1 | //===- llvm/CodeGen/GlobalISel/IRTranslator.cpp - IRTranslator ---*- C++ -*-==// |
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 | /// This file implements the IRTranslator class. |
10 | //===----------------------------------------------------------------------===// |
11 | |
12 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
13 | #include "llvm/ADT/PostOrderIterator.h" |
14 | #include "llvm/ADT/STLExtras.h" |
15 | #include "llvm/ADT/ScopeExit.h" |
16 | #include "llvm/ADT/SmallSet.h" |
17 | #include "llvm/ADT/SmallVector.h" |
18 | #include "llvm/Analysis/AliasAnalysis.h" |
19 | #include "llvm/Analysis/AssumptionCache.h" |
20 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
21 | #include "llvm/Analysis/Loads.h" |
22 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
23 | #include "llvm/Analysis/ValueTracking.h" |
24 | #include "llvm/Analysis/VectorUtils.h" |
25 | #include "llvm/CodeGen/Analysis.h" |
26 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
27 | #include "llvm/CodeGen/GlobalISel/CSEMIRBuilder.h" |
28 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
29 | #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" |
30 | #include "llvm/CodeGen/GlobalISel/InlineAsmLowering.h" |
31 | #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h" |
32 | #include "llvm/CodeGen/LowLevelTypeUtils.h" |
33 | #include "llvm/CodeGen/MachineBasicBlock.h" |
34 | #include "llvm/CodeGen/MachineFrameInfo.h" |
35 | #include "llvm/CodeGen/MachineFunction.h" |
36 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
37 | #include "llvm/CodeGen/MachineMemOperand.h" |
38 | #include "llvm/CodeGen/MachineModuleInfo.h" |
39 | #include "llvm/CodeGen/MachineOperand.h" |
40 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
41 | #include "llvm/CodeGen/StackProtector.h" |
42 | #include "llvm/CodeGen/SwitchLoweringUtils.h" |
43 | #include "llvm/CodeGen/TargetFrameLowering.h" |
44 | #include "llvm/CodeGen/TargetInstrInfo.h" |
45 | #include "llvm/CodeGen/TargetLowering.h" |
46 | #include "llvm/CodeGen/TargetOpcodes.h" |
47 | #include "llvm/CodeGen/TargetPassConfig.h" |
48 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
49 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
50 | #include "llvm/CodeGenTypes/LowLevelType.h" |
51 | #include "llvm/IR/BasicBlock.h" |
52 | #include "llvm/IR/CFG.h" |
53 | #include "llvm/IR/Constant.h" |
54 | #include "llvm/IR/Constants.h" |
55 | #include "llvm/IR/DataLayout.h" |
56 | #include "llvm/IR/DerivedTypes.h" |
57 | #include "llvm/IR/DiagnosticInfo.h" |
58 | #include "llvm/IR/Function.h" |
59 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
60 | #include "llvm/IR/InlineAsm.h" |
61 | #include "llvm/IR/InstrTypes.h" |
62 | #include "llvm/IR/Instructions.h" |
63 | #include "llvm/IR/IntrinsicInst.h" |
64 | #include "llvm/IR/Intrinsics.h" |
65 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
66 | #include "llvm/IR/LLVMContext.h" |
67 | #include "llvm/IR/Metadata.h" |
68 | #include "llvm/IR/PatternMatch.h" |
69 | #include "llvm/IR/Statepoint.h" |
70 | #include "llvm/IR/Type.h" |
71 | #include "llvm/IR/User.h" |
72 | #include "llvm/IR/Value.h" |
73 | #include "llvm/InitializePasses.h" |
74 | #include "llvm/MC/MCContext.h" |
75 | #include "llvm/Pass.h" |
76 | #include "llvm/Support/Casting.h" |
77 | #include "llvm/Support/CodeGen.h" |
78 | #include "llvm/Support/Debug.h" |
79 | #include "llvm/Support/ErrorHandling.h" |
80 | #include "llvm/Support/MathExtras.h" |
81 | #include "llvm/Support/raw_ostream.h" |
82 | #include "llvm/Target/TargetMachine.h" |
83 | #include "llvm/Transforms/Utils/Local.h" |
84 | #include "llvm/Transforms/Utils/MemoryOpRemark.h" |
85 | #include <algorithm> |
86 | #include <cassert> |
87 | #include <cstdint> |
88 | #include <iterator> |
89 | #include <optional> |
90 | #include <string> |
91 | #include <utility> |
92 | #include <vector> |
93 | |
94 | #define DEBUG_TYPE "irtranslator" |
95 | |
96 | using namespace llvm; |
97 | |
98 | static cl::opt<bool> |
99 | EnableCSEInIRTranslator("enable-cse-in-irtranslator" , |
100 | cl::desc("Should enable CSE in irtranslator" ), |
101 | cl::Optional, cl::init(Val: false)); |
102 | char IRTranslator::ID = 0; |
103 | |
104 | INITIALIZE_PASS_BEGIN(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI" , |
105 | false, false) |
106 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
107 | INITIALIZE_PASS_DEPENDENCY(GISelCSEAnalysisWrapperPass) |
108 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) |
109 | INITIALIZE_PASS_DEPENDENCY(StackProtector) |
110 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
111 | INITIALIZE_PASS_END(IRTranslator, DEBUG_TYPE, "IRTranslator LLVM IR -> MI" , |
112 | false, false) |
113 | |
114 | static void (MachineFunction &MF, |
115 | const TargetPassConfig &TPC, |
116 | OptimizationRemarkEmitter &ORE, |
117 | OptimizationRemarkMissed &R) { |
118 | MF.getProperties().setFailedISel(); |
119 | |
120 | // Print the function name explicitly if we don't have a debug location (which |
121 | // makes the diagnostic less useful) or if we're going to emit a raw error. |
122 | if (!R.getLocation().isValid() || TPC.isGlobalISelAbortEnabled()) |
123 | R << (" (in function: " + MF.getName() + ")" ).str(); |
124 | |
125 | if (TPC.isGlobalISelAbortEnabled()) |
126 | report_fatal_error(reason: Twine(R.getMsg())); |
127 | else |
128 | ORE.emit(OptDiag&: R); |
129 | } |
130 | |
131 | IRTranslator::IRTranslator(CodeGenOptLevel optlevel) |
132 | : MachineFunctionPass(ID), OptLevel(optlevel) {} |
133 | |
134 | #ifndef NDEBUG |
135 | namespace { |
136 | /// Verify that every instruction created has the same DILocation as the |
137 | /// instruction being translated. |
138 | class DILocationVerifier : public GISelChangeObserver { |
139 | const Instruction *CurrInst = nullptr; |
140 | |
141 | public: |
142 | DILocationVerifier() = default; |
143 | ~DILocationVerifier() = default; |
144 | |
145 | const Instruction *getCurrentInst() const { return CurrInst; } |
146 | void setCurrentInst(const Instruction *Inst) { CurrInst = Inst; } |
147 | |
148 | void erasingInstr(MachineInstr &MI) override {} |
149 | void changingInstr(MachineInstr &MI) override {} |
150 | void changedInstr(MachineInstr &MI) override {} |
151 | |
152 | void createdInstr(MachineInstr &MI) override { |
153 | assert(getCurrentInst() && "Inserted instruction without a current MI" ); |
154 | |
155 | // Only print the check message if we're actually checking it. |
156 | #ifndef NDEBUG |
157 | LLVM_DEBUG(dbgs() << "Checking DILocation from " << *CurrInst |
158 | << " was copied to " << MI); |
159 | #endif |
160 | // We allow insts in the entry block to have no debug loc because |
161 | // they could have originated from constants, and we don't want a jumpy |
162 | // debug experience. |
163 | assert((CurrInst->getDebugLoc() == MI.getDebugLoc() || |
164 | (MI.getParent()->isEntryBlock() && !MI.getDebugLoc()) || |
165 | (MI.isDebugInstr())) && |
166 | "Line info was not transferred to all instructions" ); |
167 | } |
168 | }; |
169 | } // namespace |
170 | #endif // ifndef NDEBUG |
171 | |
172 | |
173 | void IRTranslator::getAnalysisUsage(AnalysisUsage &AU) const { |
174 | AU.addRequired<StackProtector>(); |
175 | AU.addRequired<TargetPassConfig>(); |
176 | AU.addRequired<GISelCSEAnalysisWrapperPass>(); |
177 | AU.addRequired<AssumptionCacheTracker>(); |
178 | if (OptLevel != CodeGenOptLevel::None) { |
179 | AU.addRequired<BranchProbabilityInfoWrapperPass>(); |
180 | AU.addRequired<AAResultsWrapperPass>(); |
181 | } |
182 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
183 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
184 | getSelectionDAGFallbackAnalysisUsage(AU); |
185 | MachineFunctionPass::getAnalysisUsage(AU); |
186 | } |
187 | |
188 | IRTranslator::ValueToVRegInfo::VRegListT & |
189 | IRTranslator::allocateVRegs(const Value &Val) { |
190 | auto VRegsIt = VMap.findVRegs(V: Val); |
191 | if (VRegsIt != VMap.vregs_end()) |
192 | return *VRegsIt->second; |
193 | auto *Regs = VMap.getVRegs(V: Val); |
194 | auto *Offsets = VMap.getOffsets(V: Val); |
195 | SmallVector<LLT, 4> SplitTys; |
196 | computeValueLLTs(DL: *DL, Ty&: *Val.getType(), ValueTys&: SplitTys, |
197 | Offsets: Offsets->empty() ? Offsets : nullptr); |
198 | for (unsigned i = 0; i < SplitTys.size(); ++i) |
199 | Regs->push_back(Elt: 0); |
200 | return *Regs; |
201 | } |
202 | |
203 | ArrayRef<Register> IRTranslator::getOrCreateVRegs(const Value &Val) { |
204 | auto VRegsIt = VMap.findVRegs(V: Val); |
205 | if (VRegsIt != VMap.vregs_end()) |
206 | return *VRegsIt->second; |
207 | |
208 | if (Val.getType()->isVoidTy()) |
209 | return *VMap.getVRegs(V: Val); |
210 | |
211 | // Create entry for this type. |
212 | auto *VRegs = VMap.getVRegs(V: Val); |
213 | auto *Offsets = VMap.getOffsets(V: Val); |
214 | |
215 | if (!Val.getType()->isTokenTy()) |
216 | assert(Val.getType()->isSized() && |
217 | "Don't know how to create an empty vreg" ); |
218 | |
219 | SmallVector<LLT, 4> SplitTys; |
220 | computeValueLLTs(DL: *DL, Ty&: *Val.getType(), ValueTys&: SplitTys, |
221 | Offsets: Offsets->empty() ? Offsets : nullptr); |
222 | |
223 | if (!isa<Constant>(Val)) { |
224 | for (auto Ty : SplitTys) |
225 | VRegs->push_back(Elt: MRI->createGenericVirtualRegister(Ty)); |
226 | return *VRegs; |
227 | } |
228 | |
229 | if (Val.getType()->isAggregateType()) { |
230 | // UndefValue, ConstantAggregateZero |
231 | auto &C = cast<Constant>(Val); |
232 | unsigned Idx = 0; |
233 | while (auto Elt = C.getAggregateElement(Elt: Idx++)) { |
234 | auto EltRegs = getOrCreateVRegs(Val: *Elt); |
235 | llvm::append_range(C&: *VRegs, R&: EltRegs); |
236 | } |
237 | } else { |
238 | assert(SplitTys.size() == 1 && "unexpectedly split LLT" ); |
239 | VRegs->push_back(Elt: MRI->createGenericVirtualRegister(Ty: SplitTys[0])); |
240 | bool Success = translate(C: cast<Constant>(Val), Reg: VRegs->front()); |
241 | if (!Success) { |
242 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
243 | MF->getFunction().getSubprogram(), |
244 | &MF->getFunction().getEntryBlock()); |
245 | R << "unable to translate constant: " << ore::NV("Type" , Val.getType()); |
246 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
247 | return *VRegs; |
248 | } |
249 | } |
250 | |
251 | return *VRegs; |
252 | } |
253 | |
254 | int IRTranslator::getOrCreateFrameIndex(const AllocaInst &AI) { |
255 | auto [MapEntry, Inserted] = FrameIndices.try_emplace(Key: &AI); |
256 | if (!Inserted) |
257 | return MapEntry->second; |
258 | |
259 | uint64_t ElementSize = DL->getTypeAllocSize(Ty: AI.getAllocatedType()); |
260 | uint64_t Size = |
261 | ElementSize * cast<ConstantInt>(Val: AI.getArraySize())->getZExtValue(); |
262 | |
263 | // Always allocate at least one byte. |
264 | Size = std::max<uint64_t>(a: Size, b: 1u); |
265 | |
266 | int &FI = MapEntry->second; |
267 | FI = MF->getFrameInfo().CreateStackObject(Size, Alignment: AI.getAlign(), isSpillSlot: false, Alloca: &AI); |
268 | return FI; |
269 | } |
270 | |
271 | Align IRTranslator::getMemOpAlign(const Instruction &I) { |
272 | if (const StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) |
273 | return SI->getAlign(); |
274 | if (const LoadInst *LI = dyn_cast<LoadInst>(Val: &I)) |
275 | return LI->getAlign(); |
276 | if (const AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Val: &I)) |
277 | return AI->getAlign(); |
278 | if (const AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(Val: &I)) |
279 | return AI->getAlign(); |
280 | |
281 | OptimizationRemarkMissed R("gisel-irtranslator" , "" , &I); |
282 | R << "unable to translate memop: " << ore::NV("Opcode" , &I); |
283 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
284 | return Align(1); |
285 | } |
286 | |
287 | MachineBasicBlock &IRTranslator::getMBB(const BasicBlock &BB) { |
288 | MachineBasicBlock *MBB = FuncInfo.getMBB(BB: &BB); |
289 | assert(MBB && "BasicBlock was not encountered before" ); |
290 | return *MBB; |
291 | } |
292 | |
293 | void IRTranslator::addMachineCFGPred(CFGEdge Edge, MachineBasicBlock *NewPred) { |
294 | assert(NewPred && "new predecessor must be a real MachineBasicBlock" ); |
295 | MachinePreds[Edge].push_back(Elt: NewPred); |
296 | } |
297 | |
298 | static bool containsBF16Type(const User &U) { |
299 | // BF16 cannot currently be represented by LLT, to avoid miscompiles we |
300 | // prevent any instructions using them. FIXME: This can be removed once LLT |
301 | // supports bfloat. |
302 | return U.getType()->getScalarType()->isBFloatTy() || |
303 | any_of(Range: U.operands(), P: [](Value *V) { |
304 | return V->getType()->getScalarType()->isBFloatTy(); |
305 | }); |
306 | } |
307 | |
308 | bool IRTranslator::translateBinaryOp(unsigned Opcode, const User &U, |
309 | MachineIRBuilder &MIRBuilder) { |
310 | if (containsBF16Type(U)) |
311 | return false; |
312 | |
313 | // Get or create a virtual register for each value. |
314 | // Unless the value is a Constant => loadimm cst? |
315 | // or inline constant each time? |
316 | // Creation of a virtual register needs to have a size. |
317 | Register Op0 = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
318 | Register Op1 = getOrCreateVReg(Val: *U.getOperand(i: 1)); |
319 | Register Res = getOrCreateVReg(Val: U); |
320 | uint32_t Flags = 0; |
321 | if (isa<Instruction>(Val: U)) { |
322 | const Instruction &I = cast<Instruction>(Val: U); |
323 | Flags = MachineInstr::copyFlagsFromInstruction(I); |
324 | } |
325 | |
326 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {Res}, SrcOps: {Op0, Op1}, Flags); |
327 | return true; |
328 | } |
329 | |
330 | bool IRTranslator::translateUnaryOp(unsigned Opcode, const User &U, |
331 | MachineIRBuilder &MIRBuilder) { |
332 | if (containsBF16Type(U)) |
333 | return false; |
334 | |
335 | Register Op0 = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
336 | Register Res = getOrCreateVReg(Val: U); |
337 | uint32_t Flags = 0; |
338 | if (isa<Instruction>(Val: U)) { |
339 | const Instruction &I = cast<Instruction>(Val: U); |
340 | Flags = MachineInstr::copyFlagsFromInstruction(I); |
341 | } |
342 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {Res}, SrcOps: {Op0}, Flags); |
343 | return true; |
344 | } |
345 | |
346 | bool IRTranslator::translateFNeg(const User &U, MachineIRBuilder &MIRBuilder) { |
347 | return translateUnaryOp(Opcode: TargetOpcode::G_FNEG, U, MIRBuilder); |
348 | } |
349 | |
350 | bool IRTranslator::translateCompare(const User &U, |
351 | MachineIRBuilder &MIRBuilder) { |
352 | if (containsBF16Type(U)) |
353 | return false; |
354 | |
355 | auto *CI = cast<CmpInst>(Val: &U); |
356 | Register Op0 = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
357 | Register Op1 = getOrCreateVReg(Val: *U.getOperand(i: 1)); |
358 | Register Res = getOrCreateVReg(Val: U); |
359 | CmpInst::Predicate Pred = CI->getPredicate(); |
360 | uint32_t Flags = MachineInstr::copyFlagsFromInstruction(I: *CI); |
361 | if (CmpInst::isIntPredicate(P: Pred)) |
362 | MIRBuilder.buildICmp(Pred, Res, Op0, Op1, Flags); |
363 | else if (Pred == CmpInst::FCMP_FALSE) |
364 | MIRBuilder.buildCopy( |
365 | Res, Op: getOrCreateVReg(Val: *Constant::getNullValue(Ty: U.getType()))); |
366 | else if (Pred == CmpInst::FCMP_TRUE) |
367 | MIRBuilder.buildCopy( |
368 | Res, Op: getOrCreateVReg(Val: *Constant::getAllOnesValue(Ty: U.getType()))); |
369 | else |
370 | MIRBuilder.buildFCmp(Pred, Res, Op0, Op1, Flags); |
371 | |
372 | return true; |
373 | } |
374 | |
375 | bool IRTranslator::translateRet(const User &U, MachineIRBuilder &MIRBuilder) { |
376 | const ReturnInst &RI = cast<ReturnInst>(Val: U); |
377 | const Value *Ret = RI.getReturnValue(); |
378 | if (Ret && DL->getTypeStoreSize(Ty: Ret->getType()).isZero()) |
379 | Ret = nullptr; |
380 | |
381 | ArrayRef<Register> VRegs; |
382 | if (Ret) |
383 | VRegs = getOrCreateVRegs(Val: *Ret); |
384 | |
385 | Register SwiftErrorVReg = 0; |
386 | if (CLI->supportSwiftError() && SwiftError.getFunctionArg()) { |
387 | SwiftErrorVReg = SwiftError.getOrCreateVRegUseAt( |
388 | &RI, &MIRBuilder.getMBB(), SwiftError.getFunctionArg()); |
389 | } |
390 | |
391 | // The target may mess up with the insertion point, but |
392 | // this is not important as a return is the last instruction |
393 | // of the block anyway. |
394 | return CLI->lowerReturn(MIRBuilder, Val: Ret, VRegs, FLI&: FuncInfo, SwiftErrorVReg); |
395 | } |
396 | |
397 | void IRTranslator::emitBranchForMergedCondition( |
398 | const Value *Cond, MachineBasicBlock *TBB, MachineBasicBlock *FBB, |
399 | MachineBasicBlock *CurBB, MachineBasicBlock *SwitchBB, |
400 | BranchProbability TProb, BranchProbability FProb, bool InvertCond) { |
401 | // If the leaf of the tree is a comparison, merge the condition into |
402 | // the caseblock. |
403 | if (const CmpInst *BOp = dyn_cast<CmpInst>(Val: Cond)) { |
404 | CmpInst::Predicate Condition; |
405 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(Val: Cond)) { |
406 | Condition = InvertCond ? IC->getInversePredicate() : IC->getPredicate(); |
407 | } else { |
408 | const FCmpInst *FC = cast<FCmpInst>(Val: Cond); |
409 | Condition = InvertCond ? FC->getInversePredicate() : FC->getPredicate(); |
410 | } |
411 | |
412 | SwitchCG::CaseBlock CB(Condition, false, BOp->getOperand(i_nocapture: 0), |
413 | BOp->getOperand(i_nocapture: 1), nullptr, TBB, FBB, CurBB, |
414 | CurBuilder->getDebugLoc(), TProb, FProb); |
415 | SL->SwitchCases.push_back(x: CB); |
416 | return; |
417 | } |
418 | |
419 | // Create a CaseBlock record representing this branch. |
420 | CmpInst::Predicate Pred = InvertCond ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; |
421 | SwitchCG::CaseBlock CB( |
422 | Pred, false, Cond, ConstantInt::getTrue(Context&: MF->getFunction().getContext()), |
423 | nullptr, TBB, FBB, CurBB, CurBuilder->getDebugLoc(), TProb, FProb); |
424 | SL->SwitchCases.push_back(x: CB); |
425 | } |
426 | |
427 | static bool isValInBlock(const Value *V, const BasicBlock *BB) { |
428 | if (const Instruction *I = dyn_cast<Instruction>(Val: V)) |
429 | return I->getParent() == BB; |
430 | return true; |
431 | } |
432 | |
433 | void IRTranslator::findMergedConditions( |
434 | const Value *Cond, MachineBasicBlock *TBB, MachineBasicBlock *FBB, |
435 | MachineBasicBlock *CurBB, MachineBasicBlock *SwitchBB, |
436 | Instruction::BinaryOps Opc, BranchProbability TProb, |
437 | BranchProbability FProb, bool InvertCond) { |
438 | using namespace PatternMatch; |
439 | assert((Opc == Instruction::And || Opc == Instruction::Or) && |
440 | "Expected Opc to be AND/OR" ); |
441 | // Skip over not part of the tree and remember to invert op and operands at |
442 | // next level. |
443 | Value *NotCond; |
444 | if (match(V: Cond, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: NotCond)))) && |
445 | isValInBlock(V: NotCond, BB: CurBB->getBasicBlock())) { |
446 | findMergedConditions(Cond: NotCond, TBB, FBB, CurBB, SwitchBB, Opc, TProb, FProb, |
447 | InvertCond: !InvertCond); |
448 | return; |
449 | } |
450 | |
451 | const Instruction *BOp = dyn_cast<Instruction>(Val: Cond); |
452 | const Value *BOpOp0, *BOpOp1; |
453 | // Compute the effective opcode for Cond, taking into account whether it needs |
454 | // to be inverted, e.g. |
455 | // and (not (or A, B)), C |
456 | // gets lowered as |
457 | // and (and (not A, not B), C) |
458 | Instruction::BinaryOps BOpc = (Instruction::BinaryOps)0; |
459 | if (BOp) { |
460 | BOpc = match(V: BOp, P: m_LogicalAnd(L: m_Value(V&: BOpOp0), R: m_Value(V&: BOpOp1))) |
461 | ? Instruction::And |
462 | : (match(V: BOp, P: m_LogicalOr(L: m_Value(V&: BOpOp0), R: m_Value(V&: BOpOp1))) |
463 | ? Instruction::Or |
464 | : (Instruction::BinaryOps)0); |
465 | if (InvertCond) { |
466 | if (BOpc == Instruction::And) |
467 | BOpc = Instruction::Or; |
468 | else if (BOpc == Instruction::Or) |
469 | BOpc = Instruction::And; |
470 | } |
471 | } |
472 | |
473 | // If this node is not part of the or/and tree, emit it as a branch. |
474 | // Note that all nodes in the tree should have same opcode. |
475 | bool BOpIsInOrAndTree = BOpc && BOpc == Opc && BOp->hasOneUse(); |
476 | if (!BOpIsInOrAndTree || BOp->getParent() != CurBB->getBasicBlock() || |
477 | !isValInBlock(V: BOpOp0, BB: CurBB->getBasicBlock()) || |
478 | !isValInBlock(V: BOpOp1, BB: CurBB->getBasicBlock())) { |
479 | emitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB, TProb, FProb, |
480 | InvertCond); |
481 | return; |
482 | } |
483 | |
484 | // Create TmpBB after CurBB. |
485 | MachineFunction::iterator BBI(CurBB); |
486 | MachineBasicBlock *TmpBB = |
487 | MF->CreateMachineBasicBlock(BB: CurBB->getBasicBlock()); |
488 | CurBB->getParent()->insert(MBBI: ++BBI, MBB: TmpBB); |
489 | |
490 | if (Opc == Instruction::Or) { |
491 | // Codegen X | Y as: |
492 | // BB1: |
493 | // jmp_if_X TBB |
494 | // jmp TmpBB |
495 | // TmpBB: |
496 | // jmp_if_Y TBB |
497 | // jmp FBB |
498 | // |
499 | |
500 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. |
501 | // The requirement is that |
502 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) |
503 | // = TrueProb for original BB. |
504 | // Assuming the original probabilities are A and B, one choice is to set |
505 | // BB1's probabilities to A/2 and A/2+B, and set TmpBB's probabilities to |
506 | // A/(1+B) and 2B/(1+B). This choice assumes that |
507 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. |
508 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for |
509 | // TmpBB, but the math is more complicated. |
510 | |
511 | auto NewTrueProb = TProb / 2; |
512 | auto NewFalseProb = TProb / 2 + FProb; |
513 | // Emit the LHS condition. |
514 | findMergedConditions(Cond: BOpOp0, TBB, FBB: TmpBB, CurBB, SwitchBB, Opc, TProb: NewTrueProb, |
515 | FProb: NewFalseProb, InvertCond); |
516 | |
517 | // Normalize A/2 and B to get A/(1+B) and 2B/(1+B). |
518 | SmallVector<BranchProbability, 2> Probs{TProb / 2, FProb}; |
519 | BranchProbability::normalizeProbabilities(Begin: Probs.begin(), End: Probs.end()); |
520 | // Emit the RHS condition into TmpBB. |
521 | findMergedConditions(Cond: BOpOp1, TBB, FBB, CurBB: TmpBB, SwitchBB, Opc, TProb: Probs[0], |
522 | FProb: Probs[1], InvertCond); |
523 | } else { |
524 | assert(Opc == Instruction::And && "Unknown merge op!" ); |
525 | // Codegen X & Y as: |
526 | // BB1: |
527 | // jmp_if_X TmpBB |
528 | // jmp FBB |
529 | // TmpBB: |
530 | // jmp_if_Y TBB |
531 | // jmp FBB |
532 | // |
533 | // This requires creation of TmpBB after CurBB. |
534 | |
535 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. |
536 | // The requirement is that |
537 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) |
538 | // = FalseProb for original BB. |
539 | // Assuming the original probabilities are A and B, one choice is to set |
540 | // BB1's probabilities to A+B/2 and B/2, and set TmpBB's probabilities to |
541 | // 2A/(1+A) and B/(1+A). This choice assumes that FalseProb for BB1 == |
542 | // TrueProb for BB1 * FalseProb for TmpBB. |
543 | |
544 | auto NewTrueProb = TProb + FProb / 2; |
545 | auto NewFalseProb = FProb / 2; |
546 | // Emit the LHS condition. |
547 | findMergedConditions(Cond: BOpOp0, TBB: TmpBB, FBB, CurBB, SwitchBB, Opc, TProb: NewTrueProb, |
548 | FProb: NewFalseProb, InvertCond); |
549 | |
550 | // Normalize A and B/2 to get 2A/(1+A) and B/(1+A). |
551 | SmallVector<BranchProbability, 2> Probs{TProb, FProb / 2}; |
552 | BranchProbability::normalizeProbabilities(Begin: Probs.begin(), End: Probs.end()); |
553 | // Emit the RHS condition into TmpBB. |
554 | findMergedConditions(Cond: BOpOp1, TBB, FBB, CurBB: TmpBB, SwitchBB, Opc, TProb: Probs[0], |
555 | FProb: Probs[1], InvertCond); |
556 | } |
557 | } |
558 | |
559 | bool IRTranslator::shouldEmitAsBranches( |
560 | const std::vector<SwitchCG::CaseBlock> &Cases) { |
561 | // For multiple cases, it's better to emit as branches. |
562 | if (Cases.size() != 2) |
563 | return true; |
564 | |
565 | // If this is two comparisons of the same values or'd or and'd together, they |
566 | // will get folded into a single comparison, so don't emit two blocks. |
567 | if ((Cases[0].CmpLHS == Cases[1].CmpLHS && |
568 | Cases[0].CmpRHS == Cases[1].CmpRHS) || |
569 | (Cases[0].CmpRHS == Cases[1].CmpLHS && |
570 | Cases[0].CmpLHS == Cases[1].CmpRHS)) { |
571 | return false; |
572 | } |
573 | |
574 | // Handle: (X != null) | (Y != null) --> (X|Y) != 0 |
575 | // Handle: (X == null) & (Y == null) --> (X|Y) == 0 |
576 | if (Cases[0].CmpRHS == Cases[1].CmpRHS && |
577 | Cases[0].PredInfo.Pred == Cases[1].PredInfo.Pred && |
578 | isa<Constant>(Val: Cases[0].CmpRHS) && |
579 | cast<Constant>(Val: Cases[0].CmpRHS)->isNullValue()) { |
580 | if (Cases[0].PredInfo.Pred == CmpInst::ICMP_EQ && |
581 | Cases[0].TrueBB == Cases[1].ThisBB) |
582 | return false; |
583 | if (Cases[0].PredInfo.Pred == CmpInst::ICMP_NE && |
584 | Cases[0].FalseBB == Cases[1].ThisBB) |
585 | return false; |
586 | } |
587 | |
588 | return true; |
589 | } |
590 | |
591 | bool IRTranslator::translateBr(const User &U, MachineIRBuilder &MIRBuilder) { |
592 | const BranchInst &BrInst = cast<BranchInst>(Val: U); |
593 | auto &CurMBB = MIRBuilder.getMBB(); |
594 | auto *Succ0MBB = &getMBB(BB: *BrInst.getSuccessor(i: 0)); |
595 | |
596 | if (BrInst.isUnconditional()) { |
597 | // If the unconditional target is the layout successor, fallthrough. |
598 | if (OptLevel == CodeGenOptLevel::None || |
599 | !CurMBB.isLayoutSuccessor(MBB: Succ0MBB)) |
600 | MIRBuilder.buildBr(Dest&: *Succ0MBB); |
601 | |
602 | // Link successors. |
603 | for (const BasicBlock *Succ : successors(I: &BrInst)) |
604 | CurMBB.addSuccessor(Succ: &getMBB(BB: *Succ)); |
605 | return true; |
606 | } |
607 | |
608 | // If this condition is one of the special cases we handle, do special stuff |
609 | // now. |
610 | const Value *CondVal = BrInst.getCondition(); |
611 | MachineBasicBlock *Succ1MBB = &getMBB(BB: *BrInst.getSuccessor(i: 1)); |
612 | |
613 | // If this is a series of conditions that are or'd or and'd together, emit |
614 | // this as a sequence of branches instead of setcc's with and/or operations. |
615 | // As long as jumps are not expensive (exceptions for multi-use logic ops, |
616 | // unpredictable branches, and vector extracts because those jumps are likely |
617 | // expensive for any target), this should improve performance. |
618 | // For example, instead of something like: |
619 | // cmp A, B |
620 | // C = seteq |
621 | // cmp D, E |
622 | // F = setle |
623 | // or C, F |
624 | // jnz foo |
625 | // Emit: |
626 | // cmp A, B |
627 | // je foo |
628 | // cmp D, E |
629 | // jle foo |
630 | using namespace PatternMatch; |
631 | const Instruction *CondI = dyn_cast<Instruction>(Val: CondVal); |
632 | if (!TLI->isJumpExpensive() && CondI && CondI->hasOneUse() && |
633 | !BrInst.hasMetadata(KindID: LLVMContext::MD_unpredictable)) { |
634 | Instruction::BinaryOps Opcode = (Instruction::BinaryOps)0; |
635 | Value *Vec; |
636 | const Value *BOp0, *BOp1; |
637 | if (match(V: CondI, P: m_LogicalAnd(L: m_Value(V&: BOp0), R: m_Value(V&: BOp1)))) |
638 | Opcode = Instruction::And; |
639 | else if (match(V: CondI, P: m_LogicalOr(L: m_Value(V&: BOp0), R: m_Value(V&: BOp1)))) |
640 | Opcode = Instruction::Or; |
641 | |
642 | if (Opcode && !(match(V: BOp0, P: m_ExtractElt(Val: m_Value(V&: Vec), Idx: m_Value())) && |
643 | match(V: BOp1, P: m_ExtractElt(Val: m_Specific(V: Vec), Idx: m_Value())))) { |
644 | findMergedConditions(Cond: CondI, TBB: Succ0MBB, FBB: Succ1MBB, CurBB: &CurMBB, SwitchBB: &CurMBB, Opc: Opcode, |
645 | TProb: getEdgeProbability(Src: &CurMBB, Dst: Succ0MBB), |
646 | FProb: getEdgeProbability(Src: &CurMBB, Dst: Succ1MBB), |
647 | /*InvertCond=*/false); |
648 | assert(SL->SwitchCases[0].ThisBB == &CurMBB && "Unexpected lowering!" ); |
649 | |
650 | // Allow some cases to be rejected. |
651 | if (shouldEmitAsBranches(Cases: SL->SwitchCases)) { |
652 | // Emit the branch for this block. |
653 | emitSwitchCase(CB&: SL->SwitchCases[0], SwitchBB: &CurMBB, MIB&: *CurBuilder); |
654 | SL->SwitchCases.erase(position: SL->SwitchCases.begin()); |
655 | return true; |
656 | } |
657 | |
658 | // Okay, we decided not to do this, remove any inserted MBB's and clear |
659 | // SwitchCases. |
660 | for (unsigned I = 1, E = SL->SwitchCases.size(); I != E; ++I) |
661 | MF->erase(MBBI: SL->SwitchCases[I].ThisBB); |
662 | |
663 | SL->SwitchCases.clear(); |
664 | } |
665 | } |
666 | |
667 | // Create a CaseBlock record representing this branch. |
668 | SwitchCG::CaseBlock CB(CmpInst::ICMP_EQ, false, CondVal, |
669 | ConstantInt::getTrue(Context&: MF->getFunction().getContext()), |
670 | nullptr, Succ0MBB, Succ1MBB, &CurMBB, |
671 | CurBuilder->getDebugLoc()); |
672 | |
673 | // Use emitSwitchCase to actually insert the fast branch sequence for this |
674 | // cond branch. |
675 | emitSwitchCase(CB, SwitchBB: &CurMBB, MIB&: *CurBuilder); |
676 | return true; |
677 | } |
678 | |
679 | void IRTranslator::addSuccessorWithProb(MachineBasicBlock *Src, |
680 | MachineBasicBlock *Dst, |
681 | BranchProbability Prob) { |
682 | if (!FuncInfo.BPI) { |
683 | Src->addSuccessorWithoutProb(Succ: Dst); |
684 | return; |
685 | } |
686 | if (Prob.isUnknown()) |
687 | Prob = getEdgeProbability(Src, Dst); |
688 | Src->addSuccessor(Succ: Dst, Prob); |
689 | } |
690 | |
691 | BranchProbability |
692 | IRTranslator::getEdgeProbability(const MachineBasicBlock *Src, |
693 | const MachineBasicBlock *Dst) const { |
694 | const BasicBlock *SrcBB = Src->getBasicBlock(); |
695 | const BasicBlock *DstBB = Dst->getBasicBlock(); |
696 | if (!FuncInfo.BPI) { |
697 | // If BPI is not available, set the default probability as 1 / N, where N is |
698 | // the number of successors. |
699 | auto SuccSize = std::max<uint32_t>(a: succ_size(BB: SrcBB), b: 1); |
700 | return BranchProbability(1, SuccSize); |
701 | } |
702 | return FuncInfo.BPI->getEdgeProbability(Src: SrcBB, Dst: DstBB); |
703 | } |
704 | |
705 | bool IRTranslator::translateSwitch(const User &U, MachineIRBuilder &MIB) { |
706 | using namespace SwitchCG; |
707 | // Extract cases from the switch. |
708 | const SwitchInst &SI = cast<SwitchInst>(Val: U); |
709 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
710 | CaseClusterVector Clusters; |
711 | Clusters.reserve(n: SI.getNumCases()); |
712 | for (const auto &I : SI.cases()) { |
713 | MachineBasicBlock *Succ = &getMBB(BB: *I.getCaseSuccessor()); |
714 | assert(Succ && "Could not find successor mbb in mapping" ); |
715 | const ConstantInt *CaseVal = I.getCaseValue(); |
716 | BranchProbability Prob = |
717 | BPI ? BPI->getEdgeProbability(Src: SI.getParent(), IndexInSuccessors: I.getSuccessorIndex()) |
718 | : BranchProbability(1, SI.getNumCases() + 1); |
719 | Clusters.push_back(x: CaseCluster::range(Low: CaseVal, High: CaseVal, MBB: Succ, Prob)); |
720 | } |
721 | |
722 | MachineBasicBlock *DefaultMBB = &getMBB(BB: *SI.getDefaultDest()); |
723 | |
724 | // Cluster adjacent cases with the same destination. We do this at all |
725 | // optimization levels because it's cheap to do and will make codegen faster |
726 | // if there are many clusters. |
727 | sortAndRangeify(Clusters); |
728 | |
729 | MachineBasicBlock *SwitchMBB = &getMBB(BB: *SI.getParent()); |
730 | |
731 | // If there is only the default destination, jump there directly. |
732 | if (Clusters.empty()) { |
733 | SwitchMBB->addSuccessor(Succ: DefaultMBB); |
734 | if (DefaultMBB != SwitchMBB->getNextNode()) |
735 | MIB.buildBr(Dest&: *DefaultMBB); |
736 | return true; |
737 | } |
738 | |
739 | SL->findJumpTables(Clusters, SI: &SI, SL: std::nullopt, DefaultMBB, PSI: nullptr, BFI: nullptr); |
740 | SL->findBitTestClusters(Clusters, SI: &SI); |
741 | |
742 | LLVM_DEBUG({ |
743 | dbgs() << "Case clusters: " ; |
744 | for (const CaseCluster &C : Clusters) { |
745 | if (C.Kind == CC_JumpTable) |
746 | dbgs() << "JT:" ; |
747 | if (C.Kind == CC_BitTests) |
748 | dbgs() << "BT:" ; |
749 | |
750 | C.Low->getValue().print(dbgs(), true); |
751 | if (C.Low != C.High) { |
752 | dbgs() << '-'; |
753 | C.High->getValue().print(dbgs(), true); |
754 | } |
755 | dbgs() << ' '; |
756 | } |
757 | dbgs() << '\n'; |
758 | }); |
759 | |
760 | assert(!Clusters.empty()); |
761 | SwitchWorkList WorkList; |
762 | CaseClusterIt First = Clusters.begin(); |
763 | CaseClusterIt Last = Clusters.end() - 1; |
764 | auto DefaultProb = getEdgeProbability(Src: SwitchMBB, Dst: DefaultMBB); |
765 | WorkList.push_back(Elt: {.MBB: SwitchMBB, .FirstCluster: First, .LastCluster: Last, .GE: nullptr, .LT: nullptr, .DefaultProb: DefaultProb}); |
766 | |
767 | while (!WorkList.empty()) { |
768 | SwitchWorkListItem W = WorkList.pop_back_val(); |
769 | |
770 | unsigned NumClusters = W.LastCluster - W.FirstCluster + 1; |
771 | // For optimized builds, lower large range as a balanced binary tree. |
772 | if (NumClusters > 3 && |
773 | MF->getTarget().getOptLevel() != CodeGenOptLevel::None && |
774 | !DefaultMBB->getParent()->getFunction().hasMinSize()) { |
775 | splitWorkItem(WorkList, W, Cond: SI.getCondition(), SwitchMBB, MIB); |
776 | continue; |
777 | } |
778 | |
779 | if (!lowerSwitchWorkItem(W, Cond: SI.getCondition(), SwitchMBB, DefaultMBB, MIB)) |
780 | return false; |
781 | } |
782 | return true; |
783 | } |
784 | |
785 | void IRTranslator::splitWorkItem(SwitchCG::SwitchWorkList &WorkList, |
786 | const SwitchCG::SwitchWorkListItem &W, |
787 | Value *Cond, MachineBasicBlock *SwitchMBB, |
788 | MachineIRBuilder &MIB) { |
789 | using namespace SwitchCG; |
790 | assert(W.FirstCluster->Low->getValue().slt(W.LastCluster->Low->getValue()) && |
791 | "Clusters not sorted?" ); |
792 | assert(W.LastCluster - W.FirstCluster + 1 >= 2 && "Too small to split!" ); |
793 | |
794 | auto [LastLeft, FirstRight, LeftProb, RightProb] = |
795 | SL->computeSplitWorkItemInfo(W); |
796 | |
797 | // Use the first element on the right as pivot since we will make less-than |
798 | // comparisons against it. |
799 | CaseClusterIt PivotCluster = FirstRight; |
800 | assert(PivotCluster > W.FirstCluster); |
801 | assert(PivotCluster <= W.LastCluster); |
802 | |
803 | CaseClusterIt FirstLeft = W.FirstCluster; |
804 | CaseClusterIt LastRight = W.LastCluster; |
805 | |
806 | const ConstantInt *Pivot = PivotCluster->Low; |
807 | |
808 | // New blocks will be inserted immediately after the current one. |
809 | MachineFunction::iterator BBI(W.MBB); |
810 | ++BBI; |
811 | |
812 | // We will branch to the LHS if Value < Pivot. If LHS is a single cluster, |
813 | // we can branch to its destination directly if it's squeezed exactly in |
814 | // between the known lower bound and Pivot - 1. |
815 | MachineBasicBlock *LeftMBB; |
816 | if (FirstLeft == LastLeft && FirstLeft->Kind == CC_Range && |
817 | FirstLeft->Low == W.GE && |
818 | (FirstLeft->High->getValue() + 1LL) == Pivot->getValue()) { |
819 | LeftMBB = FirstLeft->MBB; |
820 | } else { |
821 | LeftMBB = FuncInfo.MF->CreateMachineBasicBlock(BB: W.MBB->getBasicBlock()); |
822 | FuncInfo.MF->insert(MBBI: BBI, MBB: LeftMBB); |
823 | WorkList.push_back( |
824 | Elt: {.MBB: LeftMBB, .FirstCluster: FirstLeft, .LastCluster: LastLeft, .GE: W.GE, .LT: Pivot, .DefaultProb: W.DefaultProb / 2}); |
825 | } |
826 | |
827 | // Similarly, we will branch to the RHS if Value >= Pivot. If RHS is a |
828 | // single cluster, RHS.Low == Pivot, and we can branch to its destination |
829 | // directly if RHS.High equals the current upper bound. |
830 | MachineBasicBlock *RightMBB; |
831 | if (FirstRight == LastRight && FirstRight->Kind == CC_Range && W.LT && |
832 | (FirstRight->High->getValue() + 1ULL) == W.LT->getValue()) { |
833 | RightMBB = FirstRight->MBB; |
834 | } else { |
835 | RightMBB = FuncInfo.MF->CreateMachineBasicBlock(BB: W.MBB->getBasicBlock()); |
836 | FuncInfo.MF->insert(MBBI: BBI, MBB: RightMBB); |
837 | WorkList.push_back( |
838 | Elt: {.MBB: RightMBB, .FirstCluster: FirstRight, .LastCluster: LastRight, .GE: Pivot, .LT: W.LT, .DefaultProb: W.DefaultProb / 2}); |
839 | } |
840 | |
841 | // Create the CaseBlock record that will be used to lower the branch. |
842 | CaseBlock CB(ICmpInst::Predicate::ICMP_SLT, false, Cond, Pivot, nullptr, |
843 | LeftMBB, RightMBB, W.MBB, MIB.getDebugLoc(), LeftProb, |
844 | RightProb); |
845 | |
846 | if (W.MBB == SwitchMBB) |
847 | emitSwitchCase(CB, SwitchBB: SwitchMBB, MIB); |
848 | else |
849 | SL->SwitchCases.push_back(x: CB); |
850 | } |
851 | |
852 | void IRTranslator::emitJumpTable(SwitchCG::JumpTable &JT, |
853 | MachineBasicBlock *MBB) { |
854 | // Emit the code for the jump table |
855 | assert(JT.Reg && "Should lower JT Header first!" ); |
856 | MachineIRBuilder MIB(*MBB->getParent()); |
857 | MIB.setMBB(*MBB); |
858 | MIB.setDebugLoc(CurBuilder->getDebugLoc()); |
859 | |
860 | Type *PtrIRTy = PointerType::getUnqual(C&: MF->getFunction().getContext()); |
861 | const LLT PtrTy = getLLTForType(Ty&: *PtrIRTy, DL: *DL); |
862 | |
863 | auto Table = MIB.buildJumpTable(PtrTy, JTI: JT.JTI); |
864 | MIB.buildBrJT(TablePtr: Table.getReg(Idx: 0), JTI: JT.JTI, IndexReg: JT.Reg); |
865 | } |
866 | |
867 | bool IRTranslator::(SwitchCG::JumpTable &JT, |
868 | SwitchCG::JumpTableHeader &JTH, |
869 | MachineBasicBlock *) { |
870 | MachineIRBuilder MIB(*HeaderBB->getParent()); |
871 | MIB.setMBB(*HeaderBB); |
872 | MIB.setDebugLoc(CurBuilder->getDebugLoc()); |
873 | |
874 | const Value &SValue = *JTH.SValue; |
875 | // Subtract the lowest switch case value from the value being switched on. |
876 | const LLT SwitchTy = getLLTForType(Ty&: *SValue.getType(), DL: *DL); |
877 | Register SwitchOpReg = getOrCreateVReg(Val: SValue); |
878 | auto FirstCst = MIB.buildConstant(Res: SwitchTy, Val: JTH.First); |
879 | auto Sub = MIB.buildSub(Dst: {SwitchTy}, Src0: SwitchOpReg, Src1: FirstCst); |
880 | |
881 | // This value may be smaller or larger than the target's pointer type, and |
882 | // therefore require extension or truncating. |
883 | auto *PtrIRTy = PointerType::getUnqual(C&: SValue.getContext()); |
884 | const LLT PtrScalarTy = LLT::scalar(SizeInBits: DL->getTypeSizeInBits(Ty: PtrIRTy)); |
885 | Sub = MIB.buildZExtOrTrunc(Res: PtrScalarTy, Op: Sub); |
886 | |
887 | JT.Reg = Sub.getReg(Idx: 0); |
888 | |
889 | if (JTH.FallthroughUnreachable) { |
890 | if (JT.MBB != HeaderBB->getNextNode()) |
891 | MIB.buildBr(Dest&: *JT.MBB); |
892 | return true; |
893 | } |
894 | |
895 | // Emit the range check for the jump table, and branch to the default block |
896 | // for the switch statement if the value being switched on exceeds the |
897 | // largest case in the switch. |
898 | auto Cst = getOrCreateVReg( |
899 | Val: *ConstantInt::get(Ty: SValue.getType(), V: JTH.Last - JTH.First)); |
900 | Cst = MIB.buildZExtOrTrunc(Res: PtrScalarTy, Op: Cst).getReg(Idx: 0); |
901 | auto Cmp = MIB.buildICmp(Pred: CmpInst::ICMP_UGT, Res: LLT::scalar(SizeInBits: 1), Op0: Sub, Op1: Cst); |
902 | |
903 | auto BrCond = MIB.buildBrCond(Tst: Cmp.getReg(Idx: 0), Dest&: *JT.Default); |
904 | |
905 | // Avoid emitting unnecessary branches to the next block. |
906 | if (JT.MBB != HeaderBB->getNextNode()) |
907 | BrCond = MIB.buildBr(Dest&: *JT.MBB); |
908 | return true; |
909 | } |
910 | |
911 | void IRTranslator::emitSwitchCase(SwitchCG::CaseBlock &CB, |
912 | MachineBasicBlock *SwitchBB, |
913 | MachineIRBuilder &MIB) { |
914 | Register CondLHS = getOrCreateVReg(Val: *CB.CmpLHS); |
915 | Register Cond; |
916 | DebugLoc OldDbgLoc = MIB.getDebugLoc(); |
917 | MIB.setDebugLoc(CB.DbgLoc); |
918 | MIB.setMBB(*CB.ThisBB); |
919 | |
920 | if (CB.PredInfo.NoCmp) { |
921 | // Branch or fall through to TrueBB. |
922 | addSuccessorWithProb(Src: CB.ThisBB, Dst: CB.TrueBB, Prob: CB.TrueProb); |
923 | addMachineCFGPred(Edge: {SwitchBB->getBasicBlock(), CB.TrueBB->getBasicBlock()}, |
924 | NewPred: CB.ThisBB); |
925 | CB.ThisBB->normalizeSuccProbs(); |
926 | if (CB.TrueBB != CB.ThisBB->getNextNode()) |
927 | MIB.buildBr(Dest&: *CB.TrueBB); |
928 | MIB.setDebugLoc(OldDbgLoc); |
929 | return; |
930 | } |
931 | |
932 | const LLT i1Ty = LLT::scalar(SizeInBits: 1); |
933 | // Build the compare. |
934 | if (!CB.CmpMHS) { |
935 | const auto *CI = dyn_cast<ConstantInt>(Val: CB.CmpRHS); |
936 | // For conditional branch lowering, we might try to do something silly like |
937 | // emit an G_ICMP to compare an existing G_ICMP i1 result with true. If so, |
938 | // just re-use the existing condition vreg. |
939 | if (MRI->getType(Reg: CondLHS).getSizeInBits() == 1 && CI && CI->isOne() && |
940 | CB.PredInfo.Pred == CmpInst::ICMP_EQ) { |
941 | Cond = CondLHS; |
942 | } else { |
943 | Register CondRHS = getOrCreateVReg(Val: *CB.CmpRHS); |
944 | if (CmpInst::isFPPredicate(P: CB.PredInfo.Pred)) |
945 | Cond = |
946 | MIB.buildFCmp(Pred: CB.PredInfo.Pred, Res: i1Ty, Op0: CondLHS, Op1: CondRHS).getReg(Idx: 0); |
947 | else |
948 | Cond = |
949 | MIB.buildICmp(Pred: CB.PredInfo.Pred, Res: i1Ty, Op0: CondLHS, Op1: CondRHS).getReg(Idx: 0); |
950 | } |
951 | } else { |
952 | assert(CB.PredInfo.Pred == CmpInst::ICMP_SLE && |
953 | "Can only handle SLE ranges" ); |
954 | |
955 | const APInt& Low = cast<ConstantInt>(Val: CB.CmpLHS)->getValue(); |
956 | const APInt& High = cast<ConstantInt>(Val: CB.CmpRHS)->getValue(); |
957 | |
958 | Register CmpOpReg = getOrCreateVReg(Val: *CB.CmpMHS); |
959 | if (cast<ConstantInt>(Val: CB.CmpLHS)->isMinValue(IsSigned: true)) { |
960 | Register CondRHS = getOrCreateVReg(Val: *CB.CmpRHS); |
961 | Cond = |
962 | MIB.buildICmp(Pred: CmpInst::ICMP_SLE, Res: i1Ty, Op0: CmpOpReg, Op1: CondRHS).getReg(Idx: 0); |
963 | } else { |
964 | const LLT CmpTy = MRI->getType(Reg: CmpOpReg); |
965 | auto Sub = MIB.buildSub(Dst: {CmpTy}, Src0: CmpOpReg, Src1: CondLHS); |
966 | auto Diff = MIB.buildConstant(Res: CmpTy, Val: High - Low); |
967 | Cond = MIB.buildICmp(Pred: CmpInst::ICMP_ULE, Res: i1Ty, Op0: Sub, Op1: Diff).getReg(Idx: 0); |
968 | } |
969 | } |
970 | |
971 | // Update successor info |
972 | addSuccessorWithProb(Src: CB.ThisBB, Dst: CB.TrueBB, Prob: CB.TrueProb); |
973 | |
974 | addMachineCFGPred(Edge: {SwitchBB->getBasicBlock(), CB.TrueBB->getBasicBlock()}, |
975 | NewPred: CB.ThisBB); |
976 | |
977 | // TrueBB and FalseBB are always different unless the incoming IR is |
978 | // degenerate. This only happens when running llc on weird IR. |
979 | if (CB.TrueBB != CB.FalseBB) |
980 | addSuccessorWithProb(Src: CB.ThisBB, Dst: CB.FalseBB, Prob: CB.FalseProb); |
981 | CB.ThisBB->normalizeSuccProbs(); |
982 | |
983 | addMachineCFGPred(Edge: {SwitchBB->getBasicBlock(), CB.FalseBB->getBasicBlock()}, |
984 | NewPred: CB.ThisBB); |
985 | |
986 | MIB.buildBrCond(Tst: Cond, Dest&: *CB.TrueBB); |
987 | MIB.buildBr(Dest&: *CB.FalseBB); |
988 | MIB.setDebugLoc(OldDbgLoc); |
989 | } |
990 | |
991 | bool IRTranslator::lowerJumpTableWorkItem(SwitchCG::SwitchWorkListItem W, |
992 | MachineBasicBlock *SwitchMBB, |
993 | MachineBasicBlock *CurMBB, |
994 | MachineBasicBlock *DefaultMBB, |
995 | MachineIRBuilder &MIB, |
996 | MachineFunction::iterator BBI, |
997 | BranchProbability UnhandledProbs, |
998 | SwitchCG::CaseClusterIt I, |
999 | MachineBasicBlock *Fallthrough, |
1000 | bool FallthroughUnreachable) { |
1001 | using namespace SwitchCG; |
1002 | MachineFunction *CurMF = SwitchMBB->getParent(); |
1003 | // FIXME: Optimize away range check based on pivot comparisons. |
1004 | JumpTableHeader *JTH = &SL->JTCases[I->JTCasesIndex].first; |
1005 | SwitchCG::JumpTable *JT = &SL->JTCases[I->JTCasesIndex].second; |
1006 | BranchProbability DefaultProb = W.DefaultProb; |
1007 | |
1008 | // The jump block hasn't been inserted yet; insert it here. |
1009 | MachineBasicBlock *JumpMBB = JT->MBB; |
1010 | CurMF->insert(MBBI: BBI, MBB: JumpMBB); |
1011 | |
1012 | // Since the jump table block is separate from the switch block, we need |
1013 | // to keep track of it as a machine predecessor to the default block, |
1014 | // otherwise we lose the phi edges. |
1015 | addMachineCFGPred(Edge: {SwitchMBB->getBasicBlock(), DefaultMBB->getBasicBlock()}, |
1016 | NewPred: CurMBB); |
1017 | addMachineCFGPred(Edge: {SwitchMBB->getBasicBlock(), DefaultMBB->getBasicBlock()}, |
1018 | NewPred: JumpMBB); |
1019 | |
1020 | auto JumpProb = I->Prob; |
1021 | auto FallthroughProb = UnhandledProbs; |
1022 | |
1023 | // If the default statement is a target of the jump table, we evenly |
1024 | // distribute the default probability to successors of CurMBB. Also |
1025 | // update the probability on the edge from JumpMBB to Fallthrough. |
1026 | for (MachineBasicBlock::succ_iterator SI = JumpMBB->succ_begin(), |
1027 | SE = JumpMBB->succ_end(); |
1028 | SI != SE; ++SI) { |
1029 | if (*SI == DefaultMBB) { |
1030 | JumpProb += DefaultProb / 2; |
1031 | FallthroughProb -= DefaultProb / 2; |
1032 | JumpMBB->setSuccProbability(I: SI, Prob: DefaultProb / 2); |
1033 | JumpMBB->normalizeSuccProbs(); |
1034 | } else { |
1035 | // Also record edges from the jump table block to it's successors. |
1036 | addMachineCFGPred(Edge: {SwitchMBB->getBasicBlock(), (*SI)->getBasicBlock()}, |
1037 | NewPred: JumpMBB); |
1038 | } |
1039 | } |
1040 | |
1041 | if (FallthroughUnreachable) |
1042 | JTH->FallthroughUnreachable = true; |
1043 | |
1044 | if (!JTH->FallthroughUnreachable) |
1045 | addSuccessorWithProb(Src: CurMBB, Dst: Fallthrough, Prob: FallthroughProb); |
1046 | addSuccessorWithProb(Src: CurMBB, Dst: JumpMBB, Prob: JumpProb); |
1047 | CurMBB->normalizeSuccProbs(); |
1048 | |
1049 | // The jump table header will be inserted in our current block, do the |
1050 | // range check, and fall through to our fallthrough block. |
1051 | JTH->HeaderBB = CurMBB; |
1052 | JT->Default = Fallthrough; // FIXME: Move Default to JumpTableHeader. |
1053 | |
1054 | // If we're in the right place, emit the jump table header right now. |
1055 | if (CurMBB == SwitchMBB) { |
1056 | if (!emitJumpTableHeader(JT&: *JT, JTH&: *JTH, HeaderBB: CurMBB)) |
1057 | return false; |
1058 | JTH->Emitted = true; |
1059 | } |
1060 | return true; |
1061 | } |
1062 | bool IRTranslator::lowerSwitchRangeWorkItem(SwitchCG::CaseClusterIt I, |
1063 | Value *Cond, |
1064 | MachineBasicBlock *Fallthrough, |
1065 | bool FallthroughUnreachable, |
1066 | BranchProbability UnhandledProbs, |
1067 | MachineBasicBlock *CurMBB, |
1068 | MachineIRBuilder &MIB, |
1069 | MachineBasicBlock *SwitchMBB) { |
1070 | using namespace SwitchCG; |
1071 | const Value *RHS, *LHS, *MHS; |
1072 | CmpInst::Predicate Pred; |
1073 | if (I->Low == I->High) { |
1074 | // Check Cond == I->Low. |
1075 | Pred = CmpInst::ICMP_EQ; |
1076 | LHS = Cond; |
1077 | RHS = I->Low; |
1078 | MHS = nullptr; |
1079 | } else { |
1080 | // Check I->Low <= Cond <= I->High. |
1081 | Pred = CmpInst::ICMP_SLE; |
1082 | LHS = I->Low; |
1083 | MHS = Cond; |
1084 | RHS = I->High; |
1085 | } |
1086 | |
1087 | // If Fallthrough is unreachable, fold away the comparison. |
1088 | // The false probability is the sum of all unhandled cases. |
1089 | CaseBlock CB(Pred, FallthroughUnreachable, LHS, RHS, MHS, I->MBB, Fallthrough, |
1090 | CurMBB, MIB.getDebugLoc(), I->Prob, UnhandledProbs); |
1091 | |
1092 | emitSwitchCase(CB, SwitchBB: SwitchMBB, MIB); |
1093 | return true; |
1094 | } |
1095 | |
1096 | void IRTranslator::(SwitchCG::BitTestBlock &B, |
1097 | MachineBasicBlock *SwitchBB) { |
1098 | MachineIRBuilder &MIB = *CurBuilder; |
1099 | MIB.setMBB(*SwitchBB); |
1100 | |
1101 | // Subtract the minimum value. |
1102 | Register SwitchOpReg = getOrCreateVReg(Val: *B.SValue); |
1103 | |
1104 | LLT SwitchOpTy = MRI->getType(Reg: SwitchOpReg); |
1105 | Register MinValReg = MIB.buildConstant(Res: SwitchOpTy, Val: B.First).getReg(Idx: 0); |
1106 | auto RangeSub = MIB.buildSub(Dst: SwitchOpTy, Src0: SwitchOpReg, Src1: MinValReg); |
1107 | |
1108 | Type *PtrIRTy = PointerType::getUnqual(C&: MF->getFunction().getContext()); |
1109 | const LLT PtrTy = getLLTForType(Ty&: *PtrIRTy, DL: *DL); |
1110 | |
1111 | LLT MaskTy = SwitchOpTy; |
1112 | if (MaskTy.getSizeInBits() > PtrTy.getSizeInBits() || |
1113 | !llvm::has_single_bit<uint32_t>(Value: MaskTy.getSizeInBits())) |
1114 | MaskTy = LLT::scalar(SizeInBits: PtrTy.getSizeInBits()); |
1115 | else { |
1116 | // Ensure that the type will fit the mask value. |
1117 | for (const SwitchCG::BitTestCase &Case : B.Cases) { |
1118 | if (!isUIntN(N: SwitchOpTy.getSizeInBits(), x: Case.Mask)) { |
1119 | // Switch table case range are encoded into series of masks. |
1120 | // Just use pointer type, it's guaranteed to fit. |
1121 | MaskTy = LLT::scalar(SizeInBits: PtrTy.getSizeInBits()); |
1122 | break; |
1123 | } |
1124 | } |
1125 | } |
1126 | Register SubReg = RangeSub.getReg(Idx: 0); |
1127 | if (SwitchOpTy != MaskTy) |
1128 | SubReg = MIB.buildZExtOrTrunc(Res: MaskTy, Op: SubReg).getReg(Idx: 0); |
1129 | |
1130 | B.RegVT = getMVTForLLT(Ty: MaskTy); |
1131 | B.Reg = SubReg; |
1132 | |
1133 | MachineBasicBlock *MBB = B.Cases[0].ThisBB; |
1134 | |
1135 | if (!B.FallthroughUnreachable) |
1136 | addSuccessorWithProb(Src: SwitchBB, Dst: B.Default, Prob: B.DefaultProb); |
1137 | addSuccessorWithProb(Src: SwitchBB, Dst: MBB, Prob: B.Prob); |
1138 | |
1139 | SwitchBB->normalizeSuccProbs(); |
1140 | |
1141 | if (!B.FallthroughUnreachable) { |
1142 | // Conditional branch to the default block. |
1143 | auto RangeCst = MIB.buildConstant(Res: SwitchOpTy, Val: B.Range); |
1144 | auto RangeCmp = MIB.buildICmp(Pred: CmpInst::Predicate::ICMP_UGT, Res: LLT::scalar(SizeInBits: 1), |
1145 | Op0: RangeSub, Op1: RangeCst); |
1146 | MIB.buildBrCond(Tst: RangeCmp, Dest&: *B.Default); |
1147 | } |
1148 | |
1149 | // Avoid emitting unnecessary branches to the next block. |
1150 | if (MBB != SwitchBB->getNextNode()) |
1151 | MIB.buildBr(Dest&: *MBB); |
1152 | } |
1153 | |
1154 | void IRTranslator::emitBitTestCase(SwitchCG::BitTestBlock &BB, |
1155 | MachineBasicBlock *NextMBB, |
1156 | BranchProbability BranchProbToNext, |
1157 | Register Reg, SwitchCG::BitTestCase &B, |
1158 | MachineBasicBlock *SwitchBB) { |
1159 | MachineIRBuilder &MIB = *CurBuilder; |
1160 | MIB.setMBB(*SwitchBB); |
1161 | |
1162 | LLT SwitchTy = getLLTForMVT(Ty: BB.RegVT); |
1163 | Register Cmp; |
1164 | unsigned PopCount = llvm::popcount(Value: B.Mask); |
1165 | if (PopCount == 1) { |
1166 | // Testing for a single bit; just compare the shift count with what it |
1167 | // would need to be to shift a 1 bit in that position. |
1168 | auto MaskTrailingZeros = |
1169 | MIB.buildConstant(Res: SwitchTy, Val: llvm::countr_zero(Val: B.Mask)); |
1170 | Cmp = |
1171 | MIB.buildICmp(Pred: ICmpInst::ICMP_EQ, Res: LLT::scalar(SizeInBits: 1), Op0: Reg, Op1: MaskTrailingZeros) |
1172 | .getReg(Idx: 0); |
1173 | } else if (PopCount == BB.Range) { |
1174 | // There is only one zero bit in the range, test for it directly. |
1175 | auto MaskTrailingOnes = |
1176 | MIB.buildConstant(Res: SwitchTy, Val: llvm::countr_one(Value: B.Mask)); |
1177 | Cmp = MIB.buildICmp(Pred: CmpInst::ICMP_NE, Res: LLT::scalar(SizeInBits: 1), Op0: Reg, Op1: MaskTrailingOnes) |
1178 | .getReg(Idx: 0); |
1179 | } else { |
1180 | // Make desired shift. |
1181 | auto CstOne = MIB.buildConstant(Res: SwitchTy, Val: 1); |
1182 | auto SwitchVal = MIB.buildShl(Dst: SwitchTy, Src0: CstOne, Src1: Reg); |
1183 | |
1184 | // Emit bit tests and jumps. |
1185 | auto CstMask = MIB.buildConstant(Res: SwitchTy, Val: B.Mask); |
1186 | auto AndOp = MIB.buildAnd(Dst: SwitchTy, Src0: SwitchVal, Src1: CstMask); |
1187 | auto CstZero = MIB.buildConstant(Res: SwitchTy, Val: 0); |
1188 | Cmp = MIB.buildICmp(Pred: CmpInst::ICMP_NE, Res: LLT::scalar(SizeInBits: 1), Op0: AndOp, Op1: CstZero) |
1189 | .getReg(Idx: 0); |
1190 | } |
1191 | |
1192 | // The branch probability from SwitchBB to B.TargetBB is B.ExtraProb. |
1193 | addSuccessorWithProb(Src: SwitchBB, Dst: B.TargetBB, Prob: B.ExtraProb); |
1194 | // The branch probability from SwitchBB to NextMBB is BranchProbToNext. |
1195 | addSuccessorWithProb(Src: SwitchBB, Dst: NextMBB, Prob: BranchProbToNext); |
1196 | // It is not guaranteed that the sum of B.ExtraProb and BranchProbToNext is |
1197 | // one as they are relative probabilities (and thus work more like weights), |
1198 | // and hence we need to normalize them to let the sum of them become one. |
1199 | SwitchBB->normalizeSuccProbs(); |
1200 | |
1201 | // Record the fact that the IR edge from the header to the bit test target |
1202 | // will go through our new block. Neeeded for PHIs to have nodes added. |
1203 | addMachineCFGPred(Edge: {BB.Parent->getBasicBlock(), B.TargetBB->getBasicBlock()}, |
1204 | NewPred: SwitchBB); |
1205 | |
1206 | MIB.buildBrCond(Tst: Cmp, Dest&: *B.TargetBB); |
1207 | |
1208 | // Avoid emitting unnecessary branches to the next block. |
1209 | if (NextMBB != SwitchBB->getNextNode()) |
1210 | MIB.buildBr(Dest&: *NextMBB); |
1211 | } |
1212 | |
1213 | bool IRTranslator::lowerBitTestWorkItem( |
1214 | SwitchCG::SwitchWorkListItem W, MachineBasicBlock *SwitchMBB, |
1215 | MachineBasicBlock *CurMBB, MachineBasicBlock *DefaultMBB, |
1216 | MachineIRBuilder &MIB, MachineFunction::iterator BBI, |
1217 | BranchProbability DefaultProb, BranchProbability UnhandledProbs, |
1218 | SwitchCG::CaseClusterIt I, MachineBasicBlock *Fallthrough, |
1219 | bool FallthroughUnreachable) { |
1220 | using namespace SwitchCG; |
1221 | MachineFunction *CurMF = SwitchMBB->getParent(); |
1222 | // FIXME: Optimize away range check based on pivot comparisons. |
1223 | BitTestBlock *BTB = &SL->BitTestCases[I->BTCasesIndex]; |
1224 | // The bit test blocks haven't been inserted yet; insert them here. |
1225 | for (BitTestCase &BTC : BTB->Cases) |
1226 | CurMF->insert(MBBI: BBI, MBB: BTC.ThisBB); |
1227 | |
1228 | // Fill in fields of the BitTestBlock. |
1229 | BTB->Parent = CurMBB; |
1230 | BTB->Default = Fallthrough; |
1231 | |
1232 | BTB->DefaultProb = UnhandledProbs; |
1233 | // If the cases in bit test don't form a contiguous range, we evenly |
1234 | // distribute the probability on the edge to Fallthrough to two |
1235 | // successors of CurMBB. |
1236 | if (!BTB->ContiguousRange) { |
1237 | BTB->Prob += DefaultProb / 2; |
1238 | BTB->DefaultProb -= DefaultProb / 2; |
1239 | } |
1240 | |
1241 | if (FallthroughUnreachable) |
1242 | BTB->FallthroughUnreachable = true; |
1243 | |
1244 | // If we're in the right place, emit the bit test header right now. |
1245 | if (CurMBB == SwitchMBB) { |
1246 | emitBitTestHeader(B&: *BTB, SwitchBB: SwitchMBB); |
1247 | BTB->Emitted = true; |
1248 | } |
1249 | return true; |
1250 | } |
1251 | |
1252 | bool IRTranslator::lowerSwitchWorkItem(SwitchCG::SwitchWorkListItem W, |
1253 | Value *Cond, |
1254 | MachineBasicBlock *SwitchMBB, |
1255 | MachineBasicBlock *DefaultMBB, |
1256 | MachineIRBuilder &MIB) { |
1257 | using namespace SwitchCG; |
1258 | MachineFunction *CurMF = FuncInfo.MF; |
1259 | MachineBasicBlock *NextMBB = nullptr; |
1260 | MachineFunction::iterator BBI(W.MBB); |
1261 | if (++BBI != FuncInfo.MF->end()) |
1262 | NextMBB = &*BBI; |
1263 | |
1264 | if (EnableOpts) { |
1265 | // Here, we order cases by probability so the most likely case will be |
1266 | // checked first. However, two clusters can have the same probability in |
1267 | // which case their relative ordering is non-deterministic. So we use Low |
1268 | // as a tie-breaker as clusters are guaranteed to never overlap. |
1269 | llvm::sort(Start: W.FirstCluster, End: W.LastCluster + 1, |
1270 | Comp: [](const CaseCluster &a, const CaseCluster &b) { |
1271 | return a.Prob != b.Prob |
1272 | ? a.Prob > b.Prob |
1273 | : a.Low->getValue().slt(RHS: b.Low->getValue()); |
1274 | }); |
1275 | |
1276 | // Rearrange the case blocks so that the last one falls through if possible |
1277 | // without changing the order of probabilities. |
1278 | for (CaseClusterIt I = W.LastCluster; I > W.FirstCluster;) { |
1279 | --I; |
1280 | if (I->Prob > W.LastCluster->Prob) |
1281 | break; |
1282 | if (I->Kind == CC_Range && I->MBB == NextMBB) { |
1283 | std::swap(a&: *I, b&: *W.LastCluster); |
1284 | break; |
1285 | } |
1286 | } |
1287 | } |
1288 | |
1289 | // Compute total probability. |
1290 | BranchProbability DefaultProb = W.DefaultProb; |
1291 | BranchProbability UnhandledProbs = DefaultProb; |
1292 | for (CaseClusterIt I = W.FirstCluster; I <= W.LastCluster; ++I) |
1293 | UnhandledProbs += I->Prob; |
1294 | |
1295 | MachineBasicBlock *CurMBB = W.MBB; |
1296 | for (CaseClusterIt I = W.FirstCluster, E = W.LastCluster; I <= E; ++I) { |
1297 | bool FallthroughUnreachable = false; |
1298 | MachineBasicBlock *Fallthrough; |
1299 | if (I == W.LastCluster) { |
1300 | // For the last cluster, fall through to the default destination. |
1301 | Fallthrough = DefaultMBB; |
1302 | FallthroughUnreachable = isa<UnreachableInst>( |
1303 | Val: DefaultMBB->getBasicBlock()->getFirstNonPHIOrDbg()); |
1304 | } else { |
1305 | Fallthrough = CurMF->CreateMachineBasicBlock(BB: CurMBB->getBasicBlock()); |
1306 | CurMF->insert(MBBI: BBI, MBB: Fallthrough); |
1307 | } |
1308 | UnhandledProbs -= I->Prob; |
1309 | |
1310 | switch (I->Kind) { |
1311 | case CC_BitTests: { |
1312 | if (!lowerBitTestWorkItem(W, SwitchMBB, CurMBB, DefaultMBB, MIB, BBI, |
1313 | DefaultProb, UnhandledProbs, I, Fallthrough, |
1314 | FallthroughUnreachable)) { |
1315 | LLVM_DEBUG(dbgs() << "Failed to lower bit test for switch" ); |
1316 | return false; |
1317 | } |
1318 | break; |
1319 | } |
1320 | |
1321 | case CC_JumpTable: { |
1322 | if (!lowerJumpTableWorkItem(W, SwitchMBB, CurMBB, DefaultMBB, MIB, BBI, |
1323 | UnhandledProbs, I, Fallthrough, |
1324 | FallthroughUnreachable)) { |
1325 | LLVM_DEBUG(dbgs() << "Failed to lower jump table" ); |
1326 | return false; |
1327 | } |
1328 | break; |
1329 | } |
1330 | case CC_Range: { |
1331 | if (!lowerSwitchRangeWorkItem(I, Cond, Fallthrough, |
1332 | FallthroughUnreachable, UnhandledProbs, |
1333 | CurMBB, MIB, SwitchMBB)) { |
1334 | LLVM_DEBUG(dbgs() << "Failed to lower switch range" ); |
1335 | return false; |
1336 | } |
1337 | break; |
1338 | } |
1339 | } |
1340 | CurMBB = Fallthrough; |
1341 | } |
1342 | |
1343 | return true; |
1344 | } |
1345 | |
1346 | bool IRTranslator::translateIndirectBr(const User &U, |
1347 | MachineIRBuilder &MIRBuilder) { |
1348 | const IndirectBrInst &BrInst = cast<IndirectBrInst>(Val: U); |
1349 | |
1350 | const Register Tgt = getOrCreateVReg(Val: *BrInst.getAddress()); |
1351 | MIRBuilder.buildBrIndirect(Tgt); |
1352 | |
1353 | // Link successors. |
1354 | SmallPtrSet<const BasicBlock *, 32> AddedSuccessors; |
1355 | MachineBasicBlock &CurBB = MIRBuilder.getMBB(); |
1356 | for (const BasicBlock *Succ : successors(I: &BrInst)) { |
1357 | // It's legal for indirectbr instructions to have duplicate blocks in the |
1358 | // destination list. We don't allow this in MIR. Skip anything that's |
1359 | // already a successor. |
1360 | if (!AddedSuccessors.insert(Ptr: Succ).second) |
1361 | continue; |
1362 | CurBB.addSuccessor(Succ: &getMBB(BB: *Succ)); |
1363 | } |
1364 | |
1365 | return true; |
1366 | } |
1367 | |
1368 | static bool isSwiftError(const Value *V) { |
1369 | if (auto Arg = dyn_cast<Argument>(Val: V)) |
1370 | return Arg->hasSwiftErrorAttr(); |
1371 | if (auto AI = dyn_cast<AllocaInst>(Val: V)) |
1372 | return AI->isSwiftError(); |
1373 | return false; |
1374 | } |
1375 | |
1376 | bool IRTranslator::translateLoad(const User &U, MachineIRBuilder &MIRBuilder) { |
1377 | const LoadInst &LI = cast<LoadInst>(Val: U); |
1378 | TypeSize StoreSize = DL->getTypeStoreSize(Ty: LI.getType()); |
1379 | if (StoreSize.isZero()) |
1380 | return true; |
1381 | |
1382 | ArrayRef<Register> Regs = getOrCreateVRegs(Val: LI); |
1383 | ArrayRef<uint64_t> Offsets = *VMap.getOffsets(V: LI); |
1384 | Register Base = getOrCreateVReg(Val: *LI.getPointerOperand()); |
1385 | AAMDNodes AAInfo = LI.getAAMetadata(); |
1386 | |
1387 | const Value *Ptr = LI.getPointerOperand(); |
1388 | Type *OffsetIRTy = DL->getIndexType(PtrTy: Ptr->getType()); |
1389 | LLT OffsetTy = getLLTForType(Ty&: *OffsetIRTy, DL: *DL); |
1390 | |
1391 | if (CLI->supportSwiftError() && isSwiftError(V: Ptr)) { |
1392 | assert(Regs.size() == 1 && "swifterror should be single pointer" ); |
1393 | Register VReg = |
1394 | SwiftError.getOrCreateVRegUseAt(&LI, &MIRBuilder.getMBB(), Ptr); |
1395 | MIRBuilder.buildCopy(Res: Regs[0], Op: VReg); |
1396 | return true; |
1397 | } |
1398 | |
1399 | MachineMemOperand::Flags Flags = |
1400 | TLI->getLoadMemOperandFlags(LI, DL: *DL, AC, LibInfo); |
1401 | if (AA && !(Flags & MachineMemOperand::MOInvariant)) { |
1402 | if (AA->pointsToConstantMemory( |
1403 | Loc: MemoryLocation(Ptr, LocationSize::precise(Value: StoreSize), AAInfo))) { |
1404 | Flags |= MachineMemOperand::MOInvariant; |
1405 | } |
1406 | } |
1407 | |
1408 | const MDNode *Ranges = |
1409 | Regs.size() == 1 ? LI.getMetadata(KindID: LLVMContext::MD_range) : nullptr; |
1410 | for (unsigned i = 0; i < Regs.size(); ++i) { |
1411 | Register Addr; |
1412 | MIRBuilder.materializePtrAdd(Res&: Addr, Op0: Base, ValueTy: OffsetTy, Value: Offsets[i] / 8); |
1413 | |
1414 | MachinePointerInfo Ptr(LI.getPointerOperand(), Offsets[i] / 8); |
1415 | Align BaseAlign = getMemOpAlign(I: LI); |
1416 | auto MMO = MF->getMachineMemOperand( |
1417 | PtrInfo: Ptr, f: Flags, MemTy: MRI->getType(Reg: Regs[i]), |
1418 | base_alignment: commonAlignment(A: BaseAlign, Offset: Offsets[i] / 8), AAInfo, Ranges, |
1419 | SSID: LI.getSyncScopeID(), Ordering: LI.getOrdering()); |
1420 | MIRBuilder.buildLoad(Res: Regs[i], Addr, MMO&: *MMO); |
1421 | } |
1422 | |
1423 | return true; |
1424 | } |
1425 | |
1426 | bool IRTranslator::translateStore(const User &U, MachineIRBuilder &MIRBuilder) { |
1427 | const StoreInst &SI = cast<StoreInst>(Val: U); |
1428 | if (DL->getTypeStoreSize(Ty: SI.getValueOperand()->getType()).isZero()) |
1429 | return true; |
1430 | |
1431 | ArrayRef<Register> Vals = getOrCreateVRegs(Val: *SI.getValueOperand()); |
1432 | ArrayRef<uint64_t> Offsets = *VMap.getOffsets(V: *SI.getValueOperand()); |
1433 | Register Base = getOrCreateVReg(Val: *SI.getPointerOperand()); |
1434 | |
1435 | Type *OffsetIRTy = DL->getIndexType(PtrTy: SI.getPointerOperandType()); |
1436 | LLT OffsetTy = getLLTForType(Ty&: *OffsetIRTy, DL: *DL); |
1437 | |
1438 | if (CLI->supportSwiftError() && isSwiftError(V: SI.getPointerOperand())) { |
1439 | assert(Vals.size() == 1 && "swifterror should be single pointer" ); |
1440 | |
1441 | Register VReg = SwiftError.getOrCreateVRegDefAt(&SI, &MIRBuilder.getMBB(), |
1442 | SI.getPointerOperand()); |
1443 | MIRBuilder.buildCopy(Res: VReg, Op: Vals[0]); |
1444 | return true; |
1445 | } |
1446 | |
1447 | MachineMemOperand::Flags Flags = TLI->getStoreMemOperandFlags(SI, DL: *DL); |
1448 | |
1449 | for (unsigned i = 0; i < Vals.size(); ++i) { |
1450 | Register Addr; |
1451 | MIRBuilder.materializePtrAdd(Res&: Addr, Op0: Base, ValueTy: OffsetTy, Value: Offsets[i] / 8); |
1452 | |
1453 | MachinePointerInfo Ptr(SI.getPointerOperand(), Offsets[i] / 8); |
1454 | Align BaseAlign = getMemOpAlign(I: SI); |
1455 | auto MMO = MF->getMachineMemOperand( |
1456 | PtrInfo: Ptr, f: Flags, MemTy: MRI->getType(Reg: Vals[i]), |
1457 | base_alignment: commonAlignment(A: BaseAlign, Offset: Offsets[i] / 8), AAInfo: SI.getAAMetadata(), Ranges: nullptr, |
1458 | SSID: SI.getSyncScopeID(), Ordering: SI.getOrdering()); |
1459 | MIRBuilder.buildStore(Val: Vals[i], Addr, MMO&: *MMO); |
1460 | } |
1461 | return true; |
1462 | } |
1463 | |
1464 | static uint64_t getOffsetFromIndices(const User &U, const DataLayout &DL) { |
1465 | const Value *Src = U.getOperand(i: 0); |
1466 | Type *Int32Ty = Type::getInt32Ty(C&: U.getContext()); |
1467 | |
1468 | // getIndexedOffsetInType is designed for GEPs, so the first index is the |
1469 | // usual array element rather than looking into the actual aggregate. |
1470 | SmallVector<Value *, 1> Indices; |
1471 | Indices.push_back(Elt: ConstantInt::get(Ty: Int32Ty, V: 0)); |
1472 | |
1473 | if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Val: &U)) { |
1474 | for (auto Idx : EVI->indices()) |
1475 | Indices.push_back(Elt: ConstantInt::get(Ty: Int32Ty, V: Idx)); |
1476 | } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Val: &U)) { |
1477 | for (auto Idx : IVI->indices()) |
1478 | Indices.push_back(Elt: ConstantInt::get(Ty: Int32Ty, V: Idx)); |
1479 | } else { |
1480 | llvm::append_range(C&: Indices, R: drop_begin(RangeOrContainer: U.operands())); |
1481 | } |
1482 | |
1483 | return 8 * static_cast<uint64_t>( |
1484 | DL.getIndexedOffsetInType(ElemTy: Src->getType(), Indices)); |
1485 | } |
1486 | |
1487 | bool IRTranslator::(const User &U, |
1488 | MachineIRBuilder &MIRBuilder) { |
1489 | const Value *Src = U.getOperand(i: 0); |
1490 | uint64_t Offset = getOffsetFromIndices(U, DL: *DL); |
1491 | ArrayRef<Register> SrcRegs = getOrCreateVRegs(Val: *Src); |
1492 | ArrayRef<uint64_t> Offsets = *VMap.getOffsets(V: *Src); |
1493 | unsigned Idx = llvm::lower_bound(Range&: Offsets, Value&: Offset) - Offsets.begin(); |
1494 | auto &DstRegs = allocateVRegs(Val: U); |
1495 | |
1496 | for (unsigned i = 0; i < DstRegs.size(); ++i) |
1497 | DstRegs[i] = SrcRegs[Idx++]; |
1498 | |
1499 | return true; |
1500 | } |
1501 | |
1502 | bool IRTranslator::translateInsertValue(const User &U, |
1503 | MachineIRBuilder &MIRBuilder) { |
1504 | const Value *Src = U.getOperand(i: 0); |
1505 | uint64_t Offset = getOffsetFromIndices(U, DL: *DL); |
1506 | auto &DstRegs = allocateVRegs(Val: U); |
1507 | ArrayRef<uint64_t> DstOffsets = *VMap.getOffsets(V: U); |
1508 | ArrayRef<Register> SrcRegs = getOrCreateVRegs(Val: *Src); |
1509 | ArrayRef<Register> InsertedRegs = getOrCreateVRegs(Val: *U.getOperand(i: 1)); |
1510 | auto *InsertedIt = InsertedRegs.begin(); |
1511 | |
1512 | for (unsigned i = 0; i < DstRegs.size(); ++i) { |
1513 | if (DstOffsets[i] >= Offset && InsertedIt != InsertedRegs.end()) |
1514 | DstRegs[i] = *InsertedIt++; |
1515 | else |
1516 | DstRegs[i] = SrcRegs[i]; |
1517 | } |
1518 | |
1519 | return true; |
1520 | } |
1521 | |
1522 | bool IRTranslator::translateSelect(const User &U, |
1523 | MachineIRBuilder &MIRBuilder) { |
1524 | Register Tst = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
1525 | ArrayRef<Register> ResRegs = getOrCreateVRegs(Val: U); |
1526 | ArrayRef<Register> Op0Regs = getOrCreateVRegs(Val: *U.getOperand(i: 1)); |
1527 | ArrayRef<Register> Op1Regs = getOrCreateVRegs(Val: *U.getOperand(i: 2)); |
1528 | |
1529 | uint32_t Flags = 0; |
1530 | if (const SelectInst *SI = dyn_cast<SelectInst>(Val: &U)) |
1531 | Flags = MachineInstr::copyFlagsFromInstruction(I: *SI); |
1532 | |
1533 | for (unsigned i = 0; i < ResRegs.size(); ++i) { |
1534 | MIRBuilder.buildSelect(Res: ResRegs[i], Tst, Op0: Op0Regs[i], Op1: Op1Regs[i], Flags); |
1535 | } |
1536 | |
1537 | return true; |
1538 | } |
1539 | |
1540 | bool IRTranslator::translateCopy(const User &U, const Value &V, |
1541 | MachineIRBuilder &MIRBuilder) { |
1542 | Register Src = getOrCreateVReg(Val: V); |
1543 | auto &Regs = *VMap.getVRegs(V: U); |
1544 | if (Regs.empty()) { |
1545 | Regs.push_back(Elt: Src); |
1546 | VMap.getOffsets(V: U)->push_back(Elt: 0); |
1547 | } else { |
1548 | // If we already assigned a vreg for this instruction, we can't change that. |
1549 | // Emit a copy to satisfy the users we already emitted. |
1550 | MIRBuilder.buildCopy(Res: Regs[0], Op: Src); |
1551 | } |
1552 | return true; |
1553 | } |
1554 | |
1555 | bool IRTranslator::translateBitCast(const User &U, |
1556 | MachineIRBuilder &MIRBuilder) { |
1557 | // If we're bitcasting to the source type, we can reuse the source vreg. |
1558 | if (getLLTForType(Ty&: *U.getOperand(i: 0)->getType(), DL: *DL) == |
1559 | getLLTForType(Ty&: *U.getType(), DL: *DL)) { |
1560 | // If the source is a ConstantInt then it was probably created by |
1561 | // ConstantHoisting and we should leave it alone. |
1562 | if (isa<ConstantInt>(Val: U.getOperand(i: 0))) |
1563 | return translateCast(Opcode: TargetOpcode::G_CONSTANT_FOLD_BARRIER, U, |
1564 | MIRBuilder); |
1565 | return translateCopy(U, V: *U.getOperand(i: 0), MIRBuilder); |
1566 | } |
1567 | |
1568 | return translateCast(Opcode: TargetOpcode::G_BITCAST, U, MIRBuilder); |
1569 | } |
1570 | |
1571 | bool IRTranslator::translateCast(unsigned Opcode, const User &U, |
1572 | MachineIRBuilder &MIRBuilder) { |
1573 | if (containsBF16Type(U)) |
1574 | return false; |
1575 | |
1576 | uint32_t Flags = 0; |
1577 | if (const Instruction *I = dyn_cast<Instruction>(Val: &U)) |
1578 | Flags = MachineInstr::copyFlagsFromInstruction(I: *I); |
1579 | |
1580 | Register Op = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
1581 | Register Res = getOrCreateVReg(Val: U); |
1582 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {Res}, SrcOps: {Op}, Flags); |
1583 | return true; |
1584 | } |
1585 | |
1586 | bool IRTranslator::translateGetElementPtr(const User &U, |
1587 | MachineIRBuilder &MIRBuilder) { |
1588 | Value &Op0 = *U.getOperand(i: 0); |
1589 | Register BaseReg = getOrCreateVReg(Val: Op0); |
1590 | Type *PtrIRTy = Op0.getType(); |
1591 | LLT PtrTy = getLLTForType(Ty&: *PtrIRTy, DL: *DL); |
1592 | Type *OffsetIRTy = DL->getIndexType(PtrTy: PtrIRTy); |
1593 | LLT OffsetTy = getLLTForType(Ty&: *OffsetIRTy, DL: *DL); |
1594 | |
1595 | uint32_t Flags = 0; |
1596 | if (const Instruction *I = dyn_cast<Instruction>(Val: &U)) |
1597 | Flags = MachineInstr::copyFlagsFromInstruction(I: *I); |
1598 | |
1599 | // Normalize Vector GEP - all scalar operands should be converted to the |
1600 | // splat vector. |
1601 | unsigned VectorWidth = 0; |
1602 | |
1603 | // True if we should use a splat vector; using VectorWidth alone is not |
1604 | // sufficient. |
1605 | bool WantSplatVector = false; |
1606 | if (auto *VT = dyn_cast<VectorType>(Val: U.getType())) { |
1607 | VectorWidth = cast<FixedVectorType>(Val: VT)->getNumElements(); |
1608 | // We don't produce 1 x N vectors; those are treated as scalars. |
1609 | WantSplatVector = VectorWidth > 1; |
1610 | } |
1611 | |
1612 | // We might need to splat the base pointer into a vector if the offsets |
1613 | // are vectors. |
1614 | if (WantSplatVector && !PtrTy.isVector()) { |
1615 | BaseReg = MIRBuilder |
1616 | .buildSplatBuildVector(Res: LLT::fixed_vector(NumElements: VectorWidth, ScalarTy: PtrTy), |
1617 | Src: BaseReg) |
1618 | .getReg(Idx: 0); |
1619 | PtrIRTy = FixedVectorType::get(ElementType: PtrIRTy, NumElts: VectorWidth); |
1620 | PtrTy = getLLTForType(Ty&: *PtrIRTy, DL: *DL); |
1621 | OffsetIRTy = DL->getIndexType(PtrTy: PtrIRTy); |
1622 | OffsetTy = getLLTForType(Ty&: *OffsetIRTy, DL: *DL); |
1623 | } |
1624 | |
1625 | int64_t Offset = 0; |
1626 | for (gep_type_iterator GTI = gep_type_begin(GEP: &U), E = gep_type_end(GEP: &U); |
1627 | GTI != E; ++GTI) { |
1628 | const Value *Idx = GTI.getOperand(); |
1629 | if (StructType *StTy = GTI.getStructTypeOrNull()) { |
1630 | unsigned Field = cast<Constant>(Val: Idx)->getUniqueInteger().getZExtValue(); |
1631 | Offset += DL->getStructLayout(Ty: StTy)->getElementOffset(Idx: Field); |
1632 | continue; |
1633 | } else { |
1634 | uint64_t ElementSize = GTI.getSequentialElementStride(DL: *DL); |
1635 | |
1636 | // If this is a scalar constant or a splat vector of constants, |
1637 | // handle it quickly. |
1638 | if (const auto *CI = dyn_cast<ConstantInt>(Val: Idx)) { |
1639 | if (std::optional<int64_t> Val = CI->getValue().trySExtValue()) { |
1640 | Offset += ElementSize * *Val; |
1641 | continue; |
1642 | } |
1643 | } |
1644 | |
1645 | if (Offset != 0) { |
1646 | auto OffsetMIB = MIRBuilder.buildConstant(Res: {OffsetTy}, Val: Offset); |
1647 | BaseReg = MIRBuilder.buildPtrAdd(Res: PtrTy, Op0: BaseReg, Op1: OffsetMIB.getReg(Idx: 0)) |
1648 | .getReg(Idx: 0); |
1649 | Offset = 0; |
1650 | } |
1651 | |
1652 | Register IdxReg = getOrCreateVReg(Val: *Idx); |
1653 | LLT IdxTy = MRI->getType(Reg: IdxReg); |
1654 | if (IdxTy != OffsetTy) { |
1655 | if (!IdxTy.isVector() && WantSplatVector) { |
1656 | IdxReg = MIRBuilder |
1657 | .buildSplatBuildVector(Res: OffsetTy.changeElementType(NewEltTy: IdxTy), |
1658 | Src: IdxReg) |
1659 | .getReg(Idx: 0); |
1660 | } |
1661 | |
1662 | IdxReg = MIRBuilder.buildSExtOrTrunc(Res: OffsetTy, Op: IdxReg).getReg(Idx: 0); |
1663 | } |
1664 | |
1665 | // N = N + Idx * ElementSize; |
1666 | // Avoid doing it for ElementSize of 1. |
1667 | Register GepOffsetReg; |
1668 | if (ElementSize != 1) { |
1669 | auto ElementSizeMIB = MIRBuilder.buildConstant( |
1670 | Res: getLLTForType(Ty&: *OffsetIRTy, DL: *DL), Val: ElementSize); |
1671 | GepOffsetReg = |
1672 | MIRBuilder.buildMul(Dst: OffsetTy, Src0: IdxReg, Src1: ElementSizeMIB).getReg(Idx: 0); |
1673 | } else |
1674 | GepOffsetReg = IdxReg; |
1675 | |
1676 | BaseReg = MIRBuilder.buildPtrAdd(Res: PtrTy, Op0: BaseReg, Op1: GepOffsetReg).getReg(Idx: 0); |
1677 | } |
1678 | } |
1679 | |
1680 | if (Offset != 0) { |
1681 | auto OffsetMIB = |
1682 | MIRBuilder.buildConstant(Res: OffsetTy, Val: Offset); |
1683 | |
1684 | if (int64_t(Offset) >= 0 && cast<GEPOperator>(Val: U).isInBounds()) |
1685 | Flags |= MachineInstr::MIFlag::NoUWrap; |
1686 | |
1687 | MIRBuilder.buildPtrAdd(Res: getOrCreateVReg(Val: U), Op0: BaseReg, Op1: OffsetMIB.getReg(Idx: 0), |
1688 | Flags); |
1689 | return true; |
1690 | } |
1691 | |
1692 | MIRBuilder.buildCopy(Res: getOrCreateVReg(Val: U), Op: BaseReg); |
1693 | return true; |
1694 | } |
1695 | |
1696 | bool IRTranslator::translateMemFunc(const CallInst &CI, |
1697 | MachineIRBuilder &MIRBuilder, |
1698 | unsigned Opcode) { |
1699 | const Value *SrcPtr = CI.getArgOperand(i: 1); |
1700 | // If the source is undef, then just emit a nop. |
1701 | if (isa<UndefValue>(Val: SrcPtr)) |
1702 | return true; |
1703 | |
1704 | SmallVector<Register, 3> SrcRegs; |
1705 | |
1706 | unsigned MinPtrSize = UINT_MAX; |
1707 | for (auto AI = CI.arg_begin(), AE = CI.arg_end(); std::next(x: AI) != AE; ++AI) { |
1708 | Register SrcReg = getOrCreateVReg(Val: **AI); |
1709 | LLT SrcTy = MRI->getType(Reg: SrcReg); |
1710 | if (SrcTy.isPointer()) |
1711 | MinPtrSize = std::min<unsigned>(a: SrcTy.getSizeInBits(), b: MinPtrSize); |
1712 | SrcRegs.push_back(Elt: SrcReg); |
1713 | } |
1714 | |
1715 | LLT SizeTy = LLT::scalar(SizeInBits: MinPtrSize); |
1716 | |
1717 | // The size operand should be the minimum of the pointer sizes. |
1718 | Register &SizeOpReg = SrcRegs[SrcRegs.size() - 1]; |
1719 | if (MRI->getType(Reg: SizeOpReg) != SizeTy) |
1720 | SizeOpReg = MIRBuilder.buildZExtOrTrunc(Res: SizeTy, Op: SizeOpReg).getReg(Idx: 0); |
1721 | |
1722 | auto ICall = MIRBuilder.buildInstr(Opcode); |
1723 | for (Register SrcReg : SrcRegs) |
1724 | ICall.addUse(RegNo: SrcReg); |
1725 | |
1726 | Align DstAlign; |
1727 | Align SrcAlign; |
1728 | unsigned IsVol = |
1729 | cast<ConstantInt>(Val: CI.getArgOperand(i: CI.arg_size() - 1))->getZExtValue(); |
1730 | |
1731 | ConstantInt *CopySize = nullptr; |
1732 | |
1733 | if (auto *MCI = dyn_cast<MemCpyInst>(Val: &CI)) { |
1734 | DstAlign = MCI->getDestAlign().valueOrOne(); |
1735 | SrcAlign = MCI->getSourceAlign().valueOrOne(); |
1736 | CopySize = dyn_cast<ConstantInt>(Val: MCI->getArgOperand(i: 2)); |
1737 | } else if (auto *MMI = dyn_cast<MemMoveInst>(Val: &CI)) { |
1738 | DstAlign = MMI->getDestAlign().valueOrOne(); |
1739 | SrcAlign = MMI->getSourceAlign().valueOrOne(); |
1740 | CopySize = dyn_cast<ConstantInt>(Val: MMI->getArgOperand(i: 2)); |
1741 | } else { |
1742 | auto *MSI = cast<MemSetInst>(Val: &CI); |
1743 | DstAlign = MSI->getDestAlign().valueOrOne(); |
1744 | } |
1745 | |
1746 | if (Opcode != TargetOpcode::G_MEMCPY_INLINE) { |
1747 | // We need to propagate the tail call flag from the IR inst as an argument. |
1748 | // Otherwise, we have to pessimize and assume later that we cannot tail call |
1749 | // any memory intrinsics. |
1750 | ICall.addImm(Val: CI.isTailCall() ? 1 : 0); |
1751 | } |
1752 | |
1753 | // Create mem operands to store the alignment and volatile info. |
1754 | MachineMemOperand::Flags LoadFlags = MachineMemOperand::MOLoad; |
1755 | MachineMemOperand::Flags StoreFlags = MachineMemOperand::MOStore; |
1756 | if (IsVol) { |
1757 | LoadFlags |= MachineMemOperand::MOVolatile; |
1758 | StoreFlags |= MachineMemOperand::MOVolatile; |
1759 | } |
1760 | |
1761 | AAMDNodes AAInfo = CI.getAAMetadata(); |
1762 | if (AA && CopySize && |
1763 | AA->pointsToConstantMemory(Loc: MemoryLocation( |
1764 | SrcPtr, LocationSize::precise(Value: CopySize->getZExtValue()), AAInfo))) { |
1765 | LoadFlags |= MachineMemOperand::MOInvariant; |
1766 | |
1767 | // FIXME: pointsToConstantMemory probably does not imply dereferenceable, |
1768 | // but the previous usage implied it did. Probably should check |
1769 | // isDereferenceableAndAlignedPointer. |
1770 | LoadFlags |= MachineMemOperand::MODereferenceable; |
1771 | } |
1772 | |
1773 | ICall.addMemOperand( |
1774 | MMO: MF->getMachineMemOperand(PtrInfo: MachinePointerInfo(CI.getArgOperand(i: 0)), |
1775 | F: StoreFlags, Size: 1, BaseAlignment: DstAlign, AAInfo)); |
1776 | if (Opcode != TargetOpcode::G_MEMSET) |
1777 | ICall.addMemOperand(MMO: MF->getMachineMemOperand( |
1778 | PtrInfo: MachinePointerInfo(SrcPtr), F: LoadFlags, Size: 1, BaseAlignment: SrcAlign, AAInfo)); |
1779 | |
1780 | return true; |
1781 | } |
1782 | |
1783 | bool IRTranslator::translateTrap(const CallInst &CI, |
1784 | MachineIRBuilder &MIRBuilder, |
1785 | unsigned Opcode) { |
1786 | StringRef TrapFuncName = |
1787 | CI.getAttributes().getFnAttr(Kind: "trap-func-name" ).getValueAsString(); |
1788 | if (TrapFuncName.empty()) { |
1789 | if (Opcode == TargetOpcode::G_UBSANTRAP) { |
1790 | uint64_t Code = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 0))->getZExtValue(); |
1791 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {}, SrcOps: ArrayRef<llvm::SrcOp>{Code}); |
1792 | } else { |
1793 | MIRBuilder.buildInstr(Opcode); |
1794 | } |
1795 | return true; |
1796 | } |
1797 | |
1798 | CallLowering::CallLoweringInfo Info; |
1799 | if (Opcode == TargetOpcode::G_UBSANTRAP) |
1800 | Info.OrigArgs.push_back(Elt: {getOrCreateVRegs(Val: *CI.getArgOperand(i: 0)), |
1801 | CI.getArgOperand(i: 0)->getType(), 0}); |
1802 | |
1803 | Info.Callee = MachineOperand::CreateES(SymName: TrapFuncName.data()); |
1804 | Info.CB = &CI; |
1805 | Info.OrigRet = {Register(), Type::getVoidTy(C&: CI.getContext()), 0}; |
1806 | return CLI->lowerCall(MIRBuilder, Info); |
1807 | } |
1808 | |
1809 | bool IRTranslator::translateVectorInterleave2Intrinsic( |
1810 | const CallInst &CI, MachineIRBuilder &MIRBuilder) { |
1811 | assert(CI.getIntrinsicID() == Intrinsic::vector_interleave2 && |
1812 | "This function can only be called on the interleave2 intrinsic!" ); |
1813 | // Canonicalize interleave2 to G_SHUFFLE_VECTOR (similar to SelectionDAG). |
1814 | Register Op0 = getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)); |
1815 | Register Op1 = getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 1)); |
1816 | Register Res = getOrCreateVReg(Val: CI); |
1817 | |
1818 | LLT OpTy = MRI->getType(Reg: Op0); |
1819 | MIRBuilder.buildShuffleVector(Res, Src1: Op0, Src2: Op1, |
1820 | Mask: createInterleaveMask(VF: OpTy.getNumElements(), NumVecs: 2)); |
1821 | |
1822 | return true; |
1823 | } |
1824 | |
1825 | bool IRTranslator::translateVectorDeinterleave2Intrinsic( |
1826 | const CallInst &CI, MachineIRBuilder &MIRBuilder) { |
1827 | assert(CI.getIntrinsicID() == Intrinsic::vector_deinterleave2 && |
1828 | "This function can only be called on the deinterleave2 intrinsic!" ); |
1829 | // Canonicalize deinterleave2 to shuffles that extract sub-vectors (similar to |
1830 | // SelectionDAG). |
1831 | Register Op = getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)); |
1832 | auto Undef = MIRBuilder.buildUndef(Res: MRI->getType(Reg: Op)); |
1833 | ArrayRef<Register> Res = getOrCreateVRegs(Val: CI); |
1834 | |
1835 | LLT ResTy = MRI->getType(Reg: Res[0]); |
1836 | MIRBuilder.buildShuffleVector(Res: Res[0], Src1: Op, Src2: Undef, |
1837 | Mask: createStrideMask(Start: 0, Stride: 2, VF: ResTy.getNumElements())); |
1838 | MIRBuilder.buildShuffleVector(Res: Res[1], Src1: Op, Src2: Undef, |
1839 | Mask: createStrideMask(Start: 1, Stride: 2, VF: ResTy.getNumElements())); |
1840 | |
1841 | return true; |
1842 | } |
1843 | |
1844 | void IRTranslator::getStackGuard(Register DstReg, |
1845 | MachineIRBuilder &MIRBuilder) { |
1846 | const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); |
1847 | MRI->setRegClass(Reg: DstReg, RC: TRI->getPointerRegClass(MF: *MF)); |
1848 | auto MIB = |
1849 | MIRBuilder.buildInstr(Opc: TargetOpcode::LOAD_STACK_GUARD, DstOps: {DstReg}, SrcOps: {}); |
1850 | |
1851 | Value *Global = TLI->getSDagStackGuard(M: *MF->getFunction().getParent()); |
1852 | if (!Global) |
1853 | return; |
1854 | |
1855 | unsigned AddrSpace = Global->getType()->getPointerAddressSpace(); |
1856 | LLT PtrTy = LLT::pointer(AddressSpace: AddrSpace, SizeInBits: DL->getPointerSizeInBits(AS: AddrSpace)); |
1857 | |
1858 | MachinePointerInfo MPInfo(Global); |
1859 | auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant | |
1860 | MachineMemOperand::MODereferenceable; |
1861 | MachineMemOperand *MemRef = MF->getMachineMemOperand( |
1862 | PtrInfo: MPInfo, f: Flags, MemTy: PtrTy, base_alignment: DL->getPointerABIAlignment(AS: AddrSpace)); |
1863 | MIB.setMemRefs({MemRef}); |
1864 | } |
1865 | |
1866 | bool IRTranslator::translateOverflowIntrinsic(const CallInst &CI, unsigned Op, |
1867 | MachineIRBuilder &MIRBuilder) { |
1868 | ArrayRef<Register> ResRegs = getOrCreateVRegs(Val: CI); |
1869 | MIRBuilder.buildInstr( |
1870 | Opc: Op, DstOps: {ResRegs[0], ResRegs[1]}, |
1871 | SrcOps: {getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)), getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 1))}); |
1872 | |
1873 | return true; |
1874 | } |
1875 | |
1876 | bool IRTranslator::translateFixedPointIntrinsic(unsigned Op, const CallInst &CI, |
1877 | MachineIRBuilder &MIRBuilder) { |
1878 | Register Dst = getOrCreateVReg(Val: CI); |
1879 | Register Src0 = getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)); |
1880 | Register Src1 = getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 1)); |
1881 | uint64_t Scale = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 2))->getZExtValue(); |
1882 | MIRBuilder.buildInstr(Opc: Op, DstOps: {Dst}, SrcOps: { Src0, Src1, Scale }); |
1883 | return true; |
1884 | } |
1885 | |
1886 | unsigned IRTranslator::getSimpleIntrinsicOpcode(Intrinsic::ID ID) { |
1887 | switch (ID) { |
1888 | default: |
1889 | break; |
1890 | case Intrinsic::acos: |
1891 | return TargetOpcode::G_FACOS; |
1892 | case Intrinsic::asin: |
1893 | return TargetOpcode::G_FASIN; |
1894 | case Intrinsic::atan: |
1895 | return TargetOpcode::G_FATAN; |
1896 | case Intrinsic::atan2: |
1897 | return TargetOpcode::G_FATAN2; |
1898 | case Intrinsic::bswap: |
1899 | return TargetOpcode::G_BSWAP; |
1900 | case Intrinsic::bitreverse: |
1901 | return TargetOpcode::G_BITREVERSE; |
1902 | case Intrinsic::fshl: |
1903 | return TargetOpcode::G_FSHL; |
1904 | case Intrinsic::fshr: |
1905 | return TargetOpcode::G_FSHR; |
1906 | case Intrinsic::ceil: |
1907 | return TargetOpcode::G_FCEIL; |
1908 | case Intrinsic::cos: |
1909 | return TargetOpcode::G_FCOS; |
1910 | case Intrinsic::cosh: |
1911 | return TargetOpcode::G_FCOSH; |
1912 | case Intrinsic::ctpop: |
1913 | return TargetOpcode::G_CTPOP; |
1914 | case Intrinsic::exp: |
1915 | return TargetOpcode::G_FEXP; |
1916 | case Intrinsic::exp2: |
1917 | return TargetOpcode::G_FEXP2; |
1918 | case Intrinsic::exp10: |
1919 | return TargetOpcode::G_FEXP10; |
1920 | case Intrinsic::fabs: |
1921 | return TargetOpcode::G_FABS; |
1922 | case Intrinsic::copysign: |
1923 | return TargetOpcode::G_FCOPYSIGN; |
1924 | case Intrinsic::minnum: |
1925 | return TargetOpcode::G_FMINNUM; |
1926 | case Intrinsic::maxnum: |
1927 | return TargetOpcode::G_FMAXNUM; |
1928 | case Intrinsic::minimum: |
1929 | return TargetOpcode::G_FMINIMUM; |
1930 | case Intrinsic::maximum: |
1931 | return TargetOpcode::G_FMAXIMUM; |
1932 | case Intrinsic::minimumnum: |
1933 | return TargetOpcode::G_FMINIMUMNUM; |
1934 | case Intrinsic::maximumnum: |
1935 | return TargetOpcode::G_FMAXIMUMNUM; |
1936 | case Intrinsic::canonicalize: |
1937 | return TargetOpcode::G_FCANONICALIZE; |
1938 | case Intrinsic::floor: |
1939 | return TargetOpcode::G_FFLOOR; |
1940 | case Intrinsic::fma: |
1941 | return TargetOpcode::G_FMA; |
1942 | case Intrinsic::log: |
1943 | return TargetOpcode::G_FLOG; |
1944 | case Intrinsic::log2: |
1945 | return TargetOpcode::G_FLOG2; |
1946 | case Intrinsic::log10: |
1947 | return TargetOpcode::G_FLOG10; |
1948 | case Intrinsic::ldexp: |
1949 | return TargetOpcode::G_FLDEXP; |
1950 | case Intrinsic::nearbyint: |
1951 | return TargetOpcode::G_FNEARBYINT; |
1952 | case Intrinsic::pow: |
1953 | return TargetOpcode::G_FPOW; |
1954 | case Intrinsic::powi: |
1955 | return TargetOpcode::G_FPOWI; |
1956 | case Intrinsic::rint: |
1957 | return TargetOpcode::G_FRINT; |
1958 | case Intrinsic::round: |
1959 | return TargetOpcode::G_INTRINSIC_ROUND; |
1960 | case Intrinsic::roundeven: |
1961 | return TargetOpcode::G_INTRINSIC_ROUNDEVEN; |
1962 | case Intrinsic::sin: |
1963 | return TargetOpcode::G_FSIN; |
1964 | case Intrinsic::sinh: |
1965 | return TargetOpcode::G_FSINH; |
1966 | case Intrinsic::sqrt: |
1967 | return TargetOpcode::G_FSQRT; |
1968 | case Intrinsic::tan: |
1969 | return TargetOpcode::G_FTAN; |
1970 | case Intrinsic::tanh: |
1971 | return TargetOpcode::G_FTANH; |
1972 | case Intrinsic::trunc: |
1973 | return TargetOpcode::G_INTRINSIC_TRUNC; |
1974 | case Intrinsic::readcyclecounter: |
1975 | return TargetOpcode::G_READCYCLECOUNTER; |
1976 | case Intrinsic::readsteadycounter: |
1977 | return TargetOpcode::G_READSTEADYCOUNTER; |
1978 | case Intrinsic::ptrmask: |
1979 | return TargetOpcode::G_PTRMASK; |
1980 | case Intrinsic::lrint: |
1981 | return TargetOpcode::G_INTRINSIC_LRINT; |
1982 | case Intrinsic::llrint: |
1983 | return TargetOpcode::G_INTRINSIC_LLRINT; |
1984 | // FADD/FMUL require checking the FMF, so are handled elsewhere. |
1985 | case Intrinsic::vector_reduce_fmin: |
1986 | return TargetOpcode::G_VECREDUCE_FMIN; |
1987 | case Intrinsic::vector_reduce_fmax: |
1988 | return TargetOpcode::G_VECREDUCE_FMAX; |
1989 | case Intrinsic::vector_reduce_fminimum: |
1990 | return TargetOpcode::G_VECREDUCE_FMINIMUM; |
1991 | case Intrinsic::vector_reduce_fmaximum: |
1992 | return TargetOpcode::G_VECREDUCE_FMAXIMUM; |
1993 | case Intrinsic::vector_reduce_add: |
1994 | return TargetOpcode::G_VECREDUCE_ADD; |
1995 | case Intrinsic::vector_reduce_mul: |
1996 | return TargetOpcode::G_VECREDUCE_MUL; |
1997 | case Intrinsic::vector_reduce_and: |
1998 | return TargetOpcode::G_VECREDUCE_AND; |
1999 | case Intrinsic::vector_reduce_or: |
2000 | return TargetOpcode::G_VECREDUCE_OR; |
2001 | case Intrinsic::vector_reduce_xor: |
2002 | return TargetOpcode::G_VECREDUCE_XOR; |
2003 | case Intrinsic::vector_reduce_smax: |
2004 | return TargetOpcode::G_VECREDUCE_SMAX; |
2005 | case Intrinsic::vector_reduce_smin: |
2006 | return TargetOpcode::G_VECREDUCE_SMIN; |
2007 | case Intrinsic::vector_reduce_umax: |
2008 | return TargetOpcode::G_VECREDUCE_UMAX; |
2009 | case Intrinsic::vector_reduce_umin: |
2010 | return TargetOpcode::G_VECREDUCE_UMIN; |
2011 | case Intrinsic::experimental_vector_compress: |
2012 | return TargetOpcode::G_VECTOR_COMPRESS; |
2013 | case Intrinsic::lround: |
2014 | return TargetOpcode::G_LROUND; |
2015 | case Intrinsic::llround: |
2016 | return TargetOpcode::G_LLROUND; |
2017 | case Intrinsic::get_fpenv: |
2018 | return TargetOpcode::G_GET_FPENV; |
2019 | case Intrinsic::get_fpmode: |
2020 | return TargetOpcode::G_GET_FPMODE; |
2021 | } |
2022 | return Intrinsic::not_intrinsic; |
2023 | } |
2024 | |
2025 | bool IRTranslator::translateSimpleIntrinsic(const CallInst &CI, |
2026 | Intrinsic::ID ID, |
2027 | MachineIRBuilder &MIRBuilder) { |
2028 | |
2029 | unsigned Op = getSimpleIntrinsicOpcode(ID); |
2030 | |
2031 | // Is this a simple intrinsic? |
2032 | if (Op == Intrinsic::not_intrinsic) |
2033 | return false; |
2034 | |
2035 | // Yes. Let's translate it. |
2036 | SmallVector<llvm::SrcOp, 4> VRegs; |
2037 | for (const auto &Arg : CI.args()) |
2038 | VRegs.push_back(Elt: getOrCreateVReg(Val: *Arg)); |
2039 | |
2040 | MIRBuilder.buildInstr(Opc: Op, DstOps: {getOrCreateVReg(Val: CI)}, SrcOps: VRegs, |
2041 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2042 | return true; |
2043 | } |
2044 | |
2045 | // TODO: Include ConstainedOps.def when all strict instructions are defined. |
2046 | static unsigned getConstrainedOpcode(Intrinsic::ID ID) { |
2047 | switch (ID) { |
2048 | case Intrinsic::experimental_constrained_fadd: |
2049 | return TargetOpcode::G_STRICT_FADD; |
2050 | case Intrinsic::experimental_constrained_fsub: |
2051 | return TargetOpcode::G_STRICT_FSUB; |
2052 | case Intrinsic::experimental_constrained_fmul: |
2053 | return TargetOpcode::G_STRICT_FMUL; |
2054 | case Intrinsic::experimental_constrained_fdiv: |
2055 | return TargetOpcode::G_STRICT_FDIV; |
2056 | case Intrinsic::experimental_constrained_frem: |
2057 | return TargetOpcode::G_STRICT_FREM; |
2058 | case Intrinsic::experimental_constrained_fma: |
2059 | return TargetOpcode::G_STRICT_FMA; |
2060 | case Intrinsic::experimental_constrained_sqrt: |
2061 | return TargetOpcode::G_STRICT_FSQRT; |
2062 | case Intrinsic::experimental_constrained_ldexp: |
2063 | return TargetOpcode::G_STRICT_FLDEXP; |
2064 | default: |
2065 | return 0; |
2066 | } |
2067 | } |
2068 | |
2069 | bool IRTranslator::translateConstrainedFPIntrinsic( |
2070 | const ConstrainedFPIntrinsic &FPI, MachineIRBuilder &MIRBuilder) { |
2071 | fp::ExceptionBehavior EB = *FPI.getExceptionBehavior(); |
2072 | |
2073 | unsigned Opcode = getConstrainedOpcode(ID: FPI.getIntrinsicID()); |
2074 | if (!Opcode) |
2075 | return false; |
2076 | |
2077 | uint32_t Flags = MachineInstr::copyFlagsFromInstruction(I: FPI); |
2078 | if (EB == fp::ExceptionBehavior::ebIgnore) |
2079 | Flags |= MachineInstr::NoFPExcept; |
2080 | |
2081 | SmallVector<llvm::SrcOp, 4> VRegs; |
2082 | for (unsigned I = 0, E = FPI.getNonMetadataArgCount(); I != E; ++I) |
2083 | VRegs.push_back(Elt: getOrCreateVReg(Val: *FPI.getArgOperand(i: I))); |
2084 | |
2085 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {getOrCreateVReg(Val: FPI)}, SrcOps: VRegs, Flags); |
2086 | return true; |
2087 | } |
2088 | |
2089 | std::optional<MCRegister> IRTranslator::getArgPhysReg(Argument &Arg) { |
2090 | auto VRegs = getOrCreateVRegs(Val: Arg); |
2091 | if (VRegs.size() != 1) |
2092 | return std::nullopt; |
2093 | |
2094 | // Arguments are lowered as a copy of a livein physical register. |
2095 | auto *VRegDef = MF->getRegInfo().getVRegDef(Reg: VRegs[0]); |
2096 | if (!VRegDef || !VRegDef->isCopy()) |
2097 | return std::nullopt; |
2098 | return VRegDef->getOperand(i: 1).getReg().asMCReg(); |
2099 | } |
2100 | |
2101 | bool IRTranslator::translateIfEntryValueArgument(bool isDeclare, Value *Val, |
2102 | const DILocalVariable *Var, |
2103 | const DIExpression *Expr, |
2104 | const DebugLoc &DL, |
2105 | MachineIRBuilder &MIRBuilder) { |
2106 | auto *Arg = dyn_cast<Argument>(Val); |
2107 | if (!Arg) |
2108 | return false; |
2109 | |
2110 | if (!Expr->isEntryValue()) |
2111 | return false; |
2112 | |
2113 | std::optional<MCRegister> PhysReg = getArgPhysReg(Arg&: *Arg); |
2114 | if (!PhysReg) { |
2115 | LLVM_DEBUG(dbgs() << "Dropping dbg." << (isDeclare ? "declare" : "value" ) |
2116 | << ": expression is entry_value but " |
2117 | << "couldn't find a physical register\n" ); |
2118 | LLVM_DEBUG(dbgs() << *Var << "\n" ); |
2119 | return true; |
2120 | } |
2121 | |
2122 | if (isDeclare) { |
2123 | // Append an op deref to account for the fact that this is a dbg_declare. |
2124 | Expr = DIExpression::append(Expr, Ops: dwarf::DW_OP_deref); |
2125 | MF->setVariableDbgInfo(Var, Expr, Reg: *PhysReg, Loc: DL); |
2126 | } else { |
2127 | MIRBuilder.buildDirectDbgValue(Reg: *PhysReg, Variable: Var, Expr); |
2128 | } |
2129 | |
2130 | return true; |
2131 | } |
2132 | |
2133 | static unsigned getConvOpcode(Intrinsic::ID ID) { |
2134 | switch (ID) { |
2135 | default: |
2136 | llvm_unreachable("Unexpected intrinsic" ); |
2137 | case Intrinsic::experimental_convergence_anchor: |
2138 | return TargetOpcode::CONVERGENCECTRL_ANCHOR; |
2139 | case Intrinsic::experimental_convergence_entry: |
2140 | return TargetOpcode::CONVERGENCECTRL_ENTRY; |
2141 | case Intrinsic::experimental_convergence_loop: |
2142 | return TargetOpcode::CONVERGENCECTRL_LOOP; |
2143 | } |
2144 | } |
2145 | |
2146 | bool IRTranslator::translateConvergenceControlIntrinsic( |
2147 | const CallInst &CI, Intrinsic::ID ID, MachineIRBuilder &MIRBuilder) { |
2148 | MachineInstrBuilder MIB = MIRBuilder.buildInstr(Opcode: getConvOpcode(ID)); |
2149 | Register OutputReg = getOrCreateConvergenceTokenVReg(Token: CI); |
2150 | MIB.addDef(RegNo: OutputReg); |
2151 | |
2152 | if (ID == Intrinsic::experimental_convergence_loop) { |
2153 | auto Bundle = CI.getOperandBundle(ID: LLVMContext::OB_convergencectrl); |
2154 | assert(Bundle && "Expected a convergence control token." ); |
2155 | Register InputReg = |
2156 | getOrCreateConvergenceTokenVReg(Token: *Bundle->Inputs[0].get()); |
2157 | MIB.addUse(RegNo: InputReg); |
2158 | } |
2159 | |
2160 | return true; |
2161 | } |
2162 | |
2163 | bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID, |
2164 | MachineIRBuilder &MIRBuilder) { |
2165 | if (auto *MI = dyn_cast<AnyMemIntrinsic>(Val: &CI)) { |
2166 | if (ORE->enabled()) { |
2167 | if (MemoryOpRemark::canHandle(I: MI, TLI: *LibInfo)) { |
2168 | MemoryOpRemark R(*ORE, "gisel-irtranslator-memsize" , *DL, *LibInfo); |
2169 | R.visit(I: MI); |
2170 | } |
2171 | } |
2172 | } |
2173 | |
2174 | // If this is a simple intrinsic (that is, we just need to add a def of |
2175 | // a vreg, and uses for each arg operand, then translate it. |
2176 | if (translateSimpleIntrinsic(CI, ID, MIRBuilder)) |
2177 | return true; |
2178 | |
2179 | switch (ID) { |
2180 | default: |
2181 | break; |
2182 | case Intrinsic::lifetime_start: |
2183 | case Intrinsic::lifetime_end: { |
2184 | // No stack colouring in O0, discard region information. |
2185 | if (MF->getTarget().getOptLevel() == CodeGenOptLevel::None || |
2186 | MF->getFunction().hasOptNone()) |
2187 | return true; |
2188 | |
2189 | unsigned Op = ID == Intrinsic::lifetime_start ? TargetOpcode::LIFETIME_START |
2190 | : TargetOpcode::LIFETIME_END; |
2191 | |
2192 | // Get the underlying objects for the location passed on the lifetime |
2193 | // marker. |
2194 | SmallVector<const Value *, 4> Allocas; |
2195 | getUnderlyingObjects(V: CI.getArgOperand(i: 1), Objects&: Allocas); |
2196 | |
2197 | // Iterate over each underlying object, creating lifetime markers for each |
2198 | // static alloca. Quit if we find a non-static alloca. |
2199 | for (const Value *V : Allocas) { |
2200 | const AllocaInst *AI = dyn_cast<AllocaInst>(Val: V); |
2201 | if (!AI) |
2202 | continue; |
2203 | |
2204 | if (!AI->isStaticAlloca()) |
2205 | return true; |
2206 | |
2207 | MIRBuilder.buildInstr(Opcode: Op).addFrameIndex(Idx: getOrCreateFrameIndex(AI: *AI)); |
2208 | } |
2209 | return true; |
2210 | } |
2211 | case Intrinsic::fake_use: { |
2212 | SmallVector<llvm::SrcOp, 4> VRegs; |
2213 | for (const auto &Arg : CI.args()) |
2214 | llvm::append_range(C&: VRegs, R: getOrCreateVRegs(Val: *Arg)); |
2215 | MIRBuilder.buildInstr(Opc: TargetOpcode::FAKE_USE, DstOps: {}, SrcOps: VRegs); |
2216 | MF->setHasFakeUses(true); |
2217 | return true; |
2218 | } |
2219 | case Intrinsic::dbg_declare: { |
2220 | const DbgDeclareInst &DI = cast<DbgDeclareInst>(Val: CI); |
2221 | assert(DI.getVariable() && "Missing variable" ); |
2222 | translateDbgDeclareRecord(Address: DI.getAddress(), HasArgList: DI.hasArgList(), Variable: DI.getVariable(), |
2223 | Expression: DI.getExpression(), DL: DI.getDebugLoc(), MIRBuilder); |
2224 | return true; |
2225 | } |
2226 | case Intrinsic::dbg_label: { |
2227 | const DbgLabelInst &DI = cast<DbgLabelInst>(Val: CI); |
2228 | assert(DI.getLabel() && "Missing label" ); |
2229 | |
2230 | assert(DI.getLabel()->isValidLocationForIntrinsic( |
2231 | MIRBuilder.getDebugLoc()) && |
2232 | "Expected inlined-at fields to agree" ); |
2233 | |
2234 | MIRBuilder.buildDbgLabel(Label: DI.getLabel()); |
2235 | return true; |
2236 | } |
2237 | case Intrinsic::vaend: |
2238 | // No target I know of cares about va_end. Certainly no in-tree target |
2239 | // does. Simplest intrinsic ever! |
2240 | return true; |
2241 | case Intrinsic::vastart: { |
2242 | Value *Ptr = CI.getArgOperand(i: 0); |
2243 | unsigned ListSize = TLI->getVaListSizeInBits(DL: *DL) / 8; |
2244 | Align Alignment = getKnownAlignment(V: Ptr, DL: *DL); |
2245 | |
2246 | MIRBuilder.buildInstr(Opc: TargetOpcode::G_VASTART, DstOps: {}, SrcOps: {getOrCreateVReg(Val: *Ptr)}) |
2247 | .addMemOperand(MMO: MF->getMachineMemOperand(PtrInfo: MachinePointerInfo(Ptr), |
2248 | F: MachineMemOperand::MOStore, |
2249 | Size: ListSize, BaseAlignment: Alignment)); |
2250 | return true; |
2251 | } |
2252 | case Intrinsic::dbg_assign: |
2253 | // A dbg.assign is a dbg.value with more information about stack locations, |
2254 | // typically produced during optimisation of variables with leaked |
2255 | // addresses. We can treat it like a normal dbg_value intrinsic here; to |
2256 | // benefit from the full analysis of stack/SSA locations, GlobalISel would |
2257 | // need to register for and use the AssignmentTrackingAnalysis pass. |
2258 | [[fallthrough]]; |
2259 | case Intrinsic::dbg_value: { |
2260 | // This form of DBG_VALUE is target-independent. |
2261 | const DbgValueInst &DI = cast<DbgValueInst>(Val: CI); |
2262 | translateDbgValueRecord(V: DI.getValue(), HasArgList: DI.hasArgList(), Variable: DI.getVariable(), |
2263 | Expression: DI.getExpression(), DL: DI.getDebugLoc(), MIRBuilder); |
2264 | return true; |
2265 | } |
2266 | case Intrinsic::uadd_with_overflow: |
2267 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_UADDO, MIRBuilder); |
2268 | case Intrinsic::sadd_with_overflow: |
2269 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_SADDO, MIRBuilder); |
2270 | case Intrinsic::usub_with_overflow: |
2271 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_USUBO, MIRBuilder); |
2272 | case Intrinsic::ssub_with_overflow: |
2273 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_SSUBO, MIRBuilder); |
2274 | case Intrinsic::umul_with_overflow: |
2275 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_UMULO, MIRBuilder); |
2276 | case Intrinsic::smul_with_overflow: |
2277 | return translateOverflowIntrinsic(CI, Op: TargetOpcode::G_SMULO, MIRBuilder); |
2278 | case Intrinsic::uadd_sat: |
2279 | return translateBinaryOp(Opcode: TargetOpcode::G_UADDSAT, U: CI, MIRBuilder); |
2280 | case Intrinsic::sadd_sat: |
2281 | return translateBinaryOp(Opcode: TargetOpcode::G_SADDSAT, U: CI, MIRBuilder); |
2282 | case Intrinsic::usub_sat: |
2283 | return translateBinaryOp(Opcode: TargetOpcode::G_USUBSAT, U: CI, MIRBuilder); |
2284 | case Intrinsic::ssub_sat: |
2285 | return translateBinaryOp(Opcode: TargetOpcode::G_SSUBSAT, U: CI, MIRBuilder); |
2286 | case Intrinsic::ushl_sat: |
2287 | return translateBinaryOp(Opcode: TargetOpcode::G_USHLSAT, U: CI, MIRBuilder); |
2288 | case Intrinsic::sshl_sat: |
2289 | return translateBinaryOp(Opcode: TargetOpcode::G_SSHLSAT, U: CI, MIRBuilder); |
2290 | case Intrinsic::umin: |
2291 | return translateBinaryOp(Opcode: TargetOpcode::G_UMIN, U: CI, MIRBuilder); |
2292 | case Intrinsic::umax: |
2293 | return translateBinaryOp(Opcode: TargetOpcode::G_UMAX, U: CI, MIRBuilder); |
2294 | case Intrinsic::smin: |
2295 | return translateBinaryOp(Opcode: TargetOpcode::G_SMIN, U: CI, MIRBuilder); |
2296 | case Intrinsic::smax: |
2297 | return translateBinaryOp(Opcode: TargetOpcode::G_SMAX, U: CI, MIRBuilder); |
2298 | case Intrinsic::abs: |
2299 | // TODO: Preserve "int min is poison" arg in GMIR? |
2300 | return translateUnaryOp(Opcode: TargetOpcode::G_ABS, U: CI, MIRBuilder); |
2301 | case Intrinsic::smul_fix: |
2302 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_SMULFIX, CI, MIRBuilder); |
2303 | case Intrinsic::umul_fix: |
2304 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_UMULFIX, CI, MIRBuilder); |
2305 | case Intrinsic::smul_fix_sat: |
2306 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_SMULFIXSAT, CI, MIRBuilder); |
2307 | case Intrinsic::umul_fix_sat: |
2308 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_UMULFIXSAT, CI, MIRBuilder); |
2309 | case Intrinsic::sdiv_fix: |
2310 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_SDIVFIX, CI, MIRBuilder); |
2311 | case Intrinsic::udiv_fix: |
2312 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_UDIVFIX, CI, MIRBuilder); |
2313 | case Intrinsic::sdiv_fix_sat: |
2314 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_SDIVFIXSAT, CI, MIRBuilder); |
2315 | case Intrinsic::udiv_fix_sat: |
2316 | return translateFixedPointIntrinsic(Op: TargetOpcode::G_UDIVFIXSAT, CI, MIRBuilder); |
2317 | case Intrinsic::fmuladd: { |
2318 | const TargetMachine &TM = MF->getTarget(); |
2319 | Register Dst = getOrCreateVReg(Val: CI); |
2320 | Register Op0 = getOrCreateVReg(Val: *CI.getArgOperand(i: 0)); |
2321 | Register Op1 = getOrCreateVReg(Val: *CI.getArgOperand(i: 1)); |
2322 | Register Op2 = getOrCreateVReg(Val: *CI.getArgOperand(i: 2)); |
2323 | if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict && |
2324 | TLI->isFMAFasterThanFMulAndFAdd(MF: *MF, |
2325 | TLI->getValueType(DL: *DL, Ty: CI.getType()))) { |
2326 | // TODO: Revisit this to see if we should move this part of the |
2327 | // lowering to the combiner. |
2328 | MIRBuilder.buildFMA(Dst, Src0: Op0, Src1: Op1, Src2: Op2, |
2329 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2330 | } else { |
2331 | LLT Ty = getLLTForType(Ty&: *CI.getType(), DL: *DL); |
2332 | auto FMul = MIRBuilder.buildFMul( |
2333 | Dst: Ty, Src0: Op0, Src1: Op1, Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2334 | MIRBuilder.buildFAdd(Dst, Src0: FMul, Src1: Op2, |
2335 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2336 | } |
2337 | return true; |
2338 | } |
2339 | case Intrinsic::convert_from_fp16: |
2340 | // FIXME: This intrinsic should probably be removed from the IR. |
2341 | MIRBuilder.buildFPExt(Res: getOrCreateVReg(Val: CI), |
2342 | Op: getOrCreateVReg(Val: *CI.getArgOperand(i: 0)), |
2343 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2344 | return true; |
2345 | case Intrinsic::convert_to_fp16: |
2346 | // FIXME: This intrinsic should probably be removed from the IR. |
2347 | MIRBuilder.buildFPTrunc(Res: getOrCreateVReg(Val: CI), |
2348 | Op: getOrCreateVReg(Val: *CI.getArgOperand(i: 0)), |
2349 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2350 | return true; |
2351 | case Intrinsic::frexp: { |
2352 | ArrayRef<Register> VRegs = getOrCreateVRegs(Val: CI); |
2353 | MIRBuilder.buildFFrexp(Fract: VRegs[0], Exp: VRegs[1], |
2354 | Src: getOrCreateVReg(Val: *CI.getArgOperand(i: 0)), |
2355 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2356 | return true; |
2357 | } |
2358 | case Intrinsic::sincos: { |
2359 | ArrayRef<Register> VRegs = getOrCreateVRegs(Val: CI); |
2360 | MIRBuilder.buildFSincos(Sin: VRegs[0], Cos: VRegs[1], |
2361 | Src: getOrCreateVReg(Val: *CI.getArgOperand(i: 0)), |
2362 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2363 | return true; |
2364 | } |
2365 | case Intrinsic::fptosi_sat: |
2366 | MIRBuilder.buildFPTOSI_SAT(Dst: getOrCreateVReg(Val: CI), |
2367 | Src0: getOrCreateVReg(Val: *CI.getArgOperand(i: 0))); |
2368 | return true; |
2369 | case Intrinsic::fptoui_sat: |
2370 | MIRBuilder.buildFPTOUI_SAT(Dst: getOrCreateVReg(Val: CI), |
2371 | Src0: getOrCreateVReg(Val: *CI.getArgOperand(i: 0))); |
2372 | return true; |
2373 | case Intrinsic::memcpy_inline: |
2374 | return translateMemFunc(CI, MIRBuilder, Opcode: TargetOpcode::G_MEMCPY_INLINE); |
2375 | case Intrinsic::memcpy: |
2376 | return translateMemFunc(CI, MIRBuilder, Opcode: TargetOpcode::G_MEMCPY); |
2377 | case Intrinsic::memmove: |
2378 | return translateMemFunc(CI, MIRBuilder, Opcode: TargetOpcode::G_MEMMOVE); |
2379 | case Intrinsic::memset: |
2380 | return translateMemFunc(CI, MIRBuilder, Opcode: TargetOpcode::G_MEMSET); |
2381 | case Intrinsic::eh_typeid_for: { |
2382 | GlobalValue *GV = ExtractTypeInfo(V: CI.getArgOperand(i: 0)); |
2383 | Register Reg = getOrCreateVReg(Val: CI); |
2384 | unsigned TypeID = MF->getTypeIDFor(TI: GV); |
2385 | MIRBuilder.buildConstant(Res: Reg, Val: TypeID); |
2386 | return true; |
2387 | } |
2388 | case Intrinsic::objectsize: |
2389 | llvm_unreachable("llvm.objectsize.* should have been lowered already" ); |
2390 | |
2391 | case Intrinsic::is_constant: |
2392 | llvm_unreachable("llvm.is.constant.* should have been lowered already" ); |
2393 | |
2394 | case Intrinsic::stackguard: |
2395 | getStackGuard(DstReg: getOrCreateVReg(Val: CI), MIRBuilder); |
2396 | return true; |
2397 | case Intrinsic::stackprotector: { |
2398 | LLT PtrTy = getLLTForType(Ty&: *CI.getArgOperand(i: 0)->getType(), DL: *DL); |
2399 | Register GuardVal; |
2400 | if (TLI->useLoadStackGuardNode(M: *CI.getModule())) { |
2401 | GuardVal = MRI->createGenericVirtualRegister(Ty: PtrTy); |
2402 | getStackGuard(DstReg: GuardVal, MIRBuilder); |
2403 | } else |
2404 | GuardVal = getOrCreateVReg(Val: *CI.getArgOperand(i: 0)); // The guard's value. |
2405 | |
2406 | AllocaInst *Slot = cast<AllocaInst>(Val: CI.getArgOperand(i: 1)); |
2407 | int FI = getOrCreateFrameIndex(AI: *Slot); |
2408 | MF->getFrameInfo().setStackProtectorIndex(FI); |
2409 | |
2410 | MIRBuilder.buildStore( |
2411 | Val: GuardVal, Addr: getOrCreateVReg(Val: *Slot), |
2412 | MMO&: *MF->getMachineMemOperand(PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), |
2413 | f: MachineMemOperand::MOStore | |
2414 | MachineMemOperand::MOVolatile, |
2415 | MemTy: PtrTy, base_alignment: Align(8))); |
2416 | return true; |
2417 | } |
2418 | case Intrinsic::stacksave: { |
2419 | MIRBuilder.buildInstr(Opc: TargetOpcode::G_STACKSAVE, DstOps: {getOrCreateVReg(Val: CI)}, SrcOps: {}); |
2420 | return true; |
2421 | } |
2422 | case Intrinsic::stackrestore: { |
2423 | MIRBuilder.buildInstr(Opc: TargetOpcode::G_STACKRESTORE, DstOps: {}, |
2424 | SrcOps: {getOrCreateVReg(Val: *CI.getArgOperand(i: 0))}); |
2425 | return true; |
2426 | } |
2427 | case Intrinsic::cttz: |
2428 | case Intrinsic::ctlz: { |
2429 | ConstantInt *Cst = cast<ConstantInt>(Val: CI.getArgOperand(i: 1)); |
2430 | bool isTrailing = ID == Intrinsic::cttz; |
2431 | unsigned Opcode = isTrailing |
2432 | ? Cst->isZero() ? TargetOpcode::G_CTTZ |
2433 | : TargetOpcode::G_CTTZ_ZERO_UNDEF |
2434 | : Cst->isZero() ? TargetOpcode::G_CTLZ |
2435 | : TargetOpcode::G_CTLZ_ZERO_UNDEF; |
2436 | MIRBuilder.buildInstr(Opc: Opcode, DstOps: {getOrCreateVReg(Val: CI)}, |
2437 | SrcOps: {getOrCreateVReg(Val: *CI.getArgOperand(i: 0))}); |
2438 | return true; |
2439 | } |
2440 | case Intrinsic::invariant_start: { |
2441 | MIRBuilder.buildUndef(Res: getOrCreateVReg(Val: CI)); |
2442 | return true; |
2443 | } |
2444 | case Intrinsic::invariant_end: |
2445 | return true; |
2446 | case Intrinsic::expect: |
2447 | case Intrinsic::expect_with_probability: |
2448 | case Intrinsic::annotation: |
2449 | case Intrinsic::ptr_annotation: |
2450 | case Intrinsic::launder_invariant_group: |
2451 | case Intrinsic::strip_invariant_group: { |
2452 | // Drop the intrinsic, but forward the value. |
2453 | MIRBuilder.buildCopy(Res: getOrCreateVReg(Val: CI), |
2454 | Op: getOrCreateVReg(Val: *CI.getArgOperand(i: 0))); |
2455 | return true; |
2456 | } |
2457 | case Intrinsic::assume: |
2458 | case Intrinsic::experimental_noalias_scope_decl: |
2459 | case Intrinsic::var_annotation: |
2460 | case Intrinsic::sideeffect: |
2461 | // Discard annotate attributes, assumptions, and artificial side-effects. |
2462 | return true; |
2463 | case Intrinsic::read_volatile_register: |
2464 | case Intrinsic::read_register: { |
2465 | Value *Arg = CI.getArgOperand(i: 0); |
2466 | MIRBuilder |
2467 | .buildInstr(Opc: TargetOpcode::G_READ_REGISTER, DstOps: {getOrCreateVReg(Val: CI)}, SrcOps: {}) |
2468 | .addMetadata(MD: cast<MDNode>(Val: cast<MetadataAsValue>(Val: Arg)->getMetadata())); |
2469 | return true; |
2470 | } |
2471 | case Intrinsic::write_register: { |
2472 | Value *Arg = CI.getArgOperand(i: 0); |
2473 | MIRBuilder.buildInstr(Opcode: TargetOpcode::G_WRITE_REGISTER) |
2474 | .addMetadata(MD: cast<MDNode>(Val: cast<MetadataAsValue>(Val: Arg)->getMetadata())) |
2475 | .addUse(RegNo: getOrCreateVReg(Val: *CI.getArgOperand(i: 1))); |
2476 | return true; |
2477 | } |
2478 | case Intrinsic::localescape: { |
2479 | MachineBasicBlock &EntryMBB = MF->front(); |
2480 | StringRef EscapedName = GlobalValue::dropLLVMManglingEscape(Name: MF->getName()); |
2481 | |
2482 | // Directly emit some LOCAL_ESCAPE machine instrs. Label assignment emission |
2483 | // is the same on all targets. |
2484 | for (unsigned Idx = 0, E = CI.arg_size(); Idx < E; ++Idx) { |
2485 | Value *Arg = CI.getArgOperand(i: Idx)->stripPointerCasts(); |
2486 | if (isa<ConstantPointerNull>(Val: Arg)) |
2487 | continue; // Skip null pointers. They represent a hole in index space. |
2488 | |
2489 | int FI = getOrCreateFrameIndex(AI: *cast<AllocaInst>(Val: Arg)); |
2490 | MCSymbol *FrameAllocSym = |
2491 | MF->getContext().getOrCreateFrameAllocSymbol(FuncName: EscapedName, Idx); |
2492 | |
2493 | // This should be inserted at the start of the entry block. |
2494 | auto LocalEscape = |
2495 | MIRBuilder.buildInstrNoInsert(Opcode: TargetOpcode::LOCAL_ESCAPE) |
2496 | .addSym(Sym: FrameAllocSym) |
2497 | .addFrameIndex(Idx: FI); |
2498 | |
2499 | EntryMBB.insert(I: EntryMBB.begin(), MI: LocalEscape); |
2500 | } |
2501 | |
2502 | return true; |
2503 | } |
2504 | case Intrinsic::vector_reduce_fadd: |
2505 | case Intrinsic::vector_reduce_fmul: { |
2506 | // Need to check for the reassoc flag to decide whether we want a |
2507 | // sequential reduction opcode or not. |
2508 | Register Dst = getOrCreateVReg(Val: CI); |
2509 | Register ScalarSrc = getOrCreateVReg(Val: *CI.getArgOperand(i: 0)); |
2510 | Register VecSrc = getOrCreateVReg(Val: *CI.getArgOperand(i: 1)); |
2511 | unsigned Opc = 0; |
2512 | if (!CI.hasAllowReassoc()) { |
2513 | // The sequential ordering case. |
2514 | Opc = ID == Intrinsic::vector_reduce_fadd |
2515 | ? TargetOpcode::G_VECREDUCE_SEQ_FADD |
2516 | : TargetOpcode::G_VECREDUCE_SEQ_FMUL; |
2517 | MIRBuilder.buildInstr(Opc, DstOps: {Dst}, SrcOps: {ScalarSrc, VecSrc}, |
2518 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2519 | return true; |
2520 | } |
2521 | // We split the operation into a separate G_FADD/G_FMUL + the reduce, |
2522 | // since the associativity doesn't matter. |
2523 | unsigned ScalarOpc; |
2524 | if (ID == Intrinsic::vector_reduce_fadd) { |
2525 | Opc = TargetOpcode::G_VECREDUCE_FADD; |
2526 | ScalarOpc = TargetOpcode::G_FADD; |
2527 | } else { |
2528 | Opc = TargetOpcode::G_VECREDUCE_FMUL; |
2529 | ScalarOpc = TargetOpcode::G_FMUL; |
2530 | } |
2531 | LLT DstTy = MRI->getType(Reg: Dst); |
2532 | auto Rdx = MIRBuilder.buildInstr( |
2533 | Opc, DstOps: {DstTy}, SrcOps: {VecSrc}, Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2534 | MIRBuilder.buildInstr(Opc: ScalarOpc, DstOps: {Dst}, SrcOps: {ScalarSrc, Rdx}, |
2535 | Flags: MachineInstr::copyFlagsFromInstruction(I: CI)); |
2536 | |
2537 | return true; |
2538 | } |
2539 | case Intrinsic::trap: |
2540 | return translateTrap(CI, MIRBuilder, Opcode: TargetOpcode::G_TRAP); |
2541 | case Intrinsic::debugtrap: |
2542 | return translateTrap(CI, MIRBuilder, Opcode: TargetOpcode::G_DEBUGTRAP); |
2543 | case Intrinsic::ubsantrap: |
2544 | return translateTrap(CI, MIRBuilder, Opcode: TargetOpcode::G_UBSANTRAP); |
2545 | case Intrinsic::allow_runtime_check: |
2546 | case Intrinsic::allow_ubsan_check: |
2547 | MIRBuilder.buildCopy(Res: getOrCreateVReg(Val: CI), |
2548 | Op: getOrCreateVReg(Val: *ConstantInt::getTrue(Ty: CI.getType()))); |
2549 | return true; |
2550 | case Intrinsic::amdgcn_cs_chain: |
2551 | return translateCallBase(CB: CI, MIRBuilder); |
2552 | case Intrinsic::fptrunc_round: { |
2553 | uint32_t Flags = MachineInstr::copyFlagsFromInstruction(I: CI); |
2554 | |
2555 | // Convert the metadata argument to a constant integer |
2556 | Metadata *MD = cast<MetadataAsValue>(Val: CI.getArgOperand(i: 1))->getMetadata(); |
2557 | std::optional<RoundingMode> RoundMode = |
2558 | convertStrToRoundingMode(cast<MDString>(Val: MD)->getString()); |
2559 | |
2560 | // Add the Rounding mode as an integer |
2561 | MIRBuilder |
2562 | .buildInstr(Opc: TargetOpcode::G_INTRINSIC_FPTRUNC_ROUND, |
2563 | DstOps: {getOrCreateVReg(Val: CI)}, |
2564 | SrcOps: {getOrCreateVReg(Val: *CI.getArgOperand(i: 0))}, Flags) |
2565 | .addImm(Val: (int)*RoundMode); |
2566 | |
2567 | return true; |
2568 | } |
2569 | case Intrinsic::is_fpclass: { |
2570 | Value *FpValue = CI.getOperand(i_nocapture: 0); |
2571 | ConstantInt *TestMaskValue = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 1)); |
2572 | |
2573 | MIRBuilder |
2574 | .buildInstr(Opc: TargetOpcode::G_IS_FPCLASS, DstOps: {getOrCreateVReg(Val: CI)}, |
2575 | SrcOps: {getOrCreateVReg(Val: *FpValue)}) |
2576 | .addImm(Val: TestMaskValue->getZExtValue()); |
2577 | |
2578 | return true; |
2579 | } |
2580 | case Intrinsic::set_fpenv: { |
2581 | Value *FPEnv = CI.getOperand(i_nocapture: 0); |
2582 | MIRBuilder.buildSetFPEnv(Src: getOrCreateVReg(Val: *FPEnv)); |
2583 | return true; |
2584 | } |
2585 | case Intrinsic::reset_fpenv: |
2586 | MIRBuilder.buildResetFPEnv(); |
2587 | return true; |
2588 | case Intrinsic::set_fpmode: { |
2589 | Value *FPState = CI.getOperand(i_nocapture: 0); |
2590 | MIRBuilder.buildSetFPMode(Src: getOrCreateVReg(Val: *FPState)); |
2591 | return true; |
2592 | } |
2593 | case Intrinsic::reset_fpmode: |
2594 | MIRBuilder.buildResetFPMode(); |
2595 | return true; |
2596 | case Intrinsic::vscale: { |
2597 | MIRBuilder.buildVScale(Res: getOrCreateVReg(Val: CI), MinElts: 1); |
2598 | return true; |
2599 | } |
2600 | case Intrinsic::scmp: |
2601 | MIRBuilder.buildSCmp(Res: getOrCreateVReg(Val: CI), |
2602 | Op0: getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)), |
2603 | Op1: getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 1))); |
2604 | return true; |
2605 | case Intrinsic::ucmp: |
2606 | MIRBuilder.buildUCmp(Res: getOrCreateVReg(Val: CI), |
2607 | Op0: getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 0)), |
2608 | Op1: getOrCreateVReg(Val: *CI.getOperand(i_nocapture: 1))); |
2609 | return true; |
2610 | case Intrinsic::vector_extract: |
2611 | return translateExtractVector(U: CI, MIRBuilder); |
2612 | case Intrinsic::vector_insert: |
2613 | return translateInsertVector(U: CI, MIRBuilder); |
2614 | case Intrinsic::stepvector: { |
2615 | MIRBuilder.buildStepVector(Res: getOrCreateVReg(Val: CI), Step: 1); |
2616 | return true; |
2617 | } |
2618 | case Intrinsic::prefetch: { |
2619 | Value *Addr = CI.getOperand(i_nocapture: 0); |
2620 | unsigned RW = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 1))->getZExtValue(); |
2621 | unsigned Locality = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 2))->getZExtValue(); |
2622 | unsigned CacheType = cast<ConstantInt>(Val: CI.getOperand(i_nocapture: 3))->getZExtValue(); |
2623 | |
2624 | auto Flags = RW ? MachineMemOperand::MOStore : MachineMemOperand::MOLoad; |
2625 | auto &MMO = *MF->getMachineMemOperand(PtrInfo: MachinePointerInfo(Addr), f: Flags, |
2626 | MemTy: LLT(), base_alignment: Align()); |
2627 | |
2628 | MIRBuilder.buildPrefetch(Addr: getOrCreateVReg(Val: *Addr), RW, Locality, CacheType, |
2629 | MMO); |
2630 | |
2631 | return true; |
2632 | } |
2633 | |
2634 | case Intrinsic::vector_interleave2: |
2635 | case Intrinsic::vector_deinterleave2: { |
2636 | // Both intrinsics have at least one operand. |
2637 | Value *Op0 = CI.getOperand(i_nocapture: 0); |
2638 | LLT ResTy = getLLTForType(Ty&: *Op0->getType(), DL: MIRBuilder.getDataLayout()); |
2639 | if (!ResTy.isFixedVector()) |
2640 | return false; |
2641 | |
2642 | if (CI.getIntrinsicID() == Intrinsic::vector_interleave2) |
2643 | return translateVectorInterleave2Intrinsic(CI, MIRBuilder); |
2644 | |
2645 | return translateVectorDeinterleave2Intrinsic(CI, MIRBuilder); |
2646 | } |
2647 | |
2648 | #define INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC) \ |
2649 | case Intrinsic::INTRINSIC: |
2650 | #include "llvm/IR/ConstrainedOps.def" |
2651 | return translateConstrainedFPIntrinsic(FPI: cast<ConstrainedFPIntrinsic>(Val: CI), |
2652 | MIRBuilder); |
2653 | case Intrinsic::experimental_convergence_anchor: |
2654 | case Intrinsic::experimental_convergence_entry: |
2655 | case Intrinsic::experimental_convergence_loop: |
2656 | return translateConvergenceControlIntrinsic(CI, ID, MIRBuilder); |
2657 | } |
2658 | return false; |
2659 | } |
2660 | |
2661 | bool IRTranslator::translateInlineAsm(const CallBase &CB, |
2662 | MachineIRBuilder &MIRBuilder) { |
2663 | if (containsBF16Type(U: CB)) |
2664 | return false; |
2665 | |
2666 | const InlineAsmLowering *ALI = MF->getSubtarget().getInlineAsmLowering(); |
2667 | |
2668 | if (!ALI) { |
2669 | LLVM_DEBUG( |
2670 | dbgs() << "Inline asm lowering is not supported for this target yet\n" ); |
2671 | return false; |
2672 | } |
2673 | |
2674 | return ALI->lowerInlineAsm( |
2675 | MIRBuilder, CB, GetOrCreateVRegs: [&](const Value &Val) { return getOrCreateVRegs(Val); }); |
2676 | } |
2677 | |
2678 | bool IRTranslator::translateCallBase(const CallBase &CB, |
2679 | MachineIRBuilder &MIRBuilder) { |
2680 | ArrayRef<Register> Res = getOrCreateVRegs(Val: CB); |
2681 | |
2682 | SmallVector<ArrayRef<Register>, 8> Args; |
2683 | Register SwiftInVReg = 0; |
2684 | Register SwiftErrorVReg = 0; |
2685 | for (const auto &Arg : CB.args()) { |
2686 | if (CLI->supportSwiftError() && isSwiftError(V: Arg)) { |
2687 | assert(SwiftInVReg == 0 && "Expected only one swift error argument" ); |
2688 | LLT Ty = getLLTForType(Ty&: *Arg->getType(), DL: *DL); |
2689 | SwiftInVReg = MRI->createGenericVirtualRegister(Ty); |
2690 | MIRBuilder.buildCopy(Res: SwiftInVReg, Op: SwiftError.getOrCreateVRegUseAt( |
2691 | &CB, &MIRBuilder.getMBB(), Arg)); |
2692 | Args.emplace_back(Args: ArrayRef(SwiftInVReg)); |
2693 | SwiftErrorVReg = |
2694 | SwiftError.getOrCreateVRegDefAt(&CB, &MIRBuilder.getMBB(), Arg); |
2695 | continue; |
2696 | } |
2697 | Args.push_back(Elt: getOrCreateVRegs(Val: *Arg)); |
2698 | } |
2699 | |
2700 | if (auto *CI = dyn_cast<CallInst>(Val: &CB)) { |
2701 | if (ORE->enabled()) { |
2702 | if (MemoryOpRemark::canHandle(I: CI, TLI: *LibInfo)) { |
2703 | MemoryOpRemark R(*ORE, "gisel-irtranslator-memsize" , *DL, *LibInfo); |
2704 | R.visit(I: CI); |
2705 | } |
2706 | } |
2707 | } |
2708 | |
2709 | std::optional<CallLowering::PtrAuthInfo> PAI; |
2710 | if (auto Bundle = CB.getOperandBundle(ID: LLVMContext::OB_ptrauth)) { |
2711 | // Functions should never be ptrauth-called directly. |
2712 | assert(!CB.getCalledFunction() && "invalid direct ptrauth call" ); |
2713 | |
2714 | const Value *Key = Bundle->Inputs[0]; |
2715 | const Value *Discriminator = Bundle->Inputs[1]; |
2716 | |
2717 | // Look through ptrauth constants to try to eliminate the matching bundle |
2718 | // and turn this into a direct call with no ptrauth. |
2719 | // CallLowering will use the raw pointer if it doesn't find the PAI. |
2720 | const auto *CalleeCPA = dyn_cast<ConstantPtrAuth>(Val: CB.getCalledOperand()); |
2721 | if (!CalleeCPA || !isa<Function>(Val: CalleeCPA->getPointer()) || |
2722 | !CalleeCPA->isKnownCompatibleWith(Key, Discriminator, DL: *DL)) { |
2723 | // If we can't make it direct, package the bundle into PAI. |
2724 | Register DiscReg = getOrCreateVReg(Val: *Discriminator); |
2725 | PAI = CallLowering::PtrAuthInfo{.Key: cast<ConstantInt>(Val: Key)->getZExtValue(), |
2726 | .Discriminator: DiscReg}; |
2727 | } |
2728 | } |
2729 | |
2730 | Register ConvergenceCtrlToken = 0; |
2731 | if (auto Bundle = CB.getOperandBundle(ID: LLVMContext::OB_convergencectrl)) { |
2732 | const auto &Token = *Bundle->Inputs[0].get(); |
2733 | ConvergenceCtrlToken = getOrCreateConvergenceTokenVReg(Token); |
2734 | } |
2735 | |
2736 | // We don't set HasCalls on MFI here yet because call lowering may decide to |
2737 | // optimize into tail calls. Instead, we defer that to selection where a final |
2738 | // scan is done to check if any instructions are calls. |
2739 | bool Success = CLI->lowerCall( |
2740 | MIRBuilder, Call: CB, ResRegs: Res, ArgRegs: Args, SwiftErrorVReg, PAI, ConvergenceCtrlToken, |
2741 | GetCalleeReg: [&]() { return getOrCreateVReg(Val: *CB.getCalledOperand()); }); |
2742 | |
2743 | // Check if we just inserted a tail call. |
2744 | if (Success) { |
2745 | assert(!HasTailCall && "Can't tail call return twice from block?" ); |
2746 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); |
2747 | HasTailCall = TII->isTailCall(Inst: *std::prev(x: MIRBuilder.getInsertPt())); |
2748 | } |
2749 | |
2750 | return Success; |
2751 | } |
2752 | |
2753 | bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) { |
2754 | if (containsBF16Type(U)) |
2755 | return false; |
2756 | |
2757 | const CallInst &CI = cast<CallInst>(Val: U); |
2758 | const Function *F = CI.getCalledFunction(); |
2759 | |
2760 | // FIXME: support Windows dllimport function calls and calls through |
2761 | // weak symbols. |
2762 | if (F && (F->hasDLLImportStorageClass() || |
2763 | (MF->getTarget().getTargetTriple().isOSWindows() && |
2764 | F->hasExternalWeakLinkage()))) |
2765 | return false; |
2766 | |
2767 | // FIXME: support control flow guard targets. |
2768 | if (CI.countOperandBundlesOfType(ID: LLVMContext::OB_cfguardtarget)) |
2769 | return false; |
2770 | |
2771 | // FIXME: support statepoints and related. |
2772 | if (isa<GCStatepointInst, GCRelocateInst, GCResultInst>(Val: U)) |
2773 | return false; |
2774 | |
2775 | if (CI.isInlineAsm()) |
2776 | return translateInlineAsm(CB: CI, MIRBuilder); |
2777 | |
2778 | diagnoseDontCall(CI); |
2779 | |
2780 | Intrinsic::ID ID = F ? F->getIntrinsicID() : Intrinsic::not_intrinsic; |
2781 | if (!F || ID == Intrinsic::not_intrinsic) |
2782 | return translateCallBase(CB: CI, MIRBuilder); |
2783 | |
2784 | assert(ID != Intrinsic::not_intrinsic && "unknown intrinsic" ); |
2785 | |
2786 | if (translateKnownIntrinsic(CI, ID, MIRBuilder)) |
2787 | return true; |
2788 | |
2789 | ArrayRef<Register> ResultRegs; |
2790 | if (!CI.getType()->isVoidTy()) |
2791 | ResultRegs = getOrCreateVRegs(Val: CI); |
2792 | |
2793 | // Ignore the callsite attributes. Backend code is most likely not expecting |
2794 | // an intrinsic to sometimes have side effects and sometimes not. |
2795 | MachineInstrBuilder MIB = MIRBuilder.buildIntrinsic(ID, Res: ResultRegs); |
2796 | if (isa<FPMathOperator>(Val: CI)) |
2797 | MIB->copyIRFlags(I: CI); |
2798 | |
2799 | for (const auto &Arg : enumerate(First: CI.args())) { |
2800 | // If this is required to be an immediate, don't materialize it in a |
2801 | // register. |
2802 | if (CI.paramHasAttr(ArgNo: Arg.index(), Kind: Attribute::ImmArg)) { |
2803 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Arg.value())) { |
2804 | // imm arguments are more convenient than cimm (and realistically |
2805 | // probably sufficient), so use them. |
2806 | assert(CI->getBitWidth() <= 64 && |
2807 | "large intrinsic immediates not handled" ); |
2808 | MIB.addImm(Val: CI->getSExtValue()); |
2809 | } else { |
2810 | MIB.addFPImm(Val: cast<ConstantFP>(Val: Arg.value())); |
2811 | } |
2812 | } else if (auto *MDVal = dyn_cast<MetadataAsValue>(Val: Arg.value())) { |
2813 | auto *MD = MDVal->getMetadata(); |
2814 | auto *MDN = dyn_cast<MDNode>(Val: MD); |
2815 | if (!MDN) { |
2816 | if (auto *ConstMD = dyn_cast<ConstantAsMetadata>(Val: MD)) |
2817 | MDN = MDNode::get(Context&: MF->getFunction().getContext(), MDs: ConstMD); |
2818 | else // This was probably an MDString. |
2819 | return false; |
2820 | } |
2821 | MIB.addMetadata(MD: MDN); |
2822 | } else { |
2823 | ArrayRef<Register> VRegs = getOrCreateVRegs(Val: *Arg.value()); |
2824 | if (VRegs.size() > 1) |
2825 | return false; |
2826 | MIB.addUse(RegNo: VRegs[0]); |
2827 | } |
2828 | } |
2829 | |
2830 | // Add a MachineMemOperand if it is a target mem intrinsic. |
2831 | TargetLowering::IntrinsicInfo Info; |
2832 | // TODO: Add a GlobalISel version of getTgtMemIntrinsic. |
2833 | if (TLI->getTgtMemIntrinsic(Info, CI, *MF, ID)) { |
2834 | Align Alignment = Info.align.value_or( |
2835 | u: DL->getABITypeAlign(Ty: Info.memVT.getTypeForEVT(Context&: F->getContext()))); |
2836 | LLT MemTy = Info.memVT.isSimple() |
2837 | ? getLLTForMVT(Ty: Info.memVT.getSimpleVT()) |
2838 | : LLT::scalar(SizeInBits: Info.memVT.getStoreSizeInBits()); |
2839 | |
2840 | // TODO: We currently just fallback to address space 0 if getTgtMemIntrinsic |
2841 | // didn't yield anything useful. |
2842 | MachinePointerInfo MPI; |
2843 | if (Info.ptrVal) |
2844 | MPI = MachinePointerInfo(Info.ptrVal, Info.offset); |
2845 | else if (Info.fallbackAddressSpace) |
2846 | MPI = MachinePointerInfo(*Info.fallbackAddressSpace); |
2847 | MIB.addMemOperand(MMO: MF->getMachineMemOperand( |
2848 | PtrInfo: MPI, f: Info.flags, MemTy, base_alignment: Alignment, AAInfo: CI.getAAMetadata(), |
2849 | /*Ranges=*/nullptr, SSID: Info.ssid, Ordering: Info.order, FailureOrdering: Info.failureOrder)); |
2850 | } |
2851 | |
2852 | if (CI.isConvergent()) { |
2853 | if (auto Bundle = CI.getOperandBundle(ID: LLVMContext::OB_convergencectrl)) { |
2854 | auto *Token = Bundle->Inputs[0].get(); |
2855 | Register TokenReg = getOrCreateVReg(Val: *Token); |
2856 | MIB.addUse(RegNo: TokenReg, Flags: RegState::Implicit); |
2857 | } |
2858 | } |
2859 | |
2860 | return true; |
2861 | } |
2862 | |
2863 | bool IRTranslator::findUnwindDestinations( |
2864 | const BasicBlock *EHPadBB, |
2865 | BranchProbability Prob, |
2866 | SmallVectorImpl<std::pair<MachineBasicBlock *, BranchProbability>> |
2867 | &UnwindDests) { |
2868 | EHPersonality Personality = classifyEHPersonality( |
2869 | Pers: EHPadBB->getParent()->getFunction().getPersonalityFn()); |
2870 | bool IsMSVCCXX = Personality == EHPersonality::MSVC_CXX; |
2871 | bool IsCoreCLR = Personality == EHPersonality::CoreCLR; |
2872 | bool IsWasmCXX = Personality == EHPersonality::Wasm_CXX; |
2873 | bool IsSEH = isAsynchronousEHPersonality(Pers: Personality); |
2874 | |
2875 | if (IsWasmCXX) { |
2876 | // Ignore this for now. |
2877 | return false; |
2878 | } |
2879 | |
2880 | while (EHPadBB) { |
2881 | BasicBlock::const_iterator Pad = EHPadBB->getFirstNonPHIIt(); |
2882 | BasicBlock *NewEHPadBB = nullptr; |
2883 | if (isa<LandingPadInst>(Val: Pad)) { |
2884 | // Stop on landingpads. They are not funclets. |
2885 | UnwindDests.emplace_back(Args: &getMBB(BB: *EHPadBB), Args&: Prob); |
2886 | break; |
2887 | } |
2888 | if (isa<CleanupPadInst>(Val: Pad)) { |
2889 | // Stop on cleanup pads. Cleanups are always funclet entries for all known |
2890 | // personalities. |
2891 | UnwindDests.emplace_back(Args: &getMBB(BB: *EHPadBB), Args&: Prob); |
2892 | UnwindDests.back().first->setIsEHScopeEntry(); |
2893 | UnwindDests.back().first->setIsEHFuncletEntry(); |
2894 | break; |
2895 | } |
2896 | if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Val&: Pad)) { |
2897 | // Add the catchpad handlers to the possible destinations. |
2898 | for (const BasicBlock *CatchPadBB : CatchSwitch->handlers()) { |
2899 | UnwindDests.emplace_back(Args: &getMBB(BB: *CatchPadBB), Args&: Prob); |
2900 | // For MSVC++ and the CLR, catchblocks are funclets and need prologues. |
2901 | if (IsMSVCCXX || IsCoreCLR) |
2902 | UnwindDests.back().first->setIsEHFuncletEntry(); |
2903 | if (!IsSEH) |
2904 | UnwindDests.back().first->setIsEHScopeEntry(); |
2905 | } |
2906 | NewEHPadBB = CatchSwitch->getUnwindDest(); |
2907 | } else { |
2908 | continue; |
2909 | } |
2910 | |
2911 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
2912 | if (BPI && NewEHPadBB) |
2913 | Prob *= BPI->getEdgeProbability(Src: EHPadBB, Dst: NewEHPadBB); |
2914 | EHPadBB = NewEHPadBB; |
2915 | } |
2916 | return true; |
2917 | } |
2918 | |
2919 | bool IRTranslator::translateInvoke(const User &U, |
2920 | MachineIRBuilder &MIRBuilder) { |
2921 | const InvokeInst &I = cast<InvokeInst>(Val: U); |
2922 | MCContext &Context = MF->getContext(); |
2923 | |
2924 | const BasicBlock *ReturnBB = I.getSuccessor(i: 0); |
2925 | const BasicBlock *EHPadBB = I.getSuccessor(i: 1); |
2926 | |
2927 | const Function *Fn = I.getCalledFunction(); |
2928 | |
2929 | // FIXME: support invoking patchpoint and statepoint intrinsics. |
2930 | if (Fn && Fn->isIntrinsic()) |
2931 | return false; |
2932 | |
2933 | // FIXME: support whatever these are. |
2934 | if (I.hasDeoptState()) |
2935 | return false; |
2936 | |
2937 | // FIXME: support control flow guard targets. |
2938 | if (I.countOperandBundlesOfType(ID: LLVMContext::OB_cfguardtarget)) |
2939 | return false; |
2940 | |
2941 | // FIXME: support Windows exception handling. |
2942 | if (!isa<LandingPadInst>(Val: EHPadBB->getFirstNonPHIIt())) |
2943 | return false; |
2944 | |
2945 | // FIXME: support Windows dllimport function calls and calls through |
2946 | // weak symbols. |
2947 | if (Fn && (Fn->hasDLLImportStorageClass() || |
2948 | (MF->getTarget().getTargetTriple().isOSWindows() && |
2949 | Fn->hasExternalWeakLinkage()))) |
2950 | return false; |
2951 | |
2952 | bool LowerInlineAsm = I.isInlineAsm(); |
2953 | bool NeedEHLabel = true; |
2954 | |
2955 | // Emit the actual call, bracketed by EH_LABELs so that the MF knows about |
2956 | // the region covered by the try. |
2957 | MCSymbol *BeginSymbol = nullptr; |
2958 | if (NeedEHLabel) { |
2959 | MIRBuilder.buildInstr(Opcode: TargetOpcode::G_INVOKE_REGION_START); |
2960 | BeginSymbol = Context.createTempSymbol(); |
2961 | MIRBuilder.buildInstr(Opcode: TargetOpcode::EH_LABEL).addSym(Sym: BeginSymbol); |
2962 | } |
2963 | |
2964 | if (LowerInlineAsm) { |
2965 | if (!translateInlineAsm(CB: I, MIRBuilder)) |
2966 | return false; |
2967 | } else if (!translateCallBase(CB: I, MIRBuilder)) |
2968 | return false; |
2969 | |
2970 | MCSymbol *EndSymbol = nullptr; |
2971 | if (NeedEHLabel) { |
2972 | EndSymbol = Context.createTempSymbol(); |
2973 | MIRBuilder.buildInstr(Opcode: TargetOpcode::EH_LABEL).addSym(Sym: EndSymbol); |
2974 | } |
2975 | |
2976 | SmallVector<std::pair<MachineBasicBlock *, BranchProbability>, 1> UnwindDests; |
2977 | BranchProbabilityInfo *BPI = FuncInfo.BPI; |
2978 | MachineBasicBlock *InvokeMBB = &MIRBuilder.getMBB(); |
2979 | BranchProbability EHPadBBProb = |
2980 | BPI ? BPI->getEdgeProbability(Src: InvokeMBB->getBasicBlock(), Dst: EHPadBB) |
2981 | : BranchProbability::getZero(); |
2982 | |
2983 | if (!findUnwindDestinations(EHPadBB, Prob: EHPadBBProb, UnwindDests)) |
2984 | return false; |
2985 | |
2986 | MachineBasicBlock &EHPadMBB = getMBB(BB: *EHPadBB), |
2987 | &ReturnMBB = getMBB(BB: *ReturnBB); |
2988 | // Update successor info. |
2989 | addSuccessorWithProb(Src: InvokeMBB, Dst: &ReturnMBB); |
2990 | for (auto &UnwindDest : UnwindDests) { |
2991 | UnwindDest.first->setIsEHPad(); |
2992 | addSuccessorWithProb(Src: InvokeMBB, Dst: UnwindDest.first, Prob: UnwindDest.second); |
2993 | } |
2994 | InvokeMBB->normalizeSuccProbs(); |
2995 | |
2996 | if (NeedEHLabel) { |
2997 | assert(BeginSymbol && "Expected a begin symbol!" ); |
2998 | assert(EndSymbol && "Expected an end symbol!" ); |
2999 | MF->addInvoke(LandingPad: &EHPadMBB, BeginLabel: BeginSymbol, EndLabel: EndSymbol); |
3000 | } |
3001 | |
3002 | MIRBuilder.buildBr(Dest&: ReturnMBB); |
3003 | return true; |
3004 | } |
3005 | |
3006 | bool IRTranslator::translateCallBr(const User &U, |
3007 | MachineIRBuilder &MIRBuilder) { |
3008 | // FIXME: Implement this. |
3009 | return false; |
3010 | } |
3011 | |
3012 | bool IRTranslator::translateLandingPad(const User &U, |
3013 | MachineIRBuilder &MIRBuilder) { |
3014 | const LandingPadInst &LP = cast<LandingPadInst>(Val: U); |
3015 | |
3016 | MachineBasicBlock &MBB = MIRBuilder.getMBB(); |
3017 | |
3018 | MBB.setIsEHPad(); |
3019 | |
3020 | // If there aren't registers to copy the values into (e.g., during SjLj |
3021 | // exceptions), then don't bother. |
3022 | const Constant *PersonalityFn = MF->getFunction().getPersonalityFn(); |
3023 | if (TLI->getExceptionPointerRegister(PersonalityFn) == 0 && |
3024 | TLI->getExceptionSelectorRegister(PersonalityFn) == 0) |
3025 | return true; |
3026 | |
3027 | // If landingpad's return type is token type, we don't create DAG nodes |
3028 | // for its exception pointer and selector value. The extraction of exception |
3029 | // pointer or selector value from token type landingpads is not currently |
3030 | // supported. |
3031 | if (LP.getType()->isTokenTy()) |
3032 | return true; |
3033 | |
3034 | // Add a label to mark the beginning of the landing pad. Deletion of the |
3035 | // landing pad can thus be detected via the MachineModuleInfo. |
3036 | MIRBuilder.buildInstr(Opcode: TargetOpcode::EH_LABEL) |
3037 | .addSym(Sym: MF->addLandingPad(LandingPad: &MBB)); |
3038 | |
3039 | // If the unwinder does not preserve all registers, ensure that the |
3040 | // function marks the clobbered registers as used. |
3041 | const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo(); |
3042 | if (auto *RegMask = TRI.getCustomEHPadPreservedMask(MF: *MF)) |
3043 | MF->getRegInfo().addPhysRegsUsedFromRegMask(RegMask); |
3044 | |
3045 | LLT Ty = getLLTForType(Ty&: *LP.getType(), DL: *DL); |
3046 | Register Undef = MRI->createGenericVirtualRegister(Ty); |
3047 | MIRBuilder.buildUndef(Res: Undef); |
3048 | |
3049 | SmallVector<LLT, 2> Tys; |
3050 | for (Type *Ty : cast<StructType>(Val: LP.getType())->elements()) |
3051 | Tys.push_back(Elt: getLLTForType(Ty&: *Ty, DL: *DL)); |
3052 | assert(Tys.size() == 2 && "Only two-valued landingpads are supported" ); |
3053 | |
3054 | // Mark exception register as live in. |
3055 | Register ExceptionReg = TLI->getExceptionPointerRegister(PersonalityFn); |
3056 | if (!ExceptionReg) |
3057 | return false; |
3058 | |
3059 | MBB.addLiveIn(PhysReg: ExceptionReg); |
3060 | ArrayRef<Register> ResRegs = getOrCreateVRegs(Val: LP); |
3061 | MIRBuilder.buildCopy(Res: ResRegs[0], Op: ExceptionReg); |
3062 | |
3063 | Register SelectorReg = TLI->getExceptionSelectorRegister(PersonalityFn); |
3064 | if (!SelectorReg) |
3065 | return false; |
3066 | |
3067 | MBB.addLiveIn(PhysReg: SelectorReg); |
3068 | Register PtrVReg = MRI->createGenericVirtualRegister(Ty: Tys[0]); |
3069 | MIRBuilder.buildCopy(Res: PtrVReg, Op: SelectorReg); |
3070 | MIRBuilder.buildCast(Dst: ResRegs[1], Src: PtrVReg); |
3071 | |
3072 | return true; |
3073 | } |
3074 | |
3075 | bool IRTranslator::translateAlloca(const User &U, |
3076 | MachineIRBuilder &MIRBuilder) { |
3077 | auto &AI = cast<AllocaInst>(Val: U); |
3078 | |
3079 | if (AI.isSwiftError()) |
3080 | return true; |
3081 | |
3082 | if (AI.isStaticAlloca()) { |
3083 | Register Res = getOrCreateVReg(Val: AI); |
3084 | int FI = getOrCreateFrameIndex(AI); |
3085 | MIRBuilder.buildFrameIndex(Res, Idx: FI); |
3086 | return true; |
3087 | } |
3088 | |
3089 | // FIXME: support stack probing for Windows. |
3090 | if (MF->getTarget().getTargetTriple().isOSWindows()) |
3091 | return false; |
3092 | |
3093 | // Now we're in the harder dynamic case. |
3094 | Register NumElts = getOrCreateVReg(Val: *AI.getArraySize()); |
3095 | Type *IntPtrIRTy = DL->getIntPtrType(AI.getType()); |
3096 | LLT IntPtrTy = getLLTForType(Ty&: *IntPtrIRTy, DL: *DL); |
3097 | if (MRI->getType(Reg: NumElts) != IntPtrTy) { |
3098 | Register ExtElts = MRI->createGenericVirtualRegister(Ty: IntPtrTy); |
3099 | MIRBuilder.buildZExtOrTrunc(Res: ExtElts, Op: NumElts); |
3100 | NumElts = ExtElts; |
3101 | } |
3102 | |
3103 | Type *Ty = AI.getAllocatedType(); |
3104 | |
3105 | Register AllocSize = MRI->createGenericVirtualRegister(Ty: IntPtrTy); |
3106 | Register TySize = |
3107 | getOrCreateVReg(Val: *ConstantInt::get(Ty: IntPtrIRTy, V: DL->getTypeAllocSize(Ty))); |
3108 | MIRBuilder.buildMul(Dst: AllocSize, Src0: NumElts, Src1: TySize); |
3109 | |
3110 | // Round the size of the allocation up to the stack alignment size |
3111 | // by add SA-1 to the size. This doesn't overflow because we're computing |
3112 | // an address inside an alloca. |
3113 | Align StackAlign = MF->getSubtarget().getFrameLowering()->getStackAlign(); |
3114 | auto SAMinusOne = MIRBuilder.buildConstant(Res: IntPtrTy, Val: StackAlign.value() - 1); |
3115 | auto AllocAdd = MIRBuilder.buildAdd(Dst: IntPtrTy, Src0: AllocSize, Src1: SAMinusOne, |
3116 | Flags: MachineInstr::NoUWrap); |
3117 | auto AlignCst = |
3118 | MIRBuilder.buildConstant(Res: IntPtrTy, Val: ~(uint64_t)(StackAlign.value() - 1)); |
3119 | auto AlignedAlloc = MIRBuilder.buildAnd(Dst: IntPtrTy, Src0: AllocAdd, Src1: AlignCst); |
3120 | |
3121 | Align Alignment = std::max(a: AI.getAlign(), b: DL->getPrefTypeAlign(Ty)); |
3122 | if (Alignment <= StackAlign) |
3123 | Alignment = Align(1); |
3124 | MIRBuilder.buildDynStackAlloc(Res: getOrCreateVReg(Val: AI), Size: AlignedAlloc, Alignment); |
3125 | |
3126 | MF->getFrameInfo().CreateVariableSizedObject(Alignment, Alloca: &AI); |
3127 | assert(MF->getFrameInfo().hasVarSizedObjects()); |
3128 | return true; |
3129 | } |
3130 | |
3131 | bool IRTranslator::translateVAArg(const User &U, MachineIRBuilder &MIRBuilder) { |
3132 | // FIXME: We may need more info about the type. Because of how LLT works, |
3133 | // we're completely discarding the i64/double distinction here (amongst |
3134 | // others). Fortunately the ABIs I know of where that matters don't use va_arg |
3135 | // anyway but that's not guaranteed. |
3136 | MIRBuilder.buildInstr(Opc: TargetOpcode::G_VAARG, DstOps: {getOrCreateVReg(Val: U)}, |
3137 | SrcOps: {getOrCreateVReg(Val: *U.getOperand(i: 0)), |
3138 | DL->getABITypeAlign(Ty: U.getType()).value()}); |
3139 | return true; |
3140 | } |
3141 | |
3142 | bool IRTranslator::translateUnreachable(const User &U, |
3143 | MachineIRBuilder &MIRBuilder) { |
3144 | auto &UI = cast<UnreachableInst>(Val: U); |
3145 | if (!UI.shouldLowerToTrap(TrapUnreachable: MF->getTarget().Options.TrapUnreachable, |
3146 | NoTrapAfterNoreturn: MF->getTarget().Options.NoTrapAfterNoreturn)) |
3147 | return true; |
3148 | |
3149 | MIRBuilder.buildTrap(); |
3150 | return true; |
3151 | } |
3152 | |
3153 | bool IRTranslator::translateInsertElement(const User &U, |
3154 | MachineIRBuilder &MIRBuilder) { |
3155 | // If it is a <1 x Ty> vector, use the scalar as it is |
3156 | // not a legal vector type in LLT. |
3157 | if (auto *FVT = dyn_cast<FixedVectorType>(Val: U.getType()); |
3158 | FVT && FVT->getNumElements() == 1) |
3159 | return translateCopy(U, V: *U.getOperand(i: 1), MIRBuilder); |
3160 | |
3161 | Register Res = getOrCreateVReg(Val: U); |
3162 | Register Val = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
3163 | Register Elt = getOrCreateVReg(Val: *U.getOperand(i: 1)); |
3164 | unsigned PreferredVecIdxWidth = TLI->getVectorIdxWidth(DL: *DL); |
3165 | Register Idx; |
3166 | if (auto *CI = dyn_cast<ConstantInt>(Val: U.getOperand(i: 2))) { |
3167 | if (CI->getBitWidth() != PreferredVecIdxWidth) { |
3168 | APInt NewIdx = CI->getValue().zextOrTrunc(width: PreferredVecIdxWidth); |
3169 | auto *NewIdxCI = ConstantInt::get(Context&: CI->getContext(), V: NewIdx); |
3170 | Idx = getOrCreateVReg(Val: *NewIdxCI); |
3171 | } |
3172 | } |
3173 | if (!Idx) |
3174 | Idx = getOrCreateVReg(Val: *U.getOperand(i: 2)); |
3175 | if (MRI->getType(Reg: Idx).getSizeInBits() != PreferredVecIdxWidth) { |
3176 | const LLT VecIdxTy = LLT::scalar(SizeInBits: PreferredVecIdxWidth); |
3177 | Idx = MIRBuilder.buildZExtOrTrunc(Res: VecIdxTy, Op: Idx).getReg(Idx: 0); |
3178 | } |
3179 | MIRBuilder.buildInsertVectorElement(Res, Val, Elt, Idx); |
3180 | return true; |
3181 | } |
3182 | |
3183 | bool IRTranslator::translateInsertVector(const User &U, |
3184 | MachineIRBuilder &MIRBuilder) { |
3185 | Register Dst = getOrCreateVReg(Val: U); |
3186 | Register Vec = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
3187 | Register Elt = getOrCreateVReg(Val: *U.getOperand(i: 1)); |
3188 | |
3189 | ConstantInt *CI = cast<ConstantInt>(Val: U.getOperand(i: 2)); |
3190 | unsigned PreferredVecIdxWidth = TLI->getVectorIdxWidth(DL: *DL); |
3191 | |
3192 | // Resize Index to preferred index width. |
3193 | if (CI->getBitWidth() != PreferredVecIdxWidth) { |
3194 | APInt NewIdx = CI->getValue().zextOrTrunc(width: PreferredVecIdxWidth); |
3195 | CI = ConstantInt::get(Context&: CI->getContext(), V: NewIdx); |
3196 | } |
3197 | |
3198 | // If it is a <1 x Ty> vector, we have to use other means. |
3199 | if (auto *ResultType = dyn_cast<FixedVectorType>(Val: U.getOperand(i: 1)->getType()); |
3200 | ResultType && ResultType->getNumElements() == 1) { |
3201 | if (auto *InputType = dyn_cast<FixedVectorType>(Val: U.getOperand(i: 0)->getType()); |
3202 | InputType && InputType->getNumElements() == 1) { |
3203 | // We are inserting an illegal fixed vector into an illegal |
3204 | // fixed vector, use the scalar as it is not a legal vector type |
3205 | // in LLT. |
3206 | return translateCopy(U, V: *U.getOperand(i: 0), MIRBuilder); |
3207 | } |
3208 | if (isa<FixedVectorType>(Val: U.getOperand(i: 0)->getType())) { |
3209 | // We are inserting an illegal fixed vector into a legal fixed |
3210 | // vector, use the scalar as it is not a legal vector type in |
3211 | // LLT. |
3212 | Register Idx = getOrCreateVReg(Val: *CI); |
3213 | MIRBuilder.buildInsertVectorElement(Res: Dst, Val: Vec, Elt, Idx); |
3214 | return true; |
3215 | } |
3216 | if (isa<ScalableVectorType>(Val: U.getOperand(i: 0)->getType())) { |
3217 | // We are inserting an illegal fixed vector into a scalable |
3218 | // vector, use a scalar element insert. |
3219 | LLT VecIdxTy = LLT::scalar(SizeInBits: PreferredVecIdxWidth); |
3220 | Register Idx = getOrCreateVReg(Val: *CI); |
3221 | auto ScaledIndex = MIRBuilder.buildMul( |
3222 | Dst: VecIdxTy, Src0: MIRBuilder.buildVScale(Res: VecIdxTy, MinElts: 1), Src1: Idx); |
3223 | MIRBuilder.buildInsertVectorElement(Res: Dst, Val: Vec, Elt, Idx: ScaledIndex); |
3224 | return true; |
3225 | } |
3226 | } |
3227 | |
3228 | MIRBuilder.buildInsertSubvector( |
3229 | Res: getOrCreateVReg(Val: U), Src0: getOrCreateVReg(Val: *U.getOperand(i: 0)), |
3230 | Src1: getOrCreateVReg(Val: *U.getOperand(i: 1)), Index: CI->getZExtValue()); |
3231 | return true; |
3232 | } |
3233 | |
3234 | bool IRTranslator::(const User &U, |
3235 | MachineIRBuilder &MIRBuilder) { |
3236 | // If it is a <1 x Ty> vector, use the scalar as it is |
3237 | // not a legal vector type in LLT. |
3238 | if (const FixedVectorType *FVT = |
3239 | dyn_cast<FixedVectorType>(Val: U.getOperand(i: 0)->getType())) |
3240 | if (FVT->getNumElements() == 1) |
3241 | return translateCopy(U, V: *U.getOperand(i: 0), MIRBuilder); |
3242 | |
3243 | Register Res = getOrCreateVReg(Val: U); |
3244 | Register Val = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
3245 | unsigned PreferredVecIdxWidth = TLI->getVectorIdxWidth(DL: *DL); |
3246 | Register Idx; |
3247 | if (auto *CI = dyn_cast<ConstantInt>(Val: U.getOperand(i: 1))) { |
3248 | if (CI->getBitWidth() != PreferredVecIdxWidth) { |
3249 | APInt NewIdx = CI->getValue().zextOrTrunc(width: PreferredVecIdxWidth); |
3250 | auto *NewIdxCI = ConstantInt::get(Context&: CI->getContext(), V: NewIdx); |
3251 | Idx = getOrCreateVReg(Val: *NewIdxCI); |
3252 | } |
3253 | } |
3254 | if (!Idx) |
3255 | Idx = getOrCreateVReg(Val: *U.getOperand(i: 1)); |
3256 | if (MRI->getType(Reg: Idx).getSizeInBits() != PreferredVecIdxWidth) { |
3257 | const LLT VecIdxTy = LLT::scalar(SizeInBits: PreferredVecIdxWidth); |
3258 | Idx = MIRBuilder.buildZExtOrTrunc(Res: VecIdxTy, Op: Idx).getReg(Idx: 0); |
3259 | } |
3260 | MIRBuilder.buildExtractVectorElement(Res, Val, Idx); |
3261 | return true; |
3262 | } |
3263 | |
3264 | bool IRTranslator::(const User &U, |
3265 | MachineIRBuilder &MIRBuilder) { |
3266 | Register Res = getOrCreateVReg(Val: U); |
3267 | Register Vec = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
3268 | ConstantInt *CI = cast<ConstantInt>(Val: U.getOperand(i: 1)); |
3269 | unsigned PreferredVecIdxWidth = TLI->getVectorIdxWidth(DL: *DL); |
3270 | |
3271 | // Resize Index to preferred index width. |
3272 | if (CI->getBitWidth() != PreferredVecIdxWidth) { |
3273 | APInt NewIdx = CI->getValue().zextOrTrunc(width: PreferredVecIdxWidth); |
3274 | CI = ConstantInt::get(Context&: CI->getContext(), V: NewIdx); |
3275 | } |
3276 | |
3277 | // If it is a <1 x Ty> vector, we have to use other means. |
3278 | if (auto *ResultType = dyn_cast<FixedVectorType>(Val: U.getType()); |
3279 | ResultType && ResultType->getNumElements() == 1) { |
3280 | if (auto *InputType = dyn_cast<FixedVectorType>(Val: U.getOperand(i: 0)->getType()); |
3281 | InputType && InputType->getNumElements() == 1) { |
3282 | // We are extracting an illegal fixed vector from an illegal fixed vector, |
3283 | // use the scalar as it is not a legal vector type in LLT. |
3284 | return translateCopy(U, V: *U.getOperand(i: 0), MIRBuilder); |
3285 | } |
3286 | if (isa<FixedVectorType>(Val: U.getOperand(i: 0)->getType())) { |
3287 | // We are extracting an illegal fixed vector from a legal fixed |
3288 | // vector, use the scalar as it is not a legal vector type in |
3289 | // LLT. |
3290 | Register Idx = getOrCreateVReg(Val: *CI); |
3291 | MIRBuilder.buildExtractVectorElement(Res, Val: Vec, Idx); |
3292 | return true; |
3293 | } |
3294 | if (isa<ScalableVectorType>(Val: U.getOperand(i: 0)->getType())) { |
3295 | // We are extracting an illegal fixed vector from a scalable |
3296 | // vector, use a scalar element extract. |
3297 | LLT VecIdxTy = LLT::scalar(SizeInBits: PreferredVecIdxWidth); |
3298 | Register Idx = getOrCreateVReg(Val: *CI); |
3299 | auto ScaledIndex = MIRBuilder.buildMul( |
3300 | Dst: VecIdxTy, Src0: MIRBuilder.buildVScale(Res: VecIdxTy, MinElts: 1), Src1: Idx); |
3301 | MIRBuilder.buildExtractVectorElement(Res, Val: Vec, Idx: ScaledIndex); |
3302 | return true; |
3303 | } |
3304 | } |
3305 | |
3306 | MIRBuilder.buildExtractSubvector(Res: getOrCreateVReg(Val: U), |
3307 | Src: getOrCreateVReg(Val: *U.getOperand(i: 0)), |
3308 | Index: CI->getZExtValue()); |
3309 | return true; |
3310 | } |
3311 | |
3312 | bool IRTranslator::translateShuffleVector(const User &U, |
3313 | MachineIRBuilder &MIRBuilder) { |
3314 | // A ShuffleVector that operates on scalable vectors is a splat vector where |
3315 | // the value of the splat vector is the 0th element of the first operand, |
3316 | // since the index mask operand is the zeroinitializer (undef and |
3317 | // poison are treated as zeroinitializer here). |
3318 | if (U.getOperand(i: 0)->getType()->isScalableTy()) { |
3319 | Register Val = getOrCreateVReg(Val: *U.getOperand(i: 0)); |
3320 | auto SplatVal = MIRBuilder.buildExtractVectorElementConstant( |
3321 | Res: MRI->getType(Reg: Val).getElementType(), Val, Idx: 0); |
3322 | MIRBuilder.buildSplatVector(Res: getOrCreateVReg(Val: U), Val: SplatVal); |
3323 | return true; |
3324 | } |
3325 | |
3326 | ArrayRef<int> Mask; |
3327 | if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: &U)) |
3328 | Mask = SVI->getShuffleMask(); |
3329 | else |
3330 | Mask = cast<ConstantExpr>(Val: U).getShuffleMask(); |
3331 | ArrayRef<int> MaskAlloc = MF->allocateShuffleMask(Mask); |
3332 | MIRBuilder |
3333 | .buildInstr(Opc: TargetOpcode::G_SHUFFLE_VECTOR, DstOps: {getOrCreateVReg(Val: U)}, |
3334 | SrcOps: {getOrCreateVReg(Val: *U.getOperand(i: 0)), |
3335 | getOrCreateVReg(Val: *U.getOperand(i: 1))}) |
3336 | .addShuffleMask(Val: MaskAlloc); |
3337 | return true; |
3338 | } |
3339 | |
3340 | bool IRTranslator::translatePHI(const User &U, MachineIRBuilder &MIRBuilder) { |
3341 | const PHINode &PI = cast<PHINode>(Val: U); |
3342 | |
3343 | SmallVector<MachineInstr *, 4> Insts; |
3344 | for (auto Reg : getOrCreateVRegs(Val: PI)) { |
3345 | auto MIB = MIRBuilder.buildInstr(Opc: TargetOpcode::G_PHI, DstOps: {Reg}, SrcOps: {}); |
3346 | Insts.push_back(Elt: MIB.getInstr()); |
3347 | } |
3348 | |
3349 | PendingPHIs.emplace_back(Args: &PI, Args: std::move(Insts)); |
3350 | return true; |
3351 | } |
3352 | |
3353 | bool IRTranslator::translateAtomicCmpXchg(const User &U, |
3354 | MachineIRBuilder &MIRBuilder) { |
3355 | const AtomicCmpXchgInst &I = cast<AtomicCmpXchgInst>(Val: U); |
3356 | |
3357 | auto Flags = TLI->getAtomicMemOperandFlags(AI: I, DL: *DL); |
3358 | |
3359 | auto Res = getOrCreateVRegs(Val: I); |
3360 | Register OldValRes = Res[0]; |
3361 | Register SuccessRes = Res[1]; |
3362 | Register Addr = getOrCreateVReg(Val: *I.getPointerOperand()); |
3363 | Register Cmp = getOrCreateVReg(Val: *I.getCompareOperand()); |
3364 | Register NewVal = getOrCreateVReg(Val: *I.getNewValOperand()); |
3365 | |
3366 | MIRBuilder.buildAtomicCmpXchgWithSuccess( |
3367 | OldValRes, SuccessRes, Addr, CmpVal: Cmp, NewVal, |
3368 | MMO&: *MF->getMachineMemOperand( |
3369 | PtrInfo: MachinePointerInfo(I.getPointerOperand()), f: Flags, MemTy: MRI->getType(Reg: Cmp), |
3370 | base_alignment: getMemOpAlign(I), AAInfo: I.getAAMetadata(), Ranges: nullptr, SSID: I.getSyncScopeID(), |
3371 | Ordering: I.getSuccessOrdering(), FailureOrdering: I.getFailureOrdering())); |
3372 | return true; |
3373 | } |
3374 | |
3375 | bool IRTranslator::translateAtomicRMW(const User &U, |
3376 | MachineIRBuilder &MIRBuilder) { |
3377 | if (containsBF16Type(U)) |
3378 | return false; |
3379 | |
3380 | const AtomicRMWInst &I = cast<AtomicRMWInst>(Val: U); |
3381 | auto Flags = TLI->getAtomicMemOperandFlags(AI: I, DL: *DL); |
3382 | |
3383 | Register Res = getOrCreateVReg(Val: I); |
3384 | Register Addr = getOrCreateVReg(Val: *I.getPointerOperand()); |
3385 | Register Val = getOrCreateVReg(Val: *I.getValOperand()); |
3386 | |
3387 | unsigned Opcode = 0; |
3388 | switch (I.getOperation()) { |
3389 | default: |
3390 | return false; |
3391 | case AtomicRMWInst::Xchg: |
3392 | Opcode = TargetOpcode::G_ATOMICRMW_XCHG; |
3393 | break; |
3394 | case AtomicRMWInst::Add: |
3395 | Opcode = TargetOpcode::G_ATOMICRMW_ADD; |
3396 | break; |
3397 | case AtomicRMWInst::Sub: |
3398 | Opcode = TargetOpcode::G_ATOMICRMW_SUB; |
3399 | break; |
3400 | case AtomicRMWInst::And: |
3401 | Opcode = TargetOpcode::G_ATOMICRMW_AND; |
3402 | break; |
3403 | case AtomicRMWInst::Nand: |
3404 | Opcode = TargetOpcode::G_ATOMICRMW_NAND; |
3405 | break; |
3406 | case AtomicRMWInst::Or: |
3407 | Opcode = TargetOpcode::G_ATOMICRMW_OR; |
3408 | break; |
3409 | case AtomicRMWInst::Xor: |
3410 | Opcode = TargetOpcode::G_ATOMICRMW_XOR; |
3411 | break; |
3412 | case AtomicRMWInst::Max: |
3413 | Opcode = TargetOpcode::G_ATOMICRMW_MAX; |
3414 | break; |
3415 | case AtomicRMWInst::Min: |
3416 | Opcode = TargetOpcode::G_ATOMICRMW_MIN; |
3417 | break; |
3418 | case AtomicRMWInst::UMax: |
3419 | Opcode = TargetOpcode::G_ATOMICRMW_UMAX; |
3420 | break; |
3421 | case AtomicRMWInst::UMin: |
3422 | Opcode = TargetOpcode::G_ATOMICRMW_UMIN; |
3423 | break; |
3424 | case AtomicRMWInst::FAdd: |
3425 | Opcode = TargetOpcode::G_ATOMICRMW_FADD; |
3426 | break; |
3427 | case AtomicRMWInst::FSub: |
3428 | Opcode = TargetOpcode::G_ATOMICRMW_FSUB; |
3429 | break; |
3430 | case AtomicRMWInst::FMax: |
3431 | Opcode = TargetOpcode::G_ATOMICRMW_FMAX; |
3432 | break; |
3433 | case AtomicRMWInst::FMin: |
3434 | Opcode = TargetOpcode::G_ATOMICRMW_FMIN; |
3435 | break; |
3436 | case AtomicRMWInst::FMaximum: |
3437 | Opcode = TargetOpcode::G_ATOMICRMW_FMAXIMUM; |
3438 | break; |
3439 | case AtomicRMWInst::FMinimum: |
3440 | Opcode = TargetOpcode::G_ATOMICRMW_FMINIMUM; |
3441 | break; |
3442 | case AtomicRMWInst::UIncWrap: |
3443 | Opcode = TargetOpcode::G_ATOMICRMW_UINC_WRAP; |
3444 | break; |
3445 | case AtomicRMWInst::UDecWrap: |
3446 | Opcode = TargetOpcode::G_ATOMICRMW_UDEC_WRAP; |
3447 | break; |
3448 | case AtomicRMWInst::USubCond: |
3449 | Opcode = TargetOpcode::G_ATOMICRMW_USUB_COND; |
3450 | break; |
3451 | case AtomicRMWInst::USubSat: |
3452 | Opcode = TargetOpcode::G_ATOMICRMW_USUB_SAT; |
3453 | break; |
3454 | } |
3455 | |
3456 | MIRBuilder.buildAtomicRMW( |
3457 | Opcode, OldValRes: Res, Addr, Val, |
3458 | MMO&: *MF->getMachineMemOperand(PtrInfo: MachinePointerInfo(I.getPointerOperand()), |
3459 | f: Flags, MemTy: MRI->getType(Reg: Val), base_alignment: getMemOpAlign(I), |
3460 | AAInfo: I.getAAMetadata(), Ranges: nullptr, SSID: I.getSyncScopeID(), |
3461 | Ordering: I.getOrdering())); |
3462 | return true; |
3463 | } |
3464 | |
3465 | bool IRTranslator::translateFence(const User &U, |
3466 | MachineIRBuilder &MIRBuilder) { |
3467 | const FenceInst &Fence = cast<FenceInst>(Val: U); |
3468 | MIRBuilder.buildFence(Ordering: static_cast<unsigned>(Fence.getOrdering()), |
3469 | Scope: Fence.getSyncScopeID()); |
3470 | return true; |
3471 | } |
3472 | |
3473 | bool IRTranslator::translateFreeze(const User &U, |
3474 | MachineIRBuilder &MIRBuilder) { |
3475 | const ArrayRef<Register> DstRegs = getOrCreateVRegs(Val: U); |
3476 | const ArrayRef<Register> SrcRegs = getOrCreateVRegs(Val: *U.getOperand(i: 0)); |
3477 | |
3478 | assert(DstRegs.size() == SrcRegs.size() && |
3479 | "Freeze with different source and destination type?" ); |
3480 | |
3481 | for (unsigned I = 0; I < DstRegs.size(); ++I) { |
3482 | MIRBuilder.buildFreeze(Dst: DstRegs[I], Src: SrcRegs[I]); |
3483 | } |
3484 | |
3485 | return true; |
3486 | } |
3487 | |
3488 | void IRTranslator::finishPendingPhis() { |
3489 | #ifndef NDEBUG |
3490 | DILocationVerifier Verifier; |
3491 | GISelObserverWrapper WrapperObserver(&Verifier); |
3492 | RAIIMFObsDelInstaller ObsInstall(*MF, WrapperObserver); |
3493 | #endif // ifndef NDEBUG |
3494 | for (auto &Phi : PendingPHIs) { |
3495 | const PHINode *PI = Phi.first; |
3496 | if (PI->getType()->isEmptyTy()) |
3497 | continue; |
3498 | ArrayRef<MachineInstr *> ComponentPHIs = Phi.second; |
3499 | MachineBasicBlock *PhiMBB = ComponentPHIs[0]->getParent(); |
3500 | EntryBuilder->setDebugLoc(PI->getDebugLoc()); |
3501 | #ifndef NDEBUG |
3502 | Verifier.setCurrentInst(PI); |
3503 | #endif // ifndef NDEBUG |
3504 | |
3505 | SmallSet<const MachineBasicBlock *, 16> SeenPreds; |
3506 | for (unsigned i = 0; i < PI->getNumIncomingValues(); ++i) { |
3507 | auto IRPred = PI->getIncomingBlock(i); |
3508 | ArrayRef<Register> ValRegs = getOrCreateVRegs(Val: *PI->getIncomingValue(i)); |
3509 | for (auto *Pred : getMachinePredBBs(Edge: {IRPred, PI->getParent()})) { |
3510 | if (SeenPreds.count(Ptr: Pred) || !PhiMBB->isPredecessor(MBB: Pred)) |
3511 | continue; |
3512 | SeenPreds.insert(Ptr: Pred); |
3513 | for (unsigned j = 0; j < ValRegs.size(); ++j) { |
3514 | MachineInstrBuilder MIB(*MF, ComponentPHIs[j]); |
3515 | MIB.addUse(RegNo: ValRegs[j]); |
3516 | MIB.addMBB(MBB: Pred); |
3517 | } |
3518 | } |
3519 | } |
3520 | } |
3521 | } |
3522 | |
3523 | void IRTranslator::translateDbgValueRecord(Value *V, bool HasArgList, |
3524 | const DILocalVariable *Variable, |
3525 | const DIExpression *Expression, |
3526 | const DebugLoc &DL, |
3527 | MachineIRBuilder &MIRBuilder) { |
3528 | assert(Variable->isValidLocationForIntrinsic(DL) && |
3529 | "Expected inlined-at fields to agree" ); |
3530 | // Act as if we're handling a debug intrinsic. |
3531 | MIRBuilder.setDebugLoc(DL); |
3532 | |
3533 | if (!V || HasArgList) { |
3534 | // DI cannot produce a valid DBG_VALUE, so produce an undef DBG_VALUE to |
3535 | // terminate any prior location. |
3536 | MIRBuilder.buildIndirectDbgValue(Reg: 0, Variable, Expr: Expression); |
3537 | return; |
3538 | } |
3539 | |
3540 | if (const auto *CI = dyn_cast<Constant>(Val: V)) { |
3541 | MIRBuilder.buildConstDbgValue(C: *CI, Variable, Expr: Expression); |
3542 | return; |
3543 | } |
3544 | |
3545 | if (auto *AI = dyn_cast<AllocaInst>(Val: V); |
3546 | AI && AI->isStaticAlloca() && Expression->startsWithDeref()) { |
3547 | // If the value is an alloca and the expression starts with a |
3548 | // dereference, track a stack slot instead of a register, as registers |
3549 | // may be clobbered. |
3550 | auto ExprOperands = Expression->getElements(); |
3551 | auto *ExprDerefRemoved = |
3552 | DIExpression::get(Context&: AI->getContext(), Elements: ExprOperands.drop_front()); |
3553 | MIRBuilder.buildFIDbgValue(FI: getOrCreateFrameIndex(AI: *AI), Variable, |
3554 | Expr: ExprDerefRemoved); |
3555 | return; |
3556 | } |
3557 | if (translateIfEntryValueArgument(isDeclare: false, Val: V, Var: Variable, Expr: Expression, DL, |
3558 | MIRBuilder)) |
3559 | return; |
3560 | for (Register Reg : getOrCreateVRegs(Val: *V)) { |
3561 | // FIXME: This does not handle register-indirect values at offset 0. The |
3562 | // direct/indirect thing shouldn't really be handled by something as |
3563 | // implicit as reg+noreg vs reg+imm in the first place, but it seems |
3564 | // pretty baked in right now. |
3565 | MIRBuilder.buildDirectDbgValue(Reg, Variable, Expr: Expression); |
3566 | } |
3567 | } |
3568 | |
3569 | void IRTranslator::translateDbgDeclareRecord(Value *Address, bool HasArgList, |
3570 | const DILocalVariable *Variable, |
3571 | const DIExpression *Expression, |
3572 | const DebugLoc &DL, |
3573 | MachineIRBuilder &MIRBuilder) { |
3574 | if (!Address || isa<UndefValue>(Val: Address)) { |
3575 | LLVM_DEBUG(dbgs() << "Dropping debug info for " << *Variable << "\n" ); |
3576 | return; |
3577 | } |
3578 | |
3579 | assert(Variable->isValidLocationForIntrinsic(DL) && |
3580 | "Expected inlined-at fields to agree" ); |
3581 | auto AI = dyn_cast<AllocaInst>(Val: Address); |
3582 | if (AI && AI->isStaticAlloca()) { |
3583 | // Static allocas are tracked at the MF level, no need for DBG_VALUE |
3584 | // instructions (in fact, they get ignored if they *do* exist). |
3585 | MF->setVariableDbgInfo(Var: Variable, Expr: Expression, |
3586 | Slot: getOrCreateFrameIndex(AI: *AI), Loc: DL); |
3587 | return; |
3588 | } |
3589 | |
3590 | if (translateIfEntryValueArgument(isDeclare: true, Val: Address, Var: Variable, |
3591 | Expr: Expression, DL, |
3592 | MIRBuilder)) |
3593 | return; |
3594 | |
3595 | // A dbg.declare describes the address of a source variable, so lower it |
3596 | // into an indirect DBG_VALUE. |
3597 | MIRBuilder.setDebugLoc(DL); |
3598 | MIRBuilder.buildIndirectDbgValue(Reg: getOrCreateVReg(Val: *Address), Variable, |
3599 | Expr: Expression); |
3600 | } |
3601 | |
3602 | void IRTranslator::translateDbgInfo(const Instruction &Inst, |
3603 | MachineIRBuilder &MIRBuilder) { |
3604 | for (DbgRecord &DR : Inst.getDbgRecordRange()) { |
3605 | if (DbgLabelRecord *DLR = dyn_cast<DbgLabelRecord>(Val: &DR)) { |
3606 | MIRBuilder.setDebugLoc(DLR->getDebugLoc()); |
3607 | assert(DLR->getLabel() && "Missing label" ); |
3608 | assert(DLR->getLabel()->isValidLocationForIntrinsic( |
3609 | MIRBuilder.getDebugLoc()) && |
3610 | "Expected inlined-at fields to agree" ); |
3611 | MIRBuilder.buildDbgLabel(Label: DLR->getLabel()); |
3612 | continue; |
3613 | } |
3614 | DbgVariableRecord &DVR = cast<DbgVariableRecord>(Val&: DR); |
3615 | const DILocalVariable *Variable = DVR.getVariable(); |
3616 | const DIExpression *Expression = DVR.getExpression(); |
3617 | Value *V = DVR.getVariableLocationOp(OpIdx: 0); |
3618 | if (DVR.isDbgDeclare()) |
3619 | translateDbgDeclareRecord(Address: V, HasArgList: DVR.hasArgList(), Variable, Expression, |
3620 | DL: DVR.getDebugLoc(), MIRBuilder); |
3621 | else |
3622 | translateDbgValueRecord(V, HasArgList: DVR.hasArgList(), Variable, Expression, |
3623 | DL: DVR.getDebugLoc(), MIRBuilder); |
3624 | } |
3625 | } |
3626 | |
3627 | bool IRTranslator::translate(const Instruction &Inst) { |
3628 | CurBuilder->setDebugLoc(Inst.getDebugLoc()); |
3629 | CurBuilder->setPCSections(Inst.getMetadata(KindID: LLVMContext::MD_pcsections)); |
3630 | CurBuilder->setMMRAMetadata(Inst.getMetadata(KindID: LLVMContext::MD_mmra)); |
3631 | |
3632 | if (TLI->fallBackToDAGISel(Inst)) |
3633 | return false; |
3634 | |
3635 | switch (Inst.getOpcode()) { |
3636 | #define HANDLE_INST(NUM, OPCODE, CLASS) \ |
3637 | case Instruction::OPCODE: \ |
3638 | return translate##OPCODE(Inst, *CurBuilder.get()); |
3639 | #include "llvm/IR/Instruction.def" |
3640 | default: |
3641 | return false; |
3642 | } |
3643 | } |
3644 | |
3645 | bool IRTranslator::translate(const Constant &C, Register Reg) { |
3646 | // We only emit constants into the entry block from here. To prevent jumpy |
3647 | // debug behaviour remove debug line. |
3648 | if (auto CurrInstDL = CurBuilder->getDL()) |
3649 | EntryBuilder->setDebugLoc(DebugLoc()); |
3650 | |
3651 | if (auto CI = dyn_cast<ConstantInt>(Val: &C)) { |
3652 | // buildConstant expects a to-be-splatted scalar ConstantInt. |
3653 | if (isa<VectorType>(Val: CI->getType())) |
3654 | CI = ConstantInt::get(Context&: CI->getContext(), V: CI->getValue()); |
3655 | EntryBuilder->buildConstant(Res: Reg, Val: *CI); |
3656 | } else if (auto CF = dyn_cast<ConstantFP>(Val: &C)) { |
3657 | // buildFConstant expects a to-be-splatted scalar ConstantFP. |
3658 | if (isa<VectorType>(Val: CF->getType())) |
3659 | CF = ConstantFP::get(Context&: CF->getContext(), V: CF->getValue()); |
3660 | EntryBuilder->buildFConstant(Res: Reg, Val: *CF); |
3661 | } else if (isa<UndefValue>(Val: C)) |
3662 | EntryBuilder->buildUndef(Res: Reg); |
3663 | else if (isa<ConstantPointerNull>(Val: C)) |
3664 | EntryBuilder->buildConstant(Res: Reg, Val: 0); |
3665 | else if (auto GV = dyn_cast<GlobalValue>(Val: &C)) |
3666 | EntryBuilder->buildGlobalValue(Res: Reg, GV); |
3667 | else if (auto CPA = dyn_cast<ConstantPtrAuth>(Val: &C)) { |
3668 | Register Addr = getOrCreateVReg(Val: *CPA->getPointer()); |
3669 | Register AddrDisc = getOrCreateVReg(Val: *CPA->getAddrDiscriminator()); |
3670 | EntryBuilder->buildConstantPtrAuth(Res: Reg, CPA, Addr, AddrDisc); |
3671 | } else if (auto CAZ = dyn_cast<ConstantAggregateZero>(Val: &C)) { |
3672 | Constant &Elt = *CAZ->getElementValue(Idx: 0u); |
3673 | if (isa<ScalableVectorType>(Val: CAZ->getType())) { |
3674 | EntryBuilder->buildSplatVector(Res: Reg, Val: getOrCreateVReg(Val: Elt)); |
3675 | return true; |
3676 | } |
3677 | // Return the scalar if it is a <1 x Ty> vector. |
3678 | unsigned NumElts = CAZ->getElementCount().getFixedValue(); |
3679 | if (NumElts == 1) |
3680 | return translateCopy(U: C, V: Elt, MIRBuilder&: *EntryBuilder); |
3681 | // All elements are zero so we can just use the first one. |
3682 | EntryBuilder->buildSplatBuildVector(Res: Reg, Src: getOrCreateVReg(Val: Elt)); |
3683 | } else if (auto CV = dyn_cast<ConstantDataVector>(Val: &C)) { |
3684 | // Return the scalar if it is a <1 x Ty> vector. |
3685 | if (CV->getNumElements() == 1) |
3686 | return translateCopy(U: C, V: *CV->getElementAsConstant(i: 0), MIRBuilder&: *EntryBuilder); |
3687 | SmallVector<Register, 4> Ops; |
3688 | for (unsigned i = 0; i < CV->getNumElements(); ++i) { |
3689 | Constant &Elt = *CV->getElementAsConstant(i); |
3690 | Ops.push_back(Elt: getOrCreateVReg(Val: Elt)); |
3691 | } |
3692 | EntryBuilder->buildBuildVector(Res: Reg, Ops); |
3693 | } else if (auto CE = dyn_cast<ConstantExpr>(Val: &C)) { |
3694 | switch(CE->getOpcode()) { |
3695 | #define HANDLE_INST(NUM, OPCODE, CLASS) \ |
3696 | case Instruction::OPCODE: \ |
3697 | return translate##OPCODE(*CE, *EntryBuilder.get()); |
3698 | #include "llvm/IR/Instruction.def" |
3699 | default: |
3700 | return false; |
3701 | } |
3702 | } else if (auto CV = dyn_cast<ConstantVector>(Val: &C)) { |
3703 | if (CV->getNumOperands() == 1) |
3704 | return translateCopy(U: C, V: *CV->getOperand(i_nocapture: 0), MIRBuilder&: *EntryBuilder); |
3705 | SmallVector<Register, 4> Ops; |
3706 | for (unsigned i = 0; i < CV->getNumOperands(); ++i) { |
3707 | Ops.push_back(Elt: getOrCreateVReg(Val: *CV->getOperand(i_nocapture: i))); |
3708 | } |
3709 | EntryBuilder->buildBuildVector(Res: Reg, Ops); |
3710 | } else if (auto *BA = dyn_cast<BlockAddress>(Val: &C)) { |
3711 | EntryBuilder->buildBlockAddress(Res: Reg, BA); |
3712 | } else |
3713 | return false; |
3714 | |
3715 | return true; |
3716 | } |
3717 | |
3718 | bool IRTranslator::finalizeBasicBlock(const BasicBlock &BB, |
3719 | MachineBasicBlock &MBB) { |
3720 | for (auto &BTB : SL->BitTestCases) { |
3721 | // Emit header first, if it wasn't already emitted. |
3722 | if (!BTB.Emitted) |
3723 | emitBitTestHeader(B&: BTB, SwitchBB: BTB.Parent); |
3724 | |
3725 | BranchProbability UnhandledProb = BTB.Prob; |
3726 | for (unsigned j = 0, ej = BTB.Cases.size(); j != ej; ++j) { |
3727 | UnhandledProb -= BTB.Cases[j].ExtraProb; |
3728 | // Set the current basic block to the mbb we wish to insert the code into |
3729 | MachineBasicBlock *MBB = BTB.Cases[j].ThisBB; |
3730 | // If all cases cover a contiguous range, it is not necessary to jump to |
3731 | // the default block after the last bit test fails. This is because the |
3732 | // range check during bit test header creation has guaranteed that every |
3733 | // case here doesn't go outside the range. In this case, there is no need |
3734 | // to perform the last bit test, as it will always be true. Instead, make |
3735 | // the second-to-last bit-test fall through to the target of the last bit |
3736 | // test, and delete the last bit test. |
3737 | |
3738 | MachineBasicBlock *NextMBB; |
3739 | if ((BTB.ContiguousRange || BTB.FallthroughUnreachable) && j + 2 == ej) { |
3740 | // Second-to-last bit-test with contiguous range: fall through to the |
3741 | // target of the final bit test. |
3742 | NextMBB = BTB.Cases[j + 1].TargetBB; |
3743 | } else if (j + 1 == ej) { |
3744 | // For the last bit test, fall through to Default. |
3745 | NextMBB = BTB.Default; |
3746 | } else { |
3747 | // Otherwise, fall through to the next bit test. |
3748 | NextMBB = BTB.Cases[j + 1].ThisBB; |
3749 | } |
3750 | |
3751 | emitBitTestCase(BB&: BTB, NextMBB, BranchProbToNext: UnhandledProb, Reg: BTB.Reg, B&: BTB.Cases[j], SwitchBB: MBB); |
3752 | |
3753 | if ((BTB.ContiguousRange || BTB.FallthroughUnreachable) && j + 2 == ej) { |
3754 | // We need to record the replacement phi edge here that normally |
3755 | // happens in emitBitTestCase before we delete the case, otherwise the |
3756 | // phi edge will be lost. |
3757 | addMachineCFGPred(Edge: {BTB.Parent->getBasicBlock(), |
3758 | BTB.Cases[ej - 1].TargetBB->getBasicBlock()}, |
3759 | NewPred: MBB); |
3760 | // Since we're not going to use the final bit test, remove it. |
3761 | BTB.Cases.pop_back(); |
3762 | break; |
3763 | } |
3764 | } |
3765 | // This is "default" BB. We have two jumps to it. From "header" BB and from |
3766 | // last "case" BB, unless the latter was skipped. |
3767 | CFGEdge = {BTB.Parent->getBasicBlock(), |
3768 | BTB.Default->getBasicBlock()}; |
3769 | addMachineCFGPred(Edge: HeaderToDefaultEdge, NewPred: BTB.Parent); |
3770 | if (!BTB.ContiguousRange) { |
3771 | addMachineCFGPred(Edge: HeaderToDefaultEdge, NewPred: BTB.Cases.back().ThisBB); |
3772 | } |
3773 | } |
3774 | SL->BitTestCases.clear(); |
3775 | |
3776 | for (auto &JTCase : SL->JTCases) { |
3777 | // Emit header first, if it wasn't already emitted. |
3778 | if (!JTCase.first.Emitted) |
3779 | emitJumpTableHeader(JT&: JTCase.second, JTH&: JTCase.first, HeaderBB: JTCase.first.HeaderBB); |
3780 | |
3781 | emitJumpTable(JT&: JTCase.second, MBB: JTCase.second.MBB); |
3782 | } |
3783 | SL->JTCases.clear(); |
3784 | |
3785 | for (auto &SwCase : SL->SwitchCases) |
3786 | emitSwitchCase(CB&: SwCase, SwitchBB: &CurBuilder->getMBB(), MIB&: *CurBuilder); |
3787 | SL->SwitchCases.clear(); |
3788 | |
3789 | // Check if we need to generate stack-protector guard checks. |
3790 | StackProtector &SP = getAnalysis<StackProtector>(); |
3791 | if (SP.shouldEmitSDCheck(BB)) { |
3792 | bool FunctionBasedInstrumentation = |
3793 | TLI->getSSPStackGuardCheck(M: *MF->getFunction().getParent()); |
3794 | SPDescriptor.initialize(BB: &BB, MBB: &MBB, FunctionBasedInstrumentation); |
3795 | } |
3796 | // Handle stack protector. |
3797 | if (SPDescriptor.shouldEmitFunctionBasedCheckStackProtector()) { |
3798 | LLVM_DEBUG(dbgs() << "Unimplemented stack protector case\n" ); |
3799 | return false; |
3800 | } else if (SPDescriptor.shouldEmitStackProtector()) { |
3801 | MachineBasicBlock *ParentMBB = SPDescriptor.getParentMBB(); |
3802 | MachineBasicBlock *SuccessMBB = SPDescriptor.getSuccessMBB(); |
3803 | |
3804 | // Find the split point to split the parent mbb. At the same time copy all |
3805 | // physical registers used in the tail of parent mbb into virtual registers |
3806 | // before the split point and back into physical registers after the split |
3807 | // point. This prevents us needing to deal with Live-ins and many other |
3808 | // register allocation issues caused by us splitting the parent mbb. The |
3809 | // register allocator will clean up said virtual copies later on. |
3810 | MachineBasicBlock::iterator SplitPoint = findSplitPointForStackProtector( |
3811 | BB: ParentMBB, TII: *MF->getSubtarget().getInstrInfo()); |
3812 | |
3813 | // Splice the terminator of ParentMBB into SuccessMBB. |
3814 | SuccessMBB->splice(Where: SuccessMBB->end(), Other: ParentMBB, From: SplitPoint, |
3815 | To: ParentMBB->end()); |
3816 | |
3817 | // Add compare/jump on neq/jump to the parent BB. |
3818 | if (!emitSPDescriptorParent(SPD&: SPDescriptor, ParentBB: ParentMBB)) |
3819 | return false; |
3820 | |
3821 | // CodeGen Failure MBB if we have not codegened it yet. |
3822 | MachineBasicBlock *FailureMBB = SPDescriptor.getFailureMBB(); |
3823 | if (FailureMBB->empty()) { |
3824 | if (!emitSPDescriptorFailure(SPD&: SPDescriptor, FailureBB: FailureMBB)) |
3825 | return false; |
3826 | } |
3827 | |
3828 | // Clear the Per-BB State. |
3829 | SPDescriptor.resetPerBBState(); |
3830 | } |
3831 | return true; |
3832 | } |
3833 | |
3834 | bool IRTranslator::emitSPDescriptorParent(StackProtectorDescriptor &SPD, |
3835 | MachineBasicBlock *ParentBB) { |
3836 | CurBuilder->setInsertPt(MBB&: *ParentBB, II: ParentBB->end()); |
3837 | // First create the loads to the guard/stack slot for the comparison. |
3838 | Type *PtrIRTy = PointerType::getUnqual(C&: MF->getFunction().getContext()); |
3839 | const LLT PtrTy = getLLTForType(Ty&: *PtrIRTy, DL: *DL); |
3840 | LLT PtrMemTy = getLLTForMVT(Ty: TLI->getPointerMemTy(DL: *DL)); |
3841 | |
3842 | MachineFrameInfo &MFI = ParentBB->getParent()->getFrameInfo(); |
3843 | int FI = MFI.getStackProtectorIndex(); |
3844 | |
3845 | Register Guard; |
3846 | Register StackSlotPtr = CurBuilder->buildFrameIndex(Res: PtrTy, Idx: FI).getReg(Idx: 0); |
3847 | const Module &M = *ParentBB->getParent()->getFunction().getParent(); |
3848 | Align Align = DL->getPrefTypeAlign(Ty: PointerType::getUnqual(C&: M.getContext())); |
3849 | |
3850 | // Generate code to load the content of the guard slot. |
3851 | Register GuardVal = |
3852 | CurBuilder |
3853 | ->buildLoad(Res: PtrMemTy, Addr: StackSlotPtr, |
3854 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), Alignment: Align, |
3855 | MMOFlags: MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile) |
3856 | .getReg(Idx: 0); |
3857 | |
3858 | if (TLI->useStackGuardXorFP()) { |
3859 | LLVM_DEBUG(dbgs() << "Stack protector xor'ing with FP not yet implemented" ); |
3860 | return false; |
3861 | } |
3862 | |
3863 | // Retrieve guard check function, nullptr if instrumentation is inlined. |
3864 | if (const Function *GuardCheckFn = TLI->getSSPStackGuardCheck(M)) { |
3865 | // This path is currently untestable on GlobalISel, since the only platform |
3866 | // that needs this seems to be Windows, and we fall back on that currently. |
3867 | // The code still lives here in case that changes. |
3868 | // Silence warning about unused variable until the code below that uses |
3869 | // 'GuardCheckFn' is enabled. |
3870 | (void)GuardCheckFn; |
3871 | return false; |
3872 | #if 0 |
3873 | // The target provides a guard check function to validate the guard value. |
3874 | // Generate a call to that function with the content of the guard slot as |
3875 | // argument. |
3876 | FunctionType *FnTy = GuardCheckFn->getFunctionType(); |
3877 | assert(FnTy->getNumParams() == 1 && "Invalid function signature" ); |
3878 | ISD::ArgFlagsTy Flags; |
3879 | if (GuardCheckFn->hasAttribute(1, Attribute::AttrKind::InReg)) |
3880 | Flags.setInReg(); |
3881 | CallLowering::ArgInfo GuardArgInfo( |
3882 | {GuardVal, FnTy->getParamType(0), {Flags}}); |
3883 | |
3884 | CallLowering::CallLoweringInfo Info; |
3885 | Info.OrigArgs.push_back(GuardArgInfo); |
3886 | Info.CallConv = GuardCheckFn->getCallingConv(); |
3887 | Info.Callee = MachineOperand::CreateGA(GuardCheckFn, 0); |
3888 | Info.OrigRet = {Register(), FnTy->getReturnType()}; |
3889 | if (!CLI->lowerCall(MIRBuilder, Info)) { |
3890 | LLVM_DEBUG(dbgs() << "Failed to lower call to stack protector check\n" ); |
3891 | return false; |
3892 | } |
3893 | return true; |
3894 | #endif |
3895 | } |
3896 | |
3897 | // If useLoadStackGuardNode returns true, generate LOAD_STACK_GUARD. |
3898 | // Otherwise, emit a volatile load to retrieve the stack guard value. |
3899 | if (TLI->useLoadStackGuardNode(M: *ParentBB->getBasicBlock()->getModule())) { |
3900 | Guard = |
3901 | MRI->createGenericVirtualRegister(Ty: LLT::scalar(SizeInBits: PtrTy.getSizeInBits())); |
3902 | getStackGuard(DstReg: Guard, MIRBuilder&: *CurBuilder); |
3903 | } else { |
3904 | // TODO: test using android subtarget when we support @llvm.thread.pointer. |
3905 | const Value *IRGuard = TLI->getSDagStackGuard(M); |
3906 | Register GuardPtr = getOrCreateVReg(Val: *IRGuard); |
3907 | |
3908 | Guard = CurBuilder |
3909 | ->buildLoad(Res: PtrMemTy, Addr: GuardPtr, |
3910 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), Alignment: Align, |
3911 | MMOFlags: MachineMemOperand::MOLoad | |
3912 | MachineMemOperand::MOVolatile) |
3913 | .getReg(Idx: 0); |
3914 | } |
3915 | |
3916 | // Perform the comparison. |
3917 | auto Cmp = |
3918 | CurBuilder->buildICmp(Pred: CmpInst::ICMP_NE, Res: LLT::scalar(SizeInBits: 1), Op0: Guard, Op1: GuardVal); |
3919 | // If the guard/stackslot do not equal, branch to failure MBB. |
3920 | CurBuilder->buildBrCond(Tst: Cmp, Dest&: *SPD.getFailureMBB()); |
3921 | // Otherwise branch to success MBB. |
3922 | CurBuilder->buildBr(Dest&: *SPD.getSuccessMBB()); |
3923 | return true; |
3924 | } |
3925 | |
3926 | bool IRTranslator::emitSPDescriptorFailure(StackProtectorDescriptor &SPD, |
3927 | MachineBasicBlock *FailureBB) { |
3928 | CurBuilder->setInsertPt(MBB&: *FailureBB, II: FailureBB->end()); |
3929 | |
3930 | const RTLIB::Libcall Libcall = RTLIB::STACKPROTECTOR_CHECK_FAIL; |
3931 | const char *Name = TLI->getLibcallName(Call: Libcall); |
3932 | |
3933 | CallLowering::CallLoweringInfo Info; |
3934 | Info.CallConv = TLI->getLibcallCallingConv(Call: Libcall); |
3935 | Info.Callee = MachineOperand::CreateES(SymName: Name); |
3936 | Info.OrigRet = {Register(), Type::getVoidTy(C&: MF->getFunction().getContext()), |
3937 | 0}; |
3938 | if (!CLI->lowerCall(MIRBuilder&: *CurBuilder, Info)) { |
3939 | LLVM_DEBUG(dbgs() << "Failed to lower call to stack protector fail\n" ); |
3940 | return false; |
3941 | } |
3942 | |
3943 | // Emit a trap instruction if we are required to do so. |
3944 | const TargetOptions &TargetOpts = TLI->getTargetMachine().Options; |
3945 | if (TargetOpts.TrapUnreachable && !TargetOpts.NoTrapAfterNoreturn) |
3946 | CurBuilder->buildInstr(Opcode: TargetOpcode::G_TRAP); |
3947 | |
3948 | return true; |
3949 | } |
3950 | |
3951 | void IRTranslator::finalizeFunction() { |
3952 | // Release the memory used by the different maps we |
3953 | // needed during the translation. |
3954 | PendingPHIs.clear(); |
3955 | VMap.reset(); |
3956 | FrameIndices.clear(); |
3957 | MachinePreds.clear(); |
3958 | // MachineIRBuilder::DebugLoc can outlive the DILocation it holds. Clear it |
3959 | // to avoid accessing free’d memory (in runOnMachineFunction) and to avoid |
3960 | // destroying it twice (in ~IRTranslator() and ~LLVMContext()) |
3961 | EntryBuilder.reset(); |
3962 | CurBuilder.reset(); |
3963 | FuncInfo.clear(); |
3964 | SPDescriptor.resetPerFunctionState(); |
3965 | } |
3966 | |
3967 | /// Returns true if a BasicBlock \p BB within a variadic function contains a |
3968 | /// variadic musttail call. |
3969 | static bool checkForMustTailInVarArgFn(bool IsVarArg, const BasicBlock &BB) { |
3970 | if (!IsVarArg) |
3971 | return false; |
3972 | |
3973 | // Walk the block backwards, because tail calls usually only appear at the end |
3974 | // of a block. |
3975 | return llvm::any_of(Range: llvm::reverse(C: BB), P: [](const Instruction &I) { |
3976 | const auto *CI = dyn_cast<CallInst>(Val: &I); |
3977 | return CI && CI->isMustTailCall(); |
3978 | }); |
3979 | } |
3980 | |
3981 | bool IRTranslator::runOnMachineFunction(MachineFunction &CurMF) { |
3982 | MF = &CurMF; |
3983 | const Function &F = MF->getFunction(); |
3984 | GISelCSEAnalysisWrapper &Wrapper = |
3985 | getAnalysis<GISelCSEAnalysisWrapperPass>().getCSEWrapper(); |
3986 | // Set the CSEConfig and run the analysis. |
3987 | GISelCSEInfo *CSEInfo = nullptr; |
3988 | TPC = &getAnalysis<TargetPassConfig>(); |
3989 | bool EnableCSE = EnableCSEInIRTranslator.getNumOccurrences() |
3990 | ? EnableCSEInIRTranslator |
3991 | : TPC->isGISelCSEEnabled(); |
3992 | TLI = MF->getSubtarget().getTargetLowering(); |
3993 | |
3994 | if (EnableCSE) { |
3995 | EntryBuilder = std::make_unique<CSEMIRBuilder>(args&: CurMF); |
3996 | CSEInfo = &Wrapper.get(CSEOpt: TPC->getCSEConfig()); |
3997 | EntryBuilder->setCSEInfo(CSEInfo); |
3998 | CurBuilder = std::make_unique<CSEMIRBuilder>(args&: CurMF); |
3999 | CurBuilder->setCSEInfo(CSEInfo); |
4000 | } else { |
4001 | EntryBuilder = std::make_unique<MachineIRBuilder>(); |
4002 | CurBuilder = std::make_unique<MachineIRBuilder>(); |
4003 | } |
4004 | CLI = MF->getSubtarget().getCallLowering(); |
4005 | CurBuilder->setMF(*MF); |
4006 | EntryBuilder->setMF(*MF); |
4007 | MRI = &MF->getRegInfo(); |
4008 | DL = &F.getDataLayout(); |
4009 | ORE = std::make_unique<OptimizationRemarkEmitter>(args: &F); |
4010 | const TargetMachine &TM = MF->getTarget(); |
4011 | TM.resetTargetOptions(F); |
4012 | EnableOpts = OptLevel != CodeGenOptLevel::None && !skipFunction(F); |
4013 | FuncInfo.MF = MF; |
4014 | if (EnableOpts) { |
4015 | AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
4016 | FuncInfo.BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI(); |
4017 | } else { |
4018 | AA = nullptr; |
4019 | FuncInfo.BPI = nullptr; |
4020 | } |
4021 | |
4022 | AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache( |
4023 | F&: MF->getFunction()); |
4024 | LibInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
4025 | FuncInfo.CanLowerReturn = CLI->checkReturnTypeForCallConv(MF&: *MF); |
4026 | |
4027 | SL = std::make_unique<GISelSwitchLowering>(args: this, args&: FuncInfo); |
4028 | SL->init(tli: *TLI, tm: TM, dl: *DL); |
4029 | |
4030 | assert(PendingPHIs.empty() && "stale PHIs" ); |
4031 | |
4032 | // Targets which want to use big endian can enable it using |
4033 | // enableBigEndian() |
4034 | if (!DL->isLittleEndian() && !CLI->enableBigEndian()) { |
4035 | // Currently we don't properly handle big endian code. |
4036 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
4037 | F.getSubprogram(), &F.getEntryBlock()); |
4038 | R << "unable to translate in big endian mode" ; |
4039 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
4040 | return false; |
4041 | } |
4042 | |
4043 | // Release the per-function state when we return, whether we succeeded or not. |
4044 | auto FinalizeOnReturn = make_scope_exit(F: [this]() { finalizeFunction(); }); |
4045 | |
4046 | // Setup a separate basic-block for the arguments and constants |
4047 | MachineBasicBlock *EntryBB = MF->CreateMachineBasicBlock(); |
4048 | MF->push_back(MBB: EntryBB); |
4049 | EntryBuilder->setMBB(*EntryBB); |
4050 | |
4051 | DebugLoc DbgLoc = F.getEntryBlock().getFirstNonPHIIt()->getDebugLoc(); |
4052 | SwiftError.setFunction(CurMF); |
4053 | SwiftError.createEntriesInEntryBlock(DbgLoc); |
4054 | |
4055 | bool IsVarArg = F.isVarArg(); |
4056 | bool HasMustTailInVarArgFn = false; |
4057 | |
4058 | // Create all blocks, in IR order, to preserve the layout. |
4059 | FuncInfo.MBBMap.resize(N: F.getMaxBlockNumber()); |
4060 | for (const BasicBlock &BB: F) { |
4061 | auto *&MBB = FuncInfo.MBBMap[BB.getNumber()]; |
4062 | |
4063 | MBB = MF->CreateMachineBasicBlock(BB: &BB); |
4064 | MF->push_back(MBB); |
4065 | |
4066 | if (BB.hasAddressTaken()) |
4067 | MBB->setAddressTakenIRBlock(const_cast<BasicBlock *>(&BB)); |
4068 | |
4069 | if (!HasMustTailInVarArgFn) |
4070 | HasMustTailInVarArgFn = checkForMustTailInVarArgFn(IsVarArg, BB); |
4071 | } |
4072 | |
4073 | MF->getFrameInfo().setHasMustTailInVarArgFunc(HasMustTailInVarArgFn); |
4074 | |
4075 | // Make our arguments/constants entry block fallthrough to the IR entry block. |
4076 | EntryBB->addSuccessor(Succ: &getMBB(BB: F.front())); |
4077 | |
4078 | if (CLI->fallBackToDAGISel(MF: *MF)) { |
4079 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
4080 | F.getSubprogram(), &F.getEntryBlock()); |
4081 | R << "unable to lower function: " |
4082 | << ore::NV("Prototype" , F.getFunctionType()); |
4083 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
4084 | return false; |
4085 | } |
4086 | |
4087 | // Lower the actual args into this basic block. |
4088 | SmallVector<ArrayRef<Register>, 8> VRegArgs; |
4089 | for (const Argument &Arg: F.args()) { |
4090 | if (DL->getTypeStoreSize(Ty: Arg.getType()).isZero()) |
4091 | continue; // Don't handle zero sized types. |
4092 | ArrayRef<Register> VRegs = getOrCreateVRegs(Val: Arg); |
4093 | VRegArgs.push_back(Elt: VRegs); |
4094 | |
4095 | if (Arg.hasSwiftErrorAttr()) { |
4096 | assert(VRegs.size() == 1 && "Too many vregs for Swift error" ); |
4097 | SwiftError.setCurrentVReg(MBB: EntryBB, SwiftError.getFunctionArg(), VRegs[0]); |
4098 | } |
4099 | } |
4100 | |
4101 | if (!CLI->lowerFormalArguments(MIRBuilder&: *EntryBuilder, F, VRegs: VRegArgs, FLI&: FuncInfo)) { |
4102 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
4103 | F.getSubprogram(), &F.getEntryBlock()); |
4104 | R << "unable to lower arguments: " |
4105 | << ore::NV("Prototype" , F.getFunctionType()); |
4106 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
4107 | return false; |
4108 | } |
4109 | |
4110 | // Need to visit defs before uses when translating instructions. |
4111 | GISelObserverWrapper WrapperObserver; |
4112 | if (EnableCSE && CSEInfo) |
4113 | WrapperObserver.addObserver(O: CSEInfo); |
4114 | { |
4115 | ReversePostOrderTraversal<const Function *> RPOT(&F); |
4116 | #ifndef NDEBUG |
4117 | DILocationVerifier Verifier; |
4118 | WrapperObserver.addObserver(&Verifier); |
4119 | #endif // ifndef NDEBUG |
4120 | RAIIMFObsDelInstaller ObsInstall(*MF, WrapperObserver); |
4121 | for (const BasicBlock *BB : RPOT) { |
4122 | MachineBasicBlock &MBB = getMBB(BB: *BB); |
4123 | // Set the insertion point of all the following translations to |
4124 | // the end of this basic block. |
4125 | CurBuilder->setMBB(MBB); |
4126 | HasTailCall = false; |
4127 | for (const Instruction &Inst : *BB) { |
4128 | // If we translated a tail call in the last step, then we know |
4129 | // everything after the call is either a return, or something that is |
4130 | // handled by the call itself. (E.g. a lifetime marker or assume |
4131 | // intrinsic.) In this case, we should stop translating the block and |
4132 | // move on. |
4133 | if (HasTailCall) |
4134 | break; |
4135 | #ifndef NDEBUG |
4136 | Verifier.setCurrentInst(&Inst); |
4137 | #endif // ifndef NDEBUG |
4138 | |
4139 | // Translate any debug-info attached to the instruction. |
4140 | translateDbgInfo(Inst, MIRBuilder&: *CurBuilder); |
4141 | |
4142 | if (translate(Inst)) |
4143 | continue; |
4144 | |
4145 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
4146 | Inst.getDebugLoc(), BB); |
4147 | R << "unable to translate instruction: " << ore::NV("Opcode" , &Inst); |
4148 | |
4149 | if (ORE->allowExtraAnalysis(PassName: "gisel-irtranslator" )) { |
4150 | std::string InstStrStorage; |
4151 | raw_string_ostream InstStr(InstStrStorage); |
4152 | InstStr << Inst; |
4153 | |
4154 | R << ": '" << InstStrStorage << "'" ; |
4155 | } |
4156 | |
4157 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
4158 | return false; |
4159 | } |
4160 | |
4161 | if (!finalizeBasicBlock(BB: *BB, MBB)) { |
4162 | OptimizationRemarkMissed R("gisel-irtranslator" , "GISelFailure" , |
4163 | BB->getTerminator()->getDebugLoc(), BB); |
4164 | R << "unable to translate basic block" ; |
4165 | reportTranslationError(MF&: *MF, TPC: *TPC, ORE&: *ORE, R); |
4166 | return false; |
4167 | } |
4168 | } |
4169 | #ifndef NDEBUG |
4170 | WrapperObserver.removeObserver(&Verifier); |
4171 | #endif |
4172 | } |
4173 | |
4174 | finishPendingPhis(); |
4175 | |
4176 | SwiftError.propagateVRegs(); |
4177 | |
4178 | // Merge the argument lowering and constants block with its single |
4179 | // successor, the LLVM-IR entry block. We want the basic block to |
4180 | // be maximal. |
4181 | assert(EntryBB->succ_size() == 1 && |
4182 | "Custom BB used for lowering should have only one successor" ); |
4183 | // Get the successor of the current entry block. |
4184 | MachineBasicBlock &NewEntryBB = **EntryBB->succ_begin(); |
4185 | assert(NewEntryBB.pred_size() == 1 && |
4186 | "LLVM-IR entry block has a predecessor!?" ); |
4187 | // Move all the instruction from the current entry block to the |
4188 | // new entry block. |
4189 | NewEntryBB.splice(Where: NewEntryBB.begin(), Other: EntryBB, From: EntryBB->begin(), |
4190 | To: EntryBB->end()); |
4191 | |
4192 | // Update the live-in information for the new entry block. |
4193 | for (const MachineBasicBlock::RegisterMaskPair &LiveIn : EntryBB->liveins()) |
4194 | NewEntryBB.addLiveIn(RegMaskPair: LiveIn); |
4195 | NewEntryBB.sortUniqueLiveIns(); |
4196 | |
4197 | // Get rid of the now empty basic block. |
4198 | EntryBB->removeSuccessor(Succ: &NewEntryBB); |
4199 | MF->remove(MBBI: EntryBB); |
4200 | MF->deleteMachineBasicBlock(MBB: EntryBB); |
4201 | |
4202 | assert(&MF->front() == &NewEntryBB && |
4203 | "New entry wasn't next in the list of basic block!" ); |
4204 | |
4205 | // Initialize stack protector information. |
4206 | StackProtector &SP = getAnalysis<StackProtector>(); |
4207 | SP.copyToMachineFrameInfo(MFI&: MF->getFrameInfo()); |
4208 | |
4209 | return false; |
4210 | } |
4211 | |