1 //===-- PeepholeOptimizer.cpp - Peephole Optimizations --------------------===//
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 // Perform peephole optimizations on the machine code:
12 // - Optimize Extensions
14 // Optimization of sign / zero extension instructions. It may be extended to
15 // handle other instructions with similar properties.
17 // On some targets, some instructions, e.g. X86 sign / zero extension, may
18 // leave the source value in the lower part of the result. This optimization
19 // will replace some uses of the pre-extension value with uses of the
20 // sub-register of the results.
22 // - Optimize Comparisons
24 // Optimization of comparison instructions. For instance, in this code:
30 // If the "sub" instruction all ready sets (or could be modified to set) the
31 // same flag that the "cmp" instruction sets and that "bz" uses, then we can
32 // eliminate the "cmp" instruction.
34 // - Optimize Bitcast pairs:
43 //===----------------------------------------------------------------------===//
45 #define DEBUG_TYPE "peephole-opt"
46 #include "llvm/CodeGen/Passes.h"
47 #include "llvm/CodeGen/MachineDominators.h"
48 #include "llvm/CodeGen/MachineInstrBuilder.h"
49 #include "llvm/CodeGen/MachineRegisterInfo.h"
50 #include "llvm/Target/TargetInstrInfo.h"
51 #include "llvm/Target/TargetRegisterInfo.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallPtrSet.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/ADT/Statistic.h"
59 // Optimize Extensions
61 Aggressive("aggressive-ext-opt", cl::Hidden,
62 cl::desc("Aggressive extension optimization"));
65 DisablePeephole("disable-peephole", cl::Hidden, cl::init(false),
66 cl::desc("Disable the peephole optimizer"));
68 STATISTIC(NumReuse, "Number of extension results reused");
69 STATISTIC(NumBitcasts, "Number of bitcasts eliminated");
70 STATISTIC(NumCmps, "Number of compares eliminated");
71 STATISTIC(NumImmFold, "Number of move immediate foled");
74 class PeepholeOptimizer : public MachineFunctionPass {
75 const TargetMachine *TM;
76 const TargetInstrInfo *TII;
77 MachineRegisterInfo *MRI;
78 MachineDominatorTree *DT; // Machine dominator tree
81 static char ID; // Pass identification
82 PeepholeOptimizer() : MachineFunctionPass(ID) {
83 initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry());
86 virtual bool runOnMachineFunction(MachineFunction &MF);
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
90 MachineFunctionPass::getAnalysisUsage(AU);
92 AU.addRequired<MachineDominatorTree>();
93 AU.addPreserved<MachineDominatorTree>();
98 bool OptimizeBitcastInstr(MachineInstr *MI, MachineBasicBlock *MBB);
99 bool OptimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
100 bool OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
101 SmallPtrSet<MachineInstr*, 8> &LocalMIs);
102 bool isMoveImmediate(MachineInstr *MI,
103 SmallSet<unsigned, 4> &ImmDefRegs,
104 DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
105 bool FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
106 SmallSet<unsigned, 4> &ImmDefRegs,
107 DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
111 char PeepholeOptimizer::ID = 0;
112 char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID;
113 INITIALIZE_PASS_BEGIN(PeepholeOptimizer, "peephole-opts",
114 "Peephole Optimizations", false, false)
115 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
116 INITIALIZE_PASS_END(PeepholeOptimizer, "peephole-opts",
117 "Peephole Optimizations", false, false)
119 /// OptimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
120 /// a single register and writes a single register and it does not modify the
121 /// source, and if the source value is preserved as a sub-register of the
122 /// result, then replace all reachable uses of the source with the subreg of the
125 /// Do not generate an EXTRACT that is used only in a debug use, as this changes
126 /// the code. Since this code does not currently share EXTRACTs, just ignore all
128 bool PeepholeOptimizer::
129 OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
130 SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
131 unsigned SrcReg, DstReg, SubIdx;
132 if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
135 if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
136 TargetRegisterInfo::isPhysicalRegister(SrcReg))
139 MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(SrcReg);
140 if (++UI == MRI->use_nodbg_end())
144 // The source has other uses. See if we can replace the other uses with use of
145 // the result of the extension.
146 SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
147 UI = MRI->use_nodbg_begin(DstReg);
148 for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
150 ReachedBBs.insert(UI->getParent());
152 // Uses that are in the same BB of uses of the result of the instruction.
153 SmallVector<MachineOperand*, 8> Uses;
155 // Uses that the result of the instruction can reach.
156 SmallVector<MachineOperand*, 8> ExtendedUses;
158 bool ExtendLife = true;
159 UI = MRI->use_nodbg_begin(SrcReg);
160 for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
162 MachineOperand &UseMO = UI.getOperand();
163 MachineInstr *UseMI = &*UI;
167 if (UseMI->isPHI()) {
172 // It's an error to translate this:
174 // %reg1025 = <sext> %reg1024
176 // %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
180 // %reg1025 = <sext> %reg1024
182 // %reg1027 = COPY %reg1025:4
183 // %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
185 // The problem here is that SUBREG_TO_REG is there to assert that an
186 // implicit zext occurs. It doesn't insert a zext instruction. If we allow
187 // the COPY here, it will give us the value after the <sext>, not the
188 // original value of %reg1024 before <sext>.
189 if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
192 MachineBasicBlock *UseMBB = UseMI->getParent();
194 // Local uses that come after the extension.
195 if (!LocalMIs.count(UseMI))
196 Uses.push_back(&UseMO);
197 } else if (ReachedBBs.count(UseMBB)) {
198 // Non-local uses where the result of the extension is used. Always
199 // replace these unless it's a PHI.
200 Uses.push_back(&UseMO);
201 } else if (Aggressive && DT->dominates(MBB, UseMBB)) {
202 // We may want to extend the live range of the extension result in order
203 // to replace these uses.
204 ExtendedUses.push_back(&UseMO);
206 // Both will be live out of the def MBB anyway. Don't extend live range of
207 // the extension result.
213 if (ExtendLife && !ExtendedUses.empty())
214 // Extend the liveness of the extension result.
215 std::copy(ExtendedUses.begin(), ExtendedUses.end(),
216 std::back_inserter(Uses));
218 // Now replace all uses.
219 bool Changed = false;
221 SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
223 // Look for PHI uses of the extended result, we don't want to extend the
224 // liveness of a PHI input. It breaks all kinds of assumptions down
225 // stream. A PHI use is expected to be the kill of its source values.
226 UI = MRI->use_nodbg_begin(DstReg);
227 for (MachineRegisterInfo::use_nodbg_iterator
228 UE = MRI->use_nodbg_end(); UI != UE; ++UI)
230 PHIBBs.insert(UI->getParent());
232 const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
233 for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
234 MachineOperand *UseMO = Uses[i];
235 MachineInstr *UseMI = UseMO->getParent();
236 MachineBasicBlock *UseMBB = UseMI->getParent();
237 if (PHIBBs.count(UseMBB))
240 unsigned NewVR = MRI->createVirtualRegister(RC);
241 BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
242 TII->get(TargetOpcode::COPY), NewVR)
243 .addReg(DstReg, 0, SubIdx);
245 UseMO->setReg(NewVR);
254 /// OptimizeBitcastInstr - If the instruction is a bitcast instruction A that
255 /// cannot be optimized away during isel (e.g. ARM::VMOVSR, which bitcast
256 /// a value cross register classes), and the source is defined by another
257 /// bitcast instruction B. And if the register class of source of B matches
258 /// the register class of instruction A, then it is legal to replace all uses
259 /// of the def of A with source of B. e.g.
260 /// %vreg0<def> = VMOVSR %vreg1
261 /// %vreg3<def> = VMOVRS %vreg0
262 /// Replace all uses of vreg3 with vreg1.
264 bool PeepholeOptimizer::OptimizeBitcastInstr(MachineInstr *MI,
265 MachineBasicBlock *MBB) {
266 unsigned NumDefs = MI->getDesc().getNumDefs();
267 unsigned NumSrcs = MI->getDesc().getNumOperands() - NumDefs;
273 for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
274 const MachineOperand &MO = MI->getOperand(i);
277 unsigned Reg = MO.getReg();
289 assert(Def && Src && "Malformed bitcast instruction!");
291 MachineInstr *DefMI = MRI->getVRegDef(Src);
292 if (!DefMI || !DefMI->isBitcast())
296 NumDefs = DefMI->getDesc().getNumDefs();
297 NumSrcs = DefMI->getDesc().getNumOperands() - NumDefs;
300 for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
301 const MachineOperand &MO = DefMI->getOperand(i);
302 if (!MO.isReg() || MO.isDef())
304 unsigned Reg = MO.getReg();
316 if (MRI->getRegClass(SrcSrc) != MRI->getRegClass(Def))
319 MRI->replaceRegWith(Def, SrcSrc);
320 MRI->clearKillFlags(SrcSrc);
321 MI->eraseFromParent();
326 /// OptimizeCmpInstr - If the instruction is a compare and the previous
327 /// instruction it's comparing against all ready sets (or could be modified to
328 /// set) the same flag as the compare, then we can remove the comparison and use
329 /// the flag from the previous instruction.
330 bool PeepholeOptimizer::OptimizeCmpInstr(MachineInstr *MI,
331 MachineBasicBlock *MBB) {
332 // If this instruction is a comparison against zero and isn't comparing a
333 // physical register, we can try to optimize it.
335 int CmpMask, CmpValue;
336 if (!TII->AnalyzeCompare(MI, SrcReg, CmpMask, CmpValue) ||
337 TargetRegisterInfo::isPhysicalRegister(SrcReg))
340 // Attempt to optimize the comparison instruction.
341 if (TII->OptimizeCompareInstr(MI, SrcReg, CmpMask, CmpValue, MRI)) {
349 bool PeepholeOptimizer::isMoveImmediate(MachineInstr *MI,
350 SmallSet<unsigned, 4> &ImmDefRegs,
351 DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
352 const MCInstrDesc &MCID = MI->getDesc();
353 if (!MI->isMoveImmediate())
355 if (MCID.getNumDefs() != 1)
357 unsigned Reg = MI->getOperand(0).getReg();
358 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
359 ImmDefMIs.insert(std::make_pair(Reg, MI));
360 ImmDefRegs.insert(Reg);
367 /// FoldImmediate - Try folding register operands that are defined by move
368 /// immediate instructions, i.e. a trivial constant folding optimization, if
369 /// and only if the def and use are in the same BB.
370 bool PeepholeOptimizer::FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
371 SmallSet<unsigned, 4> &ImmDefRegs,
372 DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
373 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
374 MachineOperand &MO = MI->getOperand(i);
375 if (!MO.isReg() || MO.isDef())
377 unsigned Reg = MO.getReg();
378 if (!TargetRegisterInfo::isVirtualRegister(Reg))
380 if (ImmDefRegs.count(Reg) == 0)
382 DenseMap<unsigned, MachineInstr*>::iterator II = ImmDefMIs.find(Reg);
383 assert(II != ImmDefMIs.end());
384 if (TII->FoldImmediate(MI, II->second, Reg, MRI)) {
392 bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
396 TM = &MF.getTarget();
397 TII = TM->getInstrInfo();
398 MRI = &MF.getRegInfo();
399 DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
401 bool Changed = false;
403 SmallPtrSet<MachineInstr*, 8> LocalMIs;
404 SmallSet<unsigned, 4> ImmDefRegs;
405 DenseMap<unsigned, MachineInstr*> ImmDefMIs;
406 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
407 MachineBasicBlock *MBB = &*I;
409 bool SeenMoveImm = false;
415 MachineBasicBlock::iterator PMII;
416 for (MachineBasicBlock::iterator
417 MII = I->begin(), MIE = I->end(); MII != MIE; ) {
418 MachineInstr *MI = &*MII;
421 if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
422 MI->isKill() || MI->isInlineAsm() || MI->isDebugValue() ||
423 MI->hasUnmodeledSideEffects()) {
428 if (MI->isBitcast()) {
429 if (OptimizeBitcastInstr(MI, MBB)) {
433 MII = First ? I->begin() : llvm::next(PMII);
436 } else if (MI->isCompare()) {
437 if (OptimizeCmpInstr(MI, MBB)) {
441 MII = First ? I->begin() : llvm::next(PMII);
446 if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
449 Changed |= OptimizeExtInstr(MI, MBB, LocalMIs);
451 Changed |= FoldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);