1 //===- X86InstrInfo.cpp - X86 Instruction Information -----------*- C++ -*-===//
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 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/ADT/STLExtras.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/SSARegMap.h"
24 #include "llvm/Target/TargetOptions.h"
27 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
28 : TargetInstrInfo(X86Insts, array_lengthof(X86Insts)),
29 TM(tm), RI(tm, *this) {
32 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
34 unsigned& destReg) const {
35 MachineOpCode oc = MI.getOpcode();
36 if (oc == X86::MOV8rr || oc == X86::MOV16rr ||
37 oc == X86::MOV32rr || oc == X86::MOV64rr ||
38 oc == X86::MOV16to16_ || oc == X86::MOV32to32_ ||
39 oc == X86::MOV_Fp3232 || oc == X86::MOVSSrr || oc == X86::MOVSDrr ||
40 oc == X86::MOV_Fp3264 || oc == X86::MOV_Fp6432 || oc == X86::MOV_Fp6464 ||
41 oc == X86::FsMOVAPSrr || oc == X86::FsMOVAPDrr ||
42 oc == X86::MOVAPSrr || oc == X86::MOVAPDrr ||
43 oc == X86::MOVSS2PSrr || oc == X86::MOVSD2PDrr ||
44 oc == X86::MOVPS2SSrr || oc == X86::MOVPD2SDrr ||
45 oc == X86::MMX_MOVD64rr || oc == X86::MMX_MOVQ64rr) {
46 assert(MI.getNumOperands() >= 2 &&
47 MI.getOperand(0).isRegister() &&
48 MI.getOperand(1).isRegister() &&
49 "invalid register-register move instruction");
50 sourceReg = MI.getOperand(1).getReg();
51 destReg = MI.getOperand(0).getReg();
57 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
58 int &FrameIndex) const {
59 switch (MI->getOpcode()) {
72 case X86::MMX_MOVD64rm:
73 case X86::MMX_MOVQ64rm:
74 if (MI->getOperand(1).isFrameIndex() && MI->getOperand(2).isImmediate() &&
75 MI->getOperand(3).isRegister() && MI->getOperand(4).isImmediate() &&
76 MI->getOperand(2).getImmedValue() == 1 &&
77 MI->getOperand(3).getReg() == 0 &&
78 MI->getOperand(4).getImmedValue() == 0) {
79 FrameIndex = MI->getOperand(1).getFrameIndex();
80 return MI->getOperand(0).getReg();
87 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
88 int &FrameIndex) const {
89 switch (MI->getOpcode()) {
102 case X86::MMX_MOVD64mr:
103 case X86::MMX_MOVQ64mr:
104 case X86::MMX_MOVNTQmr:
105 if (MI->getOperand(0).isFrameIndex() && MI->getOperand(1).isImmediate() &&
106 MI->getOperand(2).isRegister() && MI->getOperand(3).isImmediate() &&
107 MI->getOperand(1).getImmedValue() == 1 &&
108 MI->getOperand(2).getReg() == 0 &&
109 MI->getOperand(3).getImmedValue() == 0) {
110 FrameIndex = MI->getOperand(0).getFrameIndex();
111 return MI->getOperand(4).getReg();
119 bool X86InstrInfo::isReallyTriviallyReMaterializable(MachineInstr *MI) const {
120 switch (MI->getOpcode()) {
133 case X86::MMX_MOVD64rm:
134 case X86::MMX_MOVQ64rm:
135 // Loads from constant pools are trivially rematerializable.
136 return MI->getOperand(1).isRegister() && MI->getOperand(2).isImmediate() &&
137 MI->getOperand(3).isRegister() && MI->getOperand(4).isConstantPoolIndex() &&
138 MI->getOperand(1).getReg() == 0 &&
139 MI->getOperand(2).getImmedValue() == 1 &&
140 MI->getOperand(3).getReg() == 0;
142 // All other instructions marked M_REMATERIALIZABLE are always trivially
147 /// isReallySideEffectFree - If the M_MAY_HAVE_SIDE_EFFECTS flag is set, this
148 /// method is called to determine if the specific instance of this instruction
149 /// has side effects. This is useful in cases of instructions, like loads, which
150 /// generally always have side effects. A load from a constant pool doesn't have
151 /// side effects, though. So we need to differentiate it from the general case.
152 bool X86InstrInfo::isReallySideEffectFree(MachineInstr *MI) const {
153 switch (MI->getOpcode()) {
166 case X86::MMX_MOVD64rm:
167 case X86::MMX_MOVQ64rm:
168 // Loads from constant pools have no side effects
169 return MI->getOperand(1).isRegister() && MI->getOperand(2).isImmediate() &&
170 MI->getOperand(3).isRegister() && MI->getOperand(4).isConstantPoolIndex() &&
171 MI->getOperand(1).getReg() == 0 &&
172 MI->getOperand(2).getImmedValue() == 1 &&
173 MI->getOperand(3).getReg() == 0;
176 // All other instances of these instructions are presumed to have side
181 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
182 /// is not marked dead.
183 static bool hasLiveCondCodeDef(MachineInstr *MI) {
184 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
185 MachineOperand &MO = MI->getOperand(i);
186 if (MO.isRegister() && MO.isDef() &&
187 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
194 /// convertToThreeAddress - This method must be implemented by targets that
195 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
196 /// may be able to convert a two-address instruction into a true
197 /// three-address instruction on demand. This allows the X86 target (for
198 /// example) to convert ADD and SHL instructions into LEA instructions if they
199 /// would require register copies due to two-addressness.
201 /// This method returns a null pointer if the transformation cannot be
202 /// performed, otherwise it returns the new instruction.
205 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
206 MachineBasicBlock::iterator &MBBI,
207 LiveVariables &LV) const {
208 MachineInstr *MI = MBBI;
209 // All instructions input are two-addr instructions. Get the known operands.
210 unsigned Dest = MI->getOperand(0).getReg();
211 unsigned Src = MI->getOperand(1).getReg();
213 MachineInstr *NewMI = NULL;
214 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
215 // we have better subtarget support, enable the 16-bit LEA generation here.
216 bool DisableLEA16 = true;
218 unsigned MIOpc = MI->getOpcode();
220 case X86::SHUFPSrri: {
221 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
222 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
224 unsigned A = MI->getOperand(0).getReg();
225 unsigned B = MI->getOperand(1).getReg();
226 unsigned C = MI->getOperand(2).getReg();
227 unsigned M = MI->getOperand(3).getImm();
228 if (B != C) return 0;
229 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
233 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
234 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
235 // the flags produced by a shift yet, so this is safe.
236 unsigned Dest = MI->getOperand(0).getReg();
237 unsigned Src = MI->getOperand(1).getReg();
238 unsigned ShAmt = MI->getOperand(2).getImm();
239 if (ShAmt == 0 || ShAmt >= 4) return 0;
241 NewMI = BuildMI(get(X86::LEA64r), Dest)
242 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
246 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
247 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
248 // the flags produced by a shift yet, so this is safe.
249 unsigned Dest = MI->getOperand(0).getReg();
250 unsigned Src = MI->getOperand(1).getReg();
251 unsigned ShAmt = MI->getOperand(2).getImm();
252 if (ShAmt == 0 || ShAmt >= 4) return 0;
254 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
255 X86::LEA64_32r : X86::LEA32r;
256 NewMI = BuildMI(get(Opc), Dest)
257 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
261 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
262 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
263 // the flags produced by a shift yet, so this is safe.
264 unsigned Dest = MI->getOperand(0).getReg();
265 unsigned Src = MI->getOperand(1).getReg();
266 unsigned ShAmt = MI->getOperand(2).getImm();
267 if (ShAmt == 0 || ShAmt >= 4) return 0;
270 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
271 SSARegMap *RegMap = MFI->getParent()->getSSARegMap();
272 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
273 ? X86::LEA64_32r : X86::LEA32r;
274 unsigned leaInReg = RegMap->createVirtualRegister(&X86::GR32RegClass);
275 unsigned leaOutReg = RegMap->createVirtualRegister(&X86::GR32RegClass);
278 BuildMI(get(X86::INSERT_SUBREG), leaInReg).addReg(Src).addImm(2);
279 Ins->copyKillDeadInfo(MI);
281 NewMI = BuildMI(get(Opc), leaOutReg)
282 .addReg(0).addImm(1 << ShAmt).addReg(leaInReg).addImm(0);
285 BuildMI(get(X86::EXTRACT_SUBREG), Dest).addReg(leaOutReg).addImm(2);
286 Ext->copyKillDeadInfo(MI);
288 MFI->insert(MBBI, Ins); // Insert the insert_subreg
289 LV.instructionChanged(MI, NewMI); // Update live variables
290 LV.addVirtualRegisterKilled(leaInReg, NewMI);
291 MFI->insert(MBBI, NewMI); // Insert the new inst
292 LV.addVirtualRegisterKilled(leaOutReg, Ext);
293 MFI->insert(MBBI, Ext); // Insert the extract_subreg
296 NewMI = BuildMI(get(X86::LEA16r), Dest)
297 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
302 // The following opcodes also sets the condition code register(s). Only
303 // convert them to equivalent lea if the condition code register def's
305 if (hasLiveCondCodeDef(MI))
308 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
313 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
314 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
315 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
316 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, 1);
321 if (DisableLEA16) return 0;
322 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
323 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
327 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
328 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
329 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
330 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, -1);
335 if (DisableLEA16) return 0;
336 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
337 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
341 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
342 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
343 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
344 NewMI = addRegReg(BuildMI(get(Opc), Dest), Src,
345 MI->getOperand(2).getReg());
349 if (DisableLEA16) return 0;
350 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
351 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
352 MI->getOperand(2).getReg());
356 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
357 if (MI->getOperand(2).isImmediate())
358 NewMI = addRegOffset(BuildMI(get(X86::LEA64r), Dest), Src,
359 MI->getOperand(2).getImmedValue());
363 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
364 if (MI->getOperand(2).isImmediate()) {
365 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
366 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src,
367 MI->getOperand(2).getImmedValue());
372 if (DisableLEA16) return 0;
373 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
374 if (MI->getOperand(2).isImmediate())
375 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
376 MI->getOperand(2).getImmedValue());
379 if (DisableLEA16) return 0;
382 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImmediate() &&
383 "Unknown shl instruction!");
384 unsigned ShAmt = MI->getOperand(2).getImmedValue();
385 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
387 AM.Scale = 1 << ShAmt;
389 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
390 : (MIOpc == X86::SHL32ri
391 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
392 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
400 NewMI->copyKillDeadInfo(MI);
401 LV.instructionChanged(MI, NewMI); // Update live variables
402 MFI->insert(MBBI, NewMI); // Insert the new inst
406 /// commuteInstruction - We have a few instructions that must be hacked on to
409 MachineInstr *X86InstrInfo::commuteInstruction(MachineInstr *MI) const {
410 switch (MI->getOpcode()) {
411 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
412 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
413 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
414 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
415 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
416 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
419 switch (MI->getOpcode()) {
420 default: assert(0 && "Unreachable!");
421 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
422 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
423 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
424 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
425 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
426 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
428 unsigned Amt = MI->getOperand(3).getImmedValue();
429 unsigned A = MI->getOperand(0).getReg();
430 unsigned B = MI->getOperand(1).getReg();
431 unsigned C = MI->getOperand(2).getReg();
432 bool BisKill = MI->getOperand(1).isKill();
433 bool CisKill = MI->getOperand(2).isKill();
434 return BuildMI(get(Opc), A).addReg(C, false, false, CisKill)
435 .addReg(B, false, false, BisKill).addImm(Size-Amt);
440 case X86::CMOVAE16rr:
441 case X86::CMOVAE32rr:
442 case X86::CMOVAE64rr:
446 case X86::CMOVNE16rr:
447 case X86::CMOVNE32rr:
448 case X86::CMOVNE64rr:
449 case X86::CMOVBE16rr:
450 case X86::CMOVBE32rr:
451 case X86::CMOVBE64rr:
458 case X86::CMOVGE16rr:
459 case X86::CMOVGE32rr:
460 case X86::CMOVGE64rr:
461 case X86::CMOVLE16rr:
462 case X86::CMOVLE32rr:
463 case X86::CMOVLE64rr:
470 case X86::CMOVNS16rr:
471 case X86::CMOVNS32rr:
472 case X86::CMOVNS64rr:
476 case X86::CMOVNP16rr:
477 case X86::CMOVNP32rr:
478 case X86::CMOVNP64rr: {
480 switch (MI->getOpcode()) {
482 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
483 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
484 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
485 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
486 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
487 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
488 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
489 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
490 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
491 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
492 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
493 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
494 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
495 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
496 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
497 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
498 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
499 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
500 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
501 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
502 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
503 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
504 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
505 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
506 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
507 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
508 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
509 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
510 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
511 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
512 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
513 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
514 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
515 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
516 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
517 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
518 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
519 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
520 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
521 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
522 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
523 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
526 MI->setInstrDescriptor(get(Opc));
527 // Fallthrough intended.
530 return TargetInstrInfo::commuteInstruction(MI);
534 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
536 default: return X86::COND_INVALID;
537 case X86::JE: return X86::COND_E;
538 case X86::JNE: return X86::COND_NE;
539 case X86::JL: return X86::COND_L;
540 case X86::JLE: return X86::COND_LE;
541 case X86::JG: return X86::COND_G;
542 case X86::JGE: return X86::COND_GE;
543 case X86::JB: return X86::COND_B;
544 case X86::JBE: return X86::COND_BE;
545 case X86::JA: return X86::COND_A;
546 case X86::JAE: return X86::COND_AE;
547 case X86::JS: return X86::COND_S;
548 case X86::JNS: return X86::COND_NS;
549 case X86::JP: return X86::COND_P;
550 case X86::JNP: return X86::COND_NP;
551 case X86::JO: return X86::COND_O;
552 case X86::JNO: return X86::COND_NO;
556 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
558 default: assert(0 && "Illegal condition code!");
559 case X86::COND_E: return X86::JE;
560 case X86::COND_NE: return X86::JNE;
561 case X86::COND_L: return X86::JL;
562 case X86::COND_LE: return X86::JLE;
563 case X86::COND_G: return X86::JG;
564 case X86::COND_GE: return X86::JGE;
565 case X86::COND_B: return X86::JB;
566 case X86::COND_BE: return X86::JBE;
567 case X86::COND_A: return X86::JA;
568 case X86::COND_AE: return X86::JAE;
569 case X86::COND_S: return X86::JS;
570 case X86::COND_NS: return X86::JNS;
571 case X86::COND_P: return X86::JP;
572 case X86::COND_NP: return X86::JNP;
573 case X86::COND_O: return X86::JO;
574 case X86::COND_NO: return X86::JNO;
578 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
579 /// e.g. turning COND_E to COND_NE.
580 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
582 default: assert(0 && "Illegal condition code!");
583 case X86::COND_E: return X86::COND_NE;
584 case X86::COND_NE: return X86::COND_E;
585 case X86::COND_L: return X86::COND_GE;
586 case X86::COND_LE: return X86::COND_G;
587 case X86::COND_G: return X86::COND_LE;
588 case X86::COND_GE: return X86::COND_L;
589 case X86::COND_B: return X86::COND_AE;
590 case X86::COND_BE: return X86::COND_A;
591 case X86::COND_A: return X86::COND_BE;
592 case X86::COND_AE: return X86::COND_B;
593 case X86::COND_S: return X86::COND_NS;
594 case X86::COND_NS: return X86::COND_S;
595 case X86::COND_P: return X86::COND_NP;
596 case X86::COND_NP: return X86::COND_P;
597 case X86::COND_O: return X86::COND_NO;
598 case X86::COND_NO: return X86::COND_O;
602 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
603 const TargetInstrDescriptor *TID = MI->getInstrDescriptor();
604 if (TID->Flags & M_TERMINATOR_FLAG) {
605 // Conditional branch is a special case.
606 if ((TID->Flags & M_BRANCH_FLAG) != 0 && (TID->Flags & M_BARRIER_FLAG) == 0)
608 if ((TID->Flags & M_PREDICABLE) == 0)
610 return !isPredicated(MI);
615 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
616 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
617 const X86InstrInfo &TII) {
618 if (MI->getOpcode() == X86::FP_REG_KILL)
620 return TII.isUnpredicatedTerminator(MI);
623 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
624 MachineBasicBlock *&TBB,
625 MachineBasicBlock *&FBB,
626 std::vector<MachineOperand> &Cond) const {
627 // If the block has no terminators, it just falls into the block after it.
628 MachineBasicBlock::iterator I = MBB.end();
629 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this))
632 // Get the last instruction in the block.
633 MachineInstr *LastInst = I;
635 // If there is only one terminator instruction, process it.
636 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this)) {
637 if (!isBranch(LastInst->getOpcode()))
640 // If the block ends with a branch there are 3 possibilities:
641 // it's an unconditional, conditional, or indirect branch.
643 if (LastInst->getOpcode() == X86::JMP) {
644 TBB = LastInst->getOperand(0).getMachineBasicBlock();
647 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
648 if (BranchCode == X86::COND_INVALID)
649 return true; // Can't handle indirect branch.
651 // Otherwise, block ends with fall-through condbranch.
652 TBB = LastInst->getOperand(0).getMachineBasicBlock();
653 Cond.push_back(MachineOperand::CreateImm(BranchCode));
657 // Get the instruction before it if it's a terminator.
658 MachineInstr *SecondLastInst = I;
660 // If there are three terminators, we don't know what sort of block this is.
661 if (SecondLastInst && I != MBB.begin() &&
662 isBrAnalysisUnpredicatedTerminator(--I, *this))
665 // If the block ends with X86::JMP and a conditional branch, handle it.
666 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
667 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
668 TBB = SecondLastInst->getOperand(0).getMachineBasicBlock();
669 Cond.push_back(MachineOperand::CreateImm(BranchCode));
670 FBB = LastInst->getOperand(0).getMachineBasicBlock();
674 // If the block ends with two X86::JMPs, handle it. The second one is not
675 // executed, so remove it.
676 if (SecondLastInst->getOpcode() == X86::JMP &&
677 LastInst->getOpcode() == X86::JMP) {
678 TBB = SecondLastInst->getOperand(0).getMachineBasicBlock();
680 I->eraseFromParent();
684 // Otherwise, can't handle this.
688 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
689 MachineBasicBlock::iterator I = MBB.end();
690 if (I == MBB.begin()) return 0;
692 if (I->getOpcode() != X86::JMP &&
693 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
696 // Remove the branch.
697 I->eraseFromParent();
701 if (I == MBB.begin()) return 1;
703 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
706 // Remove the branch.
707 I->eraseFromParent();
712 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
713 MachineBasicBlock *FBB,
714 const std::vector<MachineOperand> &Cond) const {
715 // Shouldn't be a fall through.
716 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
717 assert((Cond.size() == 1 || Cond.size() == 0) &&
718 "X86 branch conditions have one component!");
720 if (FBB == 0) { // One way branch.
722 // Unconditional branch?
723 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
725 // Conditional branch.
726 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
727 BuildMI(&MBB, get(Opc)).addMBB(TBB);
732 // Two-way Conditional branch.
733 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
734 BuildMI(&MBB, get(Opc)).addMBB(TBB);
735 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
739 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
740 if (MBB.empty()) return false;
742 switch (MBB.back().getOpcode()) {
743 case X86::TCRETURNri:
744 case X86::TCRETURNdi:
745 case X86::RET: // Return.
750 case X86::JMP: // Uncond branch.
751 case X86::JMP32r: // Indirect branch.
752 case X86::JMP64r: // Indirect branch (64-bit).
753 case X86::JMP32m: // Indirect branch through mem.
754 case X86::JMP64m: // Indirect branch through mem (64-bit).
756 default: return false;
761 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
762 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
763 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
767 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
768 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
769 if (Subtarget->is64Bit())
770 return &X86::GR64RegClass;
772 return &X86::GR32RegClass;