1 //===-- PPCCTRLoops.cpp - Identify and generate CTR loops -----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass identifies loops where we can generate the PPC branch instructions
11 // that decrement and test the count register (CTR) (bdnz and friends).
12 //
13 // The pattern that defines the induction variable can changed depending on
14 // prior optimizations. For example, the IndVarSimplify phase run by 'opt'
15 // normalizes induction variables, and the Loop Strength Reduction pass
16 // run by 'llc' may also make changes to the induction variable.
17 //
18 // Criteria for CTR loops:
19 // - Countable loops (w/ ind. var for a trip count)
20 // - Try inner-most loops first
21 // - No nested CTR loops.
22 // - No function calls in loops.
23 //
24 //===----------------------------------------------------------------------===//
25
26 #include "llvm/Transforms/Scalar.h"
27 #include "PPC.h"
28 #include "PPCTargetMachine.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/ScalarEvolutionExpander.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/InlineAsm.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Module.h"
41 #include "llvm/IR/ValueHandle.h"
42 #include "llvm/PassSupport.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Utils/LoopUtils.h"
49
50 #ifndef NDEBUG
51 #include "llvm/CodeGen/MachineDominators.h"
52 #include "llvm/CodeGen/MachineFunction.h"
53 #include "llvm/CodeGen/MachineFunctionPass.h"
54 #include "llvm/CodeGen/MachineRegisterInfo.h"
55 #endif
56
57 #include <algorithm>
58 #include <vector>
59
60 using namespace llvm;
61
62 #define DEBUG_TYPE "ctrloops"
63
64 #ifndef NDEBUG
65 static cl::opt<int> CTRLoopLimit("ppc-max-ctrloop", cl::Hidden, cl::init(-1));
66 #endif
67
68 STATISTIC(NumCTRLoops, "Number of loops converted to CTR loops");
69
70 namespace llvm {
71 void initializePPCCTRLoopsPass(PassRegistry&);
72 #ifndef NDEBUG
73 void initializePPCCTRLoopsVerifyPass(PassRegistry&);
74 #endif
75 }
76
77 namespace {
78 struct PPCCTRLoops : public FunctionPass {
79
80 #ifndef NDEBUG
81 static int Counter;
82 #endif
83
84 public:
85 static char ID;
86
PPCCTRLoops__anonb8941c5d0111::PPCCTRLoops87 PPCCTRLoops() : FunctionPass(ID), TM(nullptr) {
88 initializePPCCTRLoopsPass(*PassRegistry::getPassRegistry());
89 }
PPCCTRLoops__anonb8941c5d0111::PPCCTRLoops90 PPCCTRLoops(PPCTargetMachine &TM) : FunctionPass(ID), TM(&TM) {
91 initializePPCCTRLoopsPass(*PassRegistry::getPassRegistry());
92 }
93
94 bool runOnFunction(Function &F) override;
95
getAnalysisUsage__anonb8941c5d0111::PPCCTRLoops96 void getAnalysisUsage(AnalysisUsage &AU) const override {
97 AU.addRequired<LoopInfoWrapperPass>();
98 AU.addPreserved<LoopInfoWrapperPass>();
99 AU.addRequired<DominatorTreeWrapperPass>();
100 AU.addPreserved<DominatorTreeWrapperPass>();
101 AU.addRequired<ScalarEvolution>();
102 }
103
104 private:
105 bool mightUseCTR(const Triple &TT, BasicBlock *BB);
106 bool convertToCTRLoop(Loop *L);
107
108 private:
109 PPCTargetMachine *TM;
110 LoopInfo *LI;
111 ScalarEvolution *SE;
112 const DataLayout *DL;
113 DominatorTree *DT;
114 const TargetLibraryInfo *LibInfo;
115 };
116
117 char PPCCTRLoops::ID = 0;
118 #ifndef NDEBUG
119 int PPCCTRLoops::Counter = 0;
120 #endif
121
122 #ifndef NDEBUG
123 struct PPCCTRLoopsVerify : public MachineFunctionPass {
124 public:
125 static char ID;
126
PPCCTRLoopsVerify__anonb8941c5d0111::PPCCTRLoopsVerify127 PPCCTRLoopsVerify() : MachineFunctionPass(ID) {
128 initializePPCCTRLoopsVerifyPass(*PassRegistry::getPassRegistry());
129 }
130
getAnalysisUsage__anonb8941c5d0111::PPCCTRLoopsVerify131 void getAnalysisUsage(AnalysisUsage &AU) const override {
132 AU.addRequired<MachineDominatorTree>();
133 MachineFunctionPass::getAnalysisUsage(AU);
134 }
135
136 bool runOnMachineFunction(MachineFunction &MF) override;
137
138 private:
139 MachineDominatorTree *MDT;
140 };
141
142 char PPCCTRLoopsVerify::ID = 0;
143 #endif // NDEBUG
144 } // end anonymous namespace
145
146 INITIALIZE_PASS_BEGIN(PPCCTRLoops, "ppc-ctr-loops", "PowerPC CTR Loops",
147 false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)148 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
149 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
150 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
151 INITIALIZE_PASS_END(PPCCTRLoops, "ppc-ctr-loops", "PowerPC CTR Loops",
152 false, false)
153
154 FunctionPass *llvm::createPPCCTRLoops(PPCTargetMachine &TM) {
155 return new PPCCTRLoops(TM);
156 }
157
158 #ifndef NDEBUG
159 INITIALIZE_PASS_BEGIN(PPCCTRLoopsVerify, "ppc-ctr-loops-verify",
160 "PowerPC CTR Loops Verify", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)161 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
162 INITIALIZE_PASS_END(PPCCTRLoopsVerify, "ppc-ctr-loops-verify",
163 "PowerPC CTR Loops Verify", false, false)
164
165 FunctionPass *llvm::createPPCCTRLoopsVerify() {
166 return new PPCCTRLoopsVerify();
167 }
168 #endif // NDEBUG
169
runOnFunction(Function & F)170 bool PPCCTRLoops::runOnFunction(Function &F) {
171 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
172 SE = &getAnalysis<ScalarEvolution>();
173 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
174 DL = &F.getParent()->getDataLayout();
175 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
176 LibInfo = TLIP ? &TLIP->getTLI() : nullptr;
177
178 bool MadeChange = false;
179
180 for (LoopInfo::iterator I = LI->begin(), E = LI->end();
181 I != E; ++I) {
182 Loop *L = *I;
183 if (!L->getParentLoop())
184 MadeChange |= convertToCTRLoop(L);
185 }
186
187 return MadeChange;
188 }
189
isLargeIntegerTy(bool Is32Bit,Type * Ty)190 static bool isLargeIntegerTy(bool Is32Bit, Type *Ty) {
191 if (IntegerType *ITy = dyn_cast<IntegerType>(Ty))
192 return ITy->getBitWidth() > (Is32Bit ? 32U : 64U);
193
194 return false;
195 }
196
197 // Determining the address of a TLS variable results in a function call in
198 // certain TLS models.
memAddrUsesCTR(const PPCTargetMachine * TM,const Value * MemAddr)199 static bool memAddrUsesCTR(const PPCTargetMachine *TM,
200 const Value *MemAddr) {
201 const auto *GV = dyn_cast<GlobalValue>(MemAddr);
202 if (!GV) {
203 // Recurse to check for constants that refer to TLS global variables.
204 if (const auto *CV = dyn_cast<Constant>(MemAddr))
205 for (const auto &CO : CV->operands())
206 if (memAddrUsesCTR(TM, CO))
207 return true;
208
209 return false;
210 }
211
212 if (!GV->isThreadLocal())
213 return false;
214 if (!TM)
215 return true;
216 TLSModel::Model Model = TM->getTLSModel(GV);
217 return Model == TLSModel::GeneralDynamic || Model == TLSModel::LocalDynamic;
218 }
219
mightUseCTR(const Triple & TT,BasicBlock * BB)220 bool PPCCTRLoops::mightUseCTR(const Triple &TT, BasicBlock *BB) {
221 for (BasicBlock::iterator J = BB->begin(), JE = BB->end();
222 J != JE; ++J) {
223 if (CallInst *CI = dyn_cast<CallInst>(J)) {
224 if (InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue())) {
225 // Inline ASM is okay, unless it clobbers the ctr register.
226 InlineAsm::ConstraintInfoVector CIV = IA->ParseConstraints();
227 for (unsigned i = 0, ie = CIV.size(); i < ie; ++i) {
228 InlineAsm::ConstraintInfo &C = CIV[i];
229 if (C.Type != InlineAsm::isInput)
230 for (unsigned j = 0, je = C.Codes.size(); j < je; ++j)
231 if (StringRef(C.Codes[j]).equals_lower("{ctr}"))
232 return true;
233 }
234
235 continue;
236 }
237
238 if (!TM)
239 return true;
240 const TargetLowering *TLI =
241 TM->getSubtargetImpl(*BB->getParent())->getTargetLowering();
242
243 if (Function *F = CI->getCalledFunction()) {
244 // Most intrinsics don't become function calls, but some might.
245 // sin, cos, exp and log are always calls.
246 unsigned Opcode;
247 if (F->getIntrinsicID() != Intrinsic::not_intrinsic) {
248 switch (F->getIntrinsicID()) {
249 default: continue;
250 // If we have a call to ppc_is_decremented_ctr_nonzero, or ppc_mtctr
251 // we're definitely using CTR.
252 case Intrinsic::ppc_is_decremented_ctr_nonzero:
253 case Intrinsic::ppc_mtctr:
254 return true;
255
256 // VisualStudio defines setjmp as _setjmp
257 #if defined(_MSC_VER) && defined(setjmp) && \
258 !defined(setjmp_undefined_for_msvc)
259 # pragma push_macro("setjmp")
260 # undef setjmp
261 # define setjmp_undefined_for_msvc
262 #endif
263
264 case Intrinsic::setjmp:
265
266 #if defined(_MSC_VER) && defined(setjmp_undefined_for_msvc)
267 // let's return it to _setjmp state
268 # pragma pop_macro("setjmp")
269 # undef setjmp_undefined_for_msvc
270 #endif
271
272 case Intrinsic::longjmp:
273
274 // Exclude eh_sjlj_setjmp; we don't need to exclude eh_sjlj_longjmp
275 // because, although it does clobber the counter register, the
276 // control can't then return to inside the loop unless there is also
277 // an eh_sjlj_setjmp.
278 case Intrinsic::eh_sjlj_setjmp:
279
280 case Intrinsic::memcpy:
281 case Intrinsic::memmove:
282 case Intrinsic::memset:
283 case Intrinsic::powi:
284 case Intrinsic::log:
285 case Intrinsic::log2:
286 case Intrinsic::log10:
287 case Intrinsic::exp:
288 case Intrinsic::exp2:
289 case Intrinsic::pow:
290 case Intrinsic::sin:
291 case Intrinsic::cos:
292 return true;
293 case Intrinsic::copysign:
294 if (CI->getArgOperand(0)->getType()->getScalarType()->
295 isPPC_FP128Ty())
296 return true;
297 else
298 continue; // ISD::FCOPYSIGN is never a library call.
299 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
300 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
301 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
302 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
303 case Intrinsic::rint: Opcode = ISD::FRINT; break;
304 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
305 case Intrinsic::round: Opcode = ISD::FROUND; break;
306 }
307 }
308
309 // PowerPC does not use [US]DIVREM or other library calls for
310 // operations on regular types which are not otherwise library calls
311 // (i.e. soft float or atomics). If adapting for targets that do,
312 // additional care is required here.
313
314 LibFunc::Func Func;
315 if (!F->hasLocalLinkage() && F->hasName() && LibInfo &&
316 LibInfo->getLibFunc(F->getName(), Func) &&
317 LibInfo->hasOptimizedCodeGen(Func)) {
318 // Non-read-only functions are never treated as intrinsics.
319 if (!CI->onlyReadsMemory())
320 return true;
321
322 // Conversion happens only for FP calls.
323 if (!CI->getArgOperand(0)->getType()->isFloatingPointTy())
324 return true;
325
326 switch (Func) {
327 default: return true;
328 case LibFunc::copysign:
329 case LibFunc::copysignf:
330 continue; // ISD::FCOPYSIGN is never a library call.
331 case LibFunc::copysignl:
332 return true;
333 case LibFunc::fabs:
334 case LibFunc::fabsf:
335 case LibFunc::fabsl:
336 continue; // ISD::FABS is never a library call.
337 case LibFunc::sqrt:
338 case LibFunc::sqrtf:
339 case LibFunc::sqrtl:
340 Opcode = ISD::FSQRT; break;
341 case LibFunc::floor:
342 case LibFunc::floorf:
343 case LibFunc::floorl:
344 Opcode = ISD::FFLOOR; break;
345 case LibFunc::nearbyint:
346 case LibFunc::nearbyintf:
347 case LibFunc::nearbyintl:
348 Opcode = ISD::FNEARBYINT; break;
349 case LibFunc::ceil:
350 case LibFunc::ceilf:
351 case LibFunc::ceill:
352 Opcode = ISD::FCEIL; break;
353 case LibFunc::rint:
354 case LibFunc::rintf:
355 case LibFunc::rintl:
356 Opcode = ISD::FRINT; break;
357 case LibFunc::round:
358 case LibFunc::roundf:
359 case LibFunc::roundl:
360 Opcode = ISD::FROUND; break;
361 case LibFunc::trunc:
362 case LibFunc::truncf:
363 case LibFunc::truncl:
364 Opcode = ISD::FTRUNC; break;
365 }
366
367 auto &DL = CI->getModule()->getDataLayout();
368 MVT VTy = TLI->getSimpleValueType(DL, CI->getArgOperand(0)->getType(),
369 true);
370 if (VTy == MVT::Other)
371 return true;
372
373 if (TLI->isOperationLegalOrCustom(Opcode, VTy))
374 continue;
375 else if (VTy.isVector() &&
376 TLI->isOperationLegalOrCustom(Opcode, VTy.getScalarType()))
377 continue;
378
379 return true;
380 }
381 }
382
383 return true;
384 } else if (isa<BinaryOperator>(J) &&
385 J->getType()->getScalarType()->isPPC_FP128Ty()) {
386 // Most operations on ppc_f128 values become calls.
387 return true;
388 } else if (isa<UIToFPInst>(J) || isa<SIToFPInst>(J) ||
389 isa<FPToUIInst>(J) || isa<FPToSIInst>(J)) {
390 CastInst *CI = cast<CastInst>(J);
391 if (CI->getSrcTy()->getScalarType()->isPPC_FP128Ty() ||
392 CI->getDestTy()->getScalarType()->isPPC_FP128Ty() ||
393 isLargeIntegerTy(TT.isArch32Bit(), CI->getSrcTy()->getScalarType()) ||
394 isLargeIntegerTy(TT.isArch32Bit(), CI->getDestTy()->getScalarType()))
395 return true;
396 } else if (isLargeIntegerTy(TT.isArch32Bit(),
397 J->getType()->getScalarType()) &&
398 (J->getOpcode() == Instruction::UDiv ||
399 J->getOpcode() == Instruction::SDiv ||
400 J->getOpcode() == Instruction::URem ||
401 J->getOpcode() == Instruction::SRem)) {
402 return true;
403 } else if (TT.isArch32Bit() &&
404 isLargeIntegerTy(false, J->getType()->getScalarType()) &&
405 (J->getOpcode() == Instruction::Shl ||
406 J->getOpcode() == Instruction::AShr ||
407 J->getOpcode() == Instruction::LShr)) {
408 // Only on PPC32, for 128-bit integers (specifically not 64-bit
409 // integers), these might be runtime calls.
410 return true;
411 } else if (isa<IndirectBrInst>(J) || isa<InvokeInst>(J)) {
412 // On PowerPC, indirect jumps use the counter register.
413 return true;
414 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(J)) {
415 if (!TM)
416 return true;
417 const TargetLowering *TLI =
418 TM->getSubtargetImpl(*BB->getParent())->getTargetLowering();
419
420 if (SI->getNumCases() + 1 >= (unsigned)TLI->getMinimumJumpTableEntries())
421 return true;
422 }
423 for (Value *Operand : J->operands())
424 if (memAddrUsesCTR(TM, Operand))
425 return true;
426 }
427
428 return false;
429 }
430
convertToCTRLoop(Loop * L)431 bool PPCCTRLoops::convertToCTRLoop(Loop *L) {
432 bool MadeChange = false;
433
434 const Triple TT =
435 Triple(L->getHeader()->getParent()->getParent()->getTargetTriple());
436 if (!TT.isArch32Bit() && !TT.isArch64Bit())
437 return MadeChange; // Unknown arch. type.
438
439 // Process nested loops first.
440 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
441 MadeChange |= convertToCTRLoop(*I);
442 DEBUG(dbgs() << "Nested loop converted\n");
443 }
444
445 // If a nested loop has been converted, then we can't convert this loop.
446 if (MadeChange)
447 return MadeChange;
448
449 #ifndef NDEBUG
450 // Stop trying after reaching the limit (if any).
451 int Limit = CTRLoopLimit;
452 if (Limit >= 0) {
453 if (Counter >= CTRLoopLimit)
454 return false;
455 Counter++;
456 }
457 #endif
458
459 // We don't want to spill/restore the counter register, and so we don't
460 // want to use the counter register if the loop contains calls.
461 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
462 I != IE; ++I)
463 if (mightUseCTR(TT, *I))
464 return MadeChange;
465
466 SmallVector<BasicBlock*, 4> ExitingBlocks;
467 L->getExitingBlocks(ExitingBlocks);
468
469 BasicBlock *CountedExitBlock = nullptr;
470 const SCEV *ExitCount = nullptr;
471 BranchInst *CountedExitBranch = nullptr;
472 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
473 IE = ExitingBlocks.end(); I != IE; ++I) {
474 const SCEV *EC = SE->getExitCount(L, *I);
475 DEBUG(dbgs() << "Exit Count for " << *L << " from block " <<
476 (*I)->getName() << ": " << *EC << "\n");
477 if (isa<SCEVCouldNotCompute>(EC))
478 continue;
479 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
480 if (ConstEC->getValue()->isZero())
481 continue;
482 } else if (!SE->isLoopInvariant(EC, L))
483 continue;
484
485 if (SE->getTypeSizeInBits(EC->getType()) > (TT.isArch64Bit() ? 64 : 32))
486 continue;
487
488 // We now have a loop-invariant count of loop iterations (which is not the
489 // constant zero) for which we know that this loop will not exit via this
490 // exisiting block.
491
492 // We need to make sure that this block will run on every loop iteration.
493 // For this to be true, we must dominate all blocks with backedges. Such
494 // blocks are in-loop predecessors to the header block.
495 bool NotAlways = false;
496 for (pred_iterator PI = pred_begin(L->getHeader()),
497 PIE = pred_end(L->getHeader()); PI != PIE; ++PI) {
498 if (!L->contains(*PI))
499 continue;
500
501 if (!DT->dominates(*I, *PI)) {
502 NotAlways = true;
503 break;
504 }
505 }
506
507 if (NotAlways)
508 continue;
509
510 // Make sure this blocks ends with a conditional branch.
511 Instruction *TI = (*I)->getTerminator();
512 if (!TI)
513 continue;
514
515 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
516 if (!BI->isConditional())
517 continue;
518
519 CountedExitBranch = BI;
520 } else
521 continue;
522
523 // Note that this block may not be the loop latch block, even if the loop
524 // has a latch block.
525 CountedExitBlock = *I;
526 ExitCount = EC;
527 break;
528 }
529
530 if (!CountedExitBlock)
531 return MadeChange;
532
533 BasicBlock *Preheader = L->getLoopPreheader();
534
535 // If we don't have a preheader, then insert one. If we already have a
536 // preheader, then we can use it (except if the preheader contains a use of
537 // the CTR register because some such uses might be reordered by the
538 // selection DAG after the mtctr instruction).
539 if (!Preheader || mightUseCTR(TT, Preheader))
540 Preheader = InsertPreheaderForLoop(L, this);
541 if (!Preheader)
542 return MadeChange;
543
544 DEBUG(dbgs() << "Preheader for exit count: " << Preheader->getName() << "\n");
545
546 // Insert the count into the preheader and replace the condition used by the
547 // selected branch.
548 MadeChange = true;
549
550 SCEVExpander SCEVE(*SE, Preheader->getModule()->getDataLayout(), "loopcnt");
551 LLVMContext &C = SE->getContext();
552 Type *CountType = TT.isArch64Bit() ? Type::getInt64Ty(C) :
553 Type::getInt32Ty(C);
554 if (!ExitCount->getType()->isPointerTy() &&
555 ExitCount->getType() != CountType)
556 ExitCount = SE->getZeroExtendExpr(ExitCount, CountType);
557 ExitCount = SE->getAddExpr(ExitCount,
558 SE->getConstant(CountType, 1));
559 Value *ECValue = SCEVE.expandCodeFor(ExitCount, CountType,
560 Preheader->getTerminator());
561
562 IRBuilder<> CountBuilder(Preheader->getTerminator());
563 Module *M = Preheader->getParent()->getParent();
564 Value *MTCTRFunc = Intrinsic::getDeclaration(M, Intrinsic::ppc_mtctr,
565 CountType);
566 CountBuilder.CreateCall(MTCTRFunc, ECValue);
567
568 IRBuilder<> CondBuilder(CountedExitBranch);
569 Value *DecFunc =
570 Intrinsic::getDeclaration(M, Intrinsic::ppc_is_decremented_ctr_nonzero);
571 Value *NewCond = CondBuilder.CreateCall(DecFunc, {});
572 Value *OldCond = CountedExitBranch->getCondition();
573 CountedExitBranch->setCondition(NewCond);
574
575 // The false branch must exit the loop.
576 if (!L->contains(CountedExitBranch->getSuccessor(0)))
577 CountedExitBranch->swapSuccessors();
578
579 // The old condition may be dead now, and may have even created a dead PHI
580 // (the original induction variable).
581 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
582 DeleteDeadPHIs(CountedExitBlock);
583
584 ++NumCTRLoops;
585 return MadeChange;
586 }
587
588 #ifndef NDEBUG
clobbersCTR(const MachineInstr * MI)589 static bool clobbersCTR(const MachineInstr *MI) {
590 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
591 const MachineOperand &MO = MI->getOperand(i);
592 if (MO.isReg()) {
593 if (MO.isDef() && (MO.getReg() == PPC::CTR || MO.getReg() == PPC::CTR8))
594 return true;
595 } else if (MO.isRegMask()) {
596 if (MO.clobbersPhysReg(PPC::CTR) || MO.clobbersPhysReg(PPC::CTR8))
597 return true;
598 }
599 }
600
601 return false;
602 }
603
verifyCTRBranch(MachineBasicBlock * MBB,MachineBasicBlock::iterator I)604 static bool verifyCTRBranch(MachineBasicBlock *MBB,
605 MachineBasicBlock::iterator I) {
606 MachineBasicBlock::iterator BI = I;
607 SmallSet<MachineBasicBlock *, 16> Visited;
608 SmallVector<MachineBasicBlock *, 8> Preds;
609 bool CheckPreds;
610
611 if (I == MBB->begin()) {
612 Visited.insert(MBB);
613 goto queue_preds;
614 } else
615 --I;
616
617 check_block:
618 Visited.insert(MBB);
619 if (I == MBB->end())
620 goto queue_preds;
621
622 CheckPreds = true;
623 for (MachineBasicBlock::iterator IE = MBB->begin();; --I) {
624 unsigned Opc = I->getOpcode();
625 if (Opc == PPC::MTCTRloop || Opc == PPC::MTCTR8loop) {
626 CheckPreds = false;
627 break;
628 }
629
630 if (I != BI && clobbersCTR(I)) {
631 DEBUG(dbgs() << "BB#" << MBB->getNumber() << " (" <<
632 MBB->getFullName() << ") instruction " << *I <<
633 " clobbers CTR, invalidating " << "BB#" <<
634 BI->getParent()->getNumber() << " (" <<
635 BI->getParent()->getFullName() << ") instruction " <<
636 *BI << "\n");
637 return false;
638 }
639
640 if (I == IE)
641 break;
642 }
643
644 if (!CheckPreds && Preds.empty())
645 return true;
646
647 if (CheckPreds) {
648 queue_preds:
649 if (MachineFunction::iterator(MBB) == MBB->getParent()->begin()) {
650 DEBUG(dbgs() << "Unable to find a MTCTR instruction for BB#" <<
651 BI->getParent()->getNumber() << " (" <<
652 BI->getParent()->getFullName() << ") instruction " <<
653 *BI << "\n");
654 return false;
655 }
656
657 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
658 PIE = MBB->pred_end(); PI != PIE; ++PI)
659 Preds.push_back(*PI);
660 }
661
662 do {
663 MBB = Preds.pop_back_val();
664 if (!Visited.count(MBB)) {
665 I = MBB->getLastNonDebugInstr();
666 goto check_block;
667 }
668 } while (!Preds.empty());
669
670 return true;
671 }
672
runOnMachineFunction(MachineFunction & MF)673 bool PPCCTRLoopsVerify::runOnMachineFunction(MachineFunction &MF) {
674 MDT = &getAnalysis<MachineDominatorTree>();
675
676 // Verify that all bdnz/bdz instructions are dominated by a loop mtctr before
677 // any other instructions that might clobber the ctr register.
678 for (MachineFunction::iterator I = MF.begin(), IE = MF.end();
679 I != IE; ++I) {
680 MachineBasicBlock *MBB = I;
681 if (!MDT->isReachableFromEntry(MBB))
682 continue;
683
684 for (MachineBasicBlock::iterator MII = MBB->getFirstTerminator(),
685 MIIE = MBB->end(); MII != MIIE; ++MII) {
686 unsigned Opc = MII->getOpcode();
687 if (Opc == PPC::BDNZ8 || Opc == PPC::BDNZ ||
688 Opc == PPC::BDZ8 || Opc == PPC::BDZ)
689 if (!verifyCTRBranch(MBB, MII))
690 llvm_unreachable("Invalid PPC CTR loop!");
691 }
692 }
693
694 return false;
695 }
696 #endif // NDEBUG
697
698