1 //===- X86InstrInfo.cpp - X86 Instruction Information -----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the X86 implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "X86InstrInfo.h"
16 #include "X86GenInstrInfo.inc"
17 #include "X86InstrBuilder.h"
18 #include "X86Subtarget.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/LiveVariables.h"
24 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
25 : TargetInstrInfo(X86Insts, sizeof(X86Insts)/sizeof(X86Insts[0])),
26 TM(tm), RI(tm, *this) {
29 /// getDWARF_LABELOpcode - Return the opcode of the target's DWARF_LABEL
30 /// instruction if it has one. This is used by codegen passes that update
31 /// DWARF line number info as they modify the code.
32 unsigned X86InstrInfo::getDWARF_LABELOpcode() const {
33 return X86::DWARF_LABEL;
37 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
39 unsigned& destReg) const {
40 MachineOpCode oc = MI.getOpcode();
41 if (oc == X86::MOV8rr || oc == X86::MOV16rr ||
42 oc == X86::MOV32rr || oc == X86::MOV64rr ||
43 oc == X86::MOV16to16_ || oc == X86::MOV32to32_ ||
44 oc == X86::FpMOV || oc == X86::MOVSSrr || oc == X86::MOVSDrr ||
45 oc == X86::FsMOVAPSrr || oc == X86::FsMOVAPDrr ||
46 oc == X86::MOVAPSrr || oc == X86::MOVAPDrr ||
47 oc == X86::MOVSS2PSrr || oc == X86::MOVSD2PDrr ||
48 oc == X86::MOVPS2SSrr || oc == X86::MOVPD2SDrr) {
49 assert(MI.getNumOperands() == 2 &&
50 MI.getOperand(0).isRegister() &&
51 MI.getOperand(1).isRegister() &&
52 "invalid register-register move instruction");
53 sourceReg = MI.getOperand(1).getReg();
54 destReg = MI.getOperand(0).getReg();
60 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
61 int &FrameIndex) const {
62 switch (MI->getOpcode()) {
75 if (MI->getOperand(1).isFrameIndex() && MI->getOperand(2).isImmediate() &&
76 MI->getOperand(3).isRegister() && MI->getOperand(4).isImmediate() &&
77 MI->getOperand(2).getImmedValue() == 1 &&
78 MI->getOperand(3).getReg() == 0 &&
79 MI->getOperand(4).getImmedValue() == 0) {
80 FrameIndex = MI->getOperand(1).getFrameIndex();
81 return MI->getOperand(0).getReg();
88 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
89 int &FrameIndex) const {
90 switch (MI->getOpcode()) {
103 if (MI->getOperand(0).isFrameIndex() && MI->getOperand(1).isImmediate() &&
104 MI->getOperand(2).isRegister() && MI->getOperand(3).isImmediate() &&
105 MI->getOperand(1).getImmedValue() == 1 &&
106 MI->getOperand(2).getReg() == 0 &&
107 MI->getOperand(3).getImmedValue() == 0) {
108 FrameIndex = MI->getOperand(0).getFrameIndex();
109 return MI->getOperand(4).getReg();
117 /// convertToThreeAddress - This method must be implemented by targets that
118 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
119 /// may be able to convert a two-address instruction into a true
120 /// three-address instruction on demand. This allows the X86 target (for
121 /// example) to convert ADD and SHL instructions into LEA instructions if they
122 /// would require register copies due to two-addressness.
124 /// This method returns a null pointer if the transformation cannot be
125 /// performed, otherwise it returns the new instruction.
128 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
129 MachineBasicBlock::iterator &MBBI,
130 LiveVariables &LV) const {
131 MachineInstr *MI = MBBI;
132 // All instructions input are two-addr instructions. Get the known operands.
133 unsigned Dest = MI->getOperand(0).getReg();
134 unsigned Src = MI->getOperand(1).getReg();
136 MachineInstr *NewMI = NULL;
137 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
138 // we have subtarget support, enable the 16-bit LEA generation here.
139 bool DisableLEA16 = true;
141 switch (MI->getOpcode()) {
143 case X86::SHUFPSrri: {
144 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
145 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
146 unsigned A = MI->getOperand(0).getReg();
147 unsigned B = MI->getOperand(1).getReg();
148 unsigned C = MI->getOperand(2).getReg();
149 unsigned M = MI->getOperand(3).getImmedValue();
150 if (!Subtarget->hasSSE2() || B != C) return 0;
151 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
156 // FIXME: None of these instructions are promotable to LEAs without
157 // additional information. In particular, LEA doesn't set the flags that
158 // add and inc do. :(
161 switch (MI->getOpcode()) {
164 assert(MI->getNumOperands() == 2 && "Unknown inc instruction!");
165 NewMI = addRegOffset(BuildMI(get(X86::LEA32r), Dest), Src, 1);
169 if (DisableLEA16) return 0;
170 assert(MI->getNumOperands() == 2 && "Unknown inc instruction!");
171 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
175 assert(MI->getNumOperands() == 2 && "Unknown dec instruction!");
176 NewMI = addRegOffset(BuildMI(get(X86::LEA32r), Dest), Src, -1);
180 if (DisableLEA16) return 0;
181 assert(MI->getNumOperands() == 2 && "Unknown dec instruction!");
182 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
185 assert(MI->getNumOperands() == 3 && "Unknown add instruction!");
186 NewMI = addRegReg(BuildMI(get(X86::LEA32r), Dest), Src,
187 MI->getOperand(2).getReg());
190 if (DisableLEA16) return 0;
191 assert(MI->getNumOperands() == 3 && "Unknown add instruction!");
192 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
193 MI->getOperand(2).getReg());
197 assert(MI->getNumOperands() == 3 && "Unknown add instruction!");
198 if (MI->getOperand(2).isImmediate())
199 NewMI = addRegOffset(BuildMI(get(X86::LEA32r), Dest), Src,
200 MI->getOperand(2).getImmedValue());
204 if (DisableLEA16) return 0;
205 assert(MI->getNumOperands() == 3 && "Unknown add instruction!");
206 if (MI->getOperand(2).isImmediate())
207 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
208 MI->getOperand(2).getImmedValue());
211 if (DisableLEA16) return 0;
213 assert(MI->getNumOperands() == 3 && MI->getOperand(2).isImmediate() &&
214 "Unknown shl instruction!");
215 unsigned ShAmt = MI->getOperand(2).getImmedValue();
216 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
218 AM.Scale = 1 << ShAmt;
220 unsigned Opc = MI->getOpcode() == X86::SHL32ri ? X86::LEA32r :X86::LEA16r;
221 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
228 NewMI->copyKillDeadInfo(MI);
229 LV.instructionChanged(MI, NewMI); // Update live variables
230 MFI->insert(MBBI, NewMI); // Insert the new inst
235 /// commuteInstruction - We have a few instructions that must be hacked on to
238 MachineInstr *X86InstrInfo::commuteInstruction(MachineInstr *MI) const {
239 // FIXME: Can commute cmoves by changing the condition!
240 switch (MI->getOpcode()) {
241 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
242 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
243 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
244 case X86::SHLD32rri8:{// A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
247 switch (MI->getOpcode()) {
248 default: assert(0 && "Unreachable!");
249 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
250 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
251 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
252 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
254 unsigned Amt = MI->getOperand(3).getImmedValue();
255 unsigned A = MI->getOperand(0).getReg();
256 unsigned B = MI->getOperand(1).getReg();
257 unsigned C = MI->getOperand(2).getReg();
258 bool BisKill = MI->getOperand(1).isKill();
259 bool CisKill = MI->getOperand(2).isKill();
260 return BuildMI(get(Opc), A).addReg(C, false, false, CisKill)
261 .addReg(B, false, false, BisKill).addImm(Size-Amt);
264 return TargetInstrInfo::commuteInstruction(MI);
268 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
270 default: return X86::COND_INVALID;
271 case X86::JE: return X86::COND_E;
272 case X86::JNE: return X86::COND_NE;
273 case X86::JL: return X86::COND_L;
274 case X86::JLE: return X86::COND_LE;
275 case X86::JG: return X86::COND_G;
276 case X86::JGE: return X86::COND_GE;
277 case X86::JB: return X86::COND_B;
278 case X86::JBE: return X86::COND_BE;
279 case X86::JA: return X86::COND_A;
280 case X86::JAE: return X86::COND_AE;
281 case X86::JS: return X86::COND_S;
282 case X86::JNS: return X86::COND_NS;
283 case X86::JP: return X86::COND_P;
284 case X86::JNP: return X86::COND_NP;
285 case X86::JO: return X86::COND_O;
286 case X86::JNO: return X86::COND_NO;
290 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
292 default: assert(0 && "Illegal condition code!");
293 case X86::COND_E: return X86::JE;
294 case X86::COND_NE: return X86::JNE;
295 case X86::COND_L: return X86::JL;
296 case X86::COND_LE: return X86::JLE;
297 case X86::COND_G: return X86::JG;
298 case X86::COND_GE: return X86::JGE;
299 case X86::COND_B: return X86::JB;
300 case X86::COND_BE: return X86::JBE;
301 case X86::COND_A: return X86::JA;
302 case X86::COND_AE: return X86::JAE;
303 case X86::COND_S: return X86::JS;
304 case X86::COND_NS: return X86::JNS;
305 case X86::COND_P: return X86::JP;
306 case X86::COND_NP: return X86::JNP;
307 case X86::COND_O: return X86::JO;
308 case X86::COND_NO: return X86::JNO;
312 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
313 /// e.g. turning COND_E to COND_NE.
314 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
316 default: assert(0 && "Illegal condition code!");
317 case X86::COND_E: return X86::COND_NE;
318 case X86::COND_NE: return X86::COND_E;
319 case X86::COND_L: return X86::COND_GE;
320 case X86::COND_LE: return X86::COND_G;
321 case X86::COND_G: return X86::COND_LE;
322 case X86::COND_GE: return X86::COND_L;
323 case X86::COND_B: return X86::COND_AE;
324 case X86::COND_BE: return X86::COND_A;
325 case X86::COND_A: return X86::COND_BE;
326 case X86::COND_AE: return X86::COND_B;
327 case X86::COND_S: return X86::COND_NS;
328 case X86::COND_NS: return X86::COND_S;
329 case X86::COND_P: return X86::COND_NP;
330 case X86::COND_NP: return X86::COND_P;
331 case X86::COND_O: return X86::COND_NO;
332 case X86::COND_NO: return X86::COND_O;
337 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
338 MachineBasicBlock *&TBB,
339 MachineBasicBlock *&FBB,
340 std::vector<MachineOperand> &Cond) const {
341 // TODO: If FP_REG_KILL is around, ignore it.
343 // If the block has no terminators, it just falls into the block after it.
344 MachineBasicBlock::iterator I = MBB.end();
345 if (I == MBB.begin() || !isTerminatorInstr((--I)->getOpcode()))
348 // Get the last instruction in the block.
349 MachineInstr *LastInst = I;
351 // If there is only one terminator instruction, process it.
352 if (I == MBB.begin() || !isTerminatorInstr((--I)->getOpcode())) {
353 if (!isBranch(LastInst->getOpcode()))
356 // If the block ends with a branch there are 3 possibilities:
357 // it's an unconditional, conditional, or indirect branch.
359 if (LastInst->getOpcode() == X86::JMP) {
360 TBB = LastInst->getOperand(0).getMachineBasicBlock();
363 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
364 if (BranchCode == X86::COND_INVALID)
365 return true; // Can't handle indirect branch.
367 // Otherwise, block ends with fall-through condbranch.
368 TBB = LastInst->getOperand(0).getMachineBasicBlock();
369 Cond.push_back(MachineOperand::CreateImm(BranchCode));
373 // Get the instruction before it if it's a terminator.
374 MachineInstr *SecondLastInst = I;
376 // If there are three terminators, we don't know what sort of block this is.
377 if (SecondLastInst && I != MBB.begin() &&
378 isTerminatorInstr((--I)->getOpcode()))
381 // If the block ends with X86::JMP and a conditional branch, handle it.
382 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
383 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
384 TBB = SecondLastInst->getOperand(0).getMachineBasicBlock();
385 Cond.push_back(MachineOperand::CreateImm(BranchCode));
386 FBB = LastInst->getOperand(0).getMachineBasicBlock();
390 // Otherwise, can't handle this.
394 void X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
395 MachineBasicBlock::iterator I = MBB.end();
396 if (I == MBB.begin()) return;
398 if (I->getOpcode() != X86::JMP &&
399 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
402 // Remove the branch.
403 I->eraseFromParent();
407 if (I == MBB.begin()) return;
409 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
412 // Remove the branch.
413 I->eraseFromParent();
416 void X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
417 MachineBasicBlock *FBB,
418 const std::vector<MachineOperand> &Cond) const {
419 // Shouldn't be a fall through.
420 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
421 assert((Cond.size() == 1 || Cond.size() == 0) &&
422 "X86 branch conditions have one component!");
424 if (FBB == 0) { // One way branch.
426 // Unconditional branch?
427 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
429 // Conditional branch.
430 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
431 BuildMI(&MBB, get(Opc)).addMBB(TBB);
436 // Two-way Conditional branch.
437 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
438 BuildMI(&MBB, get(Opc)).addMBB(TBB);
439 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
442 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
443 if (MBB.empty()) return false;
445 switch (MBB.back().getOpcode()) {
446 case X86::JMP: // Uncond branch.
447 case X86::JMP32r: // Indirect branch.
448 case X86::JMP32m: // Indirect branch through mem.
450 default: return false;
455 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
456 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
457 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
461 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
462 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
463 if (Subtarget->is64Bit())
464 return &X86::GR64RegClass;
466 return &X86::GR32RegClass;