConst})); |
5456 | } |
5457 | |
5458 | addLoopMetadata(Loop, Properties: LoopMetadata); |
5459 | return; |
5460 | } |
5461 | |
5462 | // Heuristically determine the unroll factor. |
5463 | if (Factor == 0) |
5464 | Factor = computeHeuristicUnrollFactor(CLI: Loop); |
5465 | |
5466 | // No change required with unroll factor 1. |
5467 | if (Factor == 1) { |
5468 | *UnrolledCLI = Loop; |
5469 | return; |
5470 | } |
5471 | |
5472 | assert(Factor >= 2 && |
5473 | "unrolling only makes sense with a factor of 2 or larger" ); |
5474 | |
5475 | Type *IndVarTy = Loop->getIndVarType(); |
5476 | |
5477 | // Apply partial unrolling by tiling the loop by the unroll-factor, then fully |
5478 | // unroll the inner loop. |
5479 | Value *FactorVal = |
5480 | ConstantInt::get(Ty: IndVarTy, V: APInt(IndVarTy->getIntegerBitWidth(), Factor, |
5481 | /*isSigned=*/false)); |
5482 | std::vector<CanonicalLoopInfo *> LoopNest = |
5483 | tileLoops(DL, Loops: {Loop}, TileSizes: {FactorVal}); |
5484 | assert(LoopNest.size() == 2 && "Expect 2 loops after tiling" ); |
5485 | *UnrolledCLI = LoopNest[0]; |
5486 | CanonicalLoopInfo *InnerLoop = LoopNest[1]; |
5487 | |
5488 | // LoopUnrollPass can only fully unroll loops with constant trip count. |
5489 | // Unroll by the unroll factor with a fallback epilog for the remainder |
5490 | // iterations if necessary. |
5491 | ConstantAsMetadata *FactorConst = ConstantAsMetadata::get( |
5492 | C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: APInt(32, Factor))); |
5493 | addLoopMetadata( |
5494 | Loop: InnerLoop, |
5495 | Properties: {MDNode::get(Context&: Ctx, MDs: MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.enable" )), |
5496 | MDNode::get( |
5497 | Context&: Ctx, MDs: {MDString::get(Context&: Ctx, Str: "llvm.loop.unroll.count" ), FactorConst})}); |
5498 | |
5499 | #ifndef NDEBUG |
5500 | (*UnrolledCLI)->assertOK(); |
5501 | #endif |
5502 | } |
5503 | |
5504 | OpenMPIRBuilder::InsertPointTy |
5505 | OpenMPIRBuilder::createCopyPrivate(const LocationDescription &Loc, |
5506 | llvm::Value *BufSize, llvm::Value *CpyBuf, |
5507 | llvm::Value *CpyFn, llvm::Value *DidIt) { |
5508 | if (!updateToLocation(Loc)) |
5509 | return Loc.IP; |
5510 | |
5511 | uint32_t SrcLocStrSize; |
5512 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5513 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5514 | Value *ThreadId = getOrCreateThreadID(Ident); |
5515 | |
5516 | llvm::Value *DidItLD = Builder.CreateLoad(Ty: Builder.getInt32Ty(), Ptr: DidIt); |
5517 | |
5518 | Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD}; |
5519 | |
5520 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_copyprivate); |
5521 | Builder.CreateCall(Callee: Fn, Args); |
5522 | |
5523 | return Builder.saveIP(); |
5524 | } |
5525 | |
5526 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSingle( |
5527 | const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, |
5528 | FinalizeCallbackTy FiniCB, bool IsNowait, ArrayRef<llvm::Value *> CPVars, |
5529 | ArrayRef<llvm::Function *> CPFuncs) { |
5530 | |
5531 | if (!updateToLocation(Loc)) |
5532 | return Loc.IP; |
5533 | |
5534 | // If needed allocate and initialize `DidIt` with 0. |
5535 | // DidIt: flag variable: 1=single thread; 0=not single thread. |
5536 | llvm::Value *DidIt = nullptr; |
5537 | if (!CPVars.empty()) { |
5538 | DidIt = Builder.CreateAlloca(Ty: llvm::Type::getInt32Ty(C&: Builder.getContext())); |
5539 | Builder.CreateStore(Val: Builder.getInt32(C: 0), Ptr: DidIt); |
5540 | } |
5541 | |
5542 | Directive OMPD = Directive::OMPD_single; |
5543 | uint32_t SrcLocStrSize; |
5544 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5545 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5546 | Value *ThreadId = getOrCreateThreadID(Ident); |
5547 | Value *Args[] = {Ident, ThreadId}; |
5548 | |
5549 | Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_single); |
5550 | Instruction *EntryCall = Builder.CreateCall(Callee: EntryRTLFn, Args); |
5551 | |
5552 | Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_single); |
5553 | Instruction *ExitCall = Builder.CreateCall(Callee: ExitRTLFn, Args); |
5554 | |
5555 | auto FiniCBWrapper = [&](InsertPointTy IP) { |
5556 | FiniCB(IP); |
5557 | |
5558 | // The thread that executes the single region must set `DidIt` to 1. |
5559 | // This is used by __kmpc_copyprivate, to know if the caller is the |
5560 | // single thread or not. |
5561 | if (DidIt) |
5562 | Builder.CreateStore(Val: Builder.getInt32(C: 1), Ptr: DidIt); |
5563 | }; |
5564 | |
5565 | // generates the following: |
5566 | // if (__kmpc_single()) { |
5567 | // .... single region ... |
5568 | // __kmpc_end_single |
5569 | // } |
5570 | // __kmpc_copyprivate |
5571 | // __kmpc_barrier |
5572 | |
5573 | EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB: FiniCBWrapper, |
5574 | /*Conditional*/ true, |
5575 | /*hasFinalize*/ HasFinalize: true); |
5576 | |
5577 | if (DidIt) { |
5578 | for (size_t I = 0, E = CPVars.size(); I < E; ++I) |
5579 | // NOTE BufSize is currently unused, so just pass 0. |
5580 | createCopyPrivate(Loc: LocationDescription(Builder.saveIP(), Loc.DL), |
5581 | /*BufSize=*/ConstantInt::get(Ty: Int64, V: 0), CpyBuf: CPVars[I], |
5582 | CpyFn: CPFuncs[I], DidIt); |
5583 | // NOTE __kmpc_copyprivate already inserts a barrier |
5584 | } else if (!IsNowait) |
5585 | createBarrier(Loc: LocationDescription(Builder.saveIP(), Loc.DL), |
5586 | Kind: omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false, |
5587 | /* CheckCancelFlag */ false); |
5588 | return Builder.saveIP(); |
5589 | } |
5590 | |
5591 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCritical( |
5592 | const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, |
5593 | FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst) { |
5594 | |
5595 | if (!updateToLocation(Loc)) |
5596 | return Loc.IP; |
5597 | |
5598 | Directive OMPD = Directive::OMPD_critical; |
5599 | uint32_t SrcLocStrSize; |
5600 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5601 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5602 | Value *ThreadId = getOrCreateThreadID(Ident); |
5603 | Value *LockVar = getOMPCriticalRegionLock(CriticalName); |
5604 | Value *Args[] = {Ident, ThreadId, LockVar}; |
5605 | |
5606 | SmallVector<llvm::Value *, 4> EnterArgs(std::begin(arr&: Args), std::end(arr&: Args)); |
5607 | Function *RTFn = nullptr; |
5608 | if (HintInst) { |
5609 | // Add Hint to entry Args and create call |
5610 | EnterArgs.push_back(Elt: HintInst); |
5611 | RTFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_critical_with_hint); |
5612 | } else { |
5613 | RTFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_critical); |
5614 | } |
5615 | Instruction *EntryCall = Builder.CreateCall(Callee: RTFn, Args: EnterArgs); |
5616 | |
5617 | Function *ExitRTLFn = |
5618 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_critical); |
5619 | Instruction *ExitCall = Builder.CreateCall(Callee: ExitRTLFn, Args); |
5620 | |
5621 | return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, |
5622 | /*Conditional*/ false, /*hasFinalize*/ HasFinalize: true); |
5623 | } |
5624 | |
5625 | OpenMPIRBuilder::InsertPointTy |
5626 | OpenMPIRBuilder::createOrderedDepend(const LocationDescription &Loc, |
5627 | InsertPointTy AllocaIP, unsigned NumLoops, |
5628 | ArrayRef<llvm::Value *> StoreValues, |
5629 | const Twine &Name, bool IsDependSource) { |
5630 | assert( |
5631 | llvm::all_of(StoreValues, |
5632 | [](Value *SV) { return SV->getType()->isIntegerTy(64); }) && |
5633 | "OpenMP runtime requires depend vec with i64 type" ); |
5634 | |
5635 | if (!updateToLocation(Loc)) |
5636 | return Loc.IP; |
5637 | |
5638 | // Allocate space for vector and generate alloc instruction. |
5639 | auto *ArrI64Ty = ArrayType::get(ElementType: Int64, NumElements: NumLoops); |
5640 | Builder.restoreIP(IP: AllocaIP); |
5641 | AllocaInst *ArgsBase = Builder.CreateAlloca(Ty: ArrI64Ty, ArraySize: nullptr, Name); |
5642 | ArgsBase->setAlignment(Align(8)); |
5643 | Builder.restoreIP(IP: Loc.IP); |
5644 | |
5645 | // Store the index value with offset in depend vector. |
5646 | for (unsigned I = 0; I < NumLoops; ++I) { |
5647 | Value *DependAddrGEPIter = Builder.CreateInBoundsGEP( |
5648 | Ty: ArrI64Ty, Ptr: ArgsBase, IdxList: {Builder.getInt64(C: 0), Builder.getInt64(C: I)}); |
5649 | StoreInst *STInst = Builder.CreateStore(Val: StoreValues[I], Ptr: DependAddrGEPIter); |
5650 | STInst->setAlignment(Align(8)); |
5651 | } |
5652 | |
5653 | Value *DependBaseAddrGEP = Builder.CreateInBoundsGEP( |
5654 | Ty: ArrI64Ty, Ptr: ArgsBase, IdxList: {Builder.getInt64(C: 0), Builder.getInt64(C: 0)}); |
5655 | |
5656 | uint32_t SrcLocStrSize; |
5657 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5658 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5659 | Value *ThreadId = getOrCreateThreadID(Ident); |
5660 | Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP}; |
5661 | |
5662 | Function *RTLFn = nullptr; |
5663 | if (IsDependSource) |
5664 | RTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_doacross_post); |
5665 | else |
5666 | RTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_doacross_wait); |
5667 | Builder.CreateCall(Callee: RTLFn, Args); |
5668 | |
5669 | return Builder.saveIP(); |
5670 | } |
5671 | |
5672 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createOrderedThreadsSimd( |
5673 | const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, |
5674 | FinalizeCallbackTy FiniCB, bool IsThreads) { |
5675 | if (!updateToLocation(Loc)) |
5676 | return Loc.IP; |
5677 | |
5678 | Directive OMPD = Directive::OMPD_ordered; |
5679 | Instruction *EntryCall = nullptr; |
5680 | Instruction *ExitCall = nullptr; |
5681 | |
5682 | if (IsThreads) { |
5683 | uint32_t SrcLocStrSize; |
5684 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5685 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5686 | Value *ThreadId = getOrCreateThreadID(Ident); |
5687 | Value *Args[] = {Ident, ThreadId}; |
5688 | |
5689 | Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_ordered); |
5690 | EntryCall = Builder.CreateCall(Callee: EntryRTLFn, Args); |
5691 | |
5692 | Function *ExitRTLFn = |
5693 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_end_ordered); |
5694 | ExitCall = Builder.CreateCall(Callee: ExitRTLFn, Args); |
5695 | } |
5696 | |
5697 | return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, |
5698 | /*Conditional*/ false, /*hasFinalize*/ HasFinalize: true); |
5699 | } |
5700 | |
5701 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::EmitOMPInlinedRegion( |
5702 | Directive OMPD, Instruction *EntryCall, Instruction *ExitCall, |
5703 | BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool Conditional, |
5704 | bool HasFinalize, bool IsCancellable) { |
5705 | |
5706 | if (HasFinalize) |
5707 | FinalizationStack.push_back(Elt: {.FiniCB: FiniCB, .DK: OMPD, .IsCancellable: IsCancellable}); |
5708 | |
5709 | // Create inlined region's entry and body blocks, in preparation |
5710 | // for conditional creation |
5711 | BasicBlock *EntryBB = Builder.GetInsertBlock(); |
5712 | Instruction *SplitPos = EntryBB->getTerminator(); |
5713 | if (!isa_and_nonnull<BranchInst>(Val: SplitPos)) |
5714 | SplitPos = new UnreachableInst(Builder.getContext(), EntryBB); |
5715 | BasicBlock *ExitBB = EntryBB->splitBasicBlock(I: SplitPos, BBName: "omp_region.end" ); |
5716 | BasicBlock *FiniBB = |
5717 | EntryBB->splitBasicBlock(I: EntryBB->getTerminator(), BBName: "omp_region.finalize" ); |
5718 | |
5719 | Builder.SetInsertPoint(EntryBB->getTerminator()); |
5720 | emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional); |
5721 | |
5722 | // generate body |
5723 | BodyGenCB(/* AllocaIP */ InsertPointTy(), |
5724 | /* CodeGenIP */ Builder.saveIP()); |
5725 | |
5726 | // emit exit call and do any needed finalization. |
5727 | auto FinIP = InsertPointTy(FiniBB, FiniBB->getFirstInsertionPt()); |
5728 | assert(FiniBB->getTerminator()->getNumSuccessors() == 1 && |
5729 | FiniBB->getTerminator()->getSuccessor(0) == ExitBB && |
5730 | "Unexpected control flow graph state!!" ); |
5731 | emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize); |
5732 | assert(FiniBB->getUniquePredecessor()->getUniqueSuccessor() == FiniBB && |
5733 | "Unexpected Control Flow State!" ); |
5734 | MergeBlockIntoPredecessor(BB: FiniBB); |
5735 | |
5736 | // If we are skipping the region of a non conditional, remove the exit |
5737 | // block, and clear the builder's insertion point. |
5738 | assert(SplitPos->getParent() == ExitBB && |
5739 | "Unexpected Insertion point location!" ); |
5740 | auto merged = MergeBlockIntoPredecessor(BB: ExitBB); |
5741 | BasicBlock *ExitPredBB = SplitPos->getParent(); |
5742 | auto InsertBB = merged ? ExitPredBB : ExitBB; |
5743 | if (!isa_and_nonnull<BranchInst>(Val: SplitPos)) |
5744 | SplitPos->eraseFromParent(); |
5745 | Builder.SetInsertPoint(InsertBB); |
5746 | |
5747 | return Builder.saveIP(); |
5748 | } |
5749 | |
5750 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveEntry( |
5751 | Directive OMPD, Value *EntryCall, BasicBlock *ExitBB, bool Conditional) { |
5752 | // if nothing to do, Return current insertion point. |
5753 | if (!Conditional || !EntryCall) |
5754 | return Builder.saveIP(); |
5755 | |
5756 | BasicBlock *EntryBB = Builder.GetInsertBlock(); |
5757 | Value *CallBool = Builder.CreateIsNotNull(Arg: EntryCall); |
5758 | auto *ThenBB = BasicBlock::Create(Context&: M.getContext(), Name: "omp_region.body" ); |
5759 | auto *UI = new UnreachableInst(Builder.getContext(), ThenBB); |
5760 | |
5761 | // Emit thenBB and set the Builder's insertion point there for |
5762 | // body generation next. Place the block after the current block. |
5763 | Function *CurFn = EntryBB->getParent(); |
5764 | CurFn->insert(Position: std::next(x: EntryBB->getIterator()), BB: ThenBB); |
5765 | |
5766 | // Move Entry branch to end of ThenBB, and replace with conditional |
5767 | // branch (If-stmt) |
5768 | Instruction *EntryBBTI = EntryBB->getTerminator(); |
5769 | Builder.CreateCondBr(Cond: CallBool, True: ThenBB, False: ExitBB); |
5770 | EntryBBTI->removeFromParent(); |
5771 | Builder.SetInsertPoint(UI); |
5772 | Builder.Insert(I: EntryBBTI); |
5773 | UI->eraseFromParent(); |
5774 | Builder.SetInsertPoint(ThenBB->getTerminator()); |
5775 | |
5776 | // return an insertion point to ExitBB. |
5777 | return IRBuilder<>::InsertPoint(ExitBB, ExitBB->getFirstInsertionPt()); |
5778 | } |
5779 | |
5780 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveExit( |
5781 | omp::Directive OMPD, InsertPointTy FinIP, Instruction *ExitCall, |
5782 | bool HasFinalize) { |
5783 | |
5784 | Builder.restoreIP(IP: FinIP); |
5785 | |
5786 | // If there is finalization to do, emit it before the exit call |
5787 | if (HasFinalize) { |
5788 | assert(!FinalizationStack.empty() && |
5789 | "Unexpected finalization stack state!" ); |
5790 | |
5791 | FinalizationInfo Fi = FinalizationStack.pop_back_val(); |
5792 | assert(Fi.DK == OMPD && "Unexpected Directive for Finalization call!" ); |
5793 | |
5794 | Fi.FiniCB(FinIP); |
5795 | |
5796 | BasicBlock *FiniBB = FinIP.getBlock(); |
5797 | Instruction *FiniBBTI = FiniBB->getTerminator(); |
5798 | |
5799 | // set Builder IP for call creation |
5800 | Builder.SetInsertPoint(FiniBBTI); |
5801 | } |
5802 | |
5803 | if (!ExitCall) |
5804 | return Builder.saveIP(); |
5805 | |
5806 | // place the Exitcall as last instruction before Finalization block terminator |
5807 | ExitCall->removeFromParent(); |
5808 | Builder.Insert(I: ExitCall); |
5809 | |
5810 | return IRBuilder<>::InsertPoint(ExitCall->getParent(), |
5811 | ExitCall->getIterator()); |
5812 | } |
5813 | |
5814 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCopyinClauseBlocks( |
5815 | InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr, |
5816 | llvm::IntegerType *IntPtrTy, bool BranchtoEnd) { |
5817 | if (!IP.isSet()) |
5818 | return IP; |
5819 | |
5820 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5821 | |
5822 | // creates the following CFG structure |
5823 | // OMP_Entry : (MasterAddr != PrivateAddr)? |
5824 | // F T |
5825 | // | \ |
5826 | // | copin.not.master |
5827 | // | / |
5828 | // v / |
5829 | // copyin.not.master.end |
5830 | // | |
5831 | // v |
5832 | // OMP.Entry.Next |
5833 | |
5834 | BasicBlock *OMP_Entry = IP.getBlock(); |
5835 | Function *CurFn = OMP_Entry->getParent(); |
5836 | BasicBlock *CopyBegin = |
5837 | BasicBlock::Create(Context&: M.getContext(), Name: "copyin.not.master" , Parent: CurFn); |
5838 | BasicBlock *CopyEnd = nullptr; |
5839 | |
5840 | // If entry block is terminated, split to preserve the branch to following |
5841 | // basic block (i.e. OMP.Entry.Next), otherwise, leave everything as is. |
5842 | if (isa_and_nonnull<BranchInst>(Val: OMP_Entry->getTerminator())) { |
5843 | CopyEnd = OMP_Entry->splitBasicBlock(I: OMP_Entry->getTerminator(), |
5844 | BBName: "copyin.not.master.end" ); |
5845 | OMP_Entry->getTerminator()->eraseFromParent(); |
5846 | } else { |
5847 | CopyEnd = |
5848 | BasicBlock::Create(Context&: M.getContext(), Name: "copyin.not.master.end" , Parent: CurFn); |
5849 | } |
5850 | |
5851 | Builder.SetInsertPoint(OMP_Entry); |
5852 | Value *MasterPtr = Builder.CreatePtrToInt(V: MasterAddr, DestTy: IntPtrTy); |
5853 | Value *PrivatePtr = Builder.CreatePtrToInt(V: PrivateAddr, DestTy: IntPtrTy); |
5854 | Value *cmp = Builder.CreateICmpNE(LHS: MasterPtr, RHS: PrivatePtr); |
5855 | Builder.CreateCondBr(Cond: cmp, True: CopyBegin, False: CopyEnd); |
5856 | |
5857 | Builder.SetInsertPoint(CopyBegin); |
5858 | if (BranchtoEnd) |
5859 | Builder.SetInsertPoint(Builder.CreateBr(Dest: CopyEnd)); |
5860 | |
5861 | return Builder.saveIP(); |
5862 | } |
5863 | |
5864 | CallInst *OpenMPIRBuilder::createOMPAlloc(const LocationDescription &Loc, |
5865 | Value *Size, Value *Allocator, |
5866 | std::string Name) { |
5867 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5868 | updateToLocation(Loc); |
5869 | |
5870 | uint32_t SrcLocStrSize; |
5871 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5872 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5873 | Value *ThreadId = getOrCreateThreadID(Ident); |
5874 | Value *Args[] = {ThreadId, Size, Allocator}; |
5875 | |
5876 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_alloc); |
5877 | |
5878 | return Builder.CreateCall(Callee: Fn, Args, Name); |
5879 | } |
5880 | |
5881 | CallInst *OpenMPIRBuilder::createOMPFree(const LocationDescription &Loc, |
5882 | Value *Addr, Value *Allocator, |
5883 | std::string Name) { |
5884 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5885 | updateToLocation(Loc); |
5886 | |
5887 | uint32_t SrcLocStrSize; |
5888 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5889 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5890 | Value *ThreadId = getOrCreateThreadID(Ident); |
5891 | Value *Args[] = {ThreadId, Addr, Allocator}; |
5892 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_free); |
5893 | return Builder.CreateCall(Callee: Fn, Args, Name); |
5894 | } |
5895 | |
5896 | CallInst *OpenMPIRBuilder::createOMPInteropInit( |
5897 | const LocationDescription &Loc, Value *InteropVar, |
5898 | omp::OMPInteropType InteropType, Value *Device, Value *NumDependences, |
5899 | Value *DependenceAddress, bool HaveNowaitClause) { |
5900 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5901 | updateToLocation(Loc); |
5902 | |
5903 | uint32_t SrcLocStrSize; |
5904 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5905 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5906 | Value *ThreadId = getOrCreateThreadID(Ident); |
5907 | if (Device == nullptr) |
5908 | Device = ConstantInt::get(Ty: Int32, V: -1); |
5909 | Constant *InteropTypeVal = ConstantInt::get(Ty: Int32, V: (int)InteropType); |
5910 | if (NumDependences == nullptr) { |
5911 | NumDependences = ConstantInt::get(Ty: Int32, V: 0); |
5912 | PointerType *PointerTypeVar = PointerType::getUnqual(C&: M.getContext()); |
5913 | DependenceAddress = ConstantPointerNull::get(T: PointerTypeVar); |
5914 | } |
5915 | Value *HaveNowaitClauseVal = ConstantInt::get(Ty: Int32, V: HaveNowaitClause); |
5916 | Value *Args[] = { |
5917 | Ident, ThreadId, InteropVar, InteropTypeVal, |
5918 | Device, NumDependences, DependenceAddress, HaveNowaitClauseVal}; |
5919 | |
5920 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___tgt_interop_init); |
5921 | |
5922 | return Builder.CreateCall(Callee: Fn, Args); |
5923 | } |
5924 | |
5925 | CallInst *OpenMPIRBuilder::createOMPInteropDestroy( |
5926 | const LocationDescription &Loc, Value *InteropVar, Value *Device, |
5927 | Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause) { |
5928 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5929 | updateToLocation(Loc); |
5930 | |
5931 | uint32_t SrcLocStrSize; |
5932 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5933 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5934 | Value *ThreadId = getOrCreateThreadID(Ident); |
5935 | if (Device == nullptr) |
5936 | Device = ConstantInt::get(Ty: Int32, V: -1); |
5937 | if (NumDependences == nullptr) { |
5938 | NumDependences = ConstantInt::get(Ty: Int32, V: 0); |
5939 | PointerType *PointerTypeVar = PointerType::getUnqual(C&: M.getContext()); |
5940 | DependenceAddress = ConstantPointerNull::get(T: PointerTypeVar); |
5941 | } |
5942 | Value *HaveNowaitClauseVal = ConstantInt::get(Ty: Int32, V: HaveNowaitClause); |
5943 | Value *Args[] = { |
5944 | Ident, ThreadId, InteropVar, Device, |
5945 | NumDependences, DependenceAddress, HaveNowaitClauseVal}; |
5946 | |
5947 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___tgt_interop_destroy); |
5948 | |
5949 | return Builder.CreateCall(Callee: Fn, Args); |
5950 | } |
5951 | |
5952 | CallInst *OpenMPIRBuilder::createOMPInteropUse(const LocationDescription &Loc, |
5953 | Value *InteropVar, Value *Device, |
5954 | Value *NumDependences, |
5955 | Value *DependenceAddress, |
5956 | bool HaveNowaitClause) { |
5957 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5958 | updateToLocation(Loc); |
5959 | uint32_t SrcLocStrSize; |
5960 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5961 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5962 | Value *ThreadId = getOrCreateThreadID(Ident); |
5963 | if (Device == nullptr) |
5964 | Device = ConstantInt::get(Ty: Int32, V: -1); |
5965 | if (NumDependences == nullptr) { |
5966 | NumDependences = ConstantInt::get(Ty: Int32, V: 0); |
5967 | PointerType *PointerTypeVar = PointerType::getUnqual(C&: M.getContext()); |
5968 | DependenceAddress = ConstantPointerNull::get(T: PointerTypeVar); |
5969 | } |
5970 | Value *HaveNowaitClauseVal = ConstantInt::get(Ty: Int32, V: HaveNowaitClause); |
5971 | Value *Args[] = { |
5972 | Ident, ThreadId, InteropVar, Device, |
5973 | NumDependences, DependenceAddress, HaveNowaitClauseVal}; |
5974 | |
5975 | Function *Fn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___tgt_interop_use); |
5976 | |
5977 | return Builder.CreateCall(Callee: Fn, Args); |
5978 | } |
5979 | |
5980 | CallInst *OpenMPIRBuilder::createCachedThreadPrivate( |
5981 | const LocationDescription &Loc, llvm::Value *Pointer, |
5982 | llvm::ConstantInt *Size, const llvm::Twine &Name) { |
5983 | IRBuilder<>::InsertPointGuard IPG(Builder); |
5984 | updateToLocation(Loc); |
5985 | |
5986 | uint32_t SrcLocStrSize; |
5987 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
5988 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
5989 | Value *ThreadId = getOrCreateThreadID(Ident); |
5990 | Constant *ThreadPrivateCache = |
5991 | getOrCreateInternalVariable(Ty: Int8PtrPtr, Name: Name.str()); |
5992 | llvm::Value *Args[] = {Ident, ThreadId, Pointer, Size, ThreadPrivateCache}; |
5993 | |
5994 | Function *Fn = |
5995 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_threadprivate_cached); |
5996 | |
5997 | return Builder.CreateCall(Callee: Fn, Args); |
5998 | } |
5999 | |
6000 | OpenMPIRBuilder::InsertPointTy |
6001 | OpenMPIRBuilder::createTargetInit(const LocationDescription &Loc, bool IsSPMD, |
6002 | int32_t MinThreadsVal, int32_t MaxThreadsVal, |
6003 | int32_t MinTeamsVal, int32_t MaxTeamsVal) { |
6004 | if (!updateToLocation(Loc)) |
6005 | return Loc.IP; |
6006 | |
6007 | uint32_t SrcLocStrSize; |
6008 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
6009 | Constant *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
6010 | Constant *IsSPMDVal = ConstantInt::getSigned( |
6011 | Ty: Int8, V: IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC); |
6012 | Constant *UseGenericStateMachineVal = ConstantInt::getSigned(Ty: Int8, V: !IsSPMD); |
6013 | Constant *MayUseNestedParallelismVal = ConstantInt::getSigned(Ty: Int8, V: true); |
6014 | Constant *DebugIndentionLevelVal = ConstantInt::getSigned(Ty: Int16, V: 0); |
6015 | |
6016 | Function *Kernel = Builder.GetInsertBlock()->getParent(); |
6017 | |
6018 | // Manifest the launch configuration in the metadata matching the kernel |
6019 | // environment. |
6020 | if (MinTeamsVal > 1 || MaxTeamsVal > 0) |
6021 | writeTeamsForKernel(T, Kernel&: *Kernel, LB: MinTeamsVal, UB: MaxTeamsVal); |
6022 | |
6023 | // For max values, < 0 means unset, == 0 means set but unknown. |
6024 | if (MaxThreadsVal < 0) |
6025 | MaxThreadsVal = std::max( |
6026 | a: int32_t(getGridValue(T, Kernel).GV_Default_WG_Size), b: MinThreadsVal); |
6027 | |
6028 | if (MaxThreadsVal > 0) |
6029 | writeThreadBoundsForKernel(T, Kernel&: *Kernel, LB: MinThreadsVal, UB: MaxThreadsVal); |
6030 | |
6031 | Constant *MinThreads = ConstantInt::getSigned(Ty: Int32, V: MinThreadsVal); |
6032 | Constant *MaxThreads = ConstantInt::getSigned(Ty: Int32, V: MaxThreadsVal); |
6033 | Constant *MinTeams = ConstantInt::getSigned(Ty: Int32, V: MinTeamsVal); |
6034 | Constant *MaxTeams = ConstantInt::getSigned(Ty: Int32, V: MaxTeamsVal); |
6035 | Constant *ReductionDataSize = ConstantInt::getSigned(Ty: Int32, V: 0); |
6036 | Constant *ReductionBufferLength = ConstantInt::getSigned(Ty: Int32, V: 0); |
6037 | |
6038 | // We need to strip the debug prefix to get the correct kernel name. |
6039 | StringRef KernelName = Kernel->getName(); |
6040 | const std::string DebugPrefix = "_debug__" ; |
6041 | if (KernelName.ends_with(Suffix: DebugPrefix)) |
6042 | KernelName = KernelName.drop_back(N: DebugPrefix.length()); |
6043 | |
6044 | Function *Fn = getOrCreateRuntimeFunctionPtr( |
6045 | FnID: omp::RuntimeFunction::OMPRTL___kmpc_target_init); |
6046 | const DataLayout &DL = Fn->getDataLayout(); |
6047 | |
6048 | Twine DynamicEnvironmentName = KernelName + "_dynamic_environment" ; |
6049 | Constant *DynamicEnvironmentInitializer = |
6050 | ConstantStruct::get(T: DynamicEnvironment, V: {DebugIndentionLevelVal}); |
6051 | GlobalVariable *DynamicEnvironmentGV = new GlobalVariable( |
6052 | M, DynamicEnvironment, /*IsConstant=*/false, GlobalValue::WeakODRLinkage, |
6053 | DynamicEnvironmentInitializer, DynamicEnvironmentName, |
6054 | /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal, |
6055 | DL.getDefaultGlobalsAddressSpace()); |
6056 | DynamicEnvironmentGV->setVisibility(GlobalValue::ProtectedVisibility); |
6057 | |
6058 | Constant *DynamicEnvironment = |
6059 | DynamicEnvironmentGV->getType() == DynamicEnvironmentPtr |
6060 | ? DynamicEnvironmentGV |
6061 | : ConstantExpr::getAddrSpaceCast(C: DynamicEnvironmentGV, |
6062 | Ty: DynamicEnvironmentPtr); |
6063 | |
6064 | Constant *ConfigurationEnvironmentInitializer = ConstantStruct::get( |
6065 | T: ConfigurationEnvironment, V: { |
6066 | UseGenericStateMachineVal, |
6067 | MayUseNestedParallelismVal, |
6068 | IsSPMDVal, |
6069 | MinThreads, |
6070 | MaxThreads, |
6071 | MinTeams, |
6072 | MaxTeams, |
6073 | ReductionDataSize, |
6074 | ReductionBufferLength, |
6075 | }); |
6076 | Constant *KernelEnvironmentInitializer = ConstantStruct::get( |
6077 | T: KernelEnvironment, V: { |
6078 | ConfigurationEnvironmentInitializer, |
6079 | Ident, |
6080 | DynamicEnvironment, |
6081 | }); |
6082 | std::string KernelEnvironmentName = |
6083 | (KernelName + "_kernel_environment" ).str(); |
6084 | GlobalVariable *KernelEnvironmentGV = new GlobalVariable( |
6085 | M, KernelEnvironment, /*IsConstant=*/true, GlobalValue::WeakODRLinkage, |
6086 | KernelEnvironmentInitializer, KernelEnvironmentName, |
6087 | /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal, |
6088 | DL.getDefaultGlobalsAddressSpace()); |
6089 | KernelEnvironmentGV->setVisibility(GlobalValue::ProtectedVisibility); |
6090 | |
6091 | Constant *KernelEnvironment = |
6092 | KernelEnvironmentGV->getType() == KernelEnvironmentPtr |
6093 | ? KernelEnvironmentGV |
6094 | : ConstantExpr::getAddrSpaceCast(C: KernelEnvironmentGV, |
6095 | Ty: KernelEnvironmentPtr); |
6096 | Value *KernelLaunchEnvironment = Kernel->getArg(i: 0); |
6097 | CallInst *ThreadKind = |
6098 | Builder.CreateCall(Callee: Fn, Args: {KernelEnvironment, KernelLaunchEnvironment}); |
6099 | |
6100 | Value *ExecUserCode = Builder.CreateICmpEQ( |
6101 | LHS: ThreadKind, RHS: ConstantInt::get(Ty: ThreadKind->getType(), V: -1), |
6102 | Name: "exec_user_code" ); |
6103 | |
6104 | // ThreadKind = __kmpc_target_init(...) |
6105 | // if (ThreadKind == -1) |
6106 | // user_code |
6107 | // else |
6108 | // return; |
6109 | |
6110 | auto *UI = Builder.CreateUnreachable(); |
6111 | BasicBlock *CheckBB = UI->getParent(); |
6112 | BasicBlock *UserCodeEntryBB = CheckBB->splitBasicBlock(I: UI, BBName: "user_code.entry" ); |
6113 | |
6114 | BasicBlock *WorkerExitBB = BasicBlock::Create( |
6115 | Context&: CheckBB->getContext(), Name: "worker.exit" , Parent: CheckBB->getParent()); |
6116 | Builder.SetInsertPoint(WorkerExitBB); |
6117 | Builder.CreateRetVoid(); |
6118 | |
6119 | auto *CheckBBTI = CheckBB->getTerminator(); |
6120 | Builder.SetInsertPoint(CheckBBTI); |
6121 | Builder.CreateCondBr(Cond: ExecUserCode, True: UI->getParent(), False: WorkerExitBB); |
6122 | |
6123 | CheckBBTI->eraseFromParent(); |
6124 | UI->eraseFromParent(); |
6125 | |
6126 | // Continue in the "user_code" block, see diagram above and in |
6127 | // openmp/libomptarget/deviceRTLs/common/include/target.h . |
6128 | return InsertPointTy(UserCodeEntryBB, UserCodeEntryBB->getFirstInsertionPt()); |
6129 | } |
6130 | |
6131 | void OpenMPIRBuilder::createTargetDeinit(const LocationDescription &Loc, |
6132 | int32_t TeamsReductionDataSize, |
6133 | int32_t TeamsReductionBufferLength) { |
6134 | if (!updateToLocation(Loc)) |
6135 | return; |
6136 | |
6137 | Function *Fn = getOrCreateRuntimeFunctionPtr( |
6138 | FnID: omp::RuntimeFunction::OMPRTL___kmpc_target_deinit); |
6139 | |
6140 | Builder.CreateCall(Callee: Fn, Args: {}); |
6141 | |
6142 | if (!TeamsReductionBufferLength || !TeamsReductionDataSize) |
6143 | return; |
6144 | |
6145 | Function *Kernel = Builder.GetInsertBlock()->getParent(); |
6146 | // We need to strip the debug prefix to get the correct kernel name. |
6147 | StringRef KernelName = Kernel->getName(); |
6148 | const std::string DebugPrefix = "_debug__" ; |
6149 | if (KernelName.ends_with(Suffix: DebugPrefix)) |
6150 | KernelName = KernelName.drop_back(N: DebugPrefix.length()); |
6151 | auto *KernelEnvironmentGV = |
6152 | M.getNamedGlobal(Name: (KernelName + "_kernel_environment" ).str()); |
6153 | assert(KernelEnvironmentGV && "Expected kernel environment global\n" ); |
6154 | auto *KernelEnvironmentInitializer = KernelEnvironmentGV->getInitializer(); |
6155 | auto *NewInitializer = ConstantFoldInsertValueInstruction( |
6156 | Agg: KernelEnvironmentInitializer, |
6157 | Val: ConstantInt::get(Ty: Int32, V: TeamsReductionDataSize), Idxs: {0, 7}); |
6158 | NewInitializer = ConstantFoldInsertValueInstruction( |
6159 | Agg: NewInitializer, Val: ConstantInt::get(Ty: Int32, V: TeamsReductionBufferLength), |
6160 | Idxs: {0, 8}); |
6161 | KernelEnvironmentGV->setInitializer(NewInitializer); |
6162 | } |
6163 | |
6164 | static MDNode *getNVPTXMDNode(Function &Kernel, StringRef Name) { |
6165 | Module &M = *Kernel.getParent(); |
6166 | NamedMDNode *MD = M.getOrInsertNamedMetadata(Name: "nvvm.annotations" ); |
6167 | for (auto *Op : MD->operands()) { |
6168 | if (Op->getNumOperands() != 3) |
6169 | continue; |
6170 | auto *KernelOp = dyn_cast<ConstantAsMetadata>(Val: Op->getOperand(I: 0)); |
6171 | if (!KernelOp || KernelOp->getValue() != &Kernel) |
6172 | continue; |
6173 | auto *Prop = dyn_cast<MDString>(Val: Op->getOperand(I: 1)); |
6174 | if (!Prop || Prop->getString() != Name) |
6175 | continue; |
6176 | return Op; |
6177 | } |
6178 | return nullptr; |
6179 | } |
6180 | |
6181 | static void updateNVPTXMetadata(Function &Kernel, StringRef Name, int32_t Value, |
6182 | bool Min) { |
6183 | // Update the "maxntidx" metadata for NVIDIA, or add it. |
6184 | MDNode *ExistingOp = getNVPTXMDNode(Kernel, Name); |
6185 | if (ExistingOp) { |
6186 | auto *OldVal = cast<ConstantAsMetadata>(Val: ExistingOp->getOperand(I: 2)); |
6187 | int32_t OldLimit = cast<ConstantInt>(Val: OldVal->getValue())->getZExtValue(); |
6188 | ExistingOp->replaceOperandWith( |
6189 | I: 2, New: ConstantAsMetadata::get(C: ConstantInt::get( |
6190 | Ty: OldVal->getValue()->getType(), |
6191 | V: Min ? std::min(a: OldLimit, b: Value) : std::max(a: OldLimit, b: Value)))); |
6192 | } else { |
6193 | LLVMContext &Ctx = Kernel.getContext(); |
6194 | Metadata *MDVals[] = {ConstantAsMetadata::get(C: &Kernel), |
6195 | MDString::get(Context&: Ctx, Str: Name), |
6196 | ConstantAsMetadata::get( |
6197 | C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: Value))}; |
6198 | // Append metadata to nvvm.annotations |
6199 | Module &M = *Kernel.getParent(); |
6200 | NamedMDNode *MD = M.getOrInsertNamedMetadata(Name: "nvvm.annotations" ); |
6201 | MD->addOperand(M: MDNode::get(Context&: Ctx, MDs: MDVals)); |
6202 | } |
6203 | } |
6204 | |
6205 | std::pair<int32_t, int32_t> |
6206 | OpenMPIRBuilder::readThreadBoundsForKernel(const Triple &T, Function &Kernel) { |
6207 | int32_t ThreadLimit = |
6208 | Kernel.getFnAttributeAsParsedInteger(Kind: "omp_target_thread_limit" ); |
6209 | |
6210 | if (T.isAMDGPU()) { |
6211 | const auto &Attr = Kernel.getFnAttribute(Kind: "amdgpu-flat-work-group-size" ); |
6212 | if (!Attr.isValid() || !Attr.isStringAttribute()) |
6213 | return {0, ThreadLimit}; |
6214 | auto [LBStr, UBStr] = Attr.getValueAsString().split(Separator: ','); |
6215 | int32_t LB, UB; |
6216 | if (!llvm::to_integer(S: UBStr, Num&: UB, Base: 10)) |
6217 | return {0, ThreadLimit}; |
6218 | UB = ThreadLimit ? std::min(a: ThreadLimit, b: UB) : UB; |
6219 | if (!llvm::to_integer(S: LBStr, Num&: LB, Base: 10)) |
6220 | return {0, UB}; |
6221 | return {LB, UB}; |
6222 | } |
6223 | |
6224 | if (MDNode *ExistingOp = getNVPTXMDNode(Kernel, Name: "maxntidx" )) { |
6225 | auto *OldVal = cast<ConstantAsMetadata>(Val: ExistingOp->getOperand(I: 2)); |
6226 | int32_t UB = cast<ConstantInt>(Val: OldVal->getValue())->getZExtValue(); |
6227 | return {0, ThreadLimit ? std::min(a: ThreadLimit, b: UB) : UB}; |
6228 | } |
6229 | return {0, ThreadLimit}; |
6230 | } |
6231 | |
6232 | void OpenMPIRBuilder::writeThreadBoundsForKernel(const Triple &T, |
6233 | Function &Kernel, int32_t LB, |
6234 | int32_t UB) { |
6235 | Kernel.addFnAttr(Kind: "omp_target_thread_limit" , Val: std::to_string(val: UB)); |
6236 | |
6237 | if (T.isAMDGPU()) { |
6238 | Kernel.addFnAttr(Kind: "amdgpu-flat-work-group-size" , |
6239 | Val: llvm::utostr(X: LB) + "," + llvm::utostr(X: UB)); |
6240 | return; |
6241 | } |
6242 | |
6243 | updateNVPTXMetadata(Kernel, Name: "maxntidx" , Value: UB, Min: true); |
6244 | } |
6245 | |
6246 | std::pair<int32_t, int32_t> |
6247 | OpenMPIRBuilder::readTeamBoundsForKernel(const Triple &, Function &Kernel) { |
6248 | // TODO: Read from backend annotations if available. |
6249 | return {0, Kernel.getFnAttributeAsParsedInteger(Kind: "omp_target_num_teams" )}; |
6250 | } |
6251 | |
6252 | void OpenMPIRBuilder::writeTeamsForKernel(const Triple &T, Function &Kernel, |
6253 | int32_t LB, int32_t UB) { |
6254 | if (T.isNVPTX()) |
6255 | if (UB > 0) |
6256 | updateNVPTXMetadata(Kernel, Name: "maxclusterrank" , Value: UB, Min: true); |
6257 | if (T.isAMDGPU()) |
6258 | Kernel.addFnAttr(Kind: "amdgpu-max-num-workgroups" , Val: llvm::utostr(X: LB) + ",1,1" ); |
6259 | |
6260 | Kernel.addFnAttr(Kind: "omp_target_num_teams" , Val: std::to_string(val: LB)); |
6261 | } |
6262 | |
6263 | void OpenMPIRBuilder::setOutlinedTargetRegionFunctionAttributes( |
6264 | Function *OutlinedFn) { |
6265 | if (Config.isTargetDevice()) { |
6266 | OutlinedFn->setLinkage(GlobalValue::WeakODRLinkage); |
6267 | // TODO: Determine if DSO local can be set to true. |
6268 | OutlinedFn->setDSOLocal(false); |
6269 | OutlinedFn->setVisibility(GlobalValue::ProtectedVisibility); |
6270 | if (T.isAMDGCN()) |
6271 | OutlinedFn->setCallingConv(CallingConv::AMDGPU_KERNEL); |
6272 | } |
6273 | } |
6274 | |
6275 | Constant *OpenMPIRBuilder::createOutlinedFunctionID(Function *OutlinedFn, |
6276 | StringRef EntryFnIDName) { |
6277 | if (Config.isTargetDevice()) { |
6278 | assert(OutlinedFn && "The outlined function must exist if embedded" ); |
6279 | return OutlinedFn; |
6280 | } |
6281 | |
6282 | return new GlobalVariable( |
6283 | M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::WeakAnyLinkage, |
6284 | Constant::getNullValue(Ty: Builder.getInt8Ty()), EntryFnIDName); |
6285 | } |
6286 | |
6287 | Constant *OpenMPIRBuilder::createTargetRegionEntryAddr(Function *OutlinedFn, |
6288 | StringRef EntryFnName) { |
6289 | if (OutlinedFn) |
6290 | return OutlinedFn; |
6291 | |
6292 | assert(!M.getGlobalVariable(EntryFnName, true) && |
6293 | "Named kernel already exists?" ); |
6294 | return new GlobalVariable( |
6295 | M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::InternalLinkage, |
6296 | Constant::getNullValue(Ty: Builder.getInt8Ty()), EntryFnName); |
6297 | } |
6298 | |
6299 | void OpenMPIRBuilder::emitTargetRegionFunction( |
6300 | TargetRegionEntryInfo &EntryInfo, |
6301 | FunctionGenCallback &GenerateFunctionCallback, bool IsOffloadEntry, |
6302 | Function *&OutlinedFn, Constant *&OutlinedFnID) { |
6303 | |
6304 | SmallString<64> EntryFnName; |
6305 | OffloadInfoManager.getTargetRegionEntryFnName(Name&: EntryFnName, EntryInfo); |
6306 | |
6307 | OutlinedFn = Config.isTargetDevice() || !Config.openMPOffloadMandatory() |
6308 | ? GenerateFunctionCallback(EntryFnName) |
6309 | : nullptr; |
6310 | |
6311 | // If this target outline function is not an offload entry, we don't need to |
6312 | // register it. This may be in the case of a false if clause, or if there are |
6313 | // no OpenMP targets. |
6314 | if (!IsOffloadEntry) |
6315 | return; |
6316 | |
6317 | std::string EntryFnIDName = |
6318 | Config.isTargetDevice() |
6319 | ? std::string(EntryFnName) |
6320 | : createPlatformSpecificName(Parts: {EntryFnName, "region_id" }); |
6321 | |
6322 | OutlinedFnID = registerTargetRegionFunction(EntryInfo, OutlinedFunction: OutlinedFn, |
6323 | EntryFnName, EntryFnIDName); |
6324 | } |
6325 | |
6326 | Constant *OpenMPIRBuilder::registerTargetRegionFunction( |
6327 | TargetRegionEntryInfo &EntryInfo, Function *OutlinedFn, |
6328 | StringRef EntryFnName, StringRef EntryFnIDName) { |
6329 | if (OutlinedFn) |
6330 | setOutlinedTargetRegionFunctionAttributes(OutlinedFn); |
6331 | auto OutlinedFnID = createOutlinedFunctionID(OutlinedFn, EntryFnIDName); |
6332 | auto EntryAddr = createTargetRegionEntryAddr(OutlinedFn, EntryFnName); |
6333 | OffloadInfoManager.registerTargetRegionEntryInfo( |
6334 | EntryInfo, Addr: EntryAddr, ID: OutlinedFnID, |
6335 | Flags: OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion); |
6336 | return OutlinedFnID; |
6337 | } |
6338 | |
6339 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetData( |
6340 | const LocationDescription &Loc, InsertPointTy AllocaIP, |
6341 | InsertPointTy CodeGenIP, Value *DeviceID, Value *IfCond, |
6342 | TargetDataInfo &Info, GenMapInfoCallbackTy GenMapInfoCB, |
6343 | omp::RuntimeFunction *MapperFunc, |
6344 | function_ref<InsertPointTy(InsertPointTy CodeGenIP, BodyGenTy BodyGenType)> |
6345 | BodyGenCB, |
6346 | function_ref<void(unsigned int, Value *)> DeviceAddrCB, |
6347 | function_ref<Value *(unsigned int)> CustomMapperCB, Value *SrcLocInfo) { |
6348 | if (!updateToLocation(Loc)) |
6349 | return InsertPointTy(); |
6350 | |
6351 | // Disable TargetData CodeGen on Device pass. |
6352 | if (Config.IsTargetDevice.value_or(u: false)) { |
6353 | if (BodyGenCB) |
6354 | Builder.restoreIP(IP: BodyGenCB(Builder.saveIP(), BodyGenTy::NoPriv)); |
6355 | return Builder.saveIP(); |
6356 | } |
6357 | |
6358 | Builder.restoreIP(IP: CodeGenIP); |
6359 | bool IsStandAlone = !BodyGenCB; |
6360 | MapInfosTy *MapInfo; |
6361 | // Generate the code for the opening of the data environment. Capture all the |
6362 | // arguments of the runtime call by reference because they are used in the |
6363 | // closing of the region. |
6364 | auto BeginThenGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP) { |
6365 | MapInfo = &GenMapInfoCB(Builder.saveIP()); |
6366 | emitOffloadingArrays(AllocaIP, CodeGenIP: Builder.saveIP(), CombinedInfo&: *MapInfo, Info, |
6367 | /*IsNonContiguous=*/true, DeviceAddrCB, |
6368 | CustomMapperCB); |
6369 | |
6370 | TargetDataRTArgs RTArgs; |
6371 | emitOffloadingArraysArgument(Builder, RTArgs, Info, |
6372 | EmitDebug: !MapInfo->Names.empty()); |
6373 | |
6374 | // Emit the number of elements in the offloading arrays. |
6375 | Value *PointerNum = Builder.getInt32(C: Info.NumberOfPtrs); |
6376 | |
6377 | // Source location for the ident struct |
6378 | if (!SrcLocInfo) { |
6379 | uint32_t SrcLocStrSize; |
6380 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
6381 | SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
6382 | } |
6383 | |
6384 | Value *OffloadingArgs[] = {SrcLocInfo, DeviceID, |
6385 | PointerNum, RTArgs.BasePointersArray, |
6386 | RTArgs.PointersArray, RTArgs.SizesArray, |
6387 | RTArgs.MapTypesArray, RTArgs.MapNamesArray, |
6388 | RTArgs.MappersArray}; |
6389 | |
6390 | if (IsStandAlone) { |
6391 | assert(MapperFunc && "MapperFunc missing for standalone target data" ); |
6392 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr(FnID: *MapperFunc), |
6393 | Args: OffloadingArgs); |
6394 | } else { |
6395 | Function *BeginMapperFunc = getOrCreateRuntimeFunctionPtr( |
6396 | FnID: omp::OMPRTL___tgt_target_data_begin_mapper); |
6397 | |
6398 | Builder.CreateCall(Callee: BeginMapperFunc, Args: OffloadingArgs); |
6399 | |
6400 | for (auto DeviceMap : Info.DevicePtrInfoMap) { |
6401 | if (isa<AllocaInst>(Val: DeviceMap.second.second)) { |
6402 | auto *LI = |
6403 | Builder.CreateLoad(Ty: Builder.getPtrTy(), Ptr: DeviceMap.second.first); |
6404 | Builder.CreateStore(Val: LI, Ptr: DeviceMap.second.second); |
6405 | } |
6406 | } |
6407 | |
6408 | // If device pointer privatization is required, emit the body of the |
6409 | // region here. It will have to be duplicated: with and without |
6410 | // privatization. |
6411 | Builder.restoreIP(IP: BodyGenCB(Builder.saveIP(), BodyGenTy::Priv)); |
6412 | } |
6413 | }; |
6414 | |
6415 | // If we need device pointer privatization, we need to emit the body of the |
6416 | // region with no privatization in the 'else' branch of the conditional. |
6417 | // Otherwise, we don't have to do anything. |
6418 | auto BeginElseGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP) { |
6419 | Builder.restoreIP(IP: BodyGenCB(Builder.saveIP(), BodyGenTy::DupNoPriv)); |
6420 | }; |
6421 | |
6422 | // Generate code for the closing of the data region. |
6423 | auto EndThenGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP) { |
6424 | TargetDataRTArgs RTArgs; |
6425 | emitOffloadingArraysArgument(Builder, RTArgs, Info, EmitDebug: !MapInfo->Names.empty(), |
6426 | /*ForEndCall=*/true); |
6427 | |
6428 | // Emit the number of elements in the offloading arrays. |
6429 | Value *PointerNum = Builder.getInt32(C: Info.NumberOfPtrs); |
6430 | |
6431 | // Source location for the ident struct |
6432 | if (!SrcLocInfo) { |
6433 | uint32_t SrcLocStrSize; |
6434 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
6435 | SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
6436 | } |
6437 | |
6438 | Value *OffloadingArgs[] = {SrcLocInfo, DeviceID, |
6439 | PointerNum, RTArgs.BasePointersArray, |
6440 | RTArgs.PointersArray, RTArgs.SizesArray, |
6441 | RTArgs.MapTypesArray, RTArgs.MapNamesArray, |
6442 | RTArgs.MappersArray}; |
6443 | Function *EndMapperFunc = |
6444 | getOrCreateRuntimeFunctionPtr(FnID: omp::OMPRTL___tgt_target_data_end_mapper); |
6445 | |
6446 | Builder.CreateCall(Callee: EndMapperFunc, Args: OffloadingArgs); |
6447 | }; |
6448 | |
6449 | // We don't have to do anything to close the region if the if clause evaluates |
6450 | // to false. |
6451 | auto EndElseGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP) {}; |
6452 | |
6453 | if (BodyGenCB) { |
6454 | if (IfCond) { |
6455 | emitIfClause(Cond: IfCond, ThenGen: BeginThenGen, ElseGen: BeginElseGen, AllocaIP); |
6456 | } else { |
6457 | BeginThenGen(AllocaIP, Builder.saveIP()); |
6458 | } |
6459 | |
6460 | // If we don't require privatization of device pointers, we emit the body in |
6461 | // between the runtime calls. This avoids duplicating the body code. |
6462 | Builder.restoreIP(IP: BodyGenCB(Builder.saveIP(), BodyGenTy::NoPriv)); |
6463 | |
6464 | if (IfCond) { |
6465 | emitIfClause(Cond: IfCond, ThenGen: EndThenGen, ElseGen: EndElseGen, AllocaIP); |
6466 | } else { |
6467 | EndThenGen(AllocaIP, Builder.saveIP()); |
6468 | } |
6469 | } else { |
6470 | if (IfCond) { |
6471 | emitIfClause(Cond: IfCond, ThenGen: BeginThenGen, ElseGen: EndElseGen, AllocaIP); |
6472 | } else { |
6473 | BeginThenGen(AllocaIP, Builder.saveIP()); |
6474 | } |
6475 | } |
6476 | |
6477 | return Builder.saveIP(); |
6478 | } |
6479 | |
6480 | FunctionCallee |
6481 | OpenMPIRBuilder::createForStaticInitFunction(unsigned IVSize, bool IVSigned, |
6482 | bool IsGPUDistribute) { |
6483 | assert((IVSize == 32 || IVSize == 64) && |
6484 | "IV size is not compatible with the omp runtime" ); |
6485 | RuntimeFunction Name; |
6486 | if (IsGPUDistribute) |
6487 | Name = IVSize == 32 |
6488 | ? (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_4 |
6489 | : omp::OMPRTL___kmpc_distribute_static_init_4u) |
6490 | : (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_8 |
6491 | : omp::OMPRTL___kmpc_distribute_static_init_8u); |
6492 | else |
6493 | Name = IVSize == 32 ? (IVSigned ? omp::OMPRTL___kmpc_for_static_init_4 |
6494 | : omp::OMPRTL___kmpc_for_static_init_4u) |
6495 | : (IVSigned ? omp::OMPRTL___kmpc_for_static_init_8 |
6496 | : omp::OMPRTL___kmpc_for_static_init_8u); |
6497 | |
6498 | return getOrCreateRuntimeFunction(M, FnID: Name); |
6499 | } |
6500 | |
6501 | FunctionCallee OpenMPIRBuilder::createDispatchInitFunction(unsigned IVSize, |
6502 | bool IVSigned) { |
6503 | assert((IVSize == 32 || IVSize == 64) && |
6504 | "IV size is not compatible with the omp runtime" ); |
6505 | RuntimeFunction Name = IVSize == 32 |
6506 | ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_4 |
6507 | : omp::OMPRTL___kmpc_dispatch_init_4u) |
6508 | : (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_8 |
6509 | : omp::OMPRTL___kmpc_dispatch_init_8u); |
6510 | |
6511 | return getOrCreateRuntimeFunction(M, FnID: Name); |
6512 | } |
6513 | |
6514 | FunctionCallee OpenMPIRBuilder::createDispatchNextFunction(unsigned IVSize, |
6515 | bool IVSigned) { |
6516 | assert((IVSize == 32 || IVSize == 64) && |
6517 | "IV size is not compatible with the omp runtime" ); |
6518 | RuntimeFunction Name = IVSize == 32 |
6519 | ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_4 |
6520 | : omp::OMPRTL___kmpc_dispatch_next_4u) |
6521 | : (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_8 |
6522 | : omp::OMPRTL___kmpc_dispatch_next_8u); |
6523 | |
6524 | return getOrCreateRuntimeFunction(M, FnID: Name); |
6525 | } |
6526 | |
6527 | FunctionCallee OpenMPIRBuilder::createDispatchFiniFunction(unsigned IVSize, |
6528 | bool IVSigned) { |
6529 | assert((IVSize == 32 || IVSize == 64) && |
6530 | "IV size is not compatible with the omp runtime" ); |
6531 | RuntimeFunction Name = IVSize == 32 |
6532 | ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_4 |
6533 | : omp::OMPRTL___kmpc_dispatch_fini_4u) |
6534 | : (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_8 |
6535 | : omp::OMPRTL___kmpc_dispatch_fini_8u); |
6536 | |
6537 | return getOrCreateRuntimeFunction(M, FnID: Name); |
6538 | } |
6539 | |
6540 | FunctionCallee OpenMPIRBuilder::createDispatchDeinitFunction() { |
6541 | return getOrCreateRuntimeFunction(M, FnID: omp::OMPRTL___kmpc_dispatch_deinit); |
6542 | } |
6543 | |
6544 | static Function *createOutlinedFunction( |
6545 | OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, StringRef FuncName, |
6546 | SmallVectorImpl<Value *> &Inputs, |
6547 | OpenMPIRBuilder::TargetBodyGenCallbackTy &CBFunc, |
6548 | OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy &ArgAccessorFuncCB) { |
6549 | SmallVector<Type *> ParameterTypes; |
6550 | if (OMPBuilder.Config.isTargetDevice()) { |
6551 | // Add the "implicit" runtime argument we use to provide launch specific |
6552 | // information for target devices. |
6553 | auto *Int8PtrTy = PointerType::getUnqual(C&: Builder.getContext()); |
6554 | ParameterTypes.push_back(Elt: Int8PtrTy); |
6555 | |
6556 | // All parameters to target devices are passed as pointers |
6557 | // or i64. This assumes 64-bit address spaces/pointers. |
6558 | for (auto &Arg : Inputs) |
6559 | ParameterTypes.push_back(Elt: Arg->getType()->isPointerTy() |
6560 | ? Arg->getType() |
6561 | : Type::getInt64Ty(C&: Builder.getContext())); |
6562 | } else { |
6563 | for (auto &Arg : Inputs) |
6564 | ParameterTypes.push_back(Elt: Arg->getType()); |
6565 | } |
6566 | |
6567 | auto FuncType = FunctionType::get(Result: Builder.getVoidTy(), Params: ParameterTypes, |
6568 | /*isVarArg*/ false); |
6569 | auto Func = Function::Create(Ty: FuncType, Linkage: GlobalValue::InternalLinkage, N: FuncName, |
6570 | M: Builder.GetInsertBlock()->getModule()); |
6571 | |
6572 | // Save insert point. |
6573 | auto OldInsertPoint = Builder.saveIP(); |
6574 | |
6575 | // Generate the region into the function. |
6576 | BasicBlock *EntryBB = BasicBlock::Create(Context&: Builder.getContext(), Name: "entry" , Parent: Func); |
6577 | Builder.SetInsertPoint(EntryBB); |
6578 | |
6579 | // Insert target init call in the device compilation pass. |
6580 | if (OMPBuilder.Config.isTargetDevice()) |
6581 | Builder.restoreIP(IP: OMPBuilder.createTargetInit(Loc: Builder, /*IsSPMD*/ false)); |
6582 | |
6583 | BasicBlock *UserCodeEntryBB = Builder.GetInsertBlock(); |
6584 | |
6585 | // As we embed the user code in the middle of our target region after we |
6586 | // generate entry code, we must move what allocas we can into the entry |
6587 | // block to avoid possible breaking optimisations for device |
6588 | if (OMPBuilder.Config.isTargetDevice()) |
6589 | OMPBuilder.ConstantAllocaRaiseCandidates.emplace_back(Args&: Func); |
6590 | |
6591 | // Insert target deinit call in the device compilation pass. |
6592 | Builder.restoreIP(IP: CBFunc(Builder.saveIP(), Builder.saveIP())); |
6593 | if (OMPBuilder.Config.isTargetDevice()) |
6594 | OMPBuilder.createTargetDeinit(Loc: Builder); |
6595 | |
6596 | // Insert return instruction. |
6597 | Builder.CreateRetVoid(); |
6598 | |
6599 | // New Alloca IP at entry point of created device function. |
6600 | Builder.SetInsertPoint(EntryBB->getFirstNonPHI()); |
6601 | auto AllocaIP = Builder.saveIP(); |
6602 | |
6603 | Builder.SetInsertPoint(UserCodeEntryBB->getFirstNonPHIOrDbg()); |
6604 | |
6605 | // Skip the artificial dyn_ptr on the device. |
6606 | const auto &ArgRange = |
6607 | OMPBuilder.Config.isTargetDevice() |
6608 | ? make_range(x: Func->arg_begin() + 1, y: Func->arg_end()) |
6609 | : Func->args(); |
6610 | |
6611 | auto ReplaceValue = [](Value *Input, Value *InputCopy, Function *Func) { |
6612 | // Things like GEP's can come in the form of Constants. Constants and |
6613 | // ConstantExpr's do not have access to the knowledge of what they're |
6614 | // contained in, so we must dig a little to find an instruction so we |
6615 | // can tell if they're used inside of the function we're outlining. We |
6616 | // also replace the original constant expression with a new instruction |
6617 | // equivalent; an instruction as it allows easy modification in the |
6618 | // following loop, as we can now know the constant (instruction) is |
6619 | // owned by our target function and replaceUsesOfWith can now be invoked |
6620 | // on it (cannot do this with constants it seems). A brand new one also |
6621 | // allows us to be cautious as it is perhaps possible the old expression |
6622 | // was used inside of the function but exists and is used externally |
6623 | // (unlikely by the nature of a Constant, but still). |
6624 | // NOTE: We cannot remove dead constants that have been rewritten to |
6625 | // instructions at this stage, we run the risk of breaking later lowering |
6626 | // by doing so as we could still be in the process of lowering the module |
6627 | // from MLIR to LLVM-IR and the MLIR lowering may still require the original |
6628 | // constants we have created rewritten versions of. |
6629 | if (auto *Const = dyn_cast<Constant>(Val: Input)) |
6630 | convertUsersOfConstantsToInstructions(Consts: Const, RestrictToFunc: Func, RemoveDeadConstants: false); |
6631 | |
6632 | // Collect all the instructions |
6633 | for (User *User : make_early_inc_range(Range: Input->users())) |
6634 | if (auto *Instr = dyn_cast<Instruction>(Val: User)) |
6635 | if (Instr->getFunction() == Func) |
6636 | Instr->replaceUsesOfWith(From: Input, To: InputCopy); |
6637 | }; |
6638 | |
6639 | SmallVector<std::pair<Value *, Value *>> DeferredReplacement; |
6640 | |
6641 | // Rewrite uses of input valus to parameters. |
6642 | for (auto InArg : zip(t&: Inputs, u: ArgRange)) { |
6643 | Value *Input = std::get<0>(t&: InArg); |
6644 | Argument &Arg = std::get<1>(t&: InArg); |
6645 | Value *InputCopy = nullptr; |
6646 | |
6647 | Builder.restoreIP( |
6648 | IP: ArgAccessorFuncCB(Arg, Input, InputCopy, AllocaIP, Builder.saveIP())); |
6649 | |
6650 | // In certain cases a Global may be set up for replacement, however, this |
6651 | // Global may be used in multiple arguments to the kernel, just segmented |
6652 | // apart, for example, if we have a global array, that is sectioned into |
6653 | // multiple mappings (technically not legal in OpenMP, but there is a case |
6654 | // in Fortran for Common Blocks where this is neccesary), we will end up |
6655 | // with GEP's into this array inside the kernel, that refer to the Global |
6656 | // but are technically seperate arguments to the kernel for all intents and |
6657 | // purposes. If we have mapped a segment that requires a GEP into the 0-th |
6658 | // index, it will fold into an referal to the Global, if we then encounter |
6659 | // this folded GEP during replacement all of the references to the |
6660 | // Global in the kernel will be replaced with the argument we have generated |
6661 | // that corresponds to it, including any other GEP's that refer to the |
6662 | // Global that may be other arguments. This will invalidate all of the other |
6663 | // preceding mapped arguments that refer to the same global that may be |
6664 | // seperate segments. To prevent this, we defer global processing until all |
6665 | // other processing has been performed. |
6666 | if (llvm::isa<llvm::GlobalValue>(Val: std::get<0>(t&: InArg)) || |
6667 | llvm::isa<llvm::GlobalObject>(Val: std::get<0>(t&: InArg)) || |
6668 | llvm::isa<llvm::GlobalVariable>(Val: std::get<0>(t&: InArg))) { |
6669 | DeferredReplacement.push_back(Elt: std::make_pair(x&: Input, y&: InputCopy)); |
6670 | continue; |
6671 | } |
6672 | |
6673 | ReplaceValue(Input, InputCopy, Func); |
6674 | } |
6675 | |
6676 | // Replace all of our deferred Input values, currently just Globals. |
6677 | for (auto Deferred : DeferredReplacement) |
6678 | ReplaceValue(std::get<0>(in&: Deferred), std::get<1>(in&: Deferred), Func); |
6679 | |
6680 | // Restore insert point. |
6681 | Builder.restoreIP(IP: OldInsertPoint); |
6682 | |
6683 | return Func; |
6684 | } |
6685 | |
6686 | /// Create an entry point for a target task with the following. |
6687 | /// It'll have the following signature |
6688 | /// void @.omp_target_task_proxy_func(i32 %thread.id, ptr %task) |
6689 | /// This function is called from emitTargetTask once the |
6690 | /// code to launch the target kernel has been outlined already. |
6691 | static Function *emitTargetTaskProxyFunction(OpenMPIRBuilder &OMPBuilder, |
6692 | IRBuilderBase &Builder, |
6693 | CallInst *StaleCI) { |
6694 | Module &M = OMPBuilder.M; |
6695 | // KernelLaunchFunction is the target launch function, i.e. |
6696 | // the function that sets up kernel arguments and calls |
6697 | // __tgt_target_kernel to launch the kernel on the device. |
6698 | // |
6699 | Function *KernelLaunchFunction = StaleCI->getCalledFunction(); |
6700 | |
6701 | // StaleCI is the CallInst which is the call to the outlined |
6702 | // target kernel launch function. If there are values that the |
6703 | // outlined function uses then these are aggregated into a structure |
6704 | // which is passed as the second argument. If not, then there's |
6705 | // only one argument, the threadID. So, StaleCI can be |
6706 | // |
6707 | // %structArg = alloca { ptr, ptr }, align 8 |
6708 | // %gep_ = getelementptr { ptr, ptr }, ptr %structArg, i32 0, i32 0 |
6709 | // store ptr %20, ptr %gep_, align 8 |
6710 | // %gep_8 = getelementptr { ptr, ptr }, ptr %structArg, i32 0, i32 1 |
6711 | // store ptr %21, ptr %gep_8, align 8 |
6712 | // call void @_QQmain..omp_par.1(i32 %global.tid.val6, ptr %structArg) |
6713 | // |
6714 | // OR |
6715 | // |
6716 | // call void @_QQmain..omp_par.1(i32 %global.tid.val6) |
6717 | OpenMPIRBuilder::InsertPointTy IP(StaleCI->getParent(), |
6718 | StaleCI->getIterator()); |
6719 | LLVMContext &Ctx = StaleCI->getParent()->getContext(); |
6720 | Type *ThreadIDTy = Type::getInt32Ty(C&: Ctx); |
6721 | Type *TaskPtrTy = OMPBuilder.TaskPtr; |
6722 | Type *TaskTy = OMPBuilder.Task; |
6723 | auto ProxyFnTy = |
6724 | FunctionType::get(Result: Builder.getVoidTy(), Params: {ThreadIDTy, TaskPtrTy}, |
6725 | /* isVarArg */ false); |
6726 | auto ProxyFn = Function::Create(Ty: ProxyFnTy, Linkage: GlobalValue::InternalLinkage, |
6727 | N: ".omp_target_task_proxy_func" , |
6728 | M: Builder.GetInsertBlock()->getModule()); |
6729 | ProxyFn->getArg(i: 0)->setName("thread.id" ); |
6730 | ProxyFn->getArg(i: 1)->setName("task" ); |
6731 | |
6732 | BasicBlock *EntryBB = |
6733 | BasicBlock::Create(Context&: Builder.getContext(), Name: "entry" , Parent: ProxyFn); |
6734 | Builder.SetInsertPoint(EntryBB); |
6735 | |
6736 | bool HasShareds = StaleCI->arg_size() > 1; |
6737 | // TODO: This is a temporary assert to prove to ourselves that |
6738 | // the outlined target launch function is always going to have |
6739 | // atmost two arguments if there is any data shared between |
6740 | // host and device. |
6741 | assert((!HasShareds || (StaleCI->arg_size() == 2)) && |
6742 | "StaleCI with shareds should have exactly two arguments." ); |
6743 | if (HasShareds) { |
6744 | auto *ArgStructAlloca = dyn_cast<AllocaInst>(Val: StaleCI->getArgOperand(i: 1)); |
6745 | assert(ArgStructAlloca && |
6746 | "Unable to find the alloca instruction corresponding to arguments " |
6747 | "for extracted function" ); |
6748 | auto *ArgStructType = |
6749 | dyn_cast<StructType>(Val: ArgStructAlloca->getAllocatedType()); |
6750 | |
6751 | AllocaInst *NewArgStructAlloca = |
6752 | Builder.CreateAlloca(Ty: ArgStructType, ArraySize: nullptr, Name: "structArg" ); |
6753 | Value *TaskT = ProxyFn->getArg(i: 1); |
6754 | Value *ThreadId = ProxyFn->getArg(i: 0); |
6755 | Value *SharedsSize = |
6756 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: ArgStructType)); |
6757 | |
6758 | Value *Shareds = Builder.CreateStructGEP(Ty: TaskTy, Ptr: TaskT, Idx: 0); |
6759 | LoadInst *LoadShared = |
6760 | Builder.CreateLoad(Ty: PointerType::getUnqual(C&: Ctx), Ptr: Shareds); |
6761 | |
6762 | Builder.CreateMemCpy( |
6763 | Dst: NewArgStructAlloca, DstAlign: NewArgStructAlloca->getAlign(), Src: LoadShared, |
6764 | SrcAlign: LoadShared->getPointerAlignment(DL: M.getDataLayout()), Size: SharedsSize); |
6765 | |
6766 | Builder.CreateCall(Callee: KernelLaunchFunction, Args: {ThreadId, NewArgStructAlloca}); |
6767 | } |
6768 | Builder.CreateRetVoid(); |
6769 | return ProxyFn; |
6770 | } |
6771 | static void emitTargetOutlinedFunction( |
6772 | OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, |
6773 | TargetRegionEntryInfo &EntryInfo, Function *&OutlinedFn, |
6774 | Constant *&OutlinedFnID, SmallVectorImpl<Value *> &Inputs, |
6775 | OpenMPIRBuilder::TargetBodyGenCallbackTy &CBFunc, |
6776 | OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy &ArgAccessorFuncCB) { |
6777 | |
6778 | OpenMPIRBuilder::FunctionGenCallback &&GenerateOutlinedFunction = |
6779 | [&OMPBuilder, &Builder, &Inputs, &CBFunc, |
6780 | &ArgAccessorFuncCB](StringRef EntryFnName) { |
6781 | return createOutlinedFunction(OMPBuilder, Builder, FuncName: EntryFnName, Inputs, |
6782 | CBFunc, ArgAccessorFuncCB); |
6783 | }; |
6784 | |
6785 | OMPBuilder.emitTargetRegionFunction(EntryInfo, GenerateFunctionCallback&: GenerateOutlinedFunction, IsOffloadEntry: true, |
6786 | OutlinedFn, OutlinedFnID); |
6787 | } |
6788 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetTask( |
6789 | Function *OutlinedFn, Value *OutlinedFnID, |
6790 | EmitFallbackCallbackTy EmitTargetCallFallbackCB, TargetKernelArgs &Args, |
6791 | Value *DeviceID, Value *RTLoc, OpenMPIRBuilder::InsertPointTy AllocaIP, |
6792 | SmallVector<llvm::OpenMPIRBuilder::DependData> &Dependencies, |
6793 | bool HasNoWait) { |
6794 | |
6795 | // When we arrive at this function, the target region itself has been |
6796 | // outlined into the function OutlinedFn. |
6797 | // So at ths point, for |
6798 | // -------------------------------------------------- |
6799 | // void user_code_that_offloads(...) { |
6800 | // omp target depend(..) map(from:a) map(to:b, c) |
6801 | // a = b + c |
6802 | // } |
6803 | // |
6804 | // -------------------------------------------------- |
6805 | // |
6806 | // we have |
6807 | // |
6808 | // -------------------------------------------------- |
6809 | // |
6810 | // void user_code_that_offloads(...) { |
6811 | // %.offload_baseptrs = alloca [3 x ptr], align 8 |
6812 | // %.offload_ptrs = alloca [3 x ptr], align 8 |
6813 | // %.offload_mappers = alloca [3 x ptr], align 8 |
6814 | // ;; target region has been outlined and now we need to |
6815 | // ;; offload to it via a target task. |
6816 | // } |
6817 | // void outlined_device_function(ptr a, ptr b, ptr c) { |
6818 | // *a = *b + *c |
6819 | // } |
6820 | // |
6821 | // We have to now do the following |
6822 | // (i) Make an offloading call to outlined_device_function using the OpenMP |
6823 | // RTL. See 'kernel_launch_function' in the pseudo code below. This is |
6824 | // emitted by emitKernelLaunch |
6825 | // (ii) Create a task entry point function that calls kernel_launch_function |
6826 | // and is the entry point for the target task. See |
6827 | // '@.omp_target_task_proxy_func in the pseudocode below. |
6828 | // (iii) Create a task with the task entry point created in (ii) |
6829 | // |
6830 | // That is we create the following |
6831 | // |
6832 | // void user_code_that_offloads(...) { |
6833 | // %.offload_baseptrs = alloca [3 x ptr], align 8 |
6834 | // %.offload_ptrs = alloca [3 x ptr], align 8 |
6835 | // %.offload_mappers = alloca [3 x ptr], align 8 |
6836 | // |
6837 | // %structArg = alloca { ptr, ptr, ptr }, align 8 |
6838 | // %strucArg[0] = %.offload_baseptrs |
6839 | // %strucArg[1] = %.offload_ptrs |
6840 | // %strucArg[2] = %.offload_mappers |
6841 | // proxy_target_task = @__kmpc_omp_task_alloc(..., |
6842 | // @.omp_target_task_proxy_func) |
6843 | // memcpy(proxy_target_task->shareds, %structArg, sizeof(structArg)) |
6844 | // dependencies_array = ... |
6845 | // ;; if nowait not present |
6846 | // call @__kmpc_omp_wait_deps(..., dependencies_array) |
6847 | // call @__kmpc_omp_task_begin_if0(...) |
6848 | // call @ @.omp_target_task_proxy_func(i32 thread_id, ptr |
6849 | // %proxy_target_task) call @__kmpc_omp_task_complete_if0(...) |
6850 | // } |
6851 | // |
6852 | // define internal void @.omp_target_task_proxy_func(i32 %thread.id, |
6853 | // ptr %task) { |
6854 | // %structArg = alloca {ptr, ptr, ptr} |
6855 | // %shared_data = load (getelementptr %task, 0, 0) |
6856 | // mempcy(%structArg, %shared_data, sizeof(structArg)) |
6857 | // kernel_launch_function(%thread.id, %structArg) |
6858 | // } |
6859 | // |
6860 | // We need the proxy function because the signature of the task entry point |
6861 | // expected by kmpc_omp_task is always the same and will be different from |
6862 | // that of the kernel_launch function. |
6863 | // |
6864 | // kernel_launch_function is generated by emitKernelLaunch and has the |
6865 | // always_inline attribute. |
6866 | // void kernel_launch_function(thread_id, |
6867 | // structArg) alwaysinline { |
6868 | // %kernel_args = alloca %struct.__tgt_kernel_arguments, align 8 |
6869 | // offload_baseptrs = load(getelementptr structArg, 0, 0) |
6870 | // offload_ptrs = load(getelementptr structArg, 0, 1) |
6871 | // offload_mappers = load(getelementptr structArg, 0, 2) |
6872 | // ; setup kernel_args using offload_baseptrs, offload_ptrs and |
6873 | // ; offload_mappers |
6874 | // call i32 @__tgt_target_kernel(..., |
6875 | // outlined_device_function, |
6876 | // ptr %kernel_args) |
6877 | // } |
6878 | // void outlined_device_function(ptr a, ptr b, ptr c) { |
6879 | // *a = *b + *c |
6880 | // } |
6881 | // |
6882 | BasicBlock *TargetTaskBodyBB = |
6883 | splitBB(Builder, /*CreateBranch=*/true, Name: "target.task.body" ); |
6884 | BasicBlock *TargetTaskAllocaBB = |
6885 | splitBB(Builder, /*CreateBranch=*/true, Name: "target.task.alloca" ); |
6886 | |
6887 | InsertPointTy TargetTaskAllocaIP(TargetTaskAllocaBB, |
6888 | TargetTaskAllocaBB->begin()); |
6889 | InsertPointTy TargetTaskBodyIP(TargetTaskBodyBB, TargetTaskBodyBB->begin()); |
6890 | |
6891 | OutlineInfo OI; |
6892 | OI.EntryBB = TargetTaskAllocaBB; |
6893 | OI.OuterAllocaBB = AllocaIP.getBlock(); |
6894 | |
6895 | // Add the thread ID argument. |
6896 | SmallVector<Instruction *, 4> ToBeDeleted; |
6897 | OI.ExcludeArgsFromAggregate.push_back(Elt: createFakeIntVal( |
6898 | Builder, OuterAllocaIP: AllocaIP, ToBeDeleted, InnerAllocaIP: TargetTaskAllocaIP, Name: "global.tid" , AsPtr: false)); |
6899 | |
6900 | Builder.restoreIP(IP: TargetTaskBodyIP); |
6901 | |
6902 | // emitKernelLaunch makes the necessary runtime call to offload the kernel. |
6903 | // We then outline all that code into a separate function |
6904 | // ('kernel_launch_function' in the pseudo code above). This function is then |
6905 | // called by the target task proxy function (see |
6906 | // '@.omp_target_task_proxy_func' in the pseudo code above) |
6907 | // "@.omp_target_task_proxy_func' is generated by emitTargetTaskProxyFunction |
6908 | Builder.restoreIP(IP: emitKernelLaunch(Loc: Builder, OutlinedFn, OutlinedFnID, |
6909 | emitTargetCallFallbackCB: EmitTargetCallFallbackCB, Args, DeviceID, |
6910 | RTLoc, AllocaIP: TargetTaskAllocaIP)); |
6911 | |
6912 | OI.ExitBB = Builder.saveIP().getBlock(); |
6913 | OI.PostOutlineCB = [this, ToBeDeleted, Dependencies, |
6914 | HasNoWait](Function &OutlinedFn) mutable { |
6915 | assert(OutlinedFn.getNumUses() == 1 && |
6916 | "there must be a single user for the outlined function" ); |
6917 | |
6918 | CallInst *StaleCI = cast<CallInst>(Val: OutlinedFn.user_back()); |
6919 | bool HasShareds = StaleCI->arg_size() > 1; |
6920 | |
6921 | Function *ProxyFn = emitTargetTaskProxyFunction(OMPBuilder&: *this, Builder, StaleCI); |
6922 | |
6923 | LLVM_DEBUG(dbgs() << "Proxy task entry function created: " << *ProxyFn |
6924 | << "\n" ); |
6925 | |
6926 | Builder.SetInsertPoint(StaleCI); |
6927 | |
6928 | // Gather the arguments for emitting the runtime call. |
6929 | uint32_t SrcLocStrSize; |
6930 | Constant *SrcLocStr = |
6931 | getOrCreateSrcLocStr(Loc: LocationDescription(Builder), SrcLocStrSize); |
6932 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
6933 | |
6934 | // @__kmpc_omp_task_alloc |
6935 | Function *TaskAllocFn = |
6936 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_alloc); |
6937 | |
6938 | // Arguments - `loc_ref` (Ident) and `gtid` (ThreadID) |
6939 | // call. |
6940 | Value *ThreadID = getOrCreateThreadID(Ident); |
6941 | |
6942 | // Argument - `sizeof_kmp_task_t` (TaskSize) |
6943 | // Tasksize refers to the size in bytes of kmp_task_t data structure |
6944 | // including private vars accessed in task. |
6945 | // TODO: add kmp_task_t_with_privates (privates) |
6946 | Value *TaskSize = |
6947 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: Task)); |
6948 | |
6949 | // Argument - `sizeof_shareds` (SharedsSize) |
6950 | // SharedsSize refers to the shareds array size in the kmp_task_t data |
6951 | // structure. |
6952 | Value *SharedsSize = Builder.getInt64(C: 0); |
6953 | if (HasShareds) { |
6954 | auto *ArgStructAlloca = dyn_cast<AllocaInst>(Val: StaleCI->getArgOperand(i: 1)); |
6955 | assert(ArgStructAlloca && |
6956 | "Unable to find the alloca instruction corresponding to arguments " |
6957 | "for extracted function" ); |
6958 | auto *ArgStructType = |
6959 | dyn_cast<StructType>(Val: ArgStructAlloca->getAllocatedType()); |
6960 | assert(ArgStructType && "Unable to find struct type corresponding to " |
6961 | "arguments for extracted function" ); |
6962 | SharedsSize = |
6963 | Builder.getInt64(C: M.getDataLayout().getTypeStoreSize(Ty: ArgStructType)); |
6964 | } |
6965 | |
6966 | // Argument - `flags` |
6967 | // Task is tied iff (Flags & 1) == 1. |
6968 | // Task is untied iff (Flags & 1) == 0. |
6969 | // Task is final iff (Flags & 2) == 2. |
6970 | // Task is not final iff (Flags & 2) == 0. |
6971 | // A target task is not final and is untied. |
6972 | Value *Flags = Builder.getInt32(C: 0); |
6973 | |
6974 | // Emit the @__kmpc_omp_task_alloc runtime call |
6975 | // The runtime call returns a pointer to an area where the task captured |
6976 | // variables must be copied before the task is run (TaskData) |
6977 | CallInst *TaskData = Builder.CreateCall( |
6978 | Callee: TaskAllocFn, Args: {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags, |
6979 | /*sizeof_task=*/TaskSize, /*sizeof_shared=*/SharedsSize, |
6980 | /*task_func=*/ProxyFn}); |
6981 | |
6982 | if (HasShareds) { |
6983 | Value *Shareds = StaleCI->getArgOperand(i: 1); |
6984 | Align Alignment = TaskData->getPointerAlignment(DL: M.getDataLayout()); |
6985 | Value *TaskShareds = Builder.CreateLoad(Ty: VoidPtr, Ptr: TaskData); |
6986 | Builder.CreateMemCpy(Dst: TaskShareds, DstAlign: Alignment, Src: Shareds, SrcAlign: Alignment, |
6987 | Size: SharedsSize); |
6988 | } |
6989 | |
6990 | Value *DepArray = emitTaskDependencies(OMPBuilder&: *this, Dependencies); |
6991 | |
6992 | // --------------------------------------------------------------- |
6993 | // V5.2 13.8 target construct |
6994 | // If the nowait clause is present, execution of the target task |
6995 | // may be deferred. If the nowait clause is not present, the target task is |
6996 | // an included task. |
6997 | // --------------------------------------------------------------- |
6998 | // The above means that the lack of a nowait on the target construct |
6999 | // translates to '#pragma omp task if(0)' |
7000 | if (!HasNoWait) { |
7001 | if (DepArray) { |
7002 | Function *TaskWaitFn = |
7003 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_wait_deps); |
7004 | Builder.CreateCall( |
7005 | Callee: TaskWaitFn, |
7006 | Args: {/*loc_ref=*/Ident, /*gtid=*/ThreadID, |
7007 | /*ndeps=*/Builder.getInt32(C: Dependencies.size()), |
7008 | /*dep_list=*/DepArray, |
7009 | /*ndeps_noalias=*/ConstantInt::get(Ty: Builder.getInt32Ty(), V: 0), |
7010 | /*noalias_dep_list=*/ |
7011 | ConstantPointerNull::get(T: PointerType::getUnqual(C&: M.getContext()))}); |
7012 | } |
7013 | // Included task. |
7014 | Function *TaskBeginFn = |
7015 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_begin_if0); |
7016 | Function *TaskCompleteFn = |
7017 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_complete_if0); |
7018 | Builder.CreateCall(Callee: TaskBeginFn, Args: {Ident, ThreadID, TaskData}); |
7019 | CallInst *CI = nullptr; |
7020 | if (HasShareds) |
7021 | CI = Builder.CreateCall(Callee: ProxyFn, Args: {ThreadID, TaskData}); |
7022 | else |
7023 | CI = Builder.CreateCall(Callee: ProxyFn, Args: {ThreadID}); |
7024 | CI->setDebugLoc(StaleCI->getDebugLoc()); |
7025 | Builder.CreateCall(Callee: TaskCompleteFn, Args: {Ident, ThreadID, TaskData}); |
7026 | } else if (DepArray) { |
7027 | // HasNoWait - meaning the task may be deferred. Call |
7028 | // __kmpc_omp_task_with_deps if there are dependencies, |
7029 | // else call __kmpc_omp_task |
7030 | Function *TaskFn = |
7031 | getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task_with_deps); |
7032 | Builder.CreateCall( |
7033 | Callee: TaskFn, |
7034 | Args: {Ident, ThreadID, TaskData, Builder.getInt32(C: Dependencies.size()), |
7035 | DepArray, ConstantInt::get(Ty: Builder.getInt32Ty(), V: 0), |
7036 | ConstantPointerNull::get(T: PointerType::getUnqual(C&: M.getContext()))}); |
7037 | } else { |
7038 | // Emit the @__kmpc_omp_task runtime call to spawn the task |
7039 | Function *TaskFn = getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_omp_task); |
7040 | Builder.CreateCall(Callee: TaskFn, Args: {Ident, ThreadID, TaskData}); |
7041 | } |
7042 | |
7043 | StaleCI->eraseFromParent(); |
7044 | llvm::for_each(Range: llvm::reverse(C&: ToBeDeleted), |
7045 | F: [](Instruction *I) { I->eraseFromParent(); }); |
7046 | }; |
7047 | addOutlineInfo(OI: std::move(OI)); |
7048 | |
7049 | LLVM_DEBUG(dbgs() << "Insert block after emitKernelLaunch = \n" |
7050 | << *(Builder.GetInsertBlock()) << "\n" ); |
7051 | LLVM_DEBUG(dbgs() << "Module after emitKernelLaunch = \n" |
7052 | << *(Builder.GetInsertBlock()->getParent()->getParent()) |
7053 | << "\n" ); |
7054 | return Builder.saveIP(); |
7055 | } |
7056 | static void emitTargetCall( |
7057 | OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, |
7058 | OpenMPIRBuilder::InsertPointTy AllocaIP, Function *OutlinedFn, |
7059 | Constant *OutlinedFnID, int32_t NumTeams, int32_t NumThreads, |
7060 | SmallVectorImpl<Value *> &Args, |
7061 | OpenMPIRBuilder::GenMapInfoCallbackTy GenMapInfoCB, |
7062 | SmallVector<llvm::OpenMPIRBuilder::DependData> Dependencies = {}) { |
7063 | |
7064 | OpenMPIRBuilder::TargetDataInfo Info( |
7065 | /*RequiresDevicePointerInfo=*/false, |
7066 | /*SeparateBeginEndCalls=*/true); |
7067 | |
7068 | OpenMPIRBuilder::MapInfosTy &MapInfo = GenMapInfoCB(Builder.saveIP()); |
7069 | OMPBuilder.emitOffloadingArrays(AllocaIP, CodeGenIP: Builder.saveIP(), CombinedInfo&: MapInfo, Info, |
7070 | /*IsNonContiguous=*/true); |
7071 | |
7072 | OpenMPIRBuilder::TargetDataRTArgs RTArgs; |
7073 | OMPBuilder.emitOffloadingArraysArgument(Builder, RTArgs, Info, |
7074 | EmitDebug: !MapInfo.Names.empty()); |
7075 | |
7076 | // emitKernelLaunch |
7077 | auto &&EmitTargetCallFallbackCB = |
7078 | [&](OpenMPIRBuilder::InsertPointTy IP) -> OpenMPIRBuilder::InsertPointTy { |
7079 | Builder.restoreIP(IP); |
7080 | Builder.CreateCall(Callee: OutlinedFn, Args); |
7081 | return Builder.saveIP(); |
7082 | }; |
7083 | |
7084 | unsigned NumTargetItems = MapInfo.BasePointers.size(); |
7085 | // TODO: Use correct device ID |
7086 | Value *DeviceID = Builder.getInt64(C: OMP_DEVICEID_UNDEF); |
7087 | Value *NumTeamsVal = Builder.getInt32(C: NumTeams); |
7088 | Value *NumThreadsVal = Builder.getInt32(C: NumThreads); |
7089 | uint32_t SrcLocStrSize; |
7090 | Constant *SrcLocStr = OMPBuilder.getOrCreateDefaultSrcLocStr(SrcLocStrSize); |
7091 | Value *RTLoc = OMPBuilder.getOrCreateIdent(SrcLocStr, SrcLocStrSize, |
7092 | LocFlags: llvm::omp::IdentFlag(0), Reserve2Flags: 0); |
7093 | // TODO: Use correct NumIterations |
7094 | Value *NumIterations = Builder.getInt64(C: 0); |
7095 | // TODO: Use correct DynCGGroupMem |
7096 | Value *DynCGGroupMem = Builder.getInt32(C: 0); |
7097 | |
7098 | bool HasNoWait = false; |
7099 | bool HasDependencies = Dependencies.size() > 0; |
7100 | bool RequiresOuterTargetTask = HasNoWait || HasDependencies; |
7101 | |
7102 | OpenMPIRBuilder::TargetKernelArgs KArgs(NumTargetItems, RTArgs, NumIterations, |
7103 | NumTeamsVal, NumThreadsVal, |
7104 | DynCGGroupMem, HasNoWait); |
7105 | |
7106 | // The presence of certain clauses on the target directive require the |
7107 | // explicit generation of the target task. |
7108 | if (RequiresOuterTargetTask) { |
7109 | Builder.restoreIP(IP: OMPBuilder.emitTargetTask( |
7110 | OutlinedFn, OutlinedFnID, EmitTargetCallFallbackCB, Args&: KArgs, DeviceID, |
7111 | RTLoc, AllocaIP, Dependencies, HasNoWait)); |
7112 | } else { |
7113 | Builder.restoreIP(IP: OMPBuilder.emitKernelLaunch( |
7114 | Loc: Builder, OutlinedFn, OutlinedFnID, emitTargetCallFallbackCB: EmitTargetCallFallbackCB, Args&: KArgs, |
7115 | DeviceID, RTLoc, AllocaIP)); |
7116 | } |
7117 | } |
7118 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTarget( |
7119 | const LocationDescription &Loc, InsertPointTy AllocaIP, |
7120 | InsertPointTy CodeGenIP, TargetRegionEntryInfo &EntryInfo, int32_t NumTeams, |
7121 | int32_t NumThreads, SmallVectorImpl<Value *> &Args, |
7122 | GenMapInfoCallbackTy GenMapInfoCB, |
7123 | OpenMPIRBuilder::TargetBodyGenCallbackTy CBFunc, |
7124 | OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy ArgAccessorFuncCB, |
7125 | SmallVector<DependData> Dependencies) { |
7126 | |
7127 | if (!updateToLocation(Loc)) |
7128 | return InsertPointTy(); |
7129 | |
7130 | Builder.restoreIP(IP: CodeGenIP); |
7131 | |
7132 | Function *OutlinedFn; |
7133 | Constant *OutlinedFnID; |
7134 | // The target region is outlined into its own function. The LLVM IR for |
7135 | // the target region itself is generated using the callbacks CBFunc |
7136 | // and ArgAccessorFuncCB |
7137 | emitTargetOutlinedFunction(OMPBuilder&: *this, Builder, EntryInfo, OutlinedFn, |
7138 | OutlinedFnID, Inputs&: Args, CBFunc, ArgAccessorFuncCB); |
7139 | |
7140 | // If we are not on the target device, then we need to generate code |
7141 | // to make a remote call (offload) to the previously outlined function |
7142 | // that represents the target region. Do that now. |
7143 | if (!Config.isTargetDevice()) |
7144 | emitTargetCall(OMPBuilder&: *this, Builder, AllocaIP, OutlinedFn, OutlinedFnID, NumTeams, |
7145 | NumThreads, Args, GenMapInfoCB, Dependencies); |
7146 | return Builder.saveIP(); |
7147 | } |
7148 | |
7149 | std::string OpenMPIRBuilder::getNameWithSeparators(ArrayRef<StringRef> Parts, |
7150 | StringRef FirstSeparator, |
7151 | StringRef Separator) { |
7152 | SmallString<128> Buffer; |
7153 | llvm::raw_svector_ostream OS(Buffer); |
7154 | StringRef Sep = FirstSeparator; |
7155 | for (StringRef Part : Parts) { |
7156 | OS << Sep << Part; |
7157 | Sep = Separator; |
7158 | } |
7159 | return OS.str().str(); |
7160 | } |
7161 | |
7162 | std::string |
7163 | OpenMPIRBuilder::createPlatformSpecificName(ArrayRef<StringRef> Parts) const { |
7164 | return OpenMPIRBuilder::getNameWithSeparators(Parts, FirstSeparator: Config.firstSeparator(), |
7165 | Separator: Config.separator()); |
7166 | } |
7167 | |
7168 | GlobalVariable * |
7169 | OpenMPIRBuilder::getOrCreateInternalVariable(Type *Ty, const StringRef &Name, |
7170 | unsigned AddressSpace) { |
7171 | auto &Elem = *InternalVars.try_emplace(Key: Name, Args: nullptr).first; |
7172 | if (Elem.second) { |
7173 | assert(Elem.second->getValueType() == Ty && |
7174 | "OMP internal variable has different type than requested" ); |
7175 | } else { |
7176 | // TODO: investigate the appropriate linkage type used for the global |
7177 | // variable for possibly changing that to internal or private, or maybe |
7178 | // create different versions of the function for different OMP internal |
7179 | // variables. |
7180 | auto Linkage = this->M.getTargetTriple().rfind(s: "wasm32" ) == 0 |
7181 | ? GlobalValue::ExternalLinkage |
7182 | : GlobalValue::CommonLinkage; |
7183 | auto *GV = new GlobalVariable(M, Ty, /*IsConstant=*/false, Linkage, |
7184 | Constant::getNullValue(Ty), Elem.first(), |
7185 | /*InsertBefore=*/nullptr, |
7186 | GlobalValue::NotThreadLocal, AddressSpace); |
7187 | const DataLayout &DL = M.getDataLayout(); |
7188 | const llvm::Align TypeAlign = DL.getABITypeAlign(Ty); |
7189 | const llvm::Align PtrAlign = DL.getPointerABIAlignment(AS: AddressSpace); |
7190 | GV->setAlignment(std::max(a: TypeAlign, b: PtrAlign)); |
7191 | Elem.second = GV; |
7192 | } |
7193 | |
7194 | return Elem.second; |
7195 | } |
7196 | |
7197 | Value *OpenMPIRBuilder::getOMPCriticalRegionLock(StringRef CriticalName) { |
7198 | std::string Prefix = Twine("gomp_critical_user_" , CriticalName).str(); |
7199 | std::string Name = getNameWithSeparators(Parts: {Prefix, "var" }, FirstSeparator: "." , Separator: "." ); |
7200 | return getOrCreateInternalVariable(Ty: KmpCriticalNameTy, Name); |
7201 | } |
7202 | |
7203 | Value *OpenMPIRBuilder::getSizeInBytes(Value *BasePtr) { |
7204 | LLVMContext &Ctx = Builder.getContext(); |
7205 | Value *Null = |
7206 | Constant::getNullValue(Ty: PointerType::getUnqual(C&: BasePtr->getContext())); |
7207 | Value *SizeGep = |
7208 | Builder.CreateGEP(Ty: BasePtr->getType(), Ptr: Null, IdxList: Builder.getInt32(C: 1)); |
7209 | Value *SizePtrToInt = Builder.CreatePtrToInt(V: SizeGep, DestTy: Type::getInt64Ty(C&: Ctx)); |
7210 | return SizePtrToInt; |
7211 | } |
7212 | |
7213 | GlobalVariable * |
7214 | OpenMPIRBuilder::createOffloadMaptypes(SmallVectorImpl<uint64_t> &Mappings, |
7215 | std::string VarName) { |
7216 | llvm::Constant *MaptypesArrayInit = |
7217 | llvm::ConstantDataArray::get(Context&: M.getContext(), Elts&: Mappings); |
7218 | auto *MaptypesArrayGlobal = new llvm::GlobalVariable( |
7219 | M, MaptypesArrayInit->getType(), |
7220 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MaptypesArrayInit, |
7221 | VarName); |
7222 | MaptypesArrayGlobal->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); |
7223 | return MaptypesArrayGlobal; |
7224 | } |
7225 | |
7226 | void OpenMPIRBuilder::createMapperAllocas(const LocationDescription &Loc, |
7227 | InsertPointTy AllocaIP, |
7228 | unsigned NumOperands, |
7229 | struct MapperAllocas &MapperAllocas) { |
7230 | if (!updateToLocation(Loc)) |
7231 | return; |
7232 | |
7233 | auto *ArrI8PtrTy = ArrayType::get(ElementType: Int8Ptr, NumElements: NumOperands); |
7234 | auto *ArrI64Ty = ArrayType::get(ElementType: Int64, NumElements: NumOperands); |
7235 | Builder.restoreIP(IP: AllocaIP); |
7236 | AllocaInst *ArgsBase = Builder.CreateAlloca( |
7237 | Ty: ArrI8PtrTy, /* ArraySize = */ nullptr, Name: ".offload_baseptrs" ); |
7238 | AllocaInst *Args = Builder.CreateAlloca(Ty: ArrI8PtrTy, /* ArraySize = */ nullptr, |
7239 | Name: ".offload_ptrs" ); |
7240 | AllocaInst *ArgSizes = Builder.CreateAlloca( |
7241 | Ty: ArrI64Ty, /* ArraySize = */ nullptr, Name: ".offload_sizes" ); |
7242 | Builder.restoreIP(IP: Loc.IP); |
7243 | MapperAllocas.ArgsBase = ArgsBase; |
7244 | MapperAllocas.Args = Args; |
7245 | MapperAllocas.ArgSizes = ArgSizes; |
7246 | } |
7247 | |
7248 | void OpenMPIRBuilder::emitMapperCall(const LocationDescription &Loc, |
7249 | Function *MapperFunc, Value *SrcLocInfo, |
7250 | Value *MaptypesArg, Value *MapnamesArg, |
7251 | struct MapperAllocas &MapperAllocas, |
7252 | int64_t DeviceID, unsigned NumOperands) { |
7253 | if (!updateToLocation(Loc)) |
7254 | return; |
7255 | |
7256 | auto *ArrI8PtrTy = ArrayType::get(ElementType: Int8Ptr, NumElements: NumOperands); |
7257 | auto *ArrI64Ty = ArrayType::get(ElementType: Int64, NumElements: NumOperands); |
7258 | Value *ArgsBaseGEP = |
7259 | Builder.CreateInBoundsGEP(Ty: ArrI8PtrTy, Ptr: MapperAllocas.ArgsBase, |
7260 | IdxList: {Builder.getInt32(C: 0), Builder.getInt32(C: 0)}); |
7261 | Value *ArgsGEP = |
7262 | Builder.CreateInBoundsGEP(Ty: ArrI8PtrTy, Ptr: MapperAllocas.Args, |
7263 | IdxList: {Builder.getInt32(C: 0), Builder.getInt32(C: 0)}); |
7264 | Value *ArgSizesGEP = |
7265 | Builder.CreateInBoundsGEP(Ty: ArrI64Ty, Ptr: MapperAllocas.ArgSizes, |
7266 | IdxList: {Builder.getInt32(C: 0), Builder.getInt32(C: 0)}); |
7267 | Value *NullPtr = |
7268 | Constant::getNullValue(Ty: PointerType::getUnqual(C&: Int8Ptr->getContext())); |
7269 | Builder.CreateCall(Callee: MapperFunc, |
7270 | Args: {SrcLocInfo, Builder.getInt64(C: DeviceID), |
7271 | Builder.getInt32(C: NumOperands), ArgsBaseGEP, ArgsGEP, |
7272 | ArgSizesGEP, MaptypesArg, MapnamesArg, NullPtr}); |
7273 | } |
7274 | |
7275 | void OpenMPIRBuilder::emitOffloadingArraysArgument(IRBuilderBase &Builder, |
7276 | TargetDataRTArgs &RTArgs, |
7277 | TargetDataInfo &Info, |
7278 | bool EmitDebug, |
7279 | bool ForEndCall) { |
7280 | assert((!ForEndCall || Info.separateBeginEndCalls()) && |
7281 | "expected region end call to runtime only when end call is separate" ); |
7282 | auto UnqualPtrTy = PointerType::getUnqual(C&: M.getContext()); |
7283 | auto VoidPtrTy = UnqualPtrTy; |
7284 | auto VoidPtrPtrTy = UnqualPtrTy; |
7285 | auto Int64Ty = Type::getInt64Ty(C&: M.getContext()); |
7286 | auto Int64PtrTy = UnqualPtrTy; |
7287 | |
7288 | if (!Info.NumberOfPtrs) { |
7289 | RTArgs.BasePointersArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7290 | RTArgs.PointersArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7291 | RTArgs.SizesArray = ConstantPointerNull::get(T: Int64PtrTy); |
7292 | RTArgs.MapTypesArray = ConstantPointerNull::get(T: Int64PtrTy); |
7293 | RTArgs.MapNamesArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7294 | RTArgs.MappersArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7295 | return; |
7296 | } |
7297 | |
7298 | RTArgs.BasePointersArray = Builder.CreateConstInBoundsGEP2_32( |
7299 | Ty: ArrayType::get(ElementType: VoidPtrTy, NumElements: Info.NumberOfPtrs), |
7300 | Ptr: Info.RTArgs.BasePointersArray, |
7301 | /*Idx0=*/0, /*Idx1=*/0); |
7302 | RTArgs.PointersArray = Builder.CreateConstInBoundsGEP2_32( |
7303 | Ty: ArrayType::get(ElementType: VoidPtrTy, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.PointersArray, |
7304 | /*Idx0=*/0, |
7305 | /*Idx1=*/0); |
7306 | RTArgs.SizesArray = Builder.CreateConstInBoundsGEP2_32( |
7307 | Ty: ArrayType::get(ElementType: Int64Ty, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.SizesArray, |
7308 | /*Idx0=*/0, /*Idx1=*/0); |
7309 | RTArgs.MapTypesArray = Builder.CreateConstInBoundsGEP2_32( |
7310 | Ty: ArrayType::get(ElementType: Int64Ty, NumElements: Info.NumberOfPtrs), |
7311 | Ptr: ForEndCall && Info.RTArgs.MapTypesArrayEnd ? Info.RTArgs.MapTypesArrayEnd |
7312 | : Info.RTArgs.MapTypesArray, |
7313 | /*Idx0=*/0, |
7314 | /*Idx1=*/0); |
7315 | |
7316 | // Only emit the mapper information arrays if debug information is |
7317 | // requested. |
7318 | if (!EmitDebug) |
7319 | RTArgs.MapNamesArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7320 | else |
7321 | RTArgs.MapNamesArray = Builder.CreateConstInBoundsGEP2_32( |
7322 | Ty: ArrayType::get(ElementType: VoidPtrTy, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.MapNamesArray, |
7323 | /*Idx0=*/0, |
7324 | /*Idx1=*/0); |
7325 | // If there is no user-defined mapper, set the mapper array to nullptr to |
7326 | // avoid an unnecessary data privatization |
7327 | if (!Info.HasMapper) |
7328 | RTArgs.MappersArray = ConstantPointerNull::get(T: VoidPtrPtrTy); |
7329 | else |
7330 | RTArgs.MappersArray = |
7331 | Builder.CreatePointerCast(V: Info.RTArgs.MappersArray, DestTy: VoidPtrPtrTy); |
7332 | } |
7333 | |
7334 | void OpenMPIRBuilder::emitNonContiguousDescriptor(InsertPointTy AllocaIP, |
7335 | InsertPointTy CodeGenIP, |
7336 | MapInfosTy &CombinedInfo, |
7337 | TargetDataInfo &Info) { |
7338 | MapInfosTy::StructNonContiguousInfo &NonContigInfo = |
7339 | CombinedInfo.NonContigInfo; |
7340 | |
7341 | // Build an array of struct descriptor_dim and then assign it to |
7342 | // offload_args. |
7343 | // |
7344 | // struct descriptor_dim { |
7345 | // uint64_t offset; |
7346 | // uint64_t count; |
7347 | // uint64_t stride |
7348 | // }; |
7349 | Type *Int64Ty = Builder.getInt64Ty(); |
7350 | StructType *DimTy = StructType::create( |
7351 | Context&: M.getContext(), Elements: ArrayRef<Type *>({Int64Ty, Int64Ty, Int64Ty}), |
7352 | Name: "struct.descriptor_dim" ); |
7353 | |
7354 | enum { OffsetFD = 0, CountFD, StrideFD }; |
7355 | // We need two index variable here since the size of "Dims" is the same as |
7356 | // the size of Components, however, the size of offset, count, and stride is |
7357 | // equal to the size of base declaration that is non-contiguous. |
7358 | for (unsigned I = 0, L = 0, E = NonContigInfo.Dims.size(); I < E; ++I) { |
7359 | // Skip emitting ir if dimension size is 1 since it cannot be |
7360 | // non-contiguous. |
7361 | if (NonContigInfo.Dims[I] == 1) |
7362 | continue; |
7363 | Builder.restoreIP(IP: AllocaIP); |
7364 | ArrayType *ArrayTy = ArrayType::get(ElementType: DimTy, NumElements: NonContigInfo.Dims[I]); |
7365 | AllocaInst *DimsAddr = |
7366 | Builder.CreateAlloca(Ty: ArrayTy, /* ArraySize = */ nullptr, Name: "dims" ); |
7367 | Builder.restoreIP(IP: CodeGenIP); |
7368 | for (unsigned II = 0, EE = NonContigInfo.Dims[I]; II < EE; ++II) { |
7369 | unsigned RevIdx = EE - II - 1; |
7370 | Value *DimsLVal = Builder.CreateInBoundsGEP( |
7371 | Ty: DimsAddr->getAllocatedType(), Ptr: DimsAddr, |
7372 | IdxList: {Builder.getInt64(C: 0), Builder.getInt64(C: II)}); |
7373 | // Offset |
7374 | Value *OffsetLVal = Builder.CreateStructGEP(Ty: DimTy, Ptr: DimsLVal, Idx: OffsetFD); |
7375 | Builder.CreateAlignedStore( |
7376 | Val: NonContigInfo.Offsets[L][RevIdx], Ptr: OffsetLVal, |
7377 | Align: M.getDataLayout().getPrefTypeAlign(Ty: OffsetLVal->getType())); |
7378 | // Count |
7379 | Value *CountLVal = Builder.CreateStructGEP(Ty: DimTy, Ptr: DimsLVal, Idx: CountFD); |
7380 | Builder.CreateAlignedStore( |
7381 | Val: NonContigInfo.Counts[L][RevIdx], Ptr: CountLVal, |
7382 | Align: M.getDataLayout().getPrefTypeAlign(Ty: CountLVal->getType())); |
7383 | // Stride |
7384 | Value *StrideLVal = Builder.CreateStructGEP(Ty: DimTy, Ptr: DimsLVal, Idx: StrideFD); |
7385 | Builder.CreateAlignedStore( |
7386 | Val: NonContigInfo.Strides[L][RevIdx], Ptr: StrideLVal, |
7387 | Align: M.getDataLayout().getPrefTypeAlign(Ty: CountLVal->getType())); |
7388 | } |
7389 | // args[I] = &dims |
7390 | Builder.restoreIP(IP: CodeGenIP); |
7391 | Value *DAddr = Builder.CreatePointerBitCastOrAddrSpaceCast( |
7392 | V: DimsAddr, DestTy: Builder.getPtrTy()); |
7393 | Value *P = Builder.CreateConstInBoundsGEP2_32( |
7394 | Ty: ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: Info.NumberOfPtrs), |
7395 | Ptr: Info.RTArgs.PointersArray, Idx0: 0, Idx1: I); |
7396 | Builder.CreateAlignedStore( |
7397 | Val: DAddr, Ptr: P, Align: M.getDataLayout().getPrefTypeAlign(Ty: Builder.getPtrTy())); |
7398 | ++L; |
7399 | } |
7400 | } |
7401 | |
7402 | void OpenMPIRBuilder::emitOffloadingArrays( |
7403 | InsertPointTy AllocaIP, InsertPointTy CodeGenIP, MapInfosTy &CombinedInfo, |
7404 | TargetDataInfo &Info, bool IsNonContiguous, |
7405 | function_ref<void(unsigned int, Value *)> DeviceAddrCB, |
7406 | function_ref<Value *(unsigned int)> CustomMapperCB) { |
7407 | |
7408 | // Reset the array information. |
7409 | Info.clearArrayInfo(); |
7410 | Info.NumberOfPtrs = CombinedInfo.BasePointers.size(); |
7411 | |
7412 | if (Info.NumberOfPtrs == 0) |
7413 | return; |
7414 | |
7415 | Builder.restoreIP(IP: AllocaIP); |
7416 | // Detect if we have any capture size requiring runtime evaluation of the |
7417 | // size so that a constant array could be eventually used. |
7418 | ArrayType *PointerArrayType = |
7419 | ArrayType::get(ElementType: Builder.getPtrTy(), NumElements: Info.NumberOfPtrs); |
7420 | |
7421 | Info.RTArgs.BasePointersArray = Builder.CreateAlloca( |
7422 | Ty: PointerArrayType, /* ArraySize = */ nullptr, Name: ".offload_baseptrs" ); |
7423 | |
7424 | Info.RTArgs.PointersArray = Builder.CreateAlloca( |
7425 | Ty: PointerArrayType, /* ArraySize = */ nullptr, Name: ".offload_ptrs" ); |
7426 | AllocaInst *MappersArray = Builder.CreateAlloca( |
7427 | Ty: PointerArrayType, /* ArraySize = */ nullptr, Name: ".offload_mappers" ); |
7428 | Info.RTArgs.MappersArray = MappersArray; |
7429 | |
7430 | // If we don't have any VLA types or other types that require runtime |
7431 | // evaluation, we can use a constant array for the map sizes, otherwise we |
7432 | // need to fill up the arrays as we do for the pointers. |
7433 | Type *Int64Ty = Builder.getInt64Ty(); |
7434 | SmallVector<Constant *> ConstSizes(CombinedInfo.Sizes.size(), |
7435 | ConstantInt::get(Ty: Int64Ty, V: 0)); |
7436 | SmallBitVector RuntimeSizes(CombinedInfo.Sizes.size()); |
7437 | for (unsigned I = 0, E = CombinedInfo.Sizes.size(); I < E; ++I) { |
7438 | if (auto *CI = dyn_cast<Constant>(Val: CombinedInfo.Sizes[I])) { |
7439 | if (!isa<ConstantExpr>(Val: CI) && !isa<GlobalValue>(Val: CI)) { |
7440 | if (IsNonContiguous && |
7441 | static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>( |
7442 | CombinedInfo.Types[I] & |
7443 | OpenMPOffloadMappingFlags::OMP_MAP_NON_CONTIG)) |
7444 | ConstSizes[I] = |
7445 | ConstantInt::get(Ty: Int64Ty, V: CombinedInfo.NonContigInfo.Dims[I]); |
7446 | else |
7447 | ConstSizes[I] = CI; |
7448 | continue; |
7449 | } |
7450 | } |
7451 | RuntimeSizes.set(I); |
7452 | } |
7453 | |
7454 | if (RuntimeSizes.all()) { |
7455 | ArrayType *SizeArrayType = ArrayType::get(ElementType: Int64Ty, NumElements: Info.NumberOfPtrs); |
7456 | Info.RTArgs.SizesArray = Builder.CreateAlloca( |
7457 | Ty: SizeArrayType, /* ArraySize = */ nullptr, Name: ".offload_sizes" ); |
7458 | Builder.restoreIP(IP: CodeGenIP); |
7459 | } else { |
7460 | auto *SizesArrayInit = ConstantArray::get( |
7461 | T: ArrayType::get(ElementType: Int64Ty, NumElements: ConstSizes.size()), V: ConstSizes); |
7462 | std::string Name = createPlatformSpecificName(Parts: {"offload_sizes" }); |
7463 | auto *SizesArrayGbl = |
7464 | new GlobalVariable(M, SizesArrayInit->getType(), /*isConstant=*/true, |
7465 | GlobalValue::PrivateLinkage, SizesArrayInit, Name); |
7466 | SizesArrayGbl->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); |
7467 | |
7468 | if (!RuntimeSizes.any()) { |
7469 | Info.RTArgs.SizesArray = SizesArrayGbl; |
7470 | } else { |
7471 | unsigned IndexSize = M.getDataLayout().getIndexSizeInBits(AS: 0); |
7472 | Align OffloadSizeAlign = M.getDataLayout().getABIIntegerTypeAlignment(BitWidth: 64); |
7473 | ArrayType *SizeArrayType = ArrayType::get(ElementType: Int64Ty, NumElements: Info.NumberOfPtrs); |
7474 | AllocaInst *Buffer = Builder.CreateAlloca( |
7475 | Ty: SizeArrayType, /* ArraySize = */ nullptr, Name: ".offload_sizes" ); |
7476 | Buffer->setAlignment(OffloadSizeAlign); |
7477 | Builder.restoreIP(IP: CodeGenIP); |
7478 | Builder.CreateMemCpy( |
7479 | Dst: Buffer, DstAlign: M.getDataLayout().getPrefTypeAlign(Ty: Buffer->getType()), |
7480 | Src: SizesArrayGbl, SrcAlign: OffloadSizeAlign, |
7481 | Size: Builder.getIntN( |
7482 | N: IndexSize, |
7483 | C: Buffer->getAllocationSize(DL: M.getDataLayout())->getFixedValue())); |
7484 | |
7485 | Info.RTArgs.SizesArray = Buffer; |
7486 | } |
7487 | Builder.restoreIP(IP: CodeGenIP); |
7488 | } |
7489 | |
7490 | // The map types are always constant so we don't need to generate code to |
7491 | // fill arrays. Instead, we create an array constant. |
7492 | SmallVector<uint64_t, 4> Mapping; |
7493 | for (auto mapFlag : CombinedInfo.Types) |
7494 | Mapping.push_back( |
7495 | Elt: static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>( |
7496 | mapFlag)); |
7497 | std::string MaptypesName = createPlatformSpecificName(Parts: {"offload_maptypes" }); |
7498 | auto *MapTypesArrayGbl = createOffloadMaptypes(Mappings&: Mapping, VarName: MaptypesName); |
7499 | Info.RTArgs.MapTypesArray = MapTypesArrayGbl; |
7500 | |
7501 | // The information types are only built if provided. |
7502 | if (!CombinedInfo.Names.empty()) { |
7503 | std::string MapnamesName = createPlatformSpecificName(Parts: {"offload_mapnames" }); |
7504 | auto *MapNamesArrayGbl = |
7505 | createOffloadMapnames(Names&: CombinedInfo.Names, VarName: MapnamesName); |
7506 | Info.RTArgs.MapNamesArray = MapNamesArrayGbl; |
7507 | } else { |
7508 | Info.RTArgs.MapNamesArray = |
7509 | Constant::getNullValue(Ty: PointerType::getUnqual(C&: Builder.getContext())); |
7510 | } |
7511 | |
7512 | // If there's a present map type modifier, it must not be applied to the end |
7513 | // of a region, so generate a separate map type array in that case. |
7514 | if (Info.separateBeginEndCalls()) { |
7515 | bool EndMapTypesDiffer = false; |
7516 | for (uint64_t &Type : Mapping) { |
7517 | if (Type & static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>( |
7518 | OpenMPOffloadMappingFlags::OMP_MAP_PRESENT)) { |
7519 | Type &= ~static_cast<std::underlying_type_t<OpenMPOffloadMappingFlags>>( |
7520 | OpenMPOffloadMappingFlags::OMP_MAP_PRESENT); |
7521 | EndMapTypesDiffer = true; |
7522 | } |
7523 | } |
7524 | if (EndMapTypesDiffer) { |
7525 | MapTypesArrayGbl = createOffloadMaptypes(Mappings&: Mapping, VarName: MaptypesName); |
7526 | Info.RTArgs.MapTypesArrayEnd = MapTypesArrayGbl; |
7527 | } |
7528 | } |
7529 | |
7530 | PointerType *PtrTy = Builder.getPtrTy(); |
7531 | for (unsigned I = 0; I < Info.NumberOfPtrs; ++I) { |
7532 | Value *BPVal = CombinedInfo.BasePointers[I]; |
7533 | Value *BP = Builder.CreateConstInBoundsGEP2_32( |
7534 | Ty: ArrayType::get(ElementType: PtrTy, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.BasePointersArray, |
7535 | Idx0: 0, Idx1: I); |
7536 | Builder.CreateAlignedStore(Val: BPVal, Ptr: BP, |
7537 | Align: M.getDataLayout().getPrefTypeAlign(Ty: PtrTy)); |
7538 | |
7539 | if (Info.requiresDevicePointerInfo()) { |
7540 | if (CombinedInfo.DevicePointers[I] == DeviceInfoTy::Pointer) { |
7541 | CodeGenIP = Builder.saveIP(); |
7542 | Builder.restoreIP(IP: AllocaIP); |
7543 | Info.DevicePtrInfoMap[BPVal] = {BP, Builder.CreateAlloca(Ty: PtrTy)}; |
7544 | Builder.restoreIP(IP: CodeGenIP); |
7545 | if (DeviceAddrCB) |
7546 | DeviceAddrCB(I, Info.DevicePtrInfoMap[BPVal].second); |
7547 | } else if (CombinedInfo.DevicePointers[I] == DeviceInfoTy::Address) { |
7548 | Info.DevicePtrInfoMap[BPVal] = {BP, BP}; |
7549 | if (DeviceAddrCB) |
7550 | DeviceAddrCB(I, BP); |
7551 | } |
7552 | } |
7553 | |
7554 | Value *PVal = CombinedInfo.Pointers[I]; |
7555 | Value *P = Builder.CreateConstInBoundsGEP2_32( |
7556 | Ty: ArrayType::get(ElementType: PtrTy, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.PointersArray, Idx0: 0, |
7557 | Idx1: I); |
7558 | // TODO: Check alignment correct. |
7559 | Builder.CreateAlignedStore(Val: PVal, Ptr: P, |
7560 | Align: M.getDataLayout().getPrefTypeAlign(Ty: PtrTy)); |
7561 | |
7562 | if (RuntimeSizes.test(Idx: I)) { |
7563 | Value *S = Builder.CreateConstInBoundsGEP2_32( |
7564 | Ty: ArrayType::get(ElementType: Int64Ty, NumElements: Info.NumberOfPtrs), Ptr: Info.RTArgs.SizesArray, |
7565 | /*Idx0=*/0, |
7566 | /*Idx1=*/I); |
7567 | Builder.CreateAlignedStore(Val: Builder.CreateIntCast(V: CombinedInfo.Sizes[I], |
7568 | DestTy: Int64Ty, |
7569 | /*isSigned=*/true), |
7570 | Ptr: S, Align: M.getDataLayout().getPrefTypeAlign(Ty: PtrTy)); |
7571 | } |
7572 | // Fill up the mapper array. |
7573 | unsigned IndexSize = M.getDataLayout().getIndexSizeInBits(AS: 0); |
7574 | Value *MFunc = ConstantPointerNull::get(T: PtrTy); |
7575 | if (CustomMapperCB) |
7576 | if (Value *CustomMFunc = CustomMapperCB(I)) |
7577 | MFunc = Builder.CreatePointerCast(V: CustomMFunc, DestTy: PtrTy); |
7578 | Value *MAddr = Builder.CreateInBoundsGEP( |
7579 | Ty: MappersArray->getAllocatedType(), Ptr: MappersArray, |
7580 | IdxList: {Builder.getIntN(N: IndexSize, C: 0), Builder.getIntN(N: IndexSize, C: I)}); |
7581 | Builder.CreateAlignedStore( |
7582 | Val: MFunc, Ptr: MAddr, Align: M.getDataLayout().getPrefTypeAlign(Ty: MAddr->getType())); |
7583 | } |
7584 | |
7585 | if (!IsNonContiguous || CombinedInfo.NonContigInfo.Offsets.empty() || |
7586 | Info.NumberOfPtrs == 0) |
7587 | return; |
7588 | emitNonContiguousDescriptor(AllocaIP, CodeGenIP, CombinedInfo, Info); |
7589 | } |
7590 | |
7591 | void OpenMPIRBuilder::emitBranch(BasicBlock *Target) { |
7592 | BasicBlock *CurBB = Builder.GetInsertBlock(); |
7593 | |
7594 | if (!CurBB || CurBB->getTerminator()) { |
7595 | // If there is no insert point or the previous block is already |
7596 | // terminated, don't touch it. |
7597 | } else { |
7598 | // Otherwise, create a fall-through branch. |
7599 | Builder.CreateBr(Dest: Target); |
7600 | } |
7601 | |
7602 | Builder.ClearInsertionPoint(); |
7603 | } |
7604 | |
7605 | void OpenMPIRBuilder::emitBlock(BasicBlock *BB, Function *CurFn, |
7606 | bool IsFinished) { |
7607 | BasicBlock *CurBB = Builder.GetInsertBlock(); |
7608 | |
7609 | // Fall out of the current block (if necessary). |
7610 | emitBranch(Target: BB); |
7611 | |
7612 | if (IsFinished && BB->use_empty()) { |
7613 | BB->eraseFromParent(); |
7614 | return; |
7615 | } |
7616 | |
7617 | // Place the block after the current block, if possible, or else at |
7618 | // the end of the function. |
7619 | if (CurBB && CurBB->getParent()) |
7620 | CurFn->insert(Position: std::next(x: CurBB->getIterator()), BB); |
7621 | else |
7622 | CurFn->insert(Position: CurFn->end(), BB); |
7623 | Builder.SetInsertPoint(BB); |
7624 | } |
7625 | |
7626 | void OpenMPIRBuilder::emitIfClause(Value *Cond, BodyGenCallbackTy ThenGen, |
7627 | BodyGenCallbackTy ElseGen, |
7628 | InsertPointTy AllocaIP) { |
7629 | // If the condition constant folds and can be elided, try to avoid emitting |
7630 | // the condition and the dead arm of the if/else. |
7631 | if (auto *CI = dyn_cast<ConstantInt>(Val: Cond)) { |
7632 | auto CondConstant = CI->getSExtValue(); |
7633 | if (CondConstant) |
7634 | ThenGen(AllocaIP, Builder.saveIP()); |
7635 | else |
7636 | ElseGen(AllocaIP, Builder.saveIP()); |
7637 | return; |
7638 | } |
7639 | |
7640 | Function *CurFn = Builder.GetInsertBlock()->getParent(); |
7641 | |
7642 | // Otherwise, the condition did not fold, or we couldn't elide it. Just |
7643 | // emit the conditional branch. |
7644 | BasicBlock *ThenBlock = BasicBlock::Create(Context&: M.getContext(), Name: "omp_if.then" ); |
7645 | BasicBlock *ElseBlock = BasicBlock::Create(Context&: M.getContext(), Name: "omp_if.else" ); |
7646 | BasicBlock *ContBlock = BasicBlock::Create(Context&: M.getContext(), Name: "omp_if.end" ); |
7647 | Builder.CreateCondBr(Cond, True: ThenBlock, False: ElseBlock); |
7648 | // Emit the 'then' code. |
7649 | emitBlock(BB: ThenBlock, CurFn); |
7650 | ThenGen(AllocaIP, Builder.saveIP()); |
7651 | emitBranch(Target: ContBlock); |
7652 | // Emit the 'else' code if present. |
7653 | // There is no need to emit line number for unconditional branch. |
7654 | emitBlock(BB: ElseBlock, CurFn); |
7655 | ElseGen(AllocaIP, Builder.saveIP()); |
7656 | // There is no need to emit line number for unconditional branch. |
7657 | emitBranch(Target: ContBlock); |
7658 | // Emit the continuation block for code after the if. |
7659 | emitBlock(BB: ContBlock, CurFn, /*IsFinished=*/true); |
7660 | } |
7661 | |
7662 | bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic( |
7663 | const LocationDescription &Loc, llvm::AtomicOrdering AO, AtomicKind AK) { |
7664 | assert(!(AO == AtomicOrdering::NotAtomic || |
7665 | AO == llvm::AtomicOrdering::Unordered) && |
7666 | "Unexpected Atomic Ordering." ); |
7667 | |
7668 | bool Flush = false; |
7669 | llvm::AtomicOrdering FlushAO = AtomicOrdering::Monotonic; |
7670 | |
7671 | switch (AK) { |
7672 | case Read: |
7673 | if (AO == AtomicOrdering::Acquire || AO == AtomicOrdering::AcquireRelease || |
7674 | AO == AtomicOrdering::SequentiallyConsistent) { |
7675 | FlushAO = AtomicOrdering::Acquire; |
7676 | Flush = true; |
7677 | } |
7678 | break; |
7679 | case Write: |
7680 | case Compare: |
7681 | case Update: |
7682 | if (AO == AtomicOrdering::Release || AO == AtomicOrdering::AcquireRelease || |
7683 | AO == AtomicOrdering::SequentiallyConsistent) { |
7684 | FlushAO = AtomicOrdering::Release; |
7685 | Flush = true; |
7686 | } |
7687 | break; |
7688 | case Capture: |
7689 | switch (AO) { |
7690 | case AtomicOrdering::Acquire: |
7691 | FlushAO = AtomicOrdering::Acquire; |
7692 | Flush = true; |
7693 | break; |
7694 | case AtomicOrdering::Release: |
7695 | FlushAO = AtomicOrdering::Release; |
7696 | Flush = true; |
7697 | break; |
7698 | case AtomicOrdering::AcquireRelease: |
7699 | case AtomicOrdering::SequentiallyConsistent: |
7700 | FlushAO = AtomicOrdering::AcquireRelease; |
7701 | Flush = true; |
7702 | break; |
7703 | default: |
7704 | // do nothing - leave silently. |
7705 | break; |
7706 | } |
7707 | } |
7708 | |
7709 | if (Flush) { |
7710 | // Currently Flush RT call still doesn't take memory_ordering, so for when |
7711 | // that happens, this tries to do the resolution of which atomic ordering |
7712 | // to use with but issue the flush call |
7713 | // TODO: pass `FlushAO` after memory ordering support is added |
7714 | (void)FlushAO; |
7715 | emitFlush(Loc); |
7716 | } |
7717 | |
7718 | // for AO == AtomicOrdering::Monotonic and all other case combinations |
7719 | // do nothing |
7720 | return Flush; |
7721 | } |
7722 | |
7723 | OpenMPIRBuilder::InsertPointTy |
7724 | OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc, |
7725 | AtomicOpValue &X, AtomicOpValue &V, |
7726 | AtomicOrdering AO) { |
7727 | if (!updateToLocation(Loc)) |
7728 | return Loc.IP; |
7729 | |
7730 | assert(X.Var->getType()->isPointerTy() && |
7731 | "OMP Atomic expects a pointer to target memory" ); |
7732 | Type *XElemTy = X.ElemTy; |
7733 | assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || |
7734 | XElemTy->isPointerTy()) && |
7735 | "OMP atomic read expected a scalar type" ); |
7736 | |
7737 | Value *XRead = nullptr; |
7738 | |
7739 | if (XElemTy->isIntegerTy()) { |
7740 | LoadInst *XLD = |
7741 | Builder.CreateLoad(Ty: XElemTy, Ptr: X.Var, isVolatile: X.IsVolatile, Name: "omp.atomic.read" ); |
7742 | XLD->setAtomic(Ordering: AO); |
7743 | XRead = cast<Value>(Val: XLD); |
7744 | } else { |
7745 | // We need to perform atomic op as integer |
7746 | IntegerType *IntCastTy = |
7747 | IntegerType::get(C&: M.getContext(), NumBits: XElemTy->getScalarSizeInBits()); |
7748 | LoadInst *XLoad = |
7749 | Builder.CreateLoad(Ty: IntCastTy, Ptr: X.Var, isVolatile: X.IsVolatile, Name: "omp.atomic.load" ); |
7750 | XLoad->setAtomic(Ordering: AO); |
7751 | if (XElemTy->isFloatingPointTy()) { |
7752 | XRead = Builder.CreateBitCast(V: XLoad, DestTy: XElemTy, Name: "atomic.flt.cast" ); |
7753 | } else { |
7754 | XRead = Builder.CreateIntToPtr(V: XLoad, DestTy: XElemTy, Name: "atomic.ptr.cast" ); |
7755 | } |
7756 | } |
7757 | checkAndEmitFlushAfterAtomic(Loc, AO, AK: AtomicKind::Read); |
7758 | Builder.CreateStore(Val: XRead, Ptr: V.Var, isVolatile: V.IsVolatile); |
7759 | return Builder.saveIP(); |
7760 | } |
7761 | |
7762 | OpenMPIRBuilder::InsertPointTy |
7763 | OpenMPIRBuilder::createAtomicWrite(const LocationDescription &Loc, |
7764 | AtomicOpValue &X, Value *Expr, |
7765 | AtomicOrdering AO) { |
7766 | if (!updateToLocation(Loc)) |
7767 | return Loc.IP; |
7768 | |
7769 | assert(X.Var->getType()->isPointerTy() && |
7770 | "OMP Atomic expects a pointer to target memory" ); |
7771 | Type *XElemTy = X.ElemTy; |
7772 | assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || |
7773 | XElemTy->isPointerTy()) && |
7774 | "OMP atomic write expected a scalar type" ); |
7775 | |
7776 | if (XElemTy->isIntegerTy()) { |
7777 | StoreInst *XSt = Builder.CreateStore(Val: Expr, Ptr: X.Var, isVolatile: X.IsVolatile); |
7778 | XSt->setAtomic(Ordering: AO); |
7779 | } else { |
7780 | // We need to bitcast and perform atomic op as integers |
7781 | IntegerType *IntCastTy = |
7782 | IntegerType::get(C&: M.getContext(), NumBits: XElemTy->getScalarSizeInBits()); |
7783 | Value *ExprCast = |
7784 | Builder.CreateBitCast(V: Expr, DestTy: IntCastTy, Name: "atomic.src.int.cast" ); |
7785 | StoreInst *XSt = Builder.CreateStore(Val: ExprCast, Ptr: X.Var, isVolatile: X.IsVolatile); |
7786 | XSt->setAtomic(Ordering: AO); |
7787 | } |
7788 | |
7789 | checkAndEmitFlushAfterAtomic(Loc, AO, AK: AtomicKind::Write); |
7790 | return Builder.saveIP(); |
7791 | } |
7792 | |
7793 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicUpdate( |
7794 | const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, |
7795 | Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, |
7796 | AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr) { |
7797 | assert(!isConflictIP(Loc.IP, AllocaIP) && "IPs must not be ambiguous" ); |
7798 | if (!updateToLocation(Loc)) |
7799 | return Loc.IP; |
7800 | |
7801 | LLVM_DEBUG({ |
7802 | Type *XTy = X.Var->getType(); |
7803 | assert(XTy->isPointerTy() && |
7804 | "OMP Atomic expects a pointer to target memory" ); |
7805 | Type *XElemTy = X.ElemTy; |
7806 | assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || |
7807 | XElemTy->isPointerTy()) && |
7808 | "OMP atomic update expected a scalar type" ); |
7809 | assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) && |
7810 | (RMWOp != AtomicRMWInst::UMax) && (RMWOp != AtomicRMWInst::UMin) && |
7811 | "OpenMP atomic does not support LT or GT operations" ); |
7812 | }); |
7813 | |
7814 | emitAtomicUpdate(AllocaIP, X: X.Var, XElemTy: X.ElemTy, Expr, AO, RMWOp, UpdateOp, |
7815 | VolatileX: X.IsVolatile, IsXBinopExpr); |
7816 | checkAndEmitFlushAfterAtomic(Loc, AO, AK: AtomicKind::Update); |
7817 | return Builder.saveIP(); |
7818 | } |
7819 | |
7820 | // FIXME: Duplicating AtomicExpand |
7821 | Value *OpenMPIRBuilder::emitRMWOpAsInstruction(Value *Src1, Value *Src2, |
7822 | AtomicRMWInst::BinOp RMWOp) { |
7823 | switch (RMWOp) { |
7824 | case AtomicRMWInst::Add: |
7825 | return Builder.CreateAdd(LHS: Src1, RHS: Src2); |
7826 | case AtomicRMWInst::Sub: |
7827 | return Builder.CreateSub(LHS: Src1, RHS: Src2); |
7828 | case AtomicRMWInst::And: |
7829 | return Builder.CreateAnd(LHS: Src1, RHS: Src2); |
7830 | case AtomicRMWInst::Nand: |
7831 | return Builder.CreateNeg(V: Builder.CreateAnd(LHS: Src1, RHS: Src2)); |
7832 | case AtomicRMWInst::Or: |
7833 | return Builder.CreateOr(LHS: Src1, RHS: Src2); |
7834 | case AtomicRMWInst::Xor: |
7835 | return Builder.CreateXor(LHS: Src1, RHS: Src2); |
7836 | case AtomicRMWInst::Xchg: |
7837 | case AtomicRMWInst::FAdd: |
7838 | case AtomicRMWInst::FSub: |
7839 | case AtomicRMWInst::BAD_BINOP: |
7840 | case AtomicRMWInst::Max: |
7841 | case AtomicRMWInst::Min: |
7842 | case AtomicRMWInst::UMax: |
7843 | case AtomicRMWInst::UMin: |
7844 | case AtomicRMWInst::FMax: |
7845 | case AtomicRMWInst::FMin: |
7846 | case AtomicRMWInst::UIncWrap: |
7847 | case AtomicRMWInst::UDecWrap: |
7848 | llvm_unreachable("Unsupported atomic update operation" ); |
7849 | } |
7850 | llvm_unreachable("Unsupported atomic update operation" ); |
7851 | } |
7852 | |
7853 | std::pair<Value *, Value *> OpenMPIRBuilder::emitAtomicUpdate( |
7854 | InsertPointTy AllocaIP, Value *X, Type *XElemTy, Value *Expr, |
7855 | AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, |
7856 | AtomicUpdateCallbackTy &UpdateOp, bool VolatileX, bool IsXBinopExpr) { |
7857 | // TODO: handle the case where XElemTy is not byte-sized or not a power of 2 |
7858 | // or a complex datatype. |
7859 | bool emitRMWOp = false; |
7860 | switch (RMWOp) { |
7861 | case AtomicRMWInst::Add: |
7862 | case AtomicRMWInst::And: |
7863 | case AtomicRMWInst::Nand: |
7864 | case AtomicRMWInst::Or: |
7865 | case AtomicRMWInst::Xor: |
7866 | case AtomicRMWInst::Xchg: |
7867 | emitRMWOp = XElemTy; |
7868 | break; |
7869 | case AtomicRMWInst::Sub: |
7870 | emitRMWOp = (IsXBinopExpr && XElemTy); |
7871 | break; |
7872 | default: |
7873 | emitRMWOp = false; |
7874 | } |
7875 | emitRMWOp &= XElemTy->isIntegerTy(); |
7876 | |
7877 | std::pair<Value *, Value *> Res; |
7878 | if (emitRMWOp) { |
7879 | Res.first = Builder.CreateAtomicRMW(Op: RMWOp, Ptr: X, Val: Expr, Align: llvm::MaybeAlign(), Ordering: AO); |
7880 | // not needed except in case of postfix captures. Generate anyway for |
7881 | // consistency with the else part. Will be removed with any DCE pass. |
7882 | // AtomicRMWInst::Xchg does not have a coressponding instruction. |
7883 | if (RMWOp == AtomicRMWInst::Xchg) |
7884 | Res.second = Res.first; |
7885 | else |
7886 | Res.second = emitRMWOpAsInstruction(Src1: Res.first, Src2: Expr, RMWOp); |
7887 | } else { |
7888 | IntegerType *IntCastTy = |
7889 | IntegerType::get(C&: M.getContext(), NumBits: XElemTy->getScalarSizeInBits()); |
7890 | LoadInst *OldVal = |
7891 | Builder.CreateLoad(Ty: IntCastTy, Ptr: X, Name: X->getName() + ".atomic.load" ); |
7892 | OldVal->setAtomic(Ordering: AO); |
7893 | // CurBB |
7894 | // | /---\ |
7895 | // ContBB | |
7896 | // | \---/ |
7897 | // ExitBB |
7898 | BasicBlock *CurBB = Builder.GetInsertBlock(); |
7899 | Instruction *CurBBTI = CurBB->getTerminator(); |
7900 | CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable(); |
7901 | BasicBlock *ExitBB = |
7902 | CurBB->splitBasicBlock(I: CurBBTI, BBName: X->getName() + ".atomic.exit" ); |
7903 | BasicBlock *ContBB = CurBB->splitBasicBlock(I: CurBB->getTerminator(), |
7904 | BBName: X->getName() + ".atomic.cont" ); |
7905 | ContBB->getTerminator()->eraseFromParent(); |
7906 | Builder.restoreIP(IP: AllocaIP); |
7907 | AllocaInst *NewAtomicAddr = Builder.CreateAlloca(Ty: XElemTy); |
7908 | NewAtomicAddr->setName(X->getName() + "x.new.val" ); |
7909 | Builder.SetInsertPoint(ContBB); |
7910 | llvm::PHINode *PHI = Builder.CreatePHI(Ty: OldVal->getType(), NumReservedValues: 2); |
7911 | PHI->addIncoming(V: OldVal, BB: CurBB); |
7912 | bool IsIntTy = XElemTy->isIntegerTy(); |
7913 | Value *OldExprVal = PHI; |
7914 | if (!IsIntTy) { |
7915 | if (XElemTy->isFloatingPointTy()) { |
7916 | OldExprVal = Builder.CreateBitCast(V: PHI, DestTy: XElemTy, |
7917 | Name: X->getName() + ".atomic.fltCast" ); |
7918 | } else { |
7919 | OldExprVal = Builder.CreateIntToPtr(V: PHI, DestTy: XElemTy, |
7920 | Name: X->getName() + ".atomic.ptrCast" ); |
7921 | } |
7922 | } |
7923 | |
7924 | Value *Upd = UpdateOp(OldExprVal, Builder); |
7925 | Builder.CreateStore(Val: Upd, Ptr: NewAtomicAddr); |
7926 | LoadInst *DesiredVal = Builder.CreateLoad(Ty: IntCastTy, Ptr: NewAtomicAddr); |
7927 | AtomicOrdering Failure = |
7928 | llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: AO); |
7929 | AtomicCmpXchgInst *Result = Builder.CreateAtomicCmpXchg( |
7930 | Ptr: X, Cmp: PHI, New: DesiredVal, Align: llvm::MaybeAlign(), SuccessOrdering: AO, FailureOrdering: Failure); |
7931 | Result->setVolatile(VolatileX); |
7932 | Value *PreviousVal = Builder.CreateExtractValue(Agg: Result, /*Idxs=*/0); |
7933 | Value *SuccessFailureVal = Builder.CreateExtractValue(Agg: Result, /*Idxs=*/1); |
7934 | PHI->addIncoming(V: PreviousVal, BB: Builder.GetInsertBlock()); |
7935 | Builder.CreateCondBr(Cond: SuccessFailureVal, True: ExitBB, False: ContBB); |
7936 | |
7937 | Res.first = OldExprVal; |
7938 | Res.second = Upd; |
7939 | |
7940 | // set Insertion point in exit block |
7941 | if (UnreachableInst *ExitTI = |
7942 | dyn_cast<UnreachableInst>(Val: ExitBB->getTerminator())) { |
7943 | CurBBTI->eraseFromParent(); |
7944 | Builder.SetInsertPoint(ExitBB); |
7945 | } else { |
7946 | Builder.SetInsertPoint(ExitTI); |
7947 | } |
7948 | } |
7949 | |
7950 | return Res; |
7951 | } |
7952 | |
7953 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCapture( |
7954 | const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, |
7955 | AtomicOpValue &V, Value *Expr, AtomicOrdering AO, |
7956 | AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp, |
7957 | bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr) { |
7958 | if (!updateToLocation(Loc)) |
7959 | return Loc.IP; |
7960 | |
7961 | LLVM_DEBUG({ |
7962 | Type *XTy = X.Var->getType(); |
7963 | assert(XTy->isPointerTy() && |
7964 | "OMP Atomic expects a pointer to target memory" ); |
7965 | Type *XElemTy = X.ElemTy; |
7966 | assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || |
7967 | XElemTy->isPointerTy()) && |
7968 | "OMP atomic capture expected a scalar type" ); |
7969 | assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) && |
7970 | "OpenMP atomic does not support LT or GT operations" ); |
7971 | }); |
7972 | |
7973 | // If UpdateExpr is 'x' updated with some `expr` not based on 'x', |
7974 | // 'x' is simply atomically rewritten with 'expr'. |
7975 | AtomicRMWInst::BinOp AtomicOp = (UpdateExpr ? RMWOp : AtomicRMWInst::Xchg); |
7976 | std::pair<Value *, Value *> Result = |
7977 | emitAtomicUpdate(AllocaIP, X: X.Var, XElemTy: X.ElemTy, Expr, AO, RMWOp: AtomicOp, UpdateOp, |
7978 | VolatileX: X.IsVolatile, IsXBinopExpr); |
7979 | |
7980 | Value *CapturedVal = (IsPostfixUpdate ? Result.first : Result.second); |
7981 | Builder.CreateStore(Val: CapturedVal, Ptr: V.Var, isVolatile: V.IsVolatile); |
7982 | |
7983 | checkAndEmitFlushAfterAtomic(Loc, AO, AK: AtomicKind::Capture); |
7984 | return Builder.saveIP(); |
7985 | } |
7986 | |
7987 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCompare( |
7988 | const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V, |
7989 | AtomicOpValue &R, Value *E, Value *D, AtomicOrdering AO, |
7990 | omp::OMPAtomicCompareOp Op, bool IsXBinopExpr, bool IsPostfixUpdate, |
7991 | bool IsFailOnly) { |
7992 | |
7993 | AtomicOrdering Failure = AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: AO); |
7994 | return createAtomicCompare(Loc, X, V, R, E, D, AO, Op, IsXBinopExpr, |
7995 | IsPostfixUpdate, IsFailOnly, Failure); |
7996 | } |
7997 | |
7998 | OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCompare( |
7999 | const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V, |
8000 | AtomicOpValue &R, Value *E, Value *D, AtomicOrdering AO, |
8001 | omp::OMPAtomicCompareOp Op, bool IsXBinopExpr, bool IsPostfixUpdate, |
8002 | bool IsFailOnly, AtomicOrdering Failure) { |
8003 | |
8004 | if (!updateToLocation(Loc)) |
8005 | return Loc.IP; |
8006 | |
8007 | assert(X.Var->getType()->isPointerTy() && |
8008 | "OMP atomic expects a pointer to target memory" ); |
8009 | // compare capture |
8010 | if (V.Var) { |
8011 | assert(V.Var->getType()->isPointerTy() && "v.var must be of pointer type" ); |
8012 | assert(V.ElemTy == X.ElemTy && "x and v must be of same type" ); |
8013 | } |
8014 | |
8015 | bool IsInteger = E->getType()->isIntegerTy(); |
8016 | |
8017 | if (Op == OMPAtomicCompareOp::EQ) { |
8018 | AtomicCmpXchgInst *Result = nullptr; |
8019 | if (!IsInteger) { |
8020 | IntegerType *IntCastTy = |
8021 | IntegerType::get(C&: M.getContext(), NumBits: X.ElemTy->getScalarSizeInBits()); |
8022 | Value *EBCast = Builder.CreateBitCast(V: E, DestTy: IntCastTy); |
8023 | Value *DBCast = Builder.CreateBitCast(V: D, DestTy: IntCastTy); |
8024 | Result = Builder.CreateAtomicCmpXchg(Ptr: X.Var, Cmp: EBCast, New: DBCast, Align: MaybeAlign(), |
8025 | SuccessOrdering: AO, FailureOrdering: Failure); |
8026 | } else { |
8027 | Result = |
8028 | Builder.CreateAtomicCmpXchg(Ptr: X.Var, Cmp: E, New: D, Align: MaybeAlign(), SuccessOrdering: AO, FailureOrdering: Failure); |
8029 | } |
8030 | |
8031 | if (V.Var) { |
8032 | Value *OldValue = Builder.CreateExtractValue(Agg: Result, /*Idxs=*/0); |
8033 | if (!IsInteger) |
8034 | OldValue = Builder.CreateBitCast(V: OldValue, DestTy: X.ElemTy); |
8035 | assert(OldValue->getType() == V.ElemTy && |
8036 | "OldValue and V must be of same type" ); |
8037 | if (IsPostfixUpdate) { |
8038 | Builder.CreateStore(Val: OldValue, Ptr: V.Var, isVolatile: V.IsVolatile); |
8039 | } else { |
8040 | Value *SuccessOrFail = Builder.CreateExtractValue(Agg: Result, /*Idxs=*/1); |
8041 | if (IsFailOnly) { |
8042 | // CurBB---- |
8043 | // | | |
8044 | // v | |
8045 | // ContBB | |
8046 | // | | |
8047 | // v | |
8048 | // ExitBB <- |
8049 | // |
8050 | // where ContBB only contains the store of old value to 'v'. |
8051 | BasicBlock *CurBB = Builder.GetInsertBlock(); |
8052 | Instruction *CurBBTI = CurBB->getTerminator(); |
8053 | CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable(); |
8054 | BasicBlock *ExitBB = CurBB->splitBasicBlock( |
8055 | I: CurBBTI, BBName: X.Var->getName() + ".atomic.exit" ); |
8056 | BasicBlock *ContBB = CurBB->splitBasicBlock( |
8057 | I: CurBB->getTerminator(), BBName: X.Var->getName() + ".atomic.cont" ); |
8058 | ContBB->getTerminator()->eraseFromParent(); |
8059 | CurBB->getTerminator()->eraseFromParent(); |
8060 | |
8061 | Builder.CreateCondBr(Cond: SuccessOrFail, True: ExitBB, False: ContBB); |
8062 | |
8063 | Builder.SetInsertPoint(ContBB); |
8064 | Builder.CreateStore(Val: OldValue, Ptr: V.Var); |
8065 | Builder.CreateBr(Dest: ExitBB); |
8066 | |
8067 | if (UnreachableInst *ExitTI = |
8068 | dyn_cast<UnreachableInst>(Val: ExitBB->getTerminator())) { |
8069 | CurBBTI->eraseFromParent(); |
8070 | Builder.SetInsertPoint(ExitBB); |
8071 | } else { |
8072 | Builder.SetInsertPoint(ExitTI); |
8073 | } |
8074 | } else { |
8075 | Value *CapturedValue = |
8076 | Builder.CreateSelect(C: SuccessOrFail, True: E, False: OldValue); |
8077 | Builder.CreateStore(Val: CapturedValue, Ptr: V.Var, isVolatile: V.IsVolatile); |
8078 | } |
8079 | } |
8080 | } |
8081 | // The comparison result has to be stored. |
8082 | if (R.Var) { |
8083 | assert(R.Var->getType()->isPointerTy() && |
8084 | "r.var must be of pointer type" ); |
8085 | assert(R.ElemTy->isIntegerTy() && "r must be of integral type" ); |
8086 | |
8087 | Value *SuccessFailureVal = Builder.CreateExtractValue(Agg: Result, /*Idxs=*/1); |
8088 | Value *ResultCast = R.IsSigned |
8089 | ? Builder.CreateSExt(V: SuccessFailureVal, DestTy: R.ElemTy) |
8090 | : Builder.CreateZExt(V: SuccessFailureVal, DestTy: R.ElemTy); |
8091 | Builder.CreateStore(Val: ResultCast, Ptr: R.Var, isVolatile: R.IsVolatile); |
8092 | } |
8093 | } else { |
8094 | assert((Op == OMPAtomicCompareOp::MAX || Op == OMPAtomicCompareOp::MIN) && |
8095 | "Op should be either max or min at this point" ); |
8096 | assert(!IsFailOnly && "IsFailOnly is only valid when the comparison is ==" ); |
8097 | |
8098 | // Reverse the ordop as the OpenMP forms are different from LLVM forms. |
8099 | // Let's take max as example. |
8100 | // OpenMP form: |
8101 | // x = x > expr ? expr : x; |
8102 | // LLVM form: |
8103 | // *ptr = *ptr > val ? *ptr : val; |
8104 | // We need to transform to LLVM form. |
8105 | // x = x <= expr ? x : expr; |
8106 | AtomicRMWInst::BinOp NewOp; |
8107 | if (IsXBinopExpr) { |
8108 | if (IsInteger) { |
8109 | if (X.IsSigned) |
8110 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Min |
8111 | : AtomicRMWInst::Max; |
8112 | else |
8113 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMin |
8114 | : AtomicRMWInst::UMax; |
8115 | } else { |
8116 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMin |
8117 | : AtomicRMWInst::FMax; |
8118 | } |
8119 | } else { |
8120 | if (IsInteger) { |
8121 | if (X.IsSigned) |
8122 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Max |
8123 | : AtomicRMWInst::Min; |
8124 | else |
8125 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMax |
8126 | : AtomicRMWInst::UMin; |
8127 | } else { |
8128 | NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMax |
8129 | : AtomicRMWInst::FMin; |
8130 | } |
8131 | } |
8132 | |
8133 | AtomicRMWInst *OldValue = |
8134 | Builder.CreateAtomicRMW(Op: NewOp, Ptr: X.Var, Val: E, Align: MaybeAlign(), Ordering: AO); |
8135 | if (V.Var) { |
8136 | Value *CapturedValue = nullptr; |
8137 | if (IsPostfixUpdate) { |
8138 | CapturedValue = OldValue; |
8139 | } else { |
8140 | CmpInst::Predicate Pred; |
8141 | switch (NewOp) { |
8142 | case AtomicRMWInst::Max: |
8143 | Pred = CmpInst::ICMP_SGT; |
8144 | break; |
8145 | case AtomicRMWInst::UMax: |
8146 | Pred = CmpInst::ICMP_UGT; |
8147 | break; |
8148 | case AtomicRMWInst::FMax: |
8149 | Pred = CmpInst::FCMP_OGT; |
8150 | break; |
8151 | case AtomicRMWInst::Min: |
8152 | Pred = CmpInst::ICMP_SLT; |
8153 | break; |
8154 | case AtomicRMWInst::UMin: |
8155 | Pred = CmpInst::ICMP_ULT; |
8156 | break; |
8157 | case AtomicRMWInst::FMin: |
8158 | Pred = CmpInst::FCMP_OLT; |
8159 | break; |
8160 | default: |
8161 | llvm_unreachable("unexpected comparison op" ); |
8162 | } |
8163 | Value *NonAtomicCmp = Builder.CreateCmp(Pred, LHS: OldValue, RHS: E); |
8164 | CapturedValue = Builder.CreateSelect(C: NonAtomicCmp, True: E, False: OldValue); |
8165 | } |
8166 | Builder.CreateStore(Val: CapturedValue, Ptr: V.Var, isVolatile: V.IsVolatile); |
8167 | } |
8168 | } |
8169 | |
8170 | checkAndEmitFlushAfterAtomic(Loc, AO, AK: AtomicKind::Compare); |
8171 | |
8172 | return Builder.saveIP(); |
8173 | } |
8174 | |
8175 | OpenMPIRBuilder::InsertPointTy |
8176 | OpenMPIRBuilder::createTeams(const LocationDescription &Loc, |
8177 | BodyGenCallbackTy BodyGenCB, Value *NumTeamsLower, |
8178 | Value *NumTeamsUpper, Value *ThreadLimit, |
8179 | Value *IfExpr) { |
8180 | if (!updateToLocation(Loc)) |
8181 | return InsertPointTy(); |
8182 | |
8183 | uint32_t SrcLocStrSize; |
8184 | Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); |
8185 | Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); |
8186 | Function *CurrentFunction = Builder.GetInsertBlock()->getParent(); |
8187 | |
8188 | // Outer allocation basicblock is the entry block of the current function. |
8189 | BasicBlock &OuterAllocaBB = CurrentFunction->getEntryBlock(); |
8190 | if (&OuterAllocaBB == Builder.GetInsertBlock()) { |
8191 | BasicBlock *BodyBB = splitBB(Builder, /*CreateBranch=*/true, Name: "teams.entry" ); |
8192 | Builder.SetInsertPoint(TheBB: BodyBB, IP: BodyBB->begin()); |
8193 | } |
8194 | |
8195 | // The current basic block is split into four basic blocks. After outlining, |
8196 | // they will be mapped as follows: |
8197 | // ``` |
8198 | // def current_fn() { |
8199 | // current_basic_block: |
8200 | // br label %teams.exit |
8201 | // teams.exit: |
8202 | // ; instructions after teams |
8203 | // } |
8204 | // |
8205 | // def outlined_fn() { |
8206 | // teams.alloca: |
8207 | // br label %teams.body |
8208 | // teams.body: |
8209 | // ; instructions within teams body |
8210 | // } |
8211 | // ``` |
8212 | BasicBlock *ExitBB = splitBB(Builder, /*CreateBranch=*/true, Name: "teams.exit" ); |
8213 | BasicBlock *BodyBB = splitBB(Builder, /*CreateBranch=*/true, Name: "teams.body" ); |
8214 | BasicBlock *AllocaBB = |
8215 | splitBB(Builder, /*CreateBranch=*/true, Name: "teams.alloca" ); |
8216 | |
8217 | bool SubClausesPresent = |
8218 | (NumTeamsLower || NumTeamsUpper || ThreadLimit || IfExpr); |
8219 | // Push num_teams |
8220 | if (!Config.isTargetDevice() && SubClausesPresent) { |
8221 | assert((NumTeamsLower == nullptr || NumTeamsUpper != nullptr) && |
8222 | "if lowerbound is non-null, then upperbound must also be non-null " |
8223 | "for bounds on num_teams" ); |
8224 | |
8225 | if (NumTeamsUpper == nullptr) |
8226 | NumTeamsUpper = Builder.getInt32(C: 0); |
8227 | |
8228 | if (NumTeamsLower == nullptr) |
8229 | NumTeamsLower = NumTeamsUpper; |
8230 | |
8231 | if (IfExpr) { |
8232 | assert(IfExpr->getType()->isIntegerTy() && |
8233 | "argument to if clause must be an integer value" ); |
8234 | |
8235 | // upper = ifexpr ? upper : 1 |
8236 | if (IfExpr->getType() != Int1) |
8237 | IfExpr = Builder.CreateICmpNE(LHS: IfExpr, |
8238 | RHS: ConstantInt::get(Ty: IfExpr->getType(), V: 0)); |
8239 | NumTeamsUpper = Builder.CreateSelect( |
8240 | C: IfExpr, True: NumTeamsUpper, False: Builder.getInt32(C: 1), Name: "numTeamsUpper" ); |
8241 | |
8242 | // lower = ifexpr ? lower : 1 |
8243 | NumTeamsLower = Builder.CreateSelect( |
8244 | C: IfExpr, True: NumTeamsLower, False: Builder.getInt32(C: 1), Name: "numTeamsLower" ); |
8245 | } |
8246 | |
8247 | if (ThreadLimit == nullptr) |
8248 | ThreadLimit = Builder.getInt32(C: 0); |
8249 | |
8250 | Value *ThreadNum = getOrCreateThreadID(Ident); |
8251 | Builder.CreateCall( |
8252 | Callee: getOrCreateRuntimeFunctionPtr(FnID: OMPRTL___kmpc_push_num_teams_51), |
8253 | Args: {Ident, ThreadNum, NumTeamsLower, NumTeamsUpper, ThreadLimit}); |
8254 | } |
8255 | // Generate the body of teams. |
8256 | InsertPointTy AllocaIP(AllocaBB, AllocaBB->begin()); |
8257 | InsertPointTy CodeGenIP(BodyBB, BodyBB->begin()); |
8258 | BodyGenCB(AllocaIP, CodeGenIP); |
8259 | |
8260 | OutlineInfo OI; |
8261 | OI.EntryBB = AllocaBB; |
8262 | OI.ExitBB = ExitBB; |
8263 | OI.OuterAllocaBB = &OuterAllocaBB; |
8264 | |
8265 | // Insert fake values for global tid and bound tid. |
8266 | SmallVector<Instruction *, 8> ToBeDeleted; |
8267 | InsertPointTy OuterAllocaIP(&OuterAllocaBB, OuterAllocaBB.begin()); |
8268 | OI.ExcludeArgsFromAggregate.push_back(Elt: createFakeIntVal( |
8269 | Builder, OuterAllocaIP, ToBeDeleted, InnerAllocaIP: AllocaIP, Name: "gid" , AsPtr: true)); |
8270 | OI.ExcludeArgsFromAggregate.push_back(Elt: createFakeIntVal( |
8271 | Builder, OuterAllocaIP, ToBeDeleted, InnerAllocaIP: AllocaIP, Name: "tid" , AsPtr: true)); |
8272 | |
8273 | auto HostPostOutlineCB = [this, Ident, |
8274 | ToBeDeleted](Function &OutlinedFn) mutable { |
8275 | // The stale call instruction will be replaced with a new call instruction |
8276 | // for runtime call with the outlined function. |
8277 | |
8278 | assert(OutlinedFn.getNumUses() == 1 && |
8279 | "there must be a single user for the outlined function" ); |
8280 | CallInst *StaleCI = cast<CallInst>(Val: OutlinedFn.user_back()); |
8281 | ToBeDeleted.push_back(Elt: StaleCI); |
8282 | |
8283 | assert((OutlinedFn.arg_size() == 2 || OutlinedFn.arg_size() == 3) && |
8284 | "Outlined function must have two or three arguments only" ); |
8285 | |
8286 | bool HasShared = OutlinedFn.arg_size() == 3; |
8287 | |
8288 | OutlinedFn.getArg(i: 0)->setName("global.tid.ptr" ); |
8289 | OutlinedFn.getArg(i: 1)->setName("bound.tid.ptr" ); |
8290 | if (HasShared) |
8291 | OutlinedFn.getArg(i: 2)->setName("data" ); |
8292 | |
8293 | // Call to the runtime function for teams in the current function. |
8294 | assert(StaleCI && "Error while outlining - no CallInst user found for the " |
8295 | "outlined function." ); |
8296 | Builder.SetInsertPoint(StaleCI); |
8297 | SmallVector<Value *> Args = { |
8298 | Ident, Builder.getInt32(C: StaleCI->arg_size() - 2), &OutlinedFn}; |
8299 | if (HasShared) |
8300 | Args.push_back(Elt: StaleCI->getArgOperand(i: 2)); |
8301 | Builder.CreateCall(Callee: getOrCreateRuntimeFunctionPtr( |
8302 | FnID: omp::RuntimeFunction::OMPRTL___kmpc_fork_teams), |
8303 | Args); |
8304 | |
8305 | llvm::for_each(Range: llvm::reverse(C&: ToBeDeleted), |
8306 | F: [](Instruction *I) { I->eraseFromParent(); }); |
8307 | |
8308 | }; |
8309 | |
8310 | if (!Config.isTargetDevice()) |
8311 | OI.PostOutlineCB = HostPostOutlineCB; |
8312 | |
8313 | addOutlineInfo(OI: std::move(OI)); |
8314 | |
8315 | Builder.SetInsertPoint(TheBB: ExitBB, IP: ExitBB->begin()); |
8316 | |
8317 | return Builder.saveIP(); |
8318 | } |
8319 | |
8320 | GlobalVariable * |
8321 | OpenMPIRBuilder::createOffloadMapnames(SmallVectorImpl<llvm::Constant *> &Names, |
8322 | std::string VarName) { |
8323 | llvm::Constant *MapNamesArrayInit = llvm::ConstantArray::get( |
8324 | T: llvm::ArrayType::get(ElementType: llvm::PointerType::getUnqual(C&: M.getContext()), |
8325 | NumElements: Names.size()), |
8326 | V: Names); |
8327 | auto *MapNamesArrayGlobal = new llvm::GlobalVariable( |
8328 | M, MapNamesArrayInit->getType(), |
8329 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MapNamesArrayInit, |
8330 | VarName); |
8331 | return MapNamesArrayGlobal; |
8332 | } |
8333 | |
8334 | // Create all simple and struct types exposed by the runtime and remember |
8335 | // the llvm::PointerTypes of them for easy access later. |
8336 | void OpenMPIRBuilder::initializeTypes(Module &M) { |
8337 | LLVMContext &Ctx = M.getContext(); |
8338 | StructType *T; |
8339 | #define OMP_TYPE(VarName, InitValue) VarName = InitValue; |
8340 | #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize) \ |
8341 | VarName##Ty = ArrayType::get(ElemTy, ArraySize); \ |
8342 | VarName##PtrTy = PointerType::getUnqual(VarName##Ty); |
8343 | #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...) \ |
8344 | VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg); \ |
8345 | VarName##Ptr = PointerType::getUnqual(VarName); |
8346 | #define OMP_STRUCT_TYPE(VarName, StructName, Packed, ...) \ |
8347 | T = StructType::getTypeByName(Ctx, StructName); \ |
8348 | if (!T) \ |
8349 | T = StructType::create(Ctx, {__VA_ARGS__}, StructName, Packed); \ |
8350 | VarName = T; \ |
8351 | VarName##Ptr = PointerType::getUnqual(T); |
8352 | #include "llvm/Frontend/OpenMP/OMPKinds.def" |
8353 | } |
8354 | |
8355 | void OpenMPIRBuilder::OutlineInfo::collectBlocks( |
8356 | SmallPtrSetImpl<BasicBlock *> &BlockSet, |
8357 | SmallVectorImpl<BasicBlock *> &BlockVector) { |
8358 | SmallVector<BasicBlock *, 32> Worklist; |
8359 | BlockSet.insert(Ptr: EntryBB); |
8360 | BlockSet.insert(Ptr: ExitBB); |
8361 | |
8362 | Worklist.push_back(Elt: EntryBB); |
8363 | while (!Worklist.empty()) { |
8364 | BasicBlock *BB = Worklist.pop_back_val(); |
8365 | BlockVector.push_back(Elt: BB); |
8366 | for (BasicBlock *SuccBB : successors(BB)) |
8367 | if (BlockSet.insert(Ptr: SuccBB).second) |
8368 | Worklist.push_back(Elt: SuccBB); |
8369 | } |
8370 | } |
8371 | |
8372 | void OpenMPIRBuilder::createOffloadEntry(Constant *ID, Constant *Addr, |
8373 | uint64_t Size, int32_t Flags, |
8374 | GlobalValue::LinkageTypes, |
8375 | StringRef Name) { |
8376 | if (!Config.isGPU()) { |
8377 | llvm::offloading::emitOffloadingEntry( |
8378 | M, Addr: ID, Name: Name.empty() ? Addr->getName() : Name, Size, Flags, /*Data=*/0, |
8379 | SectionName: "omp_offloading_entries" ); |
8380 | return; |
8381 | } |
8382 | // TODO: Add support for global variables on the device after declare target |
8383 | // support. |
8384 | Function *Fn = dyn_cast<Function>(Val: Addr); |
8385 | if (!Fn) |
8386 | return; |
8387 | |
8388 | Module &M = *(Fn->getParent()); |
8389 | LLVMContext &Ctx = M.getContext(); |
8390 | |
8391 | // Get "nvvm.annotations" metadata node. |
8392 | NamedMDNode *MD = M.getOrInsertNamedMetadata(Name: "nvvm.annotations" ); |
8393 | |
8394 | Metadata *MDVals[] = { |
8395 | ConstantAsMetadata::get(C: Fn), MDString::get(Context&: Ctx, Str: "kernel" ), |
8396 | ConstantAsMetadata::get(C: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: 1))}; |
8397 | // Append metadata to nvvm.annotations. |
8398 | MD->addOperand(M: MDNode::get(Context&: Ctx, MDs: MDVals)); |
8399 | |
8400 | // Add a function attribute for the kernel. |
8401 | Fn->addFnAttr(Attr: Attribute::get(Context&: Ctx, Kind: "kernel" )); |
8402 | if (T.isAMDGCN()) |
8403 | Fn->addFnAttr(Kind: "uniform-work-group-size" , Val: "true" ); |
8404 | Fn->addFnAttr(Kind: Attribute::MustProgress); |
8405 | } |
8406 | |
8407 | // We only generate metadata for function that contain target regions. |
8408 | void OpenMPIRBuilder::createOffloadEntriesAndInfoMetadata( |
8409 | EmitMetadataErrorReportFunctionTy &ErrorFn) { |
8410 | |
8411 | // If there are no entries, we don't need to do anything. |
8412 | if (OffloadInfoManager.empty()) |
8413 | return; |
8414 | |
8415 | LLVMContext &C = M.getContext(); |
8416 | SmallVector<std::pair<const OffloadEntriesInfoManager::OffloadEntryInfo *, |
8417 | TargetRegionEntryInfo>, |
8418 | 16> |
8419 | OrderedEntries(OffloadInfoManager.size()); |
8420 | |
8421 | // Auxiliary methods to create metadata values and strings. |
8422 | auto &&GetMDInt = [this](unsigned V) { |
8423 | return ConstantAsMetadata::get(C: ConstantInt::get(Ty: Builder.getInt32Ty(), V)); |
8424 | }; |
8425 | |
8426 | auto &&GetMDString = [&C](StringRef V) { return MDString::get(Context&: C, Str: V); }; |
8427 | |
8428 | // Create the offloading info metadata node. |
8429 | NamedMDNode *MD = M.getOrInsertNamedMetadata(Name: "omp_offload.info" ); |
8430 | auto &&TargetRegionMetadataEmitter = |
8431 | [&C, MD, &OrderedEntries, &GetMDInt, &GetMDString]( |
8432 | const TargetRegionEntryInfo &EntryInfo, |
8433 | const OffloadEntriesInfoManager::OffloadEntryInfoTargetRegion &E) { |
8434 | // Generate metadata for target regions. Each entry of this metadata |
8435 | // contains: |
8436 | // - Entry 0 -> Kind of this type of metadata (0). |
8437 | // - Entry 1 -> Device ID of the file where the entry was identified. |
8438 | // - Entry 2 -> File ID of the file where the entry was identified. |
8439 | // - Entry 3 -> Mangled name of the function where the entry was |
8440 | // identified. |
8441 | // - Entry 4 -> Line in the file where the entry was identified. |
8442 | // - Entry 5 -> Count of regions at this DeviceID/FilesID/Line. |
8443 | // - Entry 6 -> Order the entry was created. |
8444 | // The first element of the metadata node is the kind. |
8445 | Metadata *Ops[] = { |
8446 | GetMDInt(E.getKind()), GetMDInt(EntryInfo.DeviceID), |
8447 | GetMDInt(EntryInfo.FileID), GetMDString(EntryInfo.ParentName), |
8448 | GetMDInt(EntryInfo.Line), GetMDInt(EntryInfo.Count), |
8449 | GetMDInt(E.getOrder())}; |
8450 | |
8451 | // Save this entry in the right position of the ordered entries array. |
8452 | OrderedEntries[E.getOrder()] = std::make_pair(x: &E, y: EntryInfo); |
8453 | |
8454 | // Add metadata to the named metadata node. |
8455 | MD->addOperand(M: MDNode::get(Context&: C, MDs: Ops)); |
8456 | }; |
8457 | |
8458 | OffloadInfoManager.actOnTargetRegionEntriesInfo(Action: TargetRegionMetadataEmitter); |
8459 | |
8460 | // Create function that emits metadata for each device global variable entry; |
8461 | auto &&DeviceGlobalVarMetadataEmitter = |
8462 | [&C, &OrderedEntries, &GetMDInt, &GetMDString, MD]( |
8463 | StringRef MangledName, |
8464 | const OffloadEntriesInfoManager::OffloadEntryInfoDeviceGlobalVar &E) { |
8465 | // Generate metadata for global variables. Each entry of this metadata |
8466 | // contains: |
8467 | // - Entry 0 -> Kind of this type of metadata (1). |
8468 | // - Entry 1 -> Mangled name of the variable. |
8469 | // - Entry 2 -> Declare target kind. |
8470 | // - Entry 3 -> Order the entry was created. |
8471 | // The first element of the metadata node is the kind. |
8472 | Metadata *Ops[] = {GetMDInt(E.getKind()), GetMDString(MangledName), |
8473 | GetMDInt(E.getFlags()), GetMDInt(E.getOrder())}; |
8474 | |
8475 | // Save this entry in the right position of the ordered entries array. |
8476 | TargetRegionEntryInfo varInfo(MangledName, 0, 0, 0); |
8477 | OrderedEntries[E.getOrder()] = std::make_pair(x: &E, y&: varInfo); |
8478 | |
8479 | // Add metadata to the named metadata node. |
8480 | MD->addOperand(M: MDNode::get(Context&: C, MDs: Ops)); |
8481 | }; |
8482 | |
8483 | OffloadInfoManager.actOnDeviceGlobalVarEntriesInfo( |
8484 | Action: DeviceGlobalVarMetadataEmitter); |
8485 | |
8486 | for (const auto &E : OrderedEntries) { |
8487 | assert(E.first && "All ordered entries must exist!" ); |
8488 | if (const auto *CE = |
8489 | dyn_cast<OffloadEntriesInfoManager::OffloadEntryInfoTargetRegion>( |
8490 | Val: E.first)) { |
8491 | if (!CE->getID() || !CE->getAddress()) { |
8492 | // Do not blame the entry if the parent funtion is not emitted. |
8493 | TargetRegionEntryInfo EntryInfo = E.second; |
8494 | StringRef FnName = EntryInfo.ParentName; |
8495 | if (!M.getNamedValue(Name: FnName)) |
8496 | continue; |
8497 | ErrorFn(EMIT_MD_TARGET_REGION_ERROR, EntryInfo); |
8498 | continue; |
8499 | } |
8500 | createOffloadEntry(ID: CE->getID(), Addr: CE->getAddress(), |
8501 | /*Size=*/0, Flags: CE->getFlags(), |
8502 | GlobalValue::WeakAnyLinkage); |
8503 | } else if (const auto *CE = dyn_cast< |
8504 | OffloadEntriesInfoManager::OffloadEntryInfoDeviceGlobalVar>( |
8505 | Val: E.first)) { |
8506 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind Flags = |
8507 | static_cast<OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind>( |
8508 | CE->getFlags()); |
8509 | switch (Flags) { |
8510 | case OffloadEntriesInfoManager::OMPTargetGlobalVarEntryEnter: |
8511 | case OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo: |
8512 | if (Config.isTargetDevice() && Config.hasRequiresUnifiedSharedMemory()) |
8513 | continue; |
8514 | if (!CE->getAddress()) { |
8515 | ErrorFn(EMIT_MD_DECLARE_TARGET_ERROR, E.second); |
8516 | continue; |
8517 | } |
8518 | // The vaiable has no definition - no need to add the entry. |
8519 | if (CE->getVarSize() == 0) |
8520 | continue; |
8521 | break; |
8522 | case OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink: |
8523 | assert(((Config.isTargetDevice() && !CE->getAddress()) || |
8524 | (!Config.isTargetDevice() && CE->getAddress())) && |
8525 | "Declaret target link address is set." ); |
8526 | if (Config.isTargetDevice()) |
8527 | continue; |
8528 | if (!CE->getAddress()) { |
8529 | ErrorFn(EMIT_MD_GLOBAL_VAR_LINK_ERROR, TargetRegionEntryInfo()); |
8530 | continue; |
8531 | } |
8532 | break; |
8533 | default: |
8534 | break; |
8535 | } |
8536 | |
8537 | // Hidden or internal symbols on the device are not externally visible. |
8538 | // We should not attempt to register them by creating an offloading |
8539 | // entry. Indirect variables are handled separately on the device. |
8540 | if (auto *GV = dyn_cast<GlobalValue>(Val: CE->getAddress())) |
8541 | if ((GV->hasLocalLinkage() || GV->hasHiddenVisibility()) && |
8542 | Flags != OffloadEntriesInfoManager::OMPTargetGlobalVarEntryIndirect) |
8543 | continue; |
8544 | |
8545 | // Indirect globals need to use a special name that doesn't match the name |
8546 | // of the associated host global. |
8547 | if (Flags == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryIndirect) |
8548 | createOffloadEntry(ID: CE->getAddress(), Addr: CE->getAddress(), Size: CE->getVarSize(), |
8549 | Flags, CE->getLinkage(), Name: CE->getVarName()); |
8550 | else |
8551 | createOffloadEntry(ID: CE->getAddress(), Addr: CE->getAddress(), Size: CE->getVarSize(), |
8552 | Flags, CE->getLinkage()); |
8553 | |
8554 | } else { |
8555 | llvm_unreachable("Unsupported entry kind." ); |
8556 | } |
8557 | } |
8558 | |
8559 | // Emit requires directive globals to a special entry so the runtime can |
8560 | // register them when the device image is loaded. |
8561 | // TODO: This reduces the offloading entries to a 32-bit integer. Offloading |
8562 | // entries should be redesigned to better suit this use-case. |
8563 | if (Config.hasRequiresFlags() && !Config.isTargetDevice()) |
8564 | offloading::emitOffloadingEntry( |
8565 | M, Addr: Constant::getNullValue(Ty: PointerType::getUnqual(C&: M.getContext())), |
8566 | /*Name=*/"" , |
8567 | /*Size=*/0, Flags: OffloadEntriesInfoManager::OMPTargetGlobalRegisterRequires, |
8568 | Data: Config.getRequiresFlags(), SectionName: "omp_offloading_entries" ); |
8569 | } |
8570 | |
8571 | void TargetRegionEntryInfo::getTargetRegionEntryFnName( |
8572 | SmallVectorImpl<char> &Name, StringRef ParentName, unsigned DeviceID, |
8573 | unsigned FileID, unsigned Line, unsigned Count) { |
8574 | raw_svector_ostream OS(Name); |
8575 | OS << "__omp_offloading" << llvm::format(Fmt: "_%x" , Vals: DeviceID) |
8576 | << llvm::format(Fmt: "_%x_" , Vals: FileID) << ParentName << "_l" << Line; |
8577 | if (Count) |
8578 | OS << "_" << Count; |
8579 | } |
8580 | |
8581 | void OffloadEntriesInfoManager::getTargetRegionEntryFnName( |
8582 | SmallVectorImpl<char> &Name, const TargetRegionEntryInfo &EntryInfo) { |
8583 | unsigned NewCount = getTargetRegionEntryInfoCount(EntryInfo); |
8584 | TargetRegionEntryInfo::getTargetRegionEntryFnName( |
8585 | Name, ParentName: EntryInfo.ParentName, DeviceID: EntryInfo.DeviceID, FileID: EntryInfo.FileID, |
8586 | Line: EntryInfo.Line, Count: NewCount); |
8587 | } |
8588 | |
8589 | TargetRegionEntryInfo |
8590 | OpenMPIRBuilder::getTargetEntryUniqueInfo(FileIdentifierInfoCallbackTy CallBack, |
8591 | StringRef ParentName) { |
8592 | sys::fs::UniqueID ID; |
8593 | auto FileIDInfo = CallBack(); |
8594 | if (auto EC = sys::fs::getUniqueID(Path: std::get<0>(t&: FileIDInfo), Result&: ID)) { |
8595 | report_fatal_error(reason: ("Unable to get unique ID for file, during " |
8596 | "getTargetEntryUniqueInfo, error message: " + |
8597 | EC.message()) |
8598 | .c_str()); |
8599 | } |
8600 | |
8601 | return TargetRegionEntryInfo(ParentName, ID.getDevice(), ID.getFile(), |
8602 | std::get<1>(t&: FileIDInfo)); |
8603 | } |
8604 | |
8605 | unsigned OpenMPIRBuilder::getFlagMemberOffset() { |
8606 | unsigned Offset = 0; |
8607 | for (uint64_t Remain = |
8608 | static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>>( |
8609 | omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF); |
8610 | !(Remain & 1); Remain = Remain >> 1) |
8611 | Offset++; |
8612 | return Offset; |
8613 | } |
8614 | |
8615 | omp::OpenMPOffloadMappingFlags |
8616 | OpenMPIRBuilder::getMemberOfFlag(unsigned Position) { |
8617 | // Rotate by getFlagMemberOffset() bits. |
8618 | return static_cast<omp::OpenMPOffloadMappingFlags>(((uint64_t)Position + 1) |
8619 | << getFlagMemberOffset()); |
8620 | } |
8621 | |
8622 | void OpenMPIRBuilder::setCorrectMemberOfFlag( |
8623 | omp::OpenMPOffloadMappingFlags &Flags, |
8624 | omp::OpenMPOffloadMappingFlags MemberOfFlag) { |
8625 | // If the entry is PTR_AND_OBJ but has not been marked with the special |
8626 | // placeholder value 0xFFFF in the MEMBER_OF field, then it should not be |
8627 | // marked as MEMBER_OF. |
8628 | if (static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>>( |
8629 | Flags & omp::OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ) && |
8630 | static_cast<std::underlying_type_t<omp::OpenMPOffloadMappingFlags>>( |
8631 | (Flags & omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF) != |
8632 | omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF)) |
8633 | return; |
8634 | |
8635 | // Reset the placeholder value to prepare the flag for the assignment of the |
8636 | // proper MEMBER_OF value. |
8637 | Flags &= ~omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF; |
8638 | Flags |= MemberOfFlag; |
8639 | } |
8640 | |
8641 | Constant *OpenMPIRBuilder::getAddrOfDeclareTargetVar( |
8642 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind CaptureClause, |
8643 | OffloadEntriesInfoManager::OMPTargetDeviceClauseKind DeviceClause, |
8644 | bool IsDeclaration, bool IsExternallyVisible, |
8645 | TargetRegionEntryInfo EntryInfo, StringRef MangledName, |
8646 | std::vector<GlobalVariable *> &GeneratedRefs, bool OpenMPSIMD, |
8647 | std::vector<Triple> TargetTriple, Type *LlvmPtrTy, |
8648 | std::function<Constant *()> GlobalInitializer, |
8649 | std::function<GlobalValue::LinkageTypes()> VariableLinkage) { |
8650 | // TODO: convert this to utilise the IRBuilder Config rather than |
8651 | // a passed down argument. |
8652 | if (OpenMPSIMD) |
8653 | return nullptr; |
8654 | |
8655 | if (CaptureClause == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink || |
8656 | ((CaptureClause == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo || |
8657 | CaptureClause == |
8658 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryEnter) && |
8659 | Config.hasRequiresUnifiedSharedMemory())) { |
8660 | SmallString<64> PtrName; |
8661 | { |
8662 | raw_svector_ostream OS(PtrName); |
8663 | OS << MangledName; |
8664 | if (!IsExternallyVisible) |
8665 | OS << format(Fmt: "_%x" , Vals: EntryInfo.FileID); |
8666 | OS << "_decl_tgt_ref_ptr" ; |
8667 | } |
8668 | |
8669 | Value *Ptr = M.getNamedValue(Name: PtrName); |
8670 | |
8671 | if (!Ptr) { |
8672 | GlobalValue *GlobalValue = M.getNamedValue(Name: MangledName); |
8673 | Ptr = getOrCreateInternalVariable(Ty: LlvmPtrTy, Name: PtrName); |
8674 | |
8675 | auto *GV = cast<GlobalVariable>(Val: Ptr); |
8676 | GV->setLinkage(GlobalValue::WeakAnyLinkage); |
8677 | |
8678 | if (!Config.isTargetDevice()) { |
8679 | if (GlobalInitializer) |
8680 | GV->setInitializer(GlobalInitializer()); |
8681 | else |
8682 | GV->setInitializer(GlobalValue); |
8683 | } |
8684 | |
8685 | registerTargetGlobalVariable( |
8686 | CaptureClause, DeviceClause, IsDeclaration, IsExternallyVisible, |
8687 | EntryInfo, MangledName, GeneratedRefs, OpenMPSIMD, TargetTriple, |
8688 | GlobalInitializer, VariableLinkage, LlvmPtrTy, Addr: cast<Constant>(Val: Ptr)); |
8689 | } |
8690 | |
8691 | return cast<Constant>(Val: Ptr); |
8692 | } |
8693 | |
8694 | return nullptr; |
8695 | } |
8696 | |
8697 | void OpenMPIRBuilder::registerTargetGlobalVariable( |
8698 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind CaptureClause, |
8699 | OffloadEntriesInfoManager::OMPTargetDeviceClauseKind DeviceClause, |
8700 | bool IsDeclaration, bool IsExternallyVisible, |
8701 | TargetRegionEntryInfo EntryInfo, StringRef MangledName, |
8702 | std::vector<GlobalVariable *> &GeneratedRefs, bool OpenMPSIMD, |
8703 | std::vector<Triple> TargetTriple, |
8704 | std::function<Constant *()> GlobalInitializer, |
8705 | std::function<GlobalValue::LinkageTypes()> VariableLinkage, Type *LlvmPtrTy, |
8706 | Constant *Addr) { |
8707 | if (DeviceClause != OffloadEntriesInfoManager::OMPTargetDeviceClauseAny || |
8708 | (TargetTriple.empty() && !Config.isTargetDevice())) |
8709 | return; |
8710 | |
8711 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind Flags; |
8712 | StringRef VarName; |
8713 | int64_t VarSize; |
8714 | GlobalValue::LinkageTypes Linkage; |
8715 | |
8716 | if ((CaptureClause == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo || |
8717 | CaptureClause == |
8718 | OffloadEntriesInfoManager::OMPTargetGlobalVarEntryEnter) && |
8719 | !Config.hasRequiresUnifiedSharedMemory()) { |
8720 | Flags = OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo; |
8721 | VarName = MangledName; |
8722 | GlobalValue *LlvmVal = M.getNamedValue(Name: VarName); |
8723 | |
8724 | if (!IsDeclaration) |
8725 | VarSize = divideCeil( |
8726 | Numerator: M.getDataLayout().getTypeSizeInBits(Ty: LlvmVal->getValueType()), Denominator: 8); |
8727 | else |
8728 | VarSize = 0; |
8729 | Linkage = (VariableLinkage) ? VariableLinkage() : LlvmVal->getLinkage(); |
8730 | |
8731 | // This is a workaround carried over from Clang which prevents undesired |
8732 | // optimisation of internal variables. |
8733 | if (Config.isTargetDevice() && |
8734 | (!IsExternallyVisible || Linkage == GlobalValue::LinkOnceODRLinkage)) { |
8735 | // Do not create a "ref-variable" if the original is not also available |
8736 | // on the host. |
8737 | if (!OffloadInfoManager.hasDeviceGlobalVarEntryInfo(VarName)) |
8738 | return; |
8739 | |
8740 | std::string RefName = createPlatformSpecificName(Parts: {VarName, "ref" }); |
8741 | |
8742 | if (!M.getNamedValue(Name: RefName)) { |
8743 | Constant *AddrRef = |
8744 | getOrCreateInternalVariable(Ty: Addr->getType(), Name: RefName); |
8745 | auto *GvAddrRef = cast<GlobalVariable>(Val: AddrRef); |
8746 | GvAddrRef->setConstant(true); |
8747 | GvAddrRef->setLinkage(GlobalValue::InternalLinkage); |
8748 | GvAddrRef->setInitializer(Addr); |
8749 | GeneratedRefs.push_back(x: GvAddrRef); |
8750 | } |
8751 | } |
8752 | } else { |
8753 | if (CaptureClause == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink) |
8754 | Flags = OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink; |
8755 | else |
8756 | Flags = OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo; |
8757 | |
8758 | if (Config.isTargetDevice()) { |
8759 | VarName = (Addr) ? Addr->getName() : "" ; |
8760 | Addr = nullptr; |
8761 | } else { |
8762 | Addr = getAddrOfDeclareTargetVar( |
8763 | CaptureClause, DeviceClause, IsDeclaration, IsExternallyVisible, |
8764 | EntryInfo, MangledName, GeneratedRefs, OpenMPSIMD, TargetTriple, |
8765 | LlvmPtrTy, GlobalInitializer, VariableLinkage); |
8766 | VarName = (Addr) ? Addr->getName() : "" ; |
8767 | } |
8768 | VarSize = M.getDataLayout().getPointerSize(); |
8769 | Linkage = GlobalValue::WeakAnyLinkage; |
8770 | } |
8771 | |
8772 | OffloadInfoManager.registerDeviceGlobalVarEntryInfo(VarName, Addr, VarSize, |
8773 | Flags, Linkage); |
8774 | } |
8775 | |
8776 | /// Loads all the offload entries information from the host IR |
8777 | /// metadata. |
8778 | void OpenMPIRBuilder::loadOffloadInfoMetadata(Module &M) { |
8779 | // If we are in target mode, load the metadata from the host IR. This code has |
8780 | // to match the metadata creation in createOffloadEntriesAndInfoMetadata(). |
8781 | |
8782 | NamedMDNode *MD = M.getNamedMetadata(Name: ompOffloadInfoName); |
8783 | if (!MD) |
8784 | return; |
8785 | |
8786 | for (MDNode *MN : MD->operands()) { |
8787 | auto &&GetMDInt = [MN](unsigned Idx) { |
8788 | auto *V = cast<ConstantAsMetadata>(Val: MN->getOperand(I: Idx)); |
8789 | return cast<ConstantInt>(Val: V->getValue())->getZExtValue(); |
8790 | }; |
8791 | |
8792 | auto &&GetMDString = [MN](unsigned Idx) { |
8793 | auto *V = cast<MDString>(Val: MN->getOperand(I: Idx)); |
8794 | return V->getString(); |
8795 | }; |
8796 | |
8797 | switch (GetMDInt(0)) { |
8798 | default: |
8799 | llvm_unreachable("Unexpected metadata!" ); |
8800 | break; |
8801 | case OffloadEntriesInfoManager::OffloadEntryInfo:: |
8802 | OffloadingEntryInfoTargetRegion: { |
8803 | TargetRegionEntryInfo EntryInfo(/*ParentName=*/GetMDString(3), |
8804 | /*DeviceID=*/GetMDInt(1), |
8805 | /*FileID=*/GetMDInt(2), |
8806 | /*Line=*/GetMDInt(4), |
8807 | /*Count=*/GetMDInt(5)); |
8808 | OffloadInfoManager.initializeTargetRegionEntryInfo(EntryInfo, |
8809 | /*Order=*/GetMDInt(6)); |
8810 | break; |
8811 | } |
8812 | case OffloadEntriesInfoManager::OffloadEntryInfo:: |
8813 | OffloadingEntryInfoDeviceGlobalVar: |
8814 | OffloadInfoManager.initializeDeviceGlobalVarEntryInfo( |
8815 | /*MangledName=*/Name: GetMDString(1), |
8816 | Flags: static_cast<OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind>( |
8817 | /*Flags=*/GetMDInt(2)), |
8818 | /*Order=*/GetMDInt(3)); |
8819 | break; |
8820 | } |
8821 | } |
8822 | } |
8823 | |
8824 | void OpenMPIRBuilder::loadOffloadInfoMetadata(StringRef HostFilePath) { |
8825 | if (HostFilePath.empty()) |
8826 | return; |
8827 | |
8828 | auto Buf = MemoryBuffer::getFile(Filename: HostFilePath); |
8829 | if (std::error_code Err = Buf.getError()) { |
8830 | report_fatal_error(reason: ("error opening host file from host file path inside of " |
8831 | "OpenMPIRBuilder: " + |
8832 | Err.message()) |
8833 | .c_str()); |
8834 | } |
8835 | |
8836 | LLVMContext Ctx; |
8837 | auto M = expectedToErrorOrAndEmitErrors( |
8838 | Ctx, Val: parseBitcodeFile(Buffer: Buf.get()->getMemBufferRef(), Context&: Ctx)); |
8839 | if (std::error_code Err = M.getError()) { |
8840 | report_fatal_error( |
8841 | reason: ("error parsing host file inside of OpenMPIRBuilder: " + Err.message()) |
8842 | .c_str()); |
8843 | } |
8844 | |
8845 | loadOffloadInfoMetadata(M&: *M.get()); |
8846 | } |
8847 | |
8848 | //===----------------------------------------------------------------------===// |
8849 | // OffloadEntriesInfoManager |
8850 | //===----------------------------------------------------------------------===// |
8851 | |
8852 | bool OffloadEntriesInfoManager::empty() const { |
8853 | return OffloadEntriesTargetRegion.empty() && |
8854 | OffloadEntriesDeviceGlobalVar.empty(); |
8855 | } |
8856 | |
8857 | unsigned OffloadEntriesInfoManager::getTargetRegionEntryInfoCount( |
8858 | const TargetRegionEntryInfo &EntryInfo) const { |
8859 | auto It = OffloadEntriesTargetRegionCount.find( |
8860 | x: getTargetRegionEntryCountKey(EntryInfo)); |
8861 | if (It == OffloadEntriesTargetRegionCount.end()) |
8862 | return 0; |
8863 | return It->second; |
8864 | } |
8865 | |
8866 | void OffloadEntriesInfoManager::incrementTargetRegionEntryInfoCount( |
8867 | const TargetRegionEntryInfo &EntryInfo) { |
8868 | OffloadEntriesTargetRegionCount[getTargetRegionEntryCountKey(EntryInfo)] = |
8869 | EntryInfo.Count + 1; |
8870 | } |
8871 | |
8872 | /// Initialize target region entry. |
8873 | void OffloadEntriesInfoManager::initializeTargetRegionEntryInfo( |
8874 | const TargetRegionEntryInfo &EntryInfo, unsigned Order) { |
8875 | OffloadEntriesTargetRegion[EntryInfo] = |
8876 | OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr, |
8877 | OMPTargetRegionEntryTargetRegion); |
8878 | ++OffloadingEntriesNum; |
8879 | } |
8880 | |
8881 | void OffloadEntriesInfoManager::registerTargetRegionEntryInfo( |
8882 | TargetRegionEntryInfo EntryInfo, Constant *Addr, Constant *ID, |
8883 | OMPTargetRegionEntryKind Flags) { |
8884 | assert(EntryInfo.Count == 0 && "expected default EntryInfo" ); |
8885 | |
8886 | // Update the EntryInfo with the next available count for this location. |
8887 | EntryInfo.Count = getTargetRegionEntryInfoCount(EntryInfo); |
8888 | |
8889 | // If we are emitting code for a target, the entry is already initialized, |
8890 | // only has to be registered. |
8891 | if (OMPBuilder->Config.isTargetDevice()) { |
8892 | // This could happen if the device compilation is invoked standalone. |
8893 | if (!hasTargetRegionEntryInfo(EntryInfo)) { |
8894 | return; |
8895 | } |
8896 | auto &Entry = OffloadEntriesTargetRegion[EntryInfo]; |
8897 | Entry.setAddress(Addr); |
8898 | Entry.setID(ID); |
8899 | Entry.setFlags(Flags); |
8900 | } else { |
8901 | if (Flags == OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion && |
8902 | hasTargetRegionEntryInfo(EntryInfo, /*IgnoreAddressId*/ true)) |
8903 | return; |
8904 | assert(!hasTargetRegionEntryInfo(EntryInfo) && |
8905 | "Target region entry already registered!" ); |
8906 | OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum, Addr, ID, Flags); |
8907 | OffloadEntriesTargetRegion[EntryInfo] = Entry; |
8908 | ++OffloadingEntriesNum; |
8909 | } |
8910 | incrementTargetRegionEntryInfoCount(EntryInfo); |
8911 | } |
8912 | |
8913 | bool OffloadEntriesInfoManager::hasTargetRegionEntryInfo( |
8914 | TargetRegionEntryInfo EntryInfo, bool IgnoreAddressId) const { |
8915 | |
8916 | // Update the EntryInfo with the next available count for this location. |
8917 | EntryInfo.Count = getTargetRegionEntryInfoCount(EntryInfo); |
8918 | |
8919 | auto It = OffloadEntriesTargetRegion.find(x: EntryInfo); |
8920 | if (It == OffloadEntriesTargetRegion.end()) { |
8921 | return false; |
8922 | } |
8923 | // Fail if this entry is already registered. |
8924 | if (!IgnoreAddressId && (It->second.getAddress() || It->second.getID())) |
8925 | return false; |
8926 | return true; |
8927 | } |
8928 | |
8929 | void OffloadEntriesInfoManager::actOnTargetRegionEntriesInfo( |
8930 | const OffloadTargetRegionEntryInfoActTy &Action) { |
8931 | // Scan all target region entries and perform the provided action. |
8932 | for (const auto &It : OffloadEntriesTargetRegion) { |
8933 | Action(It.first, It.second); |
8934 | } |
8935 | } |
8936 | |
8937 | void OffloadEntriesInfoManager::initializeDeviceGlobalVarEntryInfo( |
8938 | StringRef Name, OMPTargetGlobalVarEntryKind Flags, unsigned Order) { |
8939 | OffloadEntriesDeviceGlobalVar.try_emplace(Key: Name, Args&: Order, Args&: Flags); |
8940 | ++OffloadingEntriesNum; |
8941 | } |
8942 | |
8943 | void OffloadEntriesInfoManager::registerDeviceGlobalVarEntryInfo( |
8944 | StringRef VarName, Constant *Addr, int64_t VarSize, |
8945 | OMPTargetGlobalVarEntryKind Flags, GlobalValue::LinkageTypes Linkage) { |
8946 | if (OMPBuilder->Config.isTargetDevice()) { |
8947 | // This could happen if the device compilation is invoked standalone. |
8948 | if (!hasDeviceGlobalVarEntryInfo(VarName)) |
8949 | return; |
8950 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; |
8951 | if (Entry.getAddress() && hasDeviceGlobalVarEntryInfo(VarName)) { |
8952 | if (Entry.getVarSize() == 0) { |
8953 | Entry.setVarSize(VarSize); |
8954 | Entry.setLinkage(Linkage); |
8955 | } |
8956 | return; |
8957 | } |
8958 | Entry.setVarSize(VarSize); |
8959 | Entry.setLinkage(Linkage); |
8960 | Entry.setAddress(Addr); |
8961 | } else { |
8962 | if (hasDeviceGlobalVarEntryInfo(VarName)) { |
8963 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; |
8964 | assert(Entry.isValid() && Entry.getFlags() == Flags && |
8965 | "Entry not initialized!" ); |
8966 | if (Entry.getVarSize() == 0) { |
8967 | Entry.setVarSize(VarSize); |
8968 | Entry.setLinkage(Linkage); |
8969 | } |
8970 | return; |
8971 | } |
8972 | if (Flags == OffloadEntriesInfoManager::OMPTargetGlobalVarEntryIndirect) |
8973 | OffloadEntriesDeviceGlobalVar.try_emplace(Key: VarName, Args&: OffloadingEntriesNum, |
8974 | Args&: Addr, Args&: VarSize, Args&: Flags, Args&: Linkage, |
8975 | Args: VarName.str()); |
8976 | else |
8977 | OffloadEntriesDeviceGlobalVar.try_emplace( |
8978 | Key: VarName, Args&: OffloadingEntriesNum, Args&: Addr, Args&: VarSize, Args&: Flags, Args&: Linkage, Args: "" ); |
8979 | ++OffloadingEntriesNum; |
8980 | } |
8981 | } |
8982 | |
8983 | void OffloadEntriesInfoManager::actOnDeviceGlobalVarEntriesInfo( |
8984 | const OffloadDeviceGlobalVarEntryInfoActTy &Action) { |
8985 | // Scan all target region entries and perform the provided action. |
8986 | for (const auto &E : OffloadEntriesDeviceGlobalVar) |
8987 | Action(E.getKey(), E.getValue()); |
8988 | } |
8989 | |
8990 | //===----------------------------------------------------------------------===// |
8991 | // CanonicalLoopInfo |
8992 | //===----------------------------------------------------------------------===// |
8993 | |
8994 | void CanonicalLoopInfo::collectControlBlocks( |
8995 | SmallVectorImpl<BasicBlock *> &BBs) { |
8996 | // We only count those BBs as control block for which we do not need to |
8997 | // reverse the CFG, i.e. not the loop body which can contain arbitrary control |
8998 | // flow. For consistency, this also means we do not add the Body block, which |
8999 | // is just the entry to the body code. |
9000 | BBs.reserve(N: BBs.size() + 6); |
9001 | BBs.append(IL: {getPreheader(), Header, Cond, Latch, Exit, getAfter()}); |
9002 | } |
9003 | |
9004 | BasicBlock *CanonicalLoopInfo::() const { |
9005 | assert(isValid() && "Requires a valid canonical loop" ); |
9006 | for (BasicBlock *Pred : predecessors(BB: Header)) { |
9007 | if (Pred != Latch) |
9008 | return Pred; |
9009 | } |
9010 | llvm_unreachable("Missing preheader" ); |
9011 | } |
9012 | |
9013 | void CanonicalLoopInfo::setTripCount(Value *TripCount) { |
9014 | assert(isValid() && "Requires a valid canonical loop" ); |
9015 | |
9016 | Instruction *CmpI = &getCond()->front(); |
9017 | assert(isa<CmpInst>(CmpI) && "First inst must compare IV with TripCount" ); |
9018 | CmpI->setOperand(i: 1, Val: TripCount); |
9019 | |
9020 | #ifndef NDEBUG |
9021 | assertOK(); |
9022 | #endif |
9023 | } |
9024 | |
9025 | void CanonicalLoopInfo::mapIndVar( |
9026 | llvm::function_ref<Value *(Instruction *)> Updater) { |
9027 | assert(isValid() && "Requires a valid canonical loop" ); |
9028 | |
9029 | Instruction *OldIV = getIndVar(); |
9030 | |
9031 | // Record all uses excluding those introduced by the updater. Uses by the |
9032 | // CanonicalLoopInfo itself to keep track of the number of iterations are |
9033 | // excluded. |
9034 | SmallVector<Use *> ReplacableUses; |
9035 | for (Use &U : OldIV->uses()) { |
9036 | auto *User = dyn_cast<Instruction>(Val: U.getUser()); |
9037 | if (!User) |
9038 | continue; |
9039 | if (User->getParent() == getCond()) |
9040 | continue; |
9041 | if (User->getParent() == getLatch()) |
9042 | continue; |
9043 | ReplacableUses.push_back(Elt: &U); |
9044 | } |
9045 | |
9046 | // Run the updater that may introduce new uses |
9047 | Value *NewIV = Updater(OldIV); |
9048 | |
9049 | // Replace the old uses with the value returned by the updater. |
9050 | for (Use *U : ReplacableUses) |
9051 | U->set(NewIV); |
9052 | |
9053 | #ifndef NDEBUG |
9054 | assertOK(); |
9055 | #endif |
9056 | } |
9057 | |
9058 | void CanonicalLoopInfo::assertOK() const { |
9059 | #ifndef NDEBUG |
9060 | // No constraints if this object currently does not describe a loop. |
9061 | if (!isValid()) |
9062 | return; |
9063 | |
9064 | BasicBlock *Preheader = getPreheader(); |
9065 | BasicBlock *Body = getBody(); |
9066 | BasicBlock *After = getAfter(); |
9067 | |
9068 | // Verify standard control-flow we use for OpenMP loops. |
9069 | assert(Preheader); |
9070 | assert(isa<BranchInst>(Preheader->getTerminator()) && |
9071 | "Preheader must terminate with unconditional branch" ); |
9072 | assert(Preheader->getSingleSuccessor() == Header && |
9073 | "Preheader must jump to header" ); |
9074 | |
9075 | assert(Header); |
9076 | assert(isa<BranchInst>(Header->getTerminator()) && |
9077 | "Header must terminate with unconditional branch" ); |
9078 | assert(Header->getSingleSuccessor() == Cond && |
9079 | "Header must jump to exiting block" ); |
9080 | |
9081 | assert(Cond); |
9082 | assert(Cond->getSinglePredecessor() == Header && |
9083 | "Exiting block only reachable from header" ); |
9084 | |
9085 | assert(isa<BranchInst>(Cond->getTerminator()) && |
9086 | "Exiting block must terminate with conditional branch" ); |
9087 | assert(size(successors(Cond)) == 2 && |
9088 | "Exiting block must have two successors" ); |
9089 | assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(0) == Body && |
9090 | "Exiting block's first successor jump to the body" ); |
9091 | assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(1) == Exit && |
9092 | "Exiting block's second successor must exit the loop" ); |
9093 | |
9094 | assert(Body); |
9095 | assert(Body->getSinglePredecessor() == Cond && |
9096 | "Body only reachable from exiting block" ); |
9097 | assert(!isa<PHINode>(Body->front())); |
9098 | |
9099 | assert(Latch); |
9100 | assert(isa<BranchInst>(Latch->getTerminator()) && |
9101 | "Latch must terminate with unconditional branch" ); |
9102 | assert(Latch->getSingleSuccessor() == Header && "Latch must jump to header" ); |
9103 | // TODO: To support simple redirecting of the end of the body code that has |
9104 | // multiple; introduce another auxiliary basic block like preheader and after. |
9105 | assert(Latch->getSinglePredecessor() != nullptr); |
9106 | assert(!isa<PHINode>(Latch->front())); |
9107 | |
9108 | assert(Exit); |
9109 | assert(isa<BranchInst>(Exit->getTerminator()) && |
9110 | "Exit block must terminate with unconditional branch" ); |
9111 | assert(Exit->getSingleSuccessor() == After && |
9112 | "Exit block must jump to after block" ); |
9113 | |
9114 | assert(After); |
9115 | assert(After->getSinglePredecessor() == Exit && |
9116 | "After block only reachable from exit block" ); |
9117 | assert(After->empty() || !isa<PHINode>(After->front())); |
9118 | |
9119 | Instruction *IndVar = getIndVar(); |
9120 | assert(IndVar && "Canonical induction variable not found?" ); |
9121 | assert(isa<IntegerType>(IndVar->getType()) && |
9122 | "Induction variable must be an integer" ); |
9123 | assert(cast<PHINode>(IndVar)->getParent() == Header && |
9124 | "Induction variable must be a PHI in the loop header" ); |
9125 | assert(cast<PHINode>(IndVar)->getIncomingBlock(0) == Preheader); |
9126 | assert( |
9127 | cast<ConstantInt>(cast<PHINode>(IndVar)->getIncomingValue(0))->isZero()); |
9128 | assert(cast<PHINode>(IndVar)->getIncomingBlock(1) == Latch); |
9129 | |
9130 | auto *NextIndVar = cast<PHINode>(IndVar)->getIncomingValue(1); |
9131 | assert(cast<Instruction>(NextIndVar)->getParent() == Latch); |
9132 | assert(cast<BinaryOperator>(NextIndVar)->getOpcode() == BinaryOperator::Add); |
9133 | assert(cast<BinaryOperator>(NextIndVar)->getOperand(0) == IndVar); |
9134 | assert(cast<ConstantInt>(cast<BinaryOperator>(NextIndVar)->getOperand(1)) |
9135 | ->isOne()); |
9136 | |
9137 | Value *TripCount = getTripCount(); |
9138 | assert(TripCount && "Loop trip count not found?" ); |
9139 | assert(IndVar->getType() == TripCount->getType() && |
9140 | "Trip count and induction variable must have the same type" ); |
9141 | |
9142 | auto *CmpI = cast<CmpInst>(&Cond->front()); |
9143 | assert(CmpI->getPredicate() == CmpInst::ICMP_ULT && |
9144 | "Exit condition must be a signed less-than comparison" ); |
9145 | assert(CmpI->getOperand(0) == IndVar && |
9146 | "Exit condition must compare the induction variable" ); |
9147 | assert(CmpI->getOperand(1) == TripCount && |
9148 | "Exit condition must compare with the trip count" ); |
9149 | #endif |
9150 | } |
9151 | |
9152 | void CanonicalLoopInfo::invalidate() { |
9153 | Header = nullptr; |
9154 | Cond = nullptr; |
9155 | Latch = nullptr; |
9156 | Exit = nullptr; |
9157 | } |
9158 | |