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 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "peephole-opt"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/CodeGen/MachineDominators.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/Target/TargetInstrInfo.h"
42 #include "llvm/Target/TargetRegisterInfo.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/ADT/SmallPtrSet.h"
45 #include "llvm/ADT/Statistic.h"
48 // Optimize Extensions
50 Aggressive("aggressive-ext-opt", cl::Hidden,
51 cl::desc("Aggressive extension optimization"));
54 DisablePeephole("disable-peephole", cl::Hidden, cl::init(false),
55 cl::desc("Disable the peephole optimizer"));
57 STATISTIC(NumReuse, "Number of extension results reused");
58 STATISTIC(NumEliminated, "Number of compares eliminated");
61 class PeepholeOptimizer : public MachineFunctionPass {
62 const TargetMachine *TM;
63 const TargetInstrInfo *TII;
64 MachineRegisterInfo *MRI;
65 MachineDominatorTree *DT; // Machine dominator tree
68 static char ID; // Pass identification
69 PeepholeOptimizer() : MachineFunctionPass(ID) {
70 initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry());
73 virtual bool runOnMachineFunction(MachineFunction &MF);
75 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 MachineFunctionPass::getAnalysisUsage(AU);
79 AU.addRequired<MachineDominatorTree>();
80 AU.addPreserved<MachineDominatorTree>();
85 bool OptimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
86 bool OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
87 SmallPtrSet<MachineInstr*, 8> &LocalMIs);
91 char PeepholeOptimizer::ID = 0;
92 INITIALIZE_PASS_BEGIN(PeepholeOptimizer, "peephole-opts",
93 "Peephole Optimizations", false, false)
94 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
95 INITIALIZE_PASS_END(PeepholeOptimizer, "peephole-opts",
96 "Peephole Optimizations", false, false)
98 FunctionPass *llvm::createPeepholeOptimizerPass() {
99 return new PeepholeOptimizer();
102 /// OptimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
103 /// a single register and writes a single register and it does not modify the
104 /// source, and if the source value is preserved as a sub-register of the
105 /// result, then replace all reachable uses of the source with the subreg of the
108 /// Do not generate an EXTRACT that is used only in a debug use, as this changes
109 /// the code. Since this code does not currently share EXTRACTs, just ignore all
111 bool PeepholeOptimizer::
112 OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
113 SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
114 unsigned SrcReg, DstReg, SubIdx;
115 if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
118 if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
119 TargetRegisterInfo::isPhysicalRegister(SrcReg))
122 MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(SrcReg);
123 if (++UI == MRI->use_nodbg_end())
127 // The source has other uses. See if we can replace the other uses with use of
128 // the result of the extension.
129 SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
130 UI = MRI->use_nodbg_begin(DstReg);
131 for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
133 ReachedBBs.insert(UI->getParent());
135 // Uses that are in the same BB of uses of the result of the instruction.
136 SmallVector<MachineOperand*, 8> Uses;
138 // Uses that the result of the instruction can reach.
139 SmallVector<MachineOperand*, 8> ExtendedUses;
141 bool ExtendLife = true;
142 UI = MRI->use_nodbg_begin(SrcReg);
143 for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
145 MachineOperand &UseMO = UI.getOperand();
146 MachineInstr *UseMI = &*UI;
150 if (UseMI->isPHI()) {
155 // It's an error to translate this:
157 // %reg1025 = <sext> %reg1024
159 // %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
163 // %reg1025 = <sext> %reg1024
165 // %reg1027 = COPY %reg1025:4
166 // %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
168 // The problem here is that SUBREG_TO_REG is there to assert that an
169 // implicit zext occurs. It doesn't insert a zext instruction. If we allow
170 // the COPY here, it will give us the value after the <sext>, not the
171 // original value of %reg1024 before <sext>.
172 if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
175 MachineBasicBlock *UseMBB = UseMI->getParent();
177 // Local uses that come after the extension.
178 if (!LocalMIs.count(UseMI))
179 Uses.push_back(&UseMO);
180 } else if (ReachedBBs.count(UseMBB)) {
181 // Non-local uses where the result of the extension is used. Always
182 // replace these unless it's a PHI.
183 Uses.push_back(&UseMO);
184 } else if (Aggressive && DT->dominates(MBB, UseMBB)) {
185 // We may want to extend the live range of the extension result in order
186 // to replace these uses.
187 ExtendedUses.push_back(&UseMO);
189 // Both will be live out of the def MBB anyway. Don't extend live range of
190 // the extension result.
196 if (ExtendLife && !ExtendedUses.empty())
197 // Extend the liveness of the extension result.
198 std::copy(ExtendedUses.begin(), ExtendedUses.end(),
199 std::back_inserter(Uses));
201 // Now replace all uses.
202 bool Changed = false;
204 SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
206 // Look for PHI uses of the extended result, we don't want to extend the
207 // liveness of a PHI input. It breaks all kinds of assumptions down
208 // stream. A PHI use is expected to be the kill of its source values.
209 UI = MRI->use_nodbg_begin(DstReg);
210 for (MachineRegisterInfo::use_nodbg_iterator
211 UE = MRI->use_nodbg_end(); UI != UE; ++UI)
213 PHIBBs.insert(UI->getParent());
215 const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
216 for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
217 MachineOperand *UseMO = Uses[i];
218 MachineInstr *UseMI = UseMO->getParent();
219 MachineBasicBlock *UseMBB = UseMI->getParent();
220 if (PHIBBs.count(UseMBB))
223 unsigned NewVR = MRI->createVirtualRegister(RC);
224 BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
225 TII->get(TargetOpcode::COPY), NewVR)
226 .addReg(DstReg, 0, SubIdx);
228 UseMO->setReg(NewVR);
237 /// OptimizeCmpInstr - If the instruction is a compare and the previous
238 /// instruction it's comparing against all ready sets (or could be modified to
239 /// set) the same flag as the compare, then we can remove the comparison and use
240 /// the flag from the previous instruction.
241 bool PeepholeOptimizer::OptimizeCmpInstr(MachineInstr *MI,
242 MachineBasicBlock *MBB){
243 // If this instruction is a comparison against zero and isn't comparing a
244 // physical register, we can try to optimize it.
246 int CmpMask, CmpValue;
247 if (!TII->AnalyzeCompare(MI, SrcReg, CmpMask, CmpValue) ||
248 TargetRegisterInfo::isPhysicalRegister(SrcReg))
251 // Attempt to optimize the comparison instruction.
252 if (TII->OptimizeCompareInstr(MI, SrcReg, CmpMask, CmpValue, MRI)) {
260 bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
264 TM = &MF.getTarget();
265 TII = TM->getInstrInfo();
266 MRI = &MF.getRegInfo();
267 DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
269 bool Changed = false;
271 SmallPtrSet<MachineInstr*, 8> LocalMIs;
272 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
273 MachineBasicBlock *MBB = &*I;
276 for (MachineBasicBlock::iterator
277 MII = I->begin(), MIE = I->end(); MII != MIE; ) {
278 MachineInstr *MI = &*MII++;
281 if (MI->getDesc().hasUnmodeledSideEffects())
284 if (MI->getDesc().isCompare()) {
285 Changed |= OptimizeCmpInstr(MI, MBB);
287 Changed |= OptimizeExtInstr(MI, MBB, LocalMIs);