1 | //===- MachineFunction.cpp ------------------------------------------------===// |
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 | // |
9 | // Collect native machine code information for a function. This allows |
10 | // target-specific information about the generated code to be stored with each |
11 | // function. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/CodeGen/MachineFunction.h" |
16 | #include "llvm/ADT/BitVector.h" |
17 | #include "llvm/ADT/DenseMap.h" |
18 | #include "llvm/ADT/DenseSet.h" |
19 | #include "llvm/ADT/STLExtras.h" |
20 | #include "llvm/ADT/SmallString.h" |
21 | #include "llvm/ADT/SmallVector.h" |
22 | #include "llvm/ADT/StringRef.h" |
23 | #include "llvm/ADT/Twine.h" |
24 | #include "llvm/Analysis/ConstantFolding.h" |
25 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
26 | #include "llvm/CodeGen/MachineBasicBlock.h" |
27 | #include "llvm/CodeGen/MachineConstantPool.h" |
28 | #include "llvm/CodeGen/MachineFrameInfo.h" |
29 | #include "llvm/CodeGen/MachineInstr.h" |
30 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
31 | #include "llvm/CodeGen/MachineMemOperand.h" |
32 | #include "llvm/CodeGen/MachineModuleInfo.h" |
33 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
34 | #include "llvm/CodeGen/PseudoSourceValue.h" |
35 | #include "llvm/CodeGen/PseudoSourceValueManager.h" |
36 | #include "llvm/CodeGen/TargetFrameLowering.h" |
37 | #include "llvm/CodeGen/TargetInstrInfo.h" |
38 | #include "llvm/CodeGen/TargetLowering.h" |
39 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
40 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
41 | #include "llvm/CodeGen/WasmEHFuncInfo.h" |
42 | #include "llvm/CodeGen/WinEHFuncInfo.h" |
43 | #include "llvm/Config/llvm-config.h" |
44 | #include "llvm/IR/Attributes.h" |
45 | #include "llvm/IR/BasicBlock.h" |
46 | #include "llvm/IR/Constant.h" |
47 | #include "llvm/IR/DataLayout.h" |
48 | #include "llvm/IR/DerivedTypes.h" |
49 | #include "llvm/IR/EHPersonalities.h" |
50 | #include "llvm/IR/Function.h" |
51 | #include "llvm/IR/GlobalValue.h" |
52 | #include "llvm/IR/Instruction.h" |
53 | #include "llvm/IR/Instructions.h" |
54 | #include "llvm/IR/Metadata.h" |
55 | #include "llvm/IR/Module.h" |
56 | #include "llvm/IR/ModuleSlotTracker.h" |
57 | #include "llvm/IR/Value.h" |
58 | #include "llvm/MC/MCContext.h" |
59 | #include "llvm/MC/MCSymbol.h" |
60 | #include "llvm/MC/SectionKind.h" |
61 | #include "llvm/Support/Casting.h" |
62 | #include "llvm/Support/CommandLine.h" |
63 | #include "llvm/Support/Compiler.h" |
64 | #include "llvm/Support/DOTGraphTraits.h" |
65 | #include "llvm/Support/ErrorHandling.h" |
66 | #include "llvm/Support/GraphWriter.h" |
67 | #include "llvm/Support/raw_ostream.h" |
68 | #include "llvm/Target/TargetMachine.h" |
69 | #include <algorithm> |
70 | #include <cassert> |
71 | #include <cstddef> |
72 | #include <cstdint> |
73 | #include <iterator> |
74 | #include <string> |
75 | #include <utility> |
76 | #include <vector> |
77 | |
78 | #include "LiveDebugValues/LiveDebugValues.h" |
79 | |
80 | using namespace llvm; |
81 | |
82 | #define DEBUG_TYPE "codegen" |
83 | |
84 | static cl::opt<unsigned> AlignAllFunctions( |
85 | "align-all-functions" , |
86 | cl::desc("Force the alignment of all functions in log2 format (e.g. 4 " |
87 | "means align on 16B boundaries)." ), |
88 | cl::init(Val: 0), cl::Hidden); |
89 | |
90 | static const char *getPropertyName(MachineFunctionProperties::Property Prop) { |
91 | using P = MachineFunctionProperties::Property; |
92 | |
93 | // clang-format off |
94 | switch(Prop) { |
95 | case P::FailedISel: return "FailedISel" ; |
96 | case P::IsSSA: return "IsSSA" ; |
97 | case P::Legalized: return "Legalized" ; |
98 | case P::NoPHIs: return "NoPHIs" ; |
99 | case P::NoVRegs: return "NoVRegs" ; |
100 | case P::RegBankSelected: return "RegBankSelected" ; |
101 | case P::Selected: return "Selected" ; |
102 | case P::TracksLiveness: return "TracksLiveness" ; |
103 | case P::TiedOpsRewritten: return "TiedOpsRewritten" ; |
104 | case P::FailsVerification: return "FailsVerification" ; |
105 | case P::FailedRegAlloc: return "FailedRegAlloc" ; |
106 | case P::TracksDebugUserValues: return "TracksDebugUserValues" ; |
107 | } |
108 | // clang-format on |
109 | llvm_unreachable("Invalid machine function property" ); |
110 | } |
111 | |
112 | void setUnsafeStackSize(const Function &F, MachineFrameInfo &FrameInfo) { |
113 | if (!F.hasFnAttribute(Kind: Attribute::SafeStack)) |
114 | return; |
115 | |
116 | auto *Existing = |
117 | dyn_cast_or_null<MDTuple>(Val: F.getMetadata(KindID: LLVMContext::MD_annotation)); |
118 | |
119 | if (!Existing || Existing->getNumOperands() != 2) |
120 | return; |
121 | |
122 | auto *MetadataName = "unsafe-stack-size" ; |
123 | if (auto &N = Existing->getOperand(I: 0)) { |
124 | if (N.equalsStr(Str: MetadataName)) { |
125 | if (auto &Op = Existing->getOperand(I: 1)) { |
126 | auto Val = mdconst::extract<ConstantInt>(MD: Op)->getZExtValue(); |
127 | FrameInfo.setUnsafeStackSize(Val); |
128 | } |
129 | } |
130 | } |
131 | } |
132 | |
133 | // Pin the vtable to this file. |
134 | void MachineFunction::Delegate::anchor() {} |
135 | |
136 | void MachineFunctionProperties::print(raw_ostream &OS) const { |
137 | const char *Separator = "" ; |
138 | for (BitVector::size_type I = 0; I < Properties.size(); ++I) { |
139 | if (!Properties[I]) |
140 | continue; |
141 | OS << Separator << getPropertyName(Prop: static_cast<Property>(I)); |
142 | Separator = ", " ; |
143 | } |
144 | } |
145 | |
146 | //===----------------------------------------------------------------------===// |
147 | // MachineFunction implementation |
148 | //===----------------------------------------------------------------------===// |
149 | |
150 | // Out-of-line virtual method. |
151 | MachineFunctionInfo::~MachineFunctionInfo() = default; |
152 | |
153 | void ilist_alloc_traits<MachineBasicBlock>::deleteNode(MachineBasicBlock *MBB) { |
154 | MBB->getParent()->deleteMachineBasicBlock(MBB); |
155 | } |
156 | |
157 | static inline Align getFnStackAlignment(const TargetSubtargetInfo *STI, |
158 | const Function &F) { |
159 | if (auto MA = F.getFnStackAlign()) |
160 | return *MA; |
161 | return STI->getFrameLowering()->getStackAlign(); |
162 | } |
163 | |
164 | MachineFunction::MachineFunction(Function &F, const TargetMachine &Target, |
165 | const TargetSubtargetInfo &STI, MCContext &Ctx, |
166 | unsigned FunctionNum) |
167 | : F(F), Target(Target), STI(&STI), Ctx(Ctx) { |
168 | FunctionNumber = FunctionNum; |
169 | init(); |
170 | } |
171 | |
172 | void MachineFunction::handleInsertion(MachineInstr &MI) { |
173 | if (TheDelegate) |
174 | TheDelegate->MF_HandleInsertion(MI); |
175 | } |
176 | |
177 | void MachineFunction::handleRemoval(MachineInstr &MI) { |
178 | if (TheDelegate) |
179 | TheDelegate->MF_HandleRemoval(MI); |
180 | } |
181 | |
182 | void MachineFunction::handleChangeDesc(MachineInstr &MI, |
183 | const MCInstrDesc &TID) { |
184 | if (TheDelegate) |
185 | TheDelegate->MF_HandleChangeDesc(MI, TID); |
186 | } |
187 | |
188 | void MachineFunction::init() { |
189 | // Assume the function starts in SSA form with correct liveness. |
190 | Properties.setIsSSA(); |
191 | Properties.setTracksLiveness(); |
192 | RegInfo = new (Allocator) MachineRegisterInfo(this); |
193 | |
194 | MFInfo = nullptr; |
195 | |
196 | // We can realign the stack if the target supports it and the user hasn't |
197 | // explicitly asked us not to. |
198 | bool CanRealignSP = STI->getFrameLowering()->isStackRealignable() && |
199 | !F.hasFnAttribute(Kind: "no-realign-stack" ); |
200 | bool ForceRealignSP = F.hasFnAttribute(Kind: Attribute::StackAlignment) || |
201 | F.hasFnAttribute(Kind: "stackrealign" ); |
202 | FrameInfo = new (Allocator) MachineFrameInfo( |
203 | getFnStackAlignment(STI, F), /*StackRealignable=*/CanRealignSP, |
204 | /*ForcedRealign=*/ForceRealignSP && CanRealignSP); |
205 | |
206 | setUnsafeStackSize(F, FrameInfo&: *FrameInfo); |
207 | |
208 | if (F.hasFnAttribute(Kind: Attribute::StackAlignment)) |
209 | FrameInfo->ensureMaxAlignment(Alignment: *F.getFnStackAlign()); |
210 | |
211 | ConstantPool = new (Allocator) MachineConstantPool(getDataLayout()); |
212 | Alignment = STI->getTargetLowering()->getMinFunctionAlignment(); |
213 | |
214 | // FIXME: Shouldn't use pref alignment if explicit alignment is set on F. |
215 | // FIXME: Use Function::hasOptSize(). |
216 | if (!F.hasFnAttribute(Kind: Attribute::OptimizeForSize)) |
217 | Alignment = std::max(a: Alignment, |
218 | b: STI->getTargetLowering()->getPrefFunctionAlignment()); |
219 | |
220 | // -fsanitize=function and -fsanitize=kcfi instrument indirect function calls |
221 | // to load a type hash before the function label. Ensure functions are aligned |
222 | // by a least 4 to avoid unaligned access, which is especially important for |
223 | // -mno-unaligned-access. |
224 | if (F.hasMetadata(KindID: LLVMContext::MD_func_sanitize) || |
225 | F.getMetadata(KindID: LLVMContext::MD_kcfi_type)) |
226 | Alignment = std::max(a: Alignment, b: Align(4)); |
227 | |
228 | if (AlignAllFunctions) |
229 | Alignment = Align(1ULL << AlignAllFunctions); |
230 | |
231 | JumpTableInfo = nullptr; |
232 | |
233 | if (isFuncletEHPersonality(Pers: classifyEHPersonality( |
234 | Pers: F.hasPersonalityFn() ? F.getPersonalityFn() : nullptr))) { |
235 | WinEHInfo = new (Allocator) WinEHFuncInfo(); |
236 | } |
237 | |
238 | if (isScopedEHPersonality(Pers: classifyEHPersonality( |
239 | Pers: F.hasPersonalityFn() ? F.getPersonalityFn() : nullptr))) { |
240 | WasmEHInfo = new (Allocator) WasmEHFuncInfo(); |
241 | } |
242 | |
243 | assert(Target.isCompatibleDataLayout(getDataLayout()) && |
244 | "Can't create a MachineFunction using a Module with a " |
245 | "Target-incompatible DataLayout attached\n" ); |
246 | |
247 | PSVManager = std::make_unique<PseudoSourceValueManager>(args: getTarget()); |
248 | } |
249 | |
250 | void MachineFunction::initTargetMachineFunctionInfo( |
251 | const TargetSubtargetInfo &STI) { |
252 | assert(!MFInfo && "MachineFunctionInfo already set" ); |
253 | MFInfo = Target.createMachineFunctionInfo(Allocator, F, STI: &STI); |
254 | } |
255 | |
256 | MachineFunction::~MachineFunction() { |
257 | clear(); |
258 | } |
259 | |
260 | void MachineFunction::clear() { |
261 | Properties.reset(); |
262 | |
263 | // Clear JumpTableInfo first. Otherwise, every MBB we delete would do a |
264 | // linear search over the jump table entries to find and erase itself. |
265 | if (JumpTableInfo) { |
266 | JumpTableInfo->~MachineJumpTableInfo(); |
267 | Allocator.Deallocate(Ptr: JumpTableInfo); |
268 | JumpTableInfo = nullptr; |
269 | } |
270 | |
271 | // Don't call destructors on MachineInstr and MachineOperand. All of their |
272 | // memory comes from the BumpPtrAllocator which is about to be purged. |
273 | // |
274 | // Do call MachineBasicBlock destructors, it contains std::vectors. |
275 | for (iterator I = begin(), E = end(); I != E; I = BasicBlocks.erase(where: I)) |
276 | I->Insts.clearAndLeakNodesUnsafely(); |
277 | MBBNumbering.clear(); |
278 | |
279 | InstructionRecycler.clear(Allocator); |
280 | OperandRecycler.clear(Allocator); |
281 | BasicBlockRecycler.clear(Allocator); |
282 | CodeViewAnnotations.clear(); |
283 | VariableDbgInfos.clear(); |
284 | if (RegInfo) { |
285 | RegInfo->~MachineRegisterInfo(); |
286 | Allocator.Deallocate(Ptr: RegInfo); |
287 | } |
288 | if (MFInfo) { |
289 | MFInfo->~MachineFunctionInfo(); |
290 | Allocator.Deallocate(Ptr: MFInfo); |
291 | } |
292 | |
293 | FrameInfo->~MachineFrameInfo(); |
294 | Allocator.Deallocate(Ptr: FrameInfo); |
295 | |
296 | ConstantPool->~MachineConstantPool(); |
297 | Allocator.Deallocate(Ptr: ConstantPool); |
298 | |
299 | if (WinEHInfo) { |
300 | WinEHInfo->~WinEHFuncInfo(); |
301 | Allocator.Deallocate(Ptr: WinEHInfo); |
302 | } |
303 | |
304 | if (WasmEHInfo) { |
305 | WasmEHInfo->~WasmEHFuncInfo(); |
306 | Allocator.Deallocate(Ptr: WasmEHInfo); |
307 | } |
308 | } |
309 | |
310 | const DataLayout &MachineFunction::getDataLayout() const { |
311 | return F.getDataLayout(); |
312 | } |
313 | |
314 | /// Get the JumpTableInfo for this function. |
315 | /// If it does not already exist, allocate one. |
316 | MachineJumpTableInfo *MachineFunction:: |
317 | getOrCreateJumpTableInfo(unsigned EntryKind) { |
318 | if (JumpTableInfo) return JumpTableInfo; |
319 | |
320 | JumpTableInfo = new (Allocator) |
321 | MachineJumpTableInfo((MachineJumpTableInfo::JTEntryKind)EntryKind); |
322 | return JumpTableInfo; |
323 | } |
324 | |
325 | DenormalMode MachineFunction::getDenormalMode(const fltSemantics &FPType) const { |
326 | return F.getDenormalMode(FPType); |
327 | } |
328 | |
329 | /// Should we be emitting segmented stack stuff for the function |
330 | bool MachineFunction::shouldSplitStack() const { |
331 | return getFunction().hasFnAttribute(Kind: "split-stack" ); |
332 | } |
333 | |
334 | [[nodiscard]] unsigned |
335 | MachineFunction::addFrameInst(const MCCFIInstruction &Inst) { |
336 | FrameInstructions.push_back(x: Inst); |
337 | return FrameInstructions.size() - 1; |
338 | } |
339 | |
340 | /// This discards all of the MachineBasicBlock numbers and recomputes them. |
341 | /// This guarantees that the MBB numbers are sequential, dense, and match the |
342 | /// ordering of the blocks within the function. If a specific MachineBasicBlock |
343 | /// is specified, only that block and those after it are renumbered. |
344 | void MachineFunction::RenumberBlocks(MachineBasicBlock *MBB) { |
345 | if (empty()) { MBBNumbering.clear(); return; } |
346 | MachineFunction::iterator MBBI, E = end(); |
347 | if (MBB == nullptr) |
348 | MBBI = begin(); |
349 | else |
350 | MBBI = MBB->getIterator(); |
351 | |
352 | // Figure out the block number this should have. |
353 | unsigned BlockNo = 0; |
354 | if (MBBI != begin()) |
355 | BlockNo = std::prev(x: MBBI)->getNumber() + 1; |
356 | |
357 | for (; MBBI != E; ++MBBI, ++BlockNo) { |
358 | if (MBBI->getNumber() != (int)BlockNo) { |
359 | // Remove use of the old number. |
360 | if (MBBI->getNumber() != -1) { |
361 | assert(MBBNumbering[MBBI->getNumber()] == &*MBBI && |
362 | "MBB number mismatch!" ); |
363 | MBBNumbering[MBBI->getNumber()] = nullptr; |
364 | } |
365 | |
366 | // If BlockNo is already taken, set that block's number to -1. |
367 | if (MBBNumbering[BlockNo]) |
368 | MBBNumbering[BlockNo]->setNumber(-1); |
369 | |
370 | MBBNumbering[BlockNo] = &*MBBI; |
371 | MBBI->setNumber(BlockNo); |
372 | } |
373 | } |
374 | |
375 | // Okay, all the blocks are renumbered. If we have compactified the block |
376 | // numbering, shrink MBBNumbering now. |
377 | assert(BlockNo <= MBBNumbering.size() && "Mismatch!" ); |
378 | MBBNumbering.resize(new_size: BlockNo); |
379 | MBBNumberingEpoch++; |
380 | } |
381 | |
382 | int64_t MachineFunction::estimateFunctionSizeInBytes() { |
383 | const TargetInstrInfo &TII = *getSubtarget().getInstrInfo(); |
384 | const Align FunctionAlignment = getAlignment(); |
385 | MachineFunction::iterator MBBI = begin(), E = end(); |
386 | /// Offset - Distance from the beginning of the function to the end |
387 | /// of the basic block. |
388 | int64_t Offset = 0; |
389 | |
390 | for (; MBBI != E; ++MBBI) { |
391 | const Align Alignment = MBBI->getAlignment(); |
392 | int64_t BlockSize = 0; |
393 | |
394 | for (auto &MI : *MBBI) { |
395 | BlockSize += TII.getInstSizeInBytes(MI); |
396 | } |
397 | |
398 | int64_t OffsetBB; |
399 | if (Alignment <= FunctionAlignment) { |
400 | OffsetBB = alignTo(Size: Offset, A: Alignment); |
401 | } else { |
402 | // The alignment of this MBB is larger than the function's alignment, so |
403 | // we can't tell whether or not it will insert nops. Assume that it will. |
404 | OffsetBB = alignTo(Size: Offset, A: Alignment) + Alignment.value() - |
405 | FunctionAlignment.value(); |
406 | } |
407 | Offset = OffsetBB + BlockSize; |
408 | } |
409 | |
410 | return Offset; |
411 | } |
412 | |
413 | /// This method iterates over the basic blocks and assigns their IsBeginSection |
414 | /// and IsEndSection fields. This must be called after MBB layout is finalized |
415 | /// and the SectionID's are assigned to MBBs. |
416 | void MachineFunction::assignBeginEndSections() { |
417 | front().setIsBeginSection(); |
418 | auto CurrentSectionID = front().getSectionID(); |
419 | for (auto MBBI = std::next(x: begin()), E = end(); MBBI != E; ++MBBI) { |
420 | if (MBBI->getSectionID() == CurrentSectionID) |
421 | continue; |
422 | MBBI->setIsBeginSection(); |
423 | std::prev(x: MBBI)->setIsEndSection(); |
424 | CurrentSectionID = MBBI->getSectionID(); |
425 | } |
426 | back().setIsEndSection(); |
427 | } |
428 | |
429 | /// Allocate a new MachineInstr. Use this instead of `new MachineInstr'. |
430 | MachineInstr *MachineFunction::CreateMachineInstr(const MCInstrDesc &MCID, |
431 | DebugLoc DL, |
432 | bool NoImplicit) { |
433 | return new (InstructionRecycler.Allocate<MachineInstr>(Allocator)) |
434 | MachineInstr(*this, MCID, std::move(DL), NoImplicit); |
435 | } |
436 | |
437 | /// Create a new MachineInstr which is a copy of the 'Orig' instruction, |
438 | /// identical in all ways except the instruction has no parent, prev, or next. |
439 | MachineInstr * |
440 | MachineFunction::CloneMachineInstr(const MachineInstr *Orig) { |
441 | return new (InstructionRecycler.Allocate<MachineInstr>(Allocator)) |
442 | MachineInstr(*this, *Orig); |
443 | } |
444 | |
445 | MachineInstr &MachineFunction::cloneMachineInstrBundle( |
446 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
447 | const MachineInstr &Orig) { |
448 | MachineInstr *FirstClone = nullptr; |
449 | MachineBasicBlock::const_instr_iterator I = Orig.getIterator(); |
450 | while (true) { |
451 | MachineInstr *Cloned = CloneMachineInstr(Orig: &*I); |
452 | MBB.insert(I: InsertBefore, MI: Cloned); |
453 | if (FirstClone == nullptr) { |
454 | FirstClone = Cloned; |
455 | } else { |
456 | Cloned->bundleWithPred(); |
457 | } |
458 | |
459 | if (!I->isBundledWithSucc()) |
460 | break; |
461 | ++I; |
462 | } |
463 | // Copy over call info to the cloned instruction if needed. If Orig is in |
464 | // a bundle, copyAdditionalCallInfo takes care of finding the call instruction |
465 | // in the bundle. |
466 | if (Orig.shouldUpdateAdditionalCallInfo()) |
467 | copyAdditionalCallInfo(Old: &Orig, New: FirstClone); |
468 | return *FirstClone; |
469 | } |
470 | |
471 | /// Delete the given MachineInstr. |
472 | /// |
473 | /// This function also serves as the MachineInstr destructor - the real |
474 | /// ~MachineInstr() destructor must be empty. |
475 | void MachineFunction::deleteMachineInstr(MachineInstr *MI) { |
476 | // Verify that a call site info is at valid state. This assertion should |
477 | // be triggered during the implementation of support for the |
478 | // call site info of a new architecture. If the assertion is triggered, |
479 | // back trace will tell where to insert a call to updateCallSiteInfo(). |
480 | assert((!MI->isCandidateForAdditionalCallInfo() || |
481 | !CallSitesInfo.contains(MI)) && |
482 | "Call site info was not updated!" ); |
483 | // Verify that the "called globals" info is in a valid state. |
484 | assert((!MI->isCandidateForAdditionalCallInfo() || |
485 | !CalledGlobalsInfo.contains(MI)) && |
486 | "Called globals info was not updated!" ); |
487 | // Strip it for parts. The operand array and the MI object itself are |
488 | // independently recyclable. |
489 | if (MI->Operands) |
490 | deallocateOperandArray(Cap: MI->CapOperands, Array: MI->Operands); |
491 | // Don't call ~MachineInstr() which must be trivial anyway because |
492 | // ~MachineFunction drops whole lists of MachineInstrs wihout calling their |
493 | // destructors. |
494 | InstructionRecycler.Deallocate(Allocator, Element: MI); |
495 | } |
496 | |
497 | /// Allocate a new MachineBasicBlock. Use this instead of |
498 | /// `new MachineBasicBlock'. |
499 | MachineBasicBlock * |
500 | MachineFunction::CreateMachineBasicBlock(const BasicBlock *BB, |
501 | std::optional<UniqueBBID> BBID) { |
502 | MachineBasicBlock *MBB = |
503 | new (BasicBlockRecycler.Allocate<MachineBasicBlock>(Allocator)) |
504 | MachineBasicBlock(*this, BB); |
505 | // Set BBID for `-basic-block-sections=list` and `-basic-block-address-map` to |
506 | // allow robust mapping of profiles to basic blocks. |
507 | if (Target.Options.BBAddrMap || |
508 | Target.getBBSectionsType() == BasicBlockSection::List) |
509 | MBB->setBBID(BBID.has_value() ? *BBID : UniqueBBID{.BaseID: NextBBID++, .CloneID: 0}); |
510 | return MBB; |
511 | } |
512 | |
513 | /// Delete the given MachineBasicBlock. |
514 | void MachineFunction::deleteMachineBasicBlock(MachineBasicBlock *MBB) { |
515 | assert(MBB->getParent() == this && "MBB parent mismatch!" ); |
516 | // Clean up any references to MBB in jump tables before deleting it. |
517 | if (JumpTableInfo) |
518 | JumpTableInfo->RemoveMBBFromJumpTables(MBB); |
519 | MBB->~MachineBasicBlock(); |
520 | BasicBlockRecycler.Deallocate(Allocator, Element: MBB); |
521 | } |
522 | |
523 | MachineMemOperand *MachineFunction::getMachineMemOperand( |
524 | MachinePointerInfo PtrInfo, MachineMemOperand::Flags F, LocationSize Size, |
525 | Align BaseAlignment, const AAMDNodes &AAInfo, const MDNode *Ranges, |
526 | SyncScope::ID SSID, AtomicOrdering Ordering, |
527 | AtomicOrdering FailureOrdering) { |
528 | assert((!Size.hasValue() || |
529 | Size.getValue().getKnownMinValue() != ~UINT64_C(0)) && |
530 | "Unexpected an unknown size to be represented using " |
531 | "LocationSize::beforeOrAfter()" ); |
532 | return new (Allocator) |
533 | MachineMemOperand(PtrInfo, F, Size, BaseAlignment, AAInfo, Ranges, SSID, |
534 | Ordering, FailureOrdering); |
535 | } |
536 | |
537 | MachineMemOperand *MachineFunction::getMachineMemOperand( |
538 | MachinePointerInfo PtrInfo, MachineMemOperand::Flags f, LLT MemTy, |
539 | Align base_alignment, const AAMDNodes &AAInfo, const MDNode *Ranges, |
540 | SyncScope::ID SSID, AtomicOrdering Ordering, |
541 | AtomicOrdering FailureOrdering) { |
542 | return new (Allocator) |
543 | MachineMemOperand(PtrInfo, f, MemTy, base_alignment, AAInfo, Ranges, SSID, |
544 | Ordering, FailureOrdering); |
545 | } |
546 | |
547 | MachineMemOperand * |
548 | MachineFunction::getMachineMemOperand(const MachineMemOperand *MMO, |
549 | const MachinePointerInfo &PtrInfo, |
550 | LocationSize Size) { |
551 | assert((!Size.hasValue() || |
552 | Size.getValue().getKnownMinValue() != ~UINT64_C(0)) && |
553 | "Unexpected an unknown size to be represented using " |
554 | "LocationSize::beforeOrAfter()" ); |
555 | return new (Allocator) |
556 | MachineMemOperand(PtrInfo, MMO->getFlags(), Size, MMO->getBaseAlign(), |
557 | AAMDNodes(), nullptr, MMO->getSyncScopeID(), |
558 | MMO->getSuccessOrdering(), MMO->getFailureOrdering()); |
559 | } |
560 | |
561 | MachineMemOperand *MachineFunction::getMachineMemOperand( |
562 | const MachineMemOperand *MMO, const MachinePointerInfo &PtrInfo, LLT Ty) { |
563 | return new (Allocator) |
564 | MachineMemOperand(PtrInfo, MMO->getFlags(), Ty, MMO->getBaseAlign(), |
565 | AAMDNodes(), nullptr, MMO->getSyncScopeID(), |
566 | MMO->getSuccessOrdering(), MMO->getFailureOrdering()); |
567 | } |
568 | |
569 | MachineMemOperand * |
570 | MachineFunction::getMachineMemOperand(const MachineMemOperand *MMO, |
571 | int64_t Offset, LLT Ty) { |
572 | const MachinePointerInfo &PtrInfo = MMO->getPointerInfo(); |
573 | |
574 | // If there is no pointer value, the offset isn't tracked so we need to adjust |
575 | // the base alignment. |
576 | Align Alignment = PtrInfo.V.isNull() |
577 | ? commonAlignment(A: MMO->getBaseAlign(), Offset) |
578 | : MMO->getBaseAlign(); |
579 | |
580 | // Do not preserve ranges, since we don't necessarily know what the high bits |
581 | // are anymore. |
582 | return new (Allocator) MachineMemOperand( |
583 | PtrInfo.getWithOffset(O: Offset), MMO->getFlags(), Ty, Alignment, |
584 | MMO->getAAInfo(), nullptr, MMO->getSyncScopeID(), |
585 | MMO->getSuccessOrdering(), MMO->getFailureOrdering()); |
586 | } |
587 | |
588 | MachineMemOperand * |
589 | MachineFunction::getMachineMemOperand(const MachineMemOperand *MMO, |
590 | const AAMDNodes &AAInfo) { |
591 | MachinePointerInfo MPI = MMO->getValue() ? |
592 | MachinePointerInfo(MMO->getValue(), MMO->getOffset()) : |
593 | MachinePointerInfo(MMO->getPseudoValue(), MMO->getOffset()); |
594 | |
595 | return new (Allocator) MachineMemOperand( |
596 | MPI, MMO->getFlags(), MMO->getSize(), MMO->getBaseAlign(), AAInfo, |
597 | MMO->getRanges(), MMO->getSyncScopeID(), MMO->getSuccessOrdering(), |
598 | MMO->getFailureOrdering()); |
599 | } |
600 | |
601 | MachineMemOperand * |
602 | MachineFunction::getMachineMemOperand(const MachineMemOperand *MMO, |
603 | MachineMemOperand::Flags Flags) { |
604 | return new (Allocator) MachineMemOperand( |
605 | MMO->getPointerInfo(), Flags, MMO->getSize(), MMO->getBaseAlign(), |
606 | MMO->getAAInfo(), MMO->getRanges(), MMO->getSyncScopeID(), |
607 | MMO->getSuccessOrdering(), MMO->getFailureOrdering()); |
608 | } |
609 | |
610 | MachineInstr::ExtraInfo *MachineFunction::createMIExtraInfo( |
611 | ArrayRef<MachineMemOperand *> MMOs, MCSymbol *PreInstrSymbol, |
612 | MCSymbol *PostInstrSymbol, MDNode *HeapAllocMarker, MDNode *PCSections, |
613 | uint32_t CFIType, MDNode *MMRAs) { |
614 | return MachineInstr::ExtraInfo::create(Allocator, MMOs, PreInstrSymbol, |
615 | PostInstrSymbol, HeapAllocMarker, |
616 | PCSections, CFIType, MMRAs); |
617 | } |
618 | |
619 | const char *MachineFunction::createExternalSymbolName(StringRef Name) { |
620 | char *Dest = Allocator.Allocate<char>(Num: Name.size() + 1); |
621 | llvm::copy(Range&: Name, Out: Dest); |
622 | Dest[Name.size()] = 0; |
623 | return Dest; |
624 | } |
625 | |
626 | uint32_t *MachineFunction::allocateRegMask() { |
627 | unsigned NumRegs = getSubtarget().getRegisterInfo()->getNumRegs(); |
628 | unsigned Size = MachineOperand::getRegMaskSize(NumRegs); |
629 | uint32_t *Mask = Allocator.Allocate<uint32_t>(Num: Size); |
630 | memset(s: Mask, c: 0, n: Size * sizeof(Mask[0])); |
631 | return Mask; |
632 | } |
633 | |
634 | ArrayRef<int> MachineFunction::allocateShuffleMask(ArrayRef<int> Mask) { |
635 | int* AllocMask = Allocator.Allocate<int>(Num: Mask.size()); |
636 | copy(Range&: Mask, Out: AllocMask); |
637 | return {AllocMask, Mask.size()}; |
638 | } |
639 | |
640 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
641 | LLVM_DUMP_METHOD void MachineFunction::dump() const { |
642 | print(dbgs()); |
643 | } |
644 | #endif |
645 | |
646 | StringRef MachineFunction::getName() const { |
647 | return getFunction().getName(); |
648 | } |
649 | |
650 | void MachineFunction::print(raw_ostream &OS, const SlotIndexes *Indexes) const { |
651 | OS << "# Machine code for function " << getName() << ": " ; |
652 | getProperties().print(OS); |
653 | OS << '\n'; |
654 | |
655 | // Print Frame Information |
656 | FrameInfo->print(MF: *this, OS); |
657 | |
658 | // Print JumpTable Information |
659 | if (JumpTableInfo) |
660 | JumpTableInfo->print(OS); |
661 | |
662 | // Print Constant Pool |
663 | ConstantPool->print(OS); |
664 | |
665 | const TargetRegisterInfo *TRI = getSubtarget().getRegisterInfo(); |
666 | |
667 | if (RegInfo && !RegInfo->livein_empty()) { |
668 | OS << "Function Live Ins: " ; |
669 | for (MachineRegisterInfo::livein_iterator |
670 | I = RegInfo->livein_begin(), E = RegInfo->livein_end(); I != E; ++I) { |
671 | OS << printReg(Reg: I->first, TRI); |
672 | if (I->second) |
673 | OS << " in " << printReg(Reg: I->second, TRI); |
674 | if (std::next(x: I) != E) |
675 | OS << ", " ; |
676 | } |
677 | OS << '\n'; |
678 | } |
679 | |
680 | ModuleSlotTracker MST(getFunction().getParent()); |
681 | MST.incorporateFunction(F: getFunction()); |
682 | for (const auto &BB : *this) { |
683 | OS << '\n'; |
684 | // If we print the whole function, print it at its most verbose level. |
685 | BB.print(OS, MST, Indexes, /*IsStandalone=*/true); |
686 | } |
687 | |
688 | OS << "\n# End machine code for function " << getName() << ".\n\n" ; |
689 | } |
690 | |
691 | /// True if this function needs frame moves for debug or exceptions. |
692 | bool MachineFunction::needsFrameMoves() const { |
693 | // TODO: Ideally, what we'd like is to have a switch that allows emitting |
694 | // synchronous (precise at call-sites only) CFA into .eh_frame. However, even |
695 | // under this switch, we'd like .debug_frame to be precise when using -g. At |
696 | // this moment, there's no way to specify that some CFI directives go into |
697 | // .eh_frame only, while others go into .debug_frame only. |
698 | return getTarget().Options.ForceDwarfFrameSection || |
699 | F.needsUnwindTableEntry() || |
700 | !F.getParent()->debug_compile_units().empty(); |
701 | } |
702 | |
703 | namespace llvm { |
704 | |
705 | template<> |
706 | struct DOTGraphTraits<const MachineFunction*> : public DefaultDOTGraphTraits { |
707 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} |
708 | |
709 | static std::string getGraphName(const MachineFunction *F) { |
710 | return ("CFG for '" + F->getName() + "' function" ).str(); |
711 | } |
712 | |
713 | std::string getNodeLabel(const MachineBasicBlock *Node, |
714 | const MachineFunction *Graph) { |
715 | std::string OutStr; |
716 | { |
717 | raw_string_ostream OSS(OutStr); |
718 | |
719 | if (isSimple()) { |
720 | OSS << printMBBReference(MBB: *Node); |
721 | if (const BasicBlock *BB = Node->getBasicBlock()) |
722 | OSS << ": " << BB->getName(); |
723 | } else |
724 | Node->print(OS&: OSS); |
725 | } |
726 | |
727 | if (OutStr[0] == '\n') OutStr.erase(position: OutStr.begin()); |
728 | |
729 | // Process string output to make it nicer... |
730 | for (unsigned i = 0; i != OutStr.length(); ++i) |
731 | if (OutStr[i] == '\n') { // Left justify |
732 | OutStr[i] = '\\'; |
733 | OutStr.insert(p: OutStr.begin()+i+1, c: 'l'); |
734 | } |
735 | return OutStr; |
736 | } |
737 | }; |
738 | |
739 | } // end namespace llvm |
740 | |
741 | void MachineFunction::viewCFG() const |
742 | { |
743 | #ifndef NDEBUG |
744 | ViewGraph(this, "mf" + getName()); |
745 | #else |
746 | errs() << "MachineFunction::viewCFG is only available in debug builds on " |
747 | << "systems with Graphviz or gv!\n" ; |
748 | #endif // NDEBUG |
749 | } |
750 | |
751 | void MachineFunction::viewCFGOnly() const |
752 | { |
753 | #ifndef NDEBUG |
754 | ViewGraph(this, "mf" + getName(), true); |
755 | #else |
756 | errs() << "MachineFunction::viewCFGOnly is only available in debug builds on " |
757 | << "systems with Graphviz or gv!\n" ; |
758 | #endif // NDEBUG |
759 | } |
760 | |
761 | /// Add the specified physical register as a live-in value and |
762 | /// create a corresponding virtual register for it. |
763 | Register MachineFunction::addLiveIn(MCRegister PReg, |
764 | const TargetRegisterClass *RC) { |
765 | MachineRegisterInfo &MRI = getRegInfo(); |
766 | Register VReg = MRI.getLiveInVirtReg(PReg); |
767 | if (VReg) { |
768 | const TargetRegisterClass *VRegRC = MRI.getRegClass(Reg: VReg); |
769 | (void)VRegRC; |
770 | // A physical register can be added several times. |
771 | // Between two calls, the register class of the related virtual register |
772 | // may have been constrained to match some operation constraints. |
773 | // In that case, check that the current register class includes the |
774 | // physical register and is a sub class of the specified RC. |
775 | assert((VRegRC == RC || (VRegRC->contains(PReg) && |
776 | RC->hasSubClassEq(VRegRC))) && |
777 | "Register class mismatch!" ); |
778 | return VReg; |
779 | } |
780 | VReg = MRI.createVirtualRegister(RegClass: RC); |
781 | MRI.addLiveIn(Reg: PReg, vreg: VReg); |
782 | return VReg; |
783 | } |
784 | |
785 | /// Return the MCSymbol for the specified non-empty jump table. |
786 | /// If isLinkerPrivate is specified, an 'l' label is returned, otherwise a |
787 | /// normal 'L' label is returned. |
788 | MCSymbol *MachineFunction::getJTISymbol(unsigned JTI, MCContext &Ctx, |
789 | bool isLinkerPrivate) const { |
790 | const DataLayout &DL = getDataLayout(); |
791 | assert(JumpTableInfo && "No jump tables" ); |
792 | assert(JTI < JumpTableInfo->getJumpTables().size() && "Invalid JTI!" ); |
793 | |
794 | StringRef Prefix = isLinkerPrivate ? DL.getLinkerPrivateGlobalPrefix() |
795 | : DL.getPrivateGlobalPrefix(); |
796 | SmallString<60> Name; |
797 | raw_svector_ostream(Name) |
798 | << Prefix << "JTI" << getFunctionNumber() << '_' << JTI; |
799 | return Ctx.getOrCreateSymbol(Name); |
800 | } |
801 | |
802 | /// Return a function-local symbol to represent the PIC base. |
803 | MCSymbol *MachineFunction::getPICBaseSymbol() const { |
804 | const DataLayout &DL = getDataLayout(); |
805 | return Ctx.getOrCreateSymbol(Name: Twine(DL.getPrivateGlobalPrefix()) + |
806 | Twine(getFunctionNumber()) + "$pb" ); |
807 | } |
808 | |
809 | /// \name Exception Handling |
810 | /// \{ |
811 | |
812 | LandingPadInfo & |
813 | MachineFunction::getOrCreateLandingPadInfo(MachineBasicBlock *LandingPad) { |
814 | unsigned N = LandingPads.size(); |
815 | for (unsigned i = 0; i < N; ++i) { |
816 | LandingPadInfo &LP = LandingPads[i]; |
817 | if (LP.LandingPadBlock == LandingPad) |
818 | return LP; |
819 | } |
820 | |
821 | LandingPads.push_back(x: LandingPadInfo(LandingPad)); |
822 | return LandingPads[N]; |
823 | } |
824 | |
825 | void MachineFunction::addInvoke(MachineBasicBlock *LandingPad, |
826 | MCSymbol *BeginLabel, MCSymbol *EndLabel) { |
827 | LandingPadInfo &LP = getOrCreateLandingPadInfo(LandingPad); |
828 | LP.BeginLabels.push_back(Elt: BeginLabel); |
829 | LP.EndLabels.push_back(Elt: EndLabel); |
830 | } |
831 | |
832 | MCSymbol *MachineFunction::addLandingPad(MachineBasicBlock *LandingPad) { |
833 | MCSymbol *LandingPadLabel = Ctx.createTempSymbol(); |
834 | LandingPadInfo &LP = getOrCreateLandingPadInfo(LandingPad); |
835 | LP.LandingPadLabel = LandingPadLabel; |
836 | |
837 | BasicBlock::const_iterator FirstI = |
838 | LandingPad->getBasicBlock()->getFirstNonPHIIt(); |
839 | if (const auto *LPI = dyn_cast<LandingPadInst>(Val&: FirstI)) { |
840 | // If there's no typeid list specified, then "cleanup" is implicit. |
841 | // Otherwise, id 0 is reserved for the cleanup action. |
842 | if (LPI->isCleanup() && LPI->getNumClauses() != 0) |
843 | LP.TypeIds.push_back(x: 0); |
844 | |
845 | // FIXME: New EH - Add the clauses in reverse order. This isn't 100% |
846 | // correct, but we need to do it this way because of how the DWARF EH |
847 | // emitter processes the clauses. |
848 | for (unsigned I = LPI->getNumClauses(); I != 0; --I) { |
849 | Value *Val = LPI->getClause(Idx: I - 1); |
850 | if (LPI->isCatch(Idx: I - 1)) { |
851 | LP.TypeIds.push_back( |
852 | x: getTypeIDFor(TI: dyn_cast<GlobalValue>(Val: Val->stripPointerCasts()))); |
853 | } else { |
854 | // Add filters in a list. |
855 | auto *CVal = cast<Constant>(Val); |
856 | SmallVector<unsigned, 4> FilterList; |
857 | for (const Use &U : CVal->operands()) |
858 | FilterList.push_back( |
859 | Elt: getTypeIDFor(TI: cast<GlobalValue>(Val: U->stripPointerCasts()))); |
860 | |
861 | LP.TypeIds.push_back(x: getFilterIDFor(TyIds: FilterList)); |
862 | } |
863 | } |
864 | |
865 | } else if (const auto *CPI = dyn_cast<CatchPadInst>(Val&: FirstI)) { |
866 | for (unsigned I = CPI->arg_size(); I != 0; --I) { |
867 | auto *TypeInfo = |
868 | dyn_cast<GlobalValue>(Val: CPI->getArgOperand(i: I - 1)->stripPointerCasts()); |
869 | LP.TypeIds.push_back(x: getTypeIDFor(TI: TypeInfo)); |
870 | } |
871 | |
872 | } else { |
873 | assert(isa<CleanupPadInst>(FirstI) && "Invalid landingpad!" ); |
874 | } |
875 | |
876 | return LandingPadLabel; |
877 | } |
878 | |
879 | void MachineFunction::setCallSiteLandingPad(MCSymbol *Sym, |
880 | ArrayRef<unsigned> Sites) { |
881 | LPadToCallSiteMap[Sym].append(in_start: Sites.begin(), in_end: Sites.end()); |
882 | } |
883 | |
884 | unsigned MachineFunction::getTypeIDFor(const GlobalValue *TI) { |
885 | for (unsigned i = 0, N = TypeInfos.size(); i != N; ++i) |
886 | if (TypeInfos[i] == TI) return i + 1; |
887 | |
888 | TypeInfos.push_back(x: TI); |
889 | return TypeInfos.size(); |
890 | } |
891 | |
892 | int MachineFunction::getFilterIDFor(ArrayRef<unsigned> TyIds) { |
893 | // If the new filter coincides with the tail of an existing filter, then |
894 | // re-use the existing filter. Folding filters more than this requires |
895 | // re-ordering filters and/or their elements - probably not worth it. |
896 | for (unsigned i : FilterEnds) { |
897 | unsigned j = TyIds.size(); |
898 | |
899 | while (i && j) |
900 | if (FilterIds[--i] != TyIds[--j]) |
901 | goto try_next; |
902 | |
903 | if (!j) |
904 | // The new filter coincides with range [i, end) of the existing filter. |
905 | return -(1 + i); |
906 | |
907 | try_next:; |
908 | } |
909 | |
910 | // Add the new filter. |
911 | int FilterID = -(1 + FilterIds.size()); |
912 | FilterIds.reserve(n: FilterIds.size() + TyIds.size() + 1); |
913 | llvm::append_range(C&: FilterIds, R&: TyIds); |
914 | FilterEnds.push_back(x: FilterIds.size()); |
915 | FilterIds.push_back(x: 0); // terminator |
916 | return FilterID; |
917 | } |
918 | |
919 | MachineFunction::CallSiteInfoMap::iterator |
920 | MachineFunction::getCallSiteInfo(const MachineInstr *MI) { |
921 | assert(MI->isCandidateForAdditionalCallInfo() && |
922 | "Call site info refers only to call (MI) candidates" ); |
923 | |
924 | if (!Target.Options.EmitCallSiteInfo) |
925 | return CallSitesInfo.end(); |
926 | return CallSitesInfo.find(Val: MI); |
927 | } |
928 | |
929 | /// Return the call machine instruction or find a call within bundle. |
930 | static const MachineInstr *getCallInstr(const MachineInstr *MI) { |
931 | if (!MI->isBundle()) |
932 | return MI; |
933 | |
934 | for (const auto &BMI : make_range(x: getBundleStart(I: MI->getIterator()), |
935 | y: getBundleEnd(I: MI->getIterator()))) |
936 | if (BMI.isCandidateForAdditionalCallInfo()) |
937 | return &BMI; |
938 | |
939 | llvm_unreachable("Unexpected bundle without a call site candidate" ); |
940 | } |
941 | |
942 | void MachineFunction::eraseAdditionalCallInfo(const MachineInstr *MI) { |
943 | assert(MI->shouldUpdateAdditionalCallInfo() && |
944 | "Call info refers only to call (MI) candidates or " |
945 | "candidates inside bundles" ); |
946 | |
947 | const MachineInstr *CallMI = getCallInstr(MI); |
948 | |
949 | CallSiteInfoMap::iterator CSIt = getCallSiteInfo(MI: CallMI); |
950 | if (CSIt != CallSitesInfo.end()) |
951 | CallSitesInfo.erase(I: CSIt); |
952 | |
953 | CalledGlobalsInfo.erase(Val: CallMI); |
954 | } |
955 | |
956 | void MachineFunction::copyAdditionalCallInfo(const MachineInstr *Old, |
957 | const MachineInstr *New) { |
958 | assert(Old->shouldUpdateAdditionalCallInfo() && |
959 | "Call info refers only to call (MI) candidates or " |
960 | "candidates inside bundles" ); |
961 | |
962 | if (!New->isCandidateForAdditionalCallInfo()) |
963 | return eraseAdditionalCallInfo(MI: Old); |
964 | |
965 | const MachineInstr *OldCallMI = getCallInstr(MI: Old); |
966 | CallSiteInfoMap::iterator CSIt = getCallSiteInfo(MI: OldCallMI); |
967 | if (CSIt != CallSitesInfo.end()) { |
968 | CallSiteInfo CSInfo = CSIt->second; |
969 | CallSitesInfo[New] = std::move(CSInfo); |
970 | } |
971 | |
972 | CalledGlobalsMap::iterator CGIt = CalledGlobalsInfo.find(Val: OldCallMI); |
973 | if (CGIt != CalledGlobalsInfo.end()) { |
974 | CalledGlobalInfo CGInfo = CGIt->second; |
975 | CalledGlobalsInfo[New] = std::move(CGInfo); |
976 | } |
977 | } |
978 | |
979 | void MachineFunction::moveAdditionalCallInfo(const MachineInstr *Old, |
980 | const MachineInstr *New) { |
981 | assert(Old->shouldUpdateAdditionalCallInfo() && |
982 | "Call info refers only to call (MI) candidates or " |
983 | "candidates inside bundles" ); |
984 | |
985 | if (!New->isCandidateForAdditionalCallInfo()) |
986 | return eraseAdditionalCallInfo(MI: Old); |
987 | |
988 | const MachineInstr *OldCallMI = getCallInstr(MI: Old); |
989 | CallSiteInfoMap::iterator CSIt = getCallSiteInfo(MI: OldCallMI); |
990 | if (CSIt != CallSitesInfo.end()) { |
991 | CallSiteInfo CSInfo = std::move(CSIt->second); |
992 | CallSitesInfo.erase(I: CSIt); |
993 | CallSitesInfo[New] = std::move(CSInfo); |
994 | } |
995 | |
996 | CalledGlobalsMap::iterator CGIt = CalledGlobalsInfo.find(Val: OldCallMI); |
997 | if (CGIt != CalledGlobalsInfo.end()) { |
998 | CalledGlobalInfo CGInfo = std::move(CGIt->second); |
999 | CalledGlobalsInfo.erase(I: CGIt); |
1000 | CalledGlobalsInfo[New] = std::move(CGInfo); |
1001 | } |
1002 | } |
1003 | |
1004 | void MachineFunction::setDebugInstrNumberingCount(unsigned Num) { |
1005 | DebugInstrNumberingCount = Num; |
1006 | } |
1007 | |
1008 | void MachineFunction::makeDebugValueSubstitution(DebugInstrOperandPair A, |
1009 | DebugInstrOperandPair B, |
1010 | unsigned Subreg) { |
1011 | // Catch any accidental self-loops. |
1012 | assert(A.first != B.first); |
1013 | // Don't allow any substitutions _from_ the memory operand number. |
1014 | assert(A.second != DebugOperandMemNumber); |
1015 | |
1016 | DebugValueSubstitutions.push_back(Elt: {A, B, Subreg}); |
1017 | } |
1018 | |
1019 | void MachineFunction::substituteDebugValuesForInst(const MachineInstr &Old, |
1020 | MachineInstr &New, |
1021 | unsigned MaxOperand) { |
1022 | // If the Old instruction wasn't tracked at all, there is no work to do. |
1023 | unsigned OldInstrNum = Old.peekDebugInstrNum(); |
1024 | if (!OldInstrNum) |
1025 | return; |
1026 | |
1027 | // Iterate over all operands looking for defs to create substitutions for. |
1028 | // Avoid creating new instr numbers unless we create a new substitution. |
1029 | // While this has no functional effect, it risks confusing someone reading |
1030 | // MIR output. |
1031 | // Examine all the operands, or the first N specified by the caller. |
1032 | MaxOperand = std::min(a: MaxOperand, b: Old.getNumOperands()); |
1033 | for (unsigned int I = 0; I < MaxOperand; ++I) { |
1034 | const auto &OldMO = Old.getOperand(i: I); |
1035 | auto &NewMO = New.getOperand(i: I); |
1036 | (void)NewMO; |
1037 | |
1038 | if (!OldMO.isReg() || !OldMO.isDef()) |
1039 | continue; |
1040 | assert(NewMO.isDef()); |
1041 | |
1042 | unsigned NewInstrNum = New.getDebugInstrNum(); |
1043 | makeDebugValueSubstitution(A: std::make_pair(x&: OldInstrNum, y&: I), |
1044 | B: std::make_pair(x&: NewInstrNum, y&: I)); |
1045 | } |
1046 | } |
1047 | |
1048 | auto MachineFunction::salvageCopySSA( |
1049 | MachineInstr &MI, DenseMap<Register, DebugInstrOperandPair> &DbgPHICache) |
1050 | -> DebugInstrOperandPair { |
1051 | const TargetInstrInfo &TII = *getSubtarget().getInstrInfo(); |
1052 | |
1053 | // Check whether this copy-like instruction has already been salvaged into |
1054 | // an operand pair. |
1055 | Register Dest; |
1056 | if (auto CopyDstSrc = TII.isCopyLikeInstr(MI)) { |
1057 | Dest = CopyDstSrc->Destination->getReg(); |
1058 | } else { |
1059 | assert(MI.isSubregToReg()); |
1060 | Dest = MI.getOperand(i: 0).getReg(); |
1061 | } |
1062 | |
1063 | auto CacheIt = DbgPHICache.find(Val: Dest); |
1064 | if (CacheIt != DbgPHICache.end()) |
1065 | return CacheIt->second; |
1066 | |
1067 | // Calculate the instruction number to use, or install a DBG_PHI. |
1068 | auto OperandPair = salvageCopySSAImpl(MI); |
1069 | DbgPHICache.insert(KV: {Dest, OperandPair}); |
1070 | return OperandPair; |
1071 | } |
1072 | |
1073 | auto MachineFunction::salvageCopySSAImpl(MachineInstr &MI) |
1074 | -> DebugInstrOperandPair { |
1075 | MachineRegisterInfo &MRI = getRegInfo(); |
1076 | const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo(); |
1077 | const TargetInstrInfo &TII = *getSubtarget().getInstrInfo(); |
1078 | |
1079 | // Chase the value read by a copy-like instruction back to the instruction |
1080 | // that ultimately _defines_ that value. This may pass: |
1081 | // * Through multiple intermediate copies, including subregister moves / |
1082 | // copies, |
1083 | // * Copies from physical registers that must then be traced back to the |
1084 | // defining instruction, |
1085 | // * Or, physical registers may be live-in to (only) the entry block, which |
1086 | // requires a DBG_PHI to be created. |
1087 | // We can pursue this problem in that order: trace back through copies, |
1088 | // optionally through a physical register, to a defining instruction. We |
1089 | // should never move from physreg to vreg. As we're still in SSA form, no need |
1090 | // to worry about partial definitions of registers. |
1091 | |
1092 | // Helper lambda to interpret a copy-like instruction. Takes instruction, |
1093 | // returns the register read and any subregister identifying which part is |
1094 | // read. |
1095 | auto GetRegAndSubreg = |
1096 | [&](const MachineInstr &Cpy) -> std::pair<Register, unsigned> { |
1097 | Register NewReg, OldReg; |
1098 | unsigned SubReg; |
1099 | if (Cpy.isCopy()) { |
1100 | OldReg = Cpy.getOperand(i: 0).getReg(); |
1101 | NewReg = Cpy.getOperand(i: 1).getReg(); |
1102 | SubReg = Cpy.getOperand(i: 1).getSubReg(); |
1103 | } else if (Cpy.isSubregToReg()) { |
1104 | OldReg = Cpy.getOperand(i: 0).getReg(); |
1105 | NewReg = Cpy.getOperand(i: 2).getReg(); |
1106 | SubReg = Cpy.getOperand(i: 3).getImm(); |
1107 | } else { |
1108 | auto CopyDetails = *TII.isCopyInstr(MI: Cpy); |
1109 | const MachineOperand &Src = *CopyDetails.Source; |
1110 | const MachineOperand &Dest = *CopyDetails.Destination; |
1111 | OldReg = Dest.getReg(); |
1112 | NewReg = Src.getReg(); |
1113 | SubReg = Src.getSubReg(); |
1114 | } |
1115 | |
1116 | return {NewReg, SubReg}; |
1117 | }; |
1118 | |
1119 | // First seek either the defining instruction, or a copy from a physreg. |
1120 | // During search, the current state is the current copy instruction, and which |
1121 | // register we've read. Accumulate qualifying subregisters into SubregsSeen; |
1122 | // deal with those later. |
1123 | auto State = GetRegAndSubreg(MI); |
1124 | auto CurInst = MI.getIterator(); |
1125 | SmallVector<unsigned, 4> SubregsSeen; |
1126 | while (true) { |
1127 | // If we've found a copy from a physreg, first portion of search is over. |
1128 | if (!State.first.isVirtual()) |
1129 | break; |
1130 | |
1131 | // Record any subregister qualifier. |
1132 | if (State.second) |
1133 | SubregsSeen.push_back(Elt: State.second); |
1134 | |
1135 | assert(MRI.hasOneDef(State.first)); |
1136 | MachineInstr &Inst = *MRI.def_begin(RegNo: State.first)->getParent(); |
1137 | CurInst = Inst.getIterator(); |
1138 | |
1139 | // Any non-copy instruction is the defining instruction we're seeking. |
1140 | if (!Inst.isCopyLike() && !TII.isCopyLikeInstr(MI: Inst)) |
1141 | break; |
1142 | State = GetRegAndSubreg(Inst); |
1143 | }; |
1144 | |
1145 | // Helper lambda to apply additional subregister substitutions to a known |
1146 | // instruction/operand pair. Adds new (fake) substitutions so that we can |
1147 | // record the subregister. FIXME: this isn't very space efficient if multiple |
1148 | // values are tracked back through the same copies; cache something later. |
1149 | auto ApplySubregisters = |
1150 | [&](DebugInstrOperandPair P) -> DebugInstrOperandPair { |
1151 | for (unsigned Subreg : reverse(C&: SubregsSeen)) { |
1152 | // Fetch a new instruction number, not attached to an actual instruction. |
1153 | unsigned NewInstrNumber = getNewDebugInstrNum(); |
1154 | // Add a substitution from the "new" number to the known one, with a |
1155 | // qualifying subreg. |
1156 | makeDebugValueSubstitution(A: {NewInstrNumber, 0}, B: P, Subreg); |
1157 | // Return the new number; to find the underlying value, consumers need to |
1158 | // deal with the qualifying subreg. |
1159 | P = {NewInstrNumber, 0}; |
1160 | } |
1161 | return P; |
1162 | }; |
1163 | |
1164 | // If we managed to find the defining instruction after COPYs, return an |
1165 | // instruction / operand pair after adding subregister qualifiers. |
1166 | if (State.first.isVirtual()) { |
1167 | // Virtual register def -- we can just look up where this happens. |
1168 | MachineInstr *Inst = MRI.def_begin(RegNo: State.first)->getParent(); |
1169 | for (auto &MO : Inst->all_defs()) { |
1170 | if (MO.getReg() != State.first) |
1171 | continue; |
1172 | return ApplySubregisters({Inst->getDebugInstrNum(), MO.getOperandNo()}); |
1173 | } |
1174 | |
1175 | llvm_unreachable("Vreg def with no corresponding operand?" ); |
1176 | } |
1177 | |
1178 | // Our search ended in a copy from a physreg: walk back up the function |
1179 | // looking for whatever defines the physreg. |
1180 | assert(CurInst->isCopyLike() || TII.isCopyInstr(*CurInst)); |
1181 | State = GetRegAndSubreg(*CurInst); |
1182 | Register RegToSeek = State.first; |
1183 | |
1184 | auto RMII = CurInst->getReverseIterator(); |
1185 | auto PrevInstrs = make_range(x: RMII, y: CurInst->getParent()->instr_rend()); |
1186 | for (auto &ToExamine : PrevInstrs) { |
1187 | for (auto &MO : ToExamine.all_defs()) { |
1188 | // Test for operand that defines something aliasing RegToSeek. |
1189 | if (!TRI.regsOverlap(RegA: RegToSeek, RegB: MO.getReg())) |
1190 | continue; |
1191 | |
1192 | return ApplySubregisters( |
1193 | {ToExamine.getDebugInstrNum(), MO.getOperandNo()}); |
1194 | } |
1195 | } |
1196 | |
1197 | MachineBasicBlock &InsertBB = *CurInst->getParent(); |
1198 | |
1199 | // We reached the start of the block before finding a defining instruction. |
1200 | // There are numerous scenarios where this can happen: |
1201 | // * Constant physical registers, |
1202 | // * Several intrinsics that allow LLVM-IR to read arbitary registers, |
1203 | // * Arguments in the entry block, |
1204 | // * Exception handling landing pads. |
1205 | // Validating all of them is too difficult, so just insert a DBG_PHI reading |
1206 | // the variable value at this position, rather than checking it makes sense. |
1207 | |
1208 | // Create DBG_PHI for specified physreg. |
1209 | auto Builder = BuildMI(BB&: InsertBB, I: InsertBB.getFirstNonPHI(), MIMD: DebugLoc(), |
1210 | MCID: TII.get(Opcode: TargetOpcode::DBG_PHI)); |
1211 | Builder.addReg(RegNo: State.first); |
1212 | unsigned NewNum = getNewDebugInstrNum(); |
1213 | Builder.addImm(Val: NewNum); |
1214 | return ApplySubregisters({NewNum, 0u}); |
1215 | } |
1216 | |
1217 | void MachineFunction::finalizeDebugInstrRefs() { |
1218 | auto *TII = getSubtarget().getInstrInfo(); |
1219 | |
1220 | auto MakeUndefDbgValue = [&](MachineInstr &MI) { |
1221 | const MCInstrDesc &RefII = TII->get(Opcode: TargetOpcode::DBG_VALUE_LIST); |
1222 | MI.setDesc(RefII); |
1223 | MI.setDebugValueUndef(); |
1224 | }; |
1225 | |
1226 | DenseMap<Register, DebugInstrOperandPair> ArgDbgPHIs; |
1227 | for (auto &MBB : *this) { |
1228 | for (auto &MI : MBB) { |
1229 | if (!MI.isDebugRef()) |
1230 | continue; |
1231 | |
1232 | bool IsValidRef = true; |
1233 | |
1234 | for (MachineOperand &MO : MI.debug_operands()) { |
1235 | if (!MO.isReg()) |
1236 | continue; |
1237 | |
1238 | Register Reg = MO.getReg(); |
1239 | |
1240 | // Some vregs can be deleted as redundant in the meantime. Mark those |
1241 | // as DBG_VALUE $noreg. Additionally, some normal instructions are |
1242 | // quickly deleted, leaving dangling references to vregs with no def. |
1243 | if (Reg == 0 || !RegInfo->hasOneDef(RegNo: Reg)) { |
1244 | IsValidRef = false; |
1245 | break; |
1246 | } |
1247 | |
1248 | assert(Reg.isVirtual()); |
1249 | MachineInstr &DefMI = *RegInfo->def_instr_begin(RegNo: Reg); |
1250 | |
1251 | // If we've found a copy-like instruction, follow it back to the |
1252 | // instruction that defines the source value, see salvageCopySSA docs |
1253 | // for why this is important. |
1254 | if (DefMI.isCopyLike() || TII->isCopyInstr(MI: DefMI)) { |
1255 | auto Result = salvageCopySSA(MI&: DefMI, DbgPHICache&: ArgDbgPHIs); |
1256 | MO.ChangeToDbgInstrRef(InstrIdx: Result.first, OpIdx: Result.second); |
1257 | } else { |
1258 | // Otherwise, identify the operand number that the VReg refers to. |
1259 | unsigned OperandIdx = 0; |
1260 | for (const auto &DefMO : DefMI.operands()) { |
1261 | if (DefMO.isReg() && DefMO.isDef() && DefMO.getReg() == Reg) |
1262 | break; |
1263 | ++OperandIdx; |
1264 | } |
1265 | assert(OperandIdx < DefMI.getNumOperands()); |
1266 | |
1267 | // Morph this instr ref to point at the given instruction and operand. |
1268 | unsigned ID = DefMI.getDebugInstrNum(); |
1269 | MO.ChangeToDbgInstrRef(InstrIdx: ID, OpIdx: OperandIdx); |
1270 | } |
1271 | } |
1272 | |
1273 | if (!IsValidRef) |
1274 | MakeUndefDbgValue(MI); |
1275 | } |
1276 | } |
1277 | } |
1278 | |
1279 | bool MachineFunction::shouldUseDebugInstrRef() const { |
1280 | // Disable instr-ref at -O0: it's very slow (in compile time). We can still |
1281 | // have optimized code inlined into this unoptimized code, however with |
1282 | // fewer and less aggressive optimizations happening, coverage and accuracy |
1283 | // should not suffer. |
1284 | if (getTarget().getOptLevel() == CodeGenOptLevel::None) |
1285 | return false; |
1286 | |
1287 | // Don't use instr-ref if this function is marked optnone. |
1288 | if (F.hasFnAttribute(Kind: Attribute::OptimizeNone)) |
1289 | return false; |
1290 | |
1291 | if (llvm::debuginfoShouldUseDebugInstrRef(T: getTarget().getTargetTriple())) |
1292 | return true; |
1293 | |
1294 | return false; |
1295 | } |
1296 | |
1297 | bool MachineFunction::useDebugInstrRef() const { |
1298 | return UseDebugInstrRef; |
1299 | } |
1300 | |
1301 | void MachineFunction::setUseDebugInstrRef(bool Use) { |
1302 | UseDebugInstrRef = Use; |
1303 | } |
1304 | |
1305 | // Use one million as a high / reserved number. |
1306 | const unsigned MachineFunction::DebugOperandMemNumber = 1000000; |
1307 | |
1308 | /// \} |
1309 | |
1310 | //===----------------------------------------------------------------------===// |
1311 | // MachineJumpTableInfo implementation |
1312 | //===----------------------------------------------------------------------===// |
1313 | |
1314 | MachineJumpTableEntry::MachineJumpTableEntry( |
1315 | const std::vector<MachineBasicBlock *> &MBBs) |
1316 | : MBBs(MBBs), Hotness(MachineFunctionDataHotness::Unknown) {} |
1317 | |
1318 | /// Return the size of each entry in the jump table. |
1319 | unsigned MachineJumpTableInfo::getEntrySize(const DataLayout &TD) const { |
1320 | // The size of a jump table entry is 4 bytes unless the entry is just the |
1321 | // address of a block, in which case it is the pointer size. |
1322 | switch (getEntryKind()) { |
1323 | case MachineJumpTableInfo::EK_BlockAddress: |
1324 | return TD.getPointerSize(); |
1325 | case MachineJumpTableInfo::EK_GPRel64BlockAddress: |
1326 | case MachineJumpTableInfo::EK_LabelDifference64: |
1327 | return 8; |
1328 | case MachineJumpTableInfo::EK_GPRel32BlockAddress: |
1329 | case MachineJumpTableInfo::EK_LabelDifference32: |
1330 | case MachineJumpTableInfo::EK_Custom32: |
1331 | return 4; |
1332 | case MachineJumpTableInfo::EK_Inline: |
1333 | return 0; |
1334 | } |
1335 | llvm_unreachable("Unknown jump table encoding!" ); |
1336 | } |
1337 | |
1338 | /// Return the alignment of each entry in the jump table. |
1339 | unsigned MachineJumpTableInfo::getEntryAlignment(const DataLayout &TD) const { |
1340 | // The alignment of a jump table entry is the alignment of int32 unless the |
1341 | // entry is just the address of a block, in which case it is the pointer |
1342 | // alignment. |
1343 | switch (getEntryKind()) { |
1344 | case MachineJumpTableInfo::EK_BlockAddress: |
1345 | return TD.getPointerABIAlignment(AS: 0).value(); |
1346 | case MachineJumpTableInfo::EK_GPRel64BlockAddress: |
1347 | case MachineJumpTableInfo::EK_LabelDifference64: |
1348 | return TD.getABIIntegerTypeAlignment(BitWidth: 64).value(); |
1349 | case MachineJumpTableInfo::EK_GPRel32BlockAddress: |
1350 | case MachineJumpTableInfo::EK_LabelDifference32: |
1351 | case MachineJumpTableInfo::EK_Custom32: |
1352 | return TD.getABIIntegerTypeAlignment(BitWidth: 32).value(); |
1353 | case MachineJumpTableInfo::EK_Inline: |
1354 | return 1; |
1355 | } |
1356 | llvm_unreachable("Unknown jump table encoding!" ); |
1357 | } |
1358 | |
1359 | /// Create a new jump table entry in the jump table info. |
1360 | unsigned MachineJumpTableInfo::createJumpTableIndex( |
1361 | const std::vector<MachineBasicBlock*> &DestBBs) { |
1362 | assert(!DestBBs.empty() && "Cannot create an empty jump table!" ); |
1363 | JumpTables.push_back(x: MachineJumpTableEntry(DestBBs)); |
1364 | return JumpTables.size()-1; |
1365 | } |
1366 | |
1367 | bool MachineJumpTableInfo::updateJumpTableEntryHotness( |
1368 | size_t JTI, MachineFunctionDataHotness Hotness) { |
1369 | assert(JTI < JumpTables.size() && "Invalid JTI!" ); |
1370 | // Record the largest hotness value. |
1371 | if (Hotness <= JumpTables[JTI].Hotness) |
1372 | return false; |
1373 | |
1374 | JumpTables[JTI].Hotness = Hotness; |
1375 | return true; |
1376 | } |
1377 | |
1378 | /// If Old is the target of any jump tables, update the jump tables to branch |
1379 | /// to New instead. |
1380 | bool MachineJumpTableInfo::ReplaceMBBInJumpTables(MachineBasicBlock *Old, |
1381 | MachineBasicBlock *New) { |
1382 | assert(Old != New && "Not making a change?" ); |
1383 | bool MadeChange = false; |
1384 | for (size_t i = 0, e = JumpTables.size(); i != e; ++i) |
1385 | ReplaceMBBInJumpTable(Idx: i, Old, New); |
1386 | return MadeChange; |
1387 | } |
1388 | |
1389 | /// If MBB is present in any jump tables, remove it. |
1390 | bool MachineJumpTableInfo::RemoveMBBFromJumpTables(MachineBasicBlock *MBB) { |
1391 | bool MadeChange = false; |
1392 | for (MachineJumpTableEntry &JTE : JumpTables) { |
1393 | auto removeBeginItr = std::remove(first: JTE.MBBs.begin(), last: JTE.MBBs.end(), value: MBB); |
1394 | MadeChange |= (removeBeginItr != JTE.MBBs.end()); |
1395 | JTE.MBBs.erase(first: removeBeginItr, last: JTE.MBBs.end()); |
1396 | } |
1397 | return MadeChange; |
1398 | } |
1399 | |
1400 | /// If Old is a target of the jump tables, update the jump table to branch to |
1401 | /// New instead. |
1402 | bool MachineJumpTableInfo::ReplaceMBBInJumpTable(unsigned Idx, |
1403 | MachineBasicBlock *Old, |
1404 | MachineBasicBlock *New) { |
1405 | assert(Old != New && "Not making a change?" ); |
1406 | bool MadeChange = false; |
1407 | MachineJumpTableEntry &JTE = JumpTables[Idx]; |
1408 | for (MachineBasicBlock *&MBB : JTE.MBBs) |
1409 | if (MBB == Old) { |
1410 | MBB = New; |
1411 | MadeChange = true; |
1412 | } |
1413 | return MadeChange; |
1414 | } |
1415 | |
1416 | void MachineJumpTableInfo::print(raw_ostream &OS) const { |
1417 | if (JumpTables.empty()) return; |
1418 | |
1419 | OS << "Jump Tables:\n" ; |
1420 | |
1421 | for (unsigned i = 0, e = JumpTables.size(); i != e; ++i) { |
1422 | OS << printJumpTableEntryReference(Idx: i) << ':'; |
1423 | for (const MachineBasicBlock *MBB : JumpTables[i].MBBs) |
1424 | OS << ' ' << printMBBReference(MBB: *MBB); |
1425 | if (i != e) |
1426 | OS << '\n'; |
1427 | } |
1428 | |
1429 | OS << '\n'; |
1430 | } |
1431 | |
1432 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1433 | LLVM_DUMP_METHOD void MachineJumpTableInfo::dump() const { print(dbgs()); } |
1434 | #endif |
1435 | |
1436 | Printable llvm::printJumpTableEntryReference(unsigned Idx) { |
1437 | return Printable([Idx](raw_ostream &OS) { OS << "%jump-table." << Idx; }); |
1438 | } |
1439 | |
1440 | //===----------------------------------------------------------------------===// |
1441 | // MachineConstantPool implementation |
1442 | //===----------------------------------------------------------------------===// |
1443 | |
1444 | void MachineConstantPoolValue::anchor() {} |
1445 | |
1446 | unsigned MachineConstantPoolValue::getSizeInBytes(const DataLayout &DL) const { |
1447 | return DL.getTypeAllocSize(Ty); |
1448 | } |
1449 | |
1450 | unsigned MachineConstantPoolEntry::getSizeInBytes(const DataLayout &DL) const { |
1451 | if (isMachineConstantPoolEntry()) |
1452 | return Val.MachineCPVal->getSizeInBytes(DL); |
1453 | return DL.getTypeAllocSize(Ty: Val.ConstVal->getType()); |
1454 | } |
1455 | |
1456 | bool MachineConstantPoolEntry::needsRelocation() const { |
1457 | if (isMachineConstantPoolEntry()) |
1458 | return true; |
1459 | return Val.ConstVal->needsDynamicRelocation(); |
1460 | } |
1461 | |
1462 | SectionKind |
1463 | MachineConstantPoolEntry::getSectionKind(const DataLayout *DL) const { |
1464 | if (needsRelocation()) |
1465 | return SectionKind::getReadOnlyWithRel(); |
1466 | switch (getSizeInBytes(DL: *DL)) { |
1467 | case 4: |
1468 | return SectionKind::getMergeableConst4(); |
1469 | case 8: |
1470 | return SectionKind::getMergeableConst8(); |
1471 | case 16: |
1472 | return SectionKind::getMergeableConst16(); |
1473 | case 32: |
1474 | return SectionKind::getMergeableConst32(); |
1475 | default: |
1476 | return SectionKind::getReadOnly(); |
1477 | } |
1478 | } |
1479 | |
1480 | MachineConstantPool::~MachineConstantPool() { |
1481 | // A constant may be a member of both Constants and MachineCPVsSharingEntries, |
1482 | // so keep track of which we've deleted to avoid double deletions. |
1483 | DenseSet<MachineConstantPoolValue*> Deleted; |
1484 | for (const MachineConstantPoolEntry &C : Constants) |
1485 | if (C.isMachineConstantPoolEntry()) { |
1486 | Deleted.insert(V: C.Val.MachineCPVal); |
1487 | delete C.Val.MachineCPVal; |
1488 | } |
1489 | for (MachineConstantPoolValue *CPV : MachineCPVsSharingEntries) { |
1490 | if (Deleted.count(V: CPV) == 0) |
1491 | delete CPV; |
1492 | } |
1493 | } |
1494 | |
1495 | /// Test whether the given two constants can be allocated the same constant pool |
1496 | /// entry referenced by \param A. |
1497 | static bool CanShareConstantPoolEntry(const Constant *A, const Constant *B, |
1498 | const DataLayout &DL) { |
1499 | // Handle the trivial case quickly. |
1500 | if (A == B) return true; |
1501 | |
1502 | // If they have the same type but weren't the same constant, quickly |
1503 | // reject them. |
1504 | if (A->getType() == B->getType()) return false; |
1505 | |
1506 | // We can't handle structs or arrays. |
1507 | if (isa<StructType>(Val: A->getType()) || isa<ArrayType>(Val: A->getType()) || |
1508 | isa<StructType>(Val: B->getType()) || isa<ArrayType>(Val: B->getType())) |
1509 | return false; |
1510 | |
1511 | // For now, only support constants with the same size. |
1512 | uint64_t StoreSize = DL.getTypeStoreSize(Ty: A->getType()); |
1513 | if (StoreSize != DL.getTypeStoreSize(Ty: B->getType()) || StoreSize > 128) |
1514 | return false; |
1515 | |
1516 | bool ContainsUndefOrPoisonA = A->containsUndefOrPoisonElement(); |
1517 | |
1518 | Type *IntTy = IntegerType::get(C&: A->getContext(), NumBits: StoreSize*8); |
1519 | |
1520 | // Try constant folding a bitcast of both instructions to an integer. If we |
1521 | // get two identical ConstantInt's, then we are good to share them. We use |
1522 | // the constant folding APIs to do this so that we get the benefit of |
1523 | // DataLayout. |
1524 | if (isa<PointerType>(Val: A->getType())) |
1525 | A = ConstantFoldCastOperand(Opcode: Instruction::PtrToInt, |
1526 | C: const_cast<Constant *>(A), DestTy: IntTy, DL); |
1527 | else if (A->getType() != IntTy) |
1528 | A = ConstantFoldCastOperand(Opcode: Instruction::BitCast, C: const_cast<Constant *>(A), |
1529 | DestTy: IntTy, DL); |
1530 | if (isa<PointerType>(Val: B->getType())) |
1531 | B = ConstantFoldCastOperand(Opcode: Instruction::PtrToInt, |
1532 | C: const_cast<Constant *>(B), DestTy: IntTy, DL); |
1533 | else if (B->getType() != IntTy) |
1534 | B = ConstantFoldCastOperand(Opcode: Instruction::BitCast, C: const_cast<Constant *>(B), |
1535 | DestTy: IntTy, DL); |
1536 | |
1537 | if (A != B) |
1538 | return false; |
1539 | |
1540 | // Constants only safely match if A doesn't contain undef/poison. |
1541 | // As we'll be reusing A, it doesn't matter if B contain undef/poison. |
1542 | // TODO: Handle cases where A and B have the same undef/poison elements. |
1543 | // TODO: Merge A and B with mismatching undef/poison elements. |
1544 | return !ContainsUndefOrPoisonA; |
1545 | } |
1546 | |
1547 | /// Create a new entry in the constant pool or return an existing one. |
1548 | /// User must specify the log2 of the minimum required alignment for the object. |
1549 | unsigned MachineConstantPool::getConstantPoolIndex(const Constant *C, |
1550 | Align Alignment) { |
1551 | if (Alignment > PoolAlignment) PoolAlignment = Alignment; |
1552 | |
1553 | // Check to see if we already have this constant. |
1554 | // |
1555 | // FIXME, this could be made much more efficient for large constant pools. |
1556 | for (unsigned i = 0, e = Constants.size(); i != e; ++i) |
1557 | if (!Constants[i].isMachineConstantPoolEntry() && |
1558 | CanShareConstantPoolEntry(A: Constants[i].Val.ConstVal, B: C, DL)) { |
1559 | if (Constants[i].getAlign() < Alignment) |
1560 | Constants[i].Alignment = Alignment; |
1561 | return i; |
1562 | } |
1563 | |
1564 | Constants.push_back(x: MachineConstantPoolEntry(C, Alignment)); |
1565 | return Constants.size()-1; |
1566 | } |
1567 | |
1568 | unsigned MachineConstantPool::getConstantPoolIndex(MachineConstantPoolValue *V, |
1569 | Align Alignment) { |
1570 | if (Alignment > PoolAlignment) PoolAlignment = Alignment; |
1571 | |
1572 | // Check to see if we already have this constant. |
1573 | // |
1574 | // FIXME, this could be made much more efficient for large constant pools. |
1575 | int Idx = V->getExistingMachineCPValue(CP: this, Alignment); |
1576 | if (Idx != -1) { |
1577 | MachineCPVsSharingEntries.insert(V); |
1578 | return (unsigned)Idx; |
1579 | } |
1580 | |
1581 | Constants.push_back(x: MachineConstantPoolEntry(V, Alignment)); |
1582 | return Constants.size()-1; |
1583 | } |
1584 | |
1585 | void MachineConstantPool::print(raw_ostream &OS) const { |
1586 | if (Constants.empty()) return; |
1587 | |
1588 | OS << "Constant Pool:\n" ; |
1589 | for (unsigned i = 0, e = Constants.size(); i != e; ++i) { |
1590 | OS << " cp#" << i << ": " ; |
1591 | if (Constants[i].isMachineConstantPoolEntry()) |
1592 | Constants[i].Val.MachineCPVal->print(O&: OS); |
1593 | else |
1594 | Constants[i].Val.ConstVal->printAsOperand(O&: OS, /*PrintType=*/false); |
1595 | OS << ", align=" << Constants[i].getAlign().value(); |
1596 | OS << "\n" ; |
1597 | } |
1598 | } |
1599 | |
1600 | //===----------------------------------------------------------------------===// |
1601 | // Template specialization for MachineFunction implementation of |
1602 | // ProfileSummaryInfo::getEntryCount(). |
1603 | //===----------------------------------------------------------------------===// |
1604 | template <> |
1605 | std::optional<Function::ProfileCount> |
1606 | ProfileSummaryInfo::getEntryCount<llvm::MachineFunction>( |
1607 | const llvm::MachineFunction *F) const { |
1608 | return F->getFunction().getEntryCount(); |
1609 | } |
1610 | |
1611 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1612 | LLVM_DUMP_METHOD void MachineConstantPool::dump() const { print(dbgs()); } |
1613 | #endif |
1614 | |