1 //===-- PPCCTRLoops.cpp - Identify and generate CTR loops -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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).
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.
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.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "ctrloops"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/PassSupport.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ValueHandle.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/LoopUtils.h"
48 #include "llvm/Target/TargetLibraryInfo.h"
49 #include "PPCTargetMachine.h"
53 #include "llvm/CodeGen/MachineDominators.h"
54 #include "llvm/CodeGen/MachineFunction.h"
55 #include "llvm/CodeGen/MachineFunctionPass.h"
56 #include "llvm/CodeGen/MachineRegisterInfo.h"
65 static cl::opt<int> CTRLoopLimit("ppc-max-ctrloop", cl::Hidden, cl::init(-1));
68 STATISTIC(NumCTRLoops, "Number of loops converted to CTR loops");
71 void initializePPCCTRLoopsPass(PassRegistry&);
73 void initializePPCCTRLoopsVerifyPass(PassRegistry&);
78 struct PPCCTRLoops : public FunctionPass {
87 PPCCTRLoops() : FunctionPass(ID), TM(0) {
88 initializePPCCTRLoopsPass(*PassRegistry::getPassRegistry());
90 PPCCTRLoops(PPCTargetMachine &TM) : FunctionPass(ID), TM(&TM) {
91 initializePPCCTRLoopsPass(*PassRegistry::getPassRegistry());
94 virtual bool runOnFunction(Function &F);
96 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
97 AU.addRequired<LoopInfo>();
98 AU.addPreserved<LoopInfo>();
99 AU.addRequired<DominatorTree>();
100 AU.addPreserved<DominatorTree>();
101 AU.addRequired<ScalarEvolution>();
105 bool mightUseCTR(const Triple &TT, BasicBlock *BB);
106 bool convertToCTRLoop(Loop *L);
109 PPCTargetMachine *TM;
114 const TargetLibraryInfo *LibInfo;
117 char PPCCTRLoops::ID = 0;
119 int PPCCTRLoops::Counter = 0;
123 struct PPCCTRLoopsVerify : public MachineFunctionPass {
127 PPCCTRLoopsVerify() : MachineFunctionPass(ID) {
128 initializePPCCTRLoopsVerifyPass(*PassRegistry::getPassRegistry());
131 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
132 AU.addRequired<MachineDominatorTree>();
133 MachineFunctionPass::getAnalysisUsage(AU);
136 virtual bool runOnMachineFunction(MachineFunction &MF);
139 MachineDominatorTree *MDT;
142 char PPCCTRLoopsVerify::ID = 0;
144 } // end anonymous namespace
146 INITIALIZE_PASS_BEGIN(PPCCTRLoops, "ppc-ctr-loops", "PowerPC CTR Loops",
148 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
149 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
150 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
151 INITIALIZE_PASS_END(PPCCTRLoops, "ppc-ctr-loops", "PowerPC CTR Loops",
154 FunctionPass *llvm::createPPCCTRLoops(PPCTargetMachine &TM) {
155 return new PPCCTRLoops(TM);
159 INITIALIZE_PASS_BEGIN(PPCCTRLoopsVerify, "ppc-ctr-loops-verify",
160 "PowerPC CTR Loops Verify", false, false)
161 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
162 INITIALIZE_PASS_END(PPCCTRLoopsVerify, "ppc-ctr-loops-verify",
163 "PowerPC CTR Loops Verify", false, false)
165 FunctionPass *llvm::createPPCCTRLoopsVerify() {
166 return new PPCCTRLoopsVerify();
170 bool PPCCTRLoops::runOnFunction(Function &F) {
171 LI = &getAnalysis<LoopInfo>();
172 SE = &getAnalysis<ScalarEvolution>();
173 DT = &getAnalysis<DominatorTree>();
174 TD = getAnalysisIfAvailable<DataLayout>();
175 LibInfo = getAnalysisIfAvailable<TargetLibraryInfo>();
177 bool MadeChange = false;
179 for (LoopInfo::iterator I = LI->begin(), E = LI->end();
182 if (!L->getParentLoop())
183 MadeChange |= convertToCTRLoop(L);
189 bool PPCCTRLoops::mightUseCTR(const Triple &TT, BasicBlock *BB) {
190 for (BasicBlock::iterator J = BB->begin(), JE = BB->end();
192 if (CallInst *CI = dyn_cast<CallInst>(J)) {
193 if (InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue())) {
194 // Inline ASM is okay, unless it clobbers the ctr register.
195 InlineAsm::ConstraintInfoVector CIV = IA->ParseConstraints();
196 for (unsigned i = 0, ie = CIV.size(); i < ie; ++i) {
197 InlineAsm::ConstraintInfo &C = CIV[i];
198 if (C.Type != InlineAsm::isInput)
199 for (unsigned j = 0, je = C.Codes.size(); j < je; ++j)
200 if (StringRef(C.Codes[j]).equals_lower("{ctr}"))
209 const TargetLowering *TLI = TM->getTargetLowering();
211 if (Function *F = CI->getCalledFunction()) {
212 // Most intrinsics don't become function calls, but some might.
213 // sin, cos, exp and log are always calls.
215 if (F->getIntrinsicID() != Intrinsic::not_intrinsic) {
216 switch (F->getIntrinsicID()) {
219 // VisualStudio defines setjmp as _setjmp
220 #if defined(_MSC_VER) && defined(setjmp) && \
221 !defined(setjmp_undefined_for_msvc)
222 # pragma push_macro("setjmp")
224 # define setjmp_undefined_for_msvc
227 case Intrinsic::setjmp:
229 #if defined(_MSC_VER) && defined(setjmp_undefined_for_msvc)
230 // let's return it to _setjmp state
231 # pragma pop_macro("setjmp")
232 # undef setjmp_undefined_for_msvc
235 case Intrinsic::longjmp:
236 case Intrinsic::memcpy:
237 case Intrinsic::memmove:
238 case Intrinsic::memset:
239 case Intrinsic::powi:
241 case Intrinsic::log2:
242 case Intrinsic::log10:
244 case Intrinsic::exp2:
249 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
250 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
251 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
252 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
253 case Intrinsic::rint: Opcode = ISD::FRINT; break;
254 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
258 // PowerPC does not use [US]DIVREM or other library calls for
259 // operations on regular types which are not otherwise library calls
260 // (i.e. soft float or atomics). If adapting for targets that do,
261 // additional care is required here.
264 if (!F->hasLocalLinkage() && F->hasName() && LibInfo &&
265 LibInfo->getLibFunc(F->getName(), Func) &&
266 LibInfo->hasOptimizedCodeGen(Func)) {
267 // Non-read-only functions are never treated as intrinsics.
268 if (!CI->onlyReadsMemory())
271 // Conversion happens only for FP calls.
272 if (!CI->getArgOperand(0)->getType()->isFloatingPointTy())
276 default: return true;
277 case LibFunc::copysign:
278 case LibFunc::copysignf:
279 case LibFunc::copysignl:
280 continue; // ISD::FCOPYSIGN is never a library call.
284 continue; // ISD::FABS is never a library call.
288 Opcode = ISD::FSQRT; break;
290 case LibFunc::floorf:
291 case LibFunc::floorl:
292 Opcode = ISD::FFLOOR; break;
293 case LibFunc::nearbyint:
294 case LibFunc::nearbyintf:
295 case LibFunc::nearbyintl:
296 Opcode = ISD::FNEARBYINT; break;
300 Opcode = ISD::FCEIL; break;
304 Opcode = ISD::FRINT; break;
306 case LibFunc::truncf:
307 case LibFunc::truncl:
308 Opcode = ISD::FTRUNC; break;
312 TLI->getSimpleValueType(CI->getArgOperand(0)->getType(), true);
313 if (VTy == MVT::Other)
316 if (TLI->isOperationLegalOrCustom(Opcode, VTy))
318 else if (VTy.isVector() &&
319 TLI->isOperationLegalOrCustom(Opcode, VTy.getScalarType()))
327 } else if (isa<BinaryOperator>(J) &&
328 J->getType()->getScalarType()->isPPC_FP128Ty()) {
329 // Most operations on ppc_f128 values become calls.
331 } else if (isa<UIToFPInst>(J) || isa<SIToFPInst>(J) ||
332 isa<FPToUIInst>(J) || isa<FPToSIInst>(J)) {
333 CastInst *CI = cast<CastInst>(J);
334 if (CI->getSrcTy()->getScalarType()->isPPC_FP128Ty() ||
335 CI->getDestTy()->getScalarType()->isPPC_FP128Ty() ||
337 (CI->getSrcTy()->getScalarType()->isIntegerTy(64) ||
338 CI->getDestTy()->getScalarType()->isIntegerTy(64))
341 } else if (TT.isArch32Bit() &&
342 J->getType()->getScalarType()->isIntegerTy(64) &&
343 (J->getOpcode() == Instruction::UDiv ||
344 J->getOpcode() == Instruction::SDiv ||
345 J->getOpcode() == Instruction::URem ||
346 J->getOpcode() == Instruction::SRem)) {
348 } else if (isa<IndirectBrInst>(J) || isa<InvokeInst>(J)) {
349 // On PowerPC, indirect jumps use the counter register.
351 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(J)) {
354 const TargetLowering *TLI = TM->getTargetLowering();
356 if (TLI->supportJumpTables() &&
357 SI->getNumCases()+1 >= (unsigned) TLI->getMinimumJumpTableEntries())
365 bool PPCCTRLoops::convertToCTRLoop(Loop *L) {
366 bool MadeChange = false;
368 Triple TT = Triple(L->getHeader()->getParent()->getParent()->
370 if (!TT.isArch32Bit() && !TT.isArch64Bit())
371 return MadeChange; // Unknown arch. type.
373 // Process nested loops first.
374 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
375 MadeChange |= convertToCTRLoop(*I);
378 // If a nested loop has been converted, then we can't convert this loop.
383 // Stop trying after reaching the limit (if any).
384 int Limit = CTRLoopLimit;
386 if (Counter >= CTRLoopLimit)
392 // We don't want to spill/restore the counter register, and so we don't
393 // want to use the counter register if the loop contains calls.
394 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
396 if (mightUseCTR(TT, *I))
399 SmallVector<BasicBlock*, 4> ExitingBlocks;
400 L->getExitingBlocks(ExitingBlocks);
402 BasicBlock *CountedExitBlock = 0;
403 const SCEV *ExitCount = 0;
404 BranchInst *CountedExitBranch = 0;
405 for (SmallVector<BasicBlock*, 4>::iterator I = ExitingBlocks.begin(),
406 IE = ExitingBlocks.end(); I != IE; ++I) {
407 const SCEV *EC = SE->getExitCount(L, *I);
408 DEBUG(dbgs() << "Exit Count for " << *L << " from block " <<
409 (*I)->getName() << ": " << *EC << "\n");
410 if (isa<SCEVCouldNotCompute>(EC))
412 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
413 if (ConstEC->getValue()->isZero())
415 } else if (!SE->isLoopInvariant(EC, L))
418 // We now have a loop-invariant count of loop iterations (which is not the
419 // constant zero) for which we know that this loop will not exit via this
422 // We need to make sure that this block will run on every loop iteration.
423 // For this to be true, we must dominate all blocks with backedges. Such
424 // blocks are in-loop predecessors to the header block.
425 bool NotAlways = false;
426 for (pred_iterator PI = pred_begin(L->getHeader()),
427 PIE = pred_end(L->getHeader()); PI != PIE; ++PI) {
428 if (!L->contains(*PI))
431 if (!DT->dominates(*I, *PI)) {
440 // Make sure this blocks ends with a conditional branch.
441 Instruction *TI = (*I)->getTerminator();
445 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
446 if (!BI->isConditional())
449 CountedExitBranch = BI;
453 // Note that this block may not be the loop latch block, even if the loop
454 // has a latch block.
455 CountedExitBlock = *I;
460 if (!CountedExitBlock)
463 BasicBlock *Preheader = L->getLoopPreheader();
465 // If we don't have a preheader, then insert one. If we already have a
466 // preheader, then we can use it (except if the preheader contains a use of
467 // the CTR register because some such uses might be reordered by the
468 // selection DAG after the mtctr instruction).
469 if (!Preheader || mightUseCTR(TT, Preheader))
470 Preheader = InsertPreheaderForLoop(L, this);
474 DEBUG(dbgs() << "Preheader for exit count: " << Preheader->getName() << "\n");
476 // Insert the count into the preheader and replace the condition used by the
480 SCEVExpander SCEVE(*SE, "loopcnt");
481 LLVMContext &C = SE->getContext();
482 Type *CountType = TT.isArch64Bit() ? Type::getInt64Ty(C) :
484 if (!ExitCount->getType()->isPointerTy() &&
485 ExitCount->getType() != CountType)
486 ExitCount = SE->getZeroExtendExpr(ExitCount, CountType);
487 ExitCount = SE->getAddExpr(ExitCount,
488 SE->getConstant(CountType, 1));
489 Value *ECValue = SCEVE.expandCodeFor(ExitCount, CountType,
490 Preheader->getTerminator());
492 IRBuilder<> CountBuilder(Preheader->getTerminator());
493 Module *M = Preheader->getParent()->getParent();
494 Value *MTCTRFunc = Intrinsic::getDeclaration(M, Intrinsic::ppc_mtctr,
496 CountBuilder.CreateCall(MTCTRFunc, ECValue);
498 IRBuilder<> CondBuilder(CountedExitBranch);
500 Intrinsic::getDeclaration(M, Intrinsic::ppc_is_decremented_ctr_nonzero);
501 Value *NewCond = CondBuilder.CreateCall(DecFunc);
502 Value *OldCond = CountedExitBranch->getCondition();
503 CountedExitBranch->setCondition(NewCond);
505 // The false branch must exit the loop.
506 if (!L->contains(CountedExitBranch->getSuccessor(0)))
507 CountedExitBranch->swapSuccessors();
509 // The old condition may be dead now, and may have even created a dead PHI
510 // (the original induction variable).
511 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
512 DeleteDeadPHIs(CountedExitBlock);
519 static bool clobbersCTR(const MachineInstr *MI) {
520 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
521 const MachineOperand &MO = MI->getOperand(i);
523 if (MO.isDef() && (MO.getReg() == PPC::CTR || MO.getReg() == PPC::CTR8))
525 } else if (MO.isRegMask()) {
526 if (MO.clobbersPhysReg(PPC::CTR) || MO.clobbersPhysReg(PPC::CTR8))
534 static bool verifyCTRBranch(MachineBasicBlock *MBB,
535 MachineBasicBlock::iterator I) {
536 MachineBasicBlock::iterator BI = I;
537 SmallSet<MachineBasicBlock *, 16> Visited;
538 SmallVector<MachineBasicBlock *, 8> Preds;
541 if (I == MBB->begin()) {
553 for (MachineBasicBlock::iterator IE = MBB->begin();; --I) {
554 unsigned Opc = I->getOpcode();
555 if (Opc == PPC::MTCTRloop || Opc == PPC::MTCTR8loop) {
560 if (I != BI && clobbersCTR(I)) {
561 DEBUG(dbgs() << "BB#" << MBB->getNumber() << " (" <<
562 MBB->getFullName() << ") instruction " << *I <<
563 " clobbers CTR, invalidating " << "BB#" <<
564 BI->getParent()->getNumber() << " (" <<
565 BI->getParent()->getFullName() << ") instruction " <<
574 if (!CheckPreds && Preds.empty())
579 if (MachineFunction::iterator(MBB) == MBB->getParent()->begin()) {
580 DEBUG(dbgs() << "Unable to find a MTCTR instruction for BB#" <<
581 BI->getParent()->getNumber() << " (" <<
582 BI->getParent()->getFullName() << ") instruction " <<
587 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
588 PIE = MBB->pred_end(); PI != PIE; ++PI)
589 Preds.push_back(*PI);
593 MBB = Preds.pop_back_val();
594 if (!Visited.count(MBB)) {
595 I = MBB->getLastNonDebugInstr();
598 } while (!Preds.empty());
603 bool PPCCTRLoopsVerify::runOnMachineFunction(MachineFunction &MF) {
604 MDT = &getAnalysis<MachineDominatorTree>();
606 // Verify that all bdnz/bdz instructions are dominated by a loop mtctr before
607 // any other instructions that might clobber the ctr register.
608 for (MachineFunction::iterator I = MF.begin(), IE = MF.end();
610 MachineBasicBlock *MBB = I;
611 if (!MDT->isReachableFromEntry(MBB))
614 for (MachineBasicBlock::iterator MII = MBB->getFirstTerminator(),
615 MIIE = MBB->end(); MII != MIIE; ++MII) {
616 unsigned Opc = MII->getOpcode();
617 if (Opc == PPC::BDNZ8 || Opc == PPC::BDNZ ||
618 Opc == PPC::BDZ8 || Opc == PPC::BDZ)
619 if (!verifyCTRBranch(MBB, MII))
620 llvm_unreachable("Invalid PPC CTR loop!");