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 "X86MachineFunctionInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/LiveVariables.h"
26 #include "llvm/Target/TargetOptions.h"
29 X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
30 : TargetInstrInfoImpl(X86Insts, array_lengthof(X86Insts)),
31 TM(tm), RI(tm, *this) {
34 bool X86InstrInfo::isMoveInstr(const MachineInstr& MI,
36 unsigned& destReg) const {
37 MachineOpCode oc = MI.getOpcode();
38 if (oc == X86::MOV8rr || oc == X86::MOV16rr ||
39 oc == X86::MOV32rr || oc == X86::MOV64rr ||
40 oc == X86::MOV16to16_ || oc == X86::MOV32to32_ ||
41 oc == X86::MOV_Fp3232 || oc == X86::MOVSSrr || oc == X86::MOVSDrr ||
42 oc == X86::MOV_Fp3264 || oc == X86::MOV_Fp6432 || oc == X86::MOV_Fp6464 ||
43 oc == X86::FsMOVAPSrr || oc == X86::FsMOVAPDrr ||
44 oc == X86::MOVAPSrr || oc == X86::MOVAPDrr ||
45 oc == X86::MOVSS2PSrr || oc == X86::MOVSD2PDrr ||
46 oc == X86::MOVPS2SSrr || oc == X86::MOVPD2SDrr ||
47 oc == X86::MMX_MOVD64rr || oc == X86::MMX_MOVQ64rr) {
48 assert(MI.getNumOperands() >= 2 &&
49 MI.getOperand(0).isRegister() &&
50 MI.getOperand(1).isRegister() &&
51 "invalid register-register move instruction");
52 sourceReg = MI.getOperand(1).getReg();
53 destReg = MI.getOperand(0).getReg();
59 unsigned X86InstrInfo::isLoadFromStackSlot(MachineInstr *MI,
60 int &FrameIndex) const {
61 switch (MI->getOpcode()) {
74 case X86::MMX_MOVD64rm:
75 case X86::MMX_MOVQ64rm:
76 if (MI->getOperand(1).isFI() && MI->getOperand(2).isImm() &&
77 MI->getOperand(3).isReg() && MI->getOperand(4).isImm() &&
78 MI->getOperand(2).getImm() == 1 &&
79 MI->getOperand(3).getReg() == 0 &&
80 MI->getOperand(4).getImm() == 0) {
81 FrameIndex = MI->getOperand(1).getIndex();
82 return MI->getOperand(0).getReg();
89 unsigned X86InstrInfo::isStoreToStackSlot(MachineInstr *MI,
90 int &FrameIndex) const {
91 switch (MI->getOpcode()) {
104 case X86::MMX_MOVD64mr:
105 case X86::MMX_MOVQ64mr:
106 case X86::MMX_MOVNTQmr:
107 if (MI->getOperand(0).isFI() && MI->getOperand(1).isImm() &&
108 MI->getOperand(2).isReg() && MI->getOperand(3).isImm() &&
109 MI->getOperand(1).getImm() == 1 &&
110 MI->getOperand(2).getReg() == 0 &&
111 MI->getOperand(3).getImm() == 0) {
112 FrameIndex = MI->getOperand(0).getIndex();
113 return MI->getOperand(4).getReg();
121 bool X86InstrInfo::isReallyTriviallyReMaterializable(MachineInstr *MI) const {
122 switch (MI->getOpcode()) {
135 case X86::MMX_MOVD64rm:
136 case X86::MMX_MOVQ64rm:
137 // Loads from constant pools are trivially rematerializable.
138 if (MI->getOperand(1).isReg() && MI->getOperand(2).isImm() &&
139 MI->getOperand(3).isReg() && MI->getOperand(4).isCPI() &&
140 MI->getOperand(1).getReg() == 0 &&
141 MI->getOperand(2).getImm() == 1 &&
142 MI->getOperand(3).getReg() == 0)
146 // All other instructions marked M_REMATERIALIZABLE are always trivially
151 /// isReallySideEffectFree - If the M_MAY_HAVE_SIDE_EFFECTS flag is set, this
152 /// method is called to determine if the specific instance of this instruction
153 /// has side effects. This is useful in cases of instructions, like loads, which
154 /// generally always have side effects. A load from a constant pool doesn't have
155 /// side effects, though. So we need to differentiate it from the general case.
156 bool X86InstrInfo::isReallySideEffectFree(MachineInstr *MI) const {
157 switch (MI->getOpcode()) {
160 if (MI->getOperand(1).isRegister()) {
161 unsigned Reg = MI->getOperand(1).getReg();
163 // Loads from global addresses which aren't redefined in the function are
165 if (Reg != 0 && MRegisterInfo::isVirtualRegister(Reg) &&
166 MI->getOperand(2).isImmediate() &&
167 MI->getOperand(3).isRegister() &&
168 MI->getOperand(4).isGlobalAddress() &&
169 MI->getOperand(2).getImm() == 1 &&
170 MI->getOperand(3).getReg() == 0)
176 // Anything that is rematerializable obviously has no side effects.
177 return isReallyTriviallyReMaterializable(MI);
180 /// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that
181 /// is not marked dead.
182 static bool hasLiveCondCodeDef(MachineInstr *MI) {
183 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
184 MachineOperand &MO = MI->getOperand(i);
185 if (MO.isRegister() && MO.isDef() &&
186 MO.getReg() == X86::EFLAGS && !MO.isDead()) {
193 /// convertToThreeAddress - This method must be implemented by targets that
194 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
195 /// may be able to convert a two-address instruction into a true
196 /// three-address instruction on demand. This allows the X86 target (for
197 /// example) to convert ADD and SHL instructions into LEA instructions if they
198 /// would require register copies due to two-addressness.
200 /// This method returns a null pointer if the transformation cannot be
201 /// performed, otherwise it returns the new instruction.
204 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
205 MachineBasicBlock::iterator &MBBI,
206 LiveVariables &LV) const {
207 MachineInstr *MI = MBBI;
208 // All instructions input are two-addr instructions. Get the known operands.
209 unsigned Dest = MI->getOperand(0).getReg();
210 unsigned Src = MI->getOperand(1).getReg();
212 MachineInstr *NewMI = NULL;
213 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When
214 // we have better subtarget support, enable the 16-bit LEA generation here.
215 bool DisableLEA16 = true;
217 unsigned MIOpc = MI->getOpcode();
219 case X86::SHUFPSrri: {
220 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!");
221 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0;
223 unsigned A = MI->getOperand(0).getReg();
224 unsigned B = MI->getOperand(1).getReg();
225 unsigned C = MI->getOperand(2).getReg();
226 unsigned M = MI->getOperand(3).getImm();
227 if (B != C) return 0;
228 NewMI = BuildMI(get(X86::PSHUFDri), A).addReg(B).addImm(M);
232 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
233 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
234 // the flags produced by a shift yet, so this is safe.
235 unsigned Dest = MI->getOperand(0).getReg();
236 unsigned Src = MI->getOperand(1).getReg();
237 unsigned ShAmt = MI->getOperand(2).getImm();
238 if (ShAmt == 0 || ShAmt >= 4) return 0;
240 NewMI = BuildMI(get(X86::LEA64r), Dest)
241 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
245 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
246 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
247 // the flags produced by a shift yet, so this is safe.
248 unsigned Dest = MI->getOperand(0).getReg();
249 unsigned Src = MI->getOperand(1).getReg();
250 unsigned ShAmt = MI->getOperand(2).getImm();
251 if (ShAmt == 0 || ShAmt >= 4) return 0;
253 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() ?
254 X86::LEA64_32r : X86::LEA32r;
255 NewMI = BuildMI(get(Opc), Dest)
256 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
260 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!");
261 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses
262 // the flags produced by a shift yet, so this is safe.
263 unsigned Dest = MI->getOperand(0).getReg();
264 unsigned Src = MI->getOperand(1).getReg();
265 unsigned ShAmt = MI->getOperand(2).getImm();
266 if (ShAmt == 0 || ShAmt >= 4) return 0;
269 // If 16-bit LEA is disabled, use 32-bit LEA via subregisters.
270 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo();
271 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit()
272 ? X86::LEA64_32r : X86::LEA32r;
273 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
274 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass);
277 BuildMI(get(X86::INSERT_SUBREG), leaInReg).addReg(Src).addImm(2);
278 Ins->copyKillDeadInfo(MI);
280 NewMI = BuildMI(get(Opc), leaOutReg)
281 .addReg(0).addImm(1 << ShAmt).addReg(leaInReg).addImm(0);
284 BuildMI(get(X86::EXTRACT_SUBREG), Dest).addReg(leaOutReg).addImm(2);
285 Ext->copyKillDeadInfo(MI);
287 MFI->insert(MBBI, Ins); // Insert the insert_subreg
288 LV.instructionChanged(MI, NewMI); // Update live variables
289 LV.addVirtualRegisterKilled(leaInReg, NewMI);
290 MFI->insert(MBBI, NewMI); // Insert the new inst
291 LV.addVirtualRegisterKilled(leaOutReg, Ext);
292 MFI->insert(MBBI, Ext); // Insert the extract_subreg
295 NewMI = BuildMI(get(X86::LEA16r), Dest)
296 .addReg(0).addImm(1 << ShAmt).addReg(Src).addImm(0);
301 // The following opcodes also sets the condition code register(s). Only
302 // convert them to equivalent lea if the condition code register def's
304 if (hasLiveCondCodeDef(MI))
307 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
312 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
313 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r
314 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
315 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, 1);
320 if (DisableLEA16) return 0;
321 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!");
322 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, 1);
326 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
327 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r
328 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
329 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src, -1);
334 if (DisableLEA16) return 0;
335 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!");
336 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src, -1);
340 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
341 unsigned Opc = MIOpc == X86::ADD64rr ? X86::LEA64r
342 : (is64Bit ? X86::LEA64_32r : X86::LEA32r);
343 NewMI = addRegReg(BuildMI(get(Opc), Dest), Src,
344 MI->getOperand(2).getReg());
348 if (DisableLEA16) return 0;
349 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
350 NewMI = addRegReg(BuildMI(get(X86::LEA16r), Dest), Src,
351 MI->getOperand(2).getReg());
355 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
356 if (MI->getOperand(2).isImmediate())
357 NewMI = addRegOffset(BuildMI(get(X86::LEA64r), Dest), Src,
358 MI->getOperand(2).getImm());
362 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
363 if (MI->getOperand(2).isImmediate()) {
364 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r;
365 NewMI = addRegOffset(BuildMI(get(Opc), Dest), Src,
366 MI->getOperand(2).getImm());
371 if (DisableLEA16) return 0;
372 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!");
373 if (MI->getOperand(2).isImmediate())
374 NewMI = addRegOffset(BuildMI(get(X86::LEA16r), Dest), Src,
375 MI->getOperand(2).getImm());
378 if (DisableLEA16) return 0;
381 assert(MI->getNumOperands() >= 3 && MI->getOperand(2).isImmediate() &&
382 "Unknown shl instruction!");
383 unsigned ShAmt = MI->getOperand(2).getImm();
384 if (ShAmt == 1 || ShAmt == 2 || ShAmt == 3) {
386 AM.Scale = 1 << ShAmt;
388 unsigned Opc = MIOpc == X86::SHL64ri ? X86::LEA64r
389 : (MIOpc == X86::SHL32ri
390 ? (is64Bit ? X86::LEA64_32r : X86::LEA32r) : X86::LEA16r);
391 NewMI = addFullAddress(BuildMI(get(Opc), Dest), AM);
399 NewMI->copyKillDeadInfo(MI);
400 LV.instructionChanged(MI, NewMI); // Update live variables
401 MFI->insert(MBBI, NewMI); // Insert the new inst
405 /// commuteInstruction - We have a few instructions that must be hacked on to
408 MachineInstr *X86InstrInfo::commuteInstruction(MachineInstr *MI) const {
409 switch (MI->getOpcode()) {
410 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I)
411 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I)
412 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I)
413 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I)
414 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I)
415 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I)
418 switch (MI->getOpcode()) {
419 default: assert(0 && "Unreachable!");
420 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break;
421 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break;
422 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break;
423 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break;
424 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break;
425 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break;
427 unsigned Amt = MI->getOperand(3).getImm();
428 unsigned A = MI->getOperand(0).getReg();
429 unsigned B = MI->getOperand(1).getReg();
430 unsigned C = MI->getOperand(2).getReg();
431 bool BisKill = MI->getOperand(1).isKill();
432 bool CisKill = MI->getOperand(2).isKill();
433 return BuildMI(get(Opc), A).addReg(C, false, false, CisKill)
434 .addReg(B, false, false, BisKill).addImm(Size-Amt);
439 case X86::CMOVAE16rr:
440 case X86::CMOVAE32rr:
441 case X86::CMOVAE64rr:
445 case X86::CMOVNE16rr:
446 case X86::CMOVNE32rr:
447 case X86::CMOVNE64rr:
448 case X86::CMOVBE16rr:
449 case X86::CMOVBE32rr:
450 case X86::CMOVBE64rr:
457 case X86::CMOVGE16rr:
458 case X86::CMOVGE32rr:
459 case X86::CMOVGE64rr:
460 case X86::CMOVLE16rr:
461 case X86::CMOVLE32rr:
462 case X86::CMOVLE64rr:
469 case X86::CMOVNS16rr:
470 case X86::CMOVNS32rr:
471 case X86::CMOVNS64rr:
475 case X86::CMOVNP16rr:
476 case X86::CMOVNP32rr:
477 case X86::CMOVNP64rr: {
479 switch (MI->getOpcode()) {
481 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break;
482 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break;
483 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break;
484 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break;
485 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break;
486 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break;
487 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break;
488 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break;
489 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break;
490 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break;
491 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break;
492 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break;
493 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break;
494 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break;
495 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break;
496 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break;
497 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break;
498 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break;
499 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break;
500 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break;
501 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break;
502 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break;
503 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break;
504 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break;
505 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break;
506 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break;
507 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break;
508 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break;
509 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break;
510 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break;
511 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break;
512 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break;
513 case X86::CMOVS64rr: Opc = X86::CMOVNS32rr; break;
514 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break;
515 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break;
516 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break;
517 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break;
518 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break;
519 case X86::CMOVP64rr: Opc = X86::CMOVNP32rr; break;
520 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break;
521 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break;
522 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break;
525 MI->setInstrDescriptor(get(Opc));
526 // Fallthrough intended.
529 return TargetInstrInfoImpl::commuteInstruction(MI);
533 static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) {
535 default: return X86::COND_INVALID;
536 case X86::JE: return X86::COND_E;
537 case X86::JNE: return X86::COND_NE;
538 case X86::JL: return X86::COND_L;
539 case X86::JLE: return X86::COND_LE;
540 case X86::JG: return X86::COND_G;
541 case X86::JGE: return X86::COND_GE;
542 case X86::JB: return X86::COND_B;
543 case X86::JBE: return X86::COND_BE;
544 case X86::JA: return X86::COND_A;
545 case X86::JAE: return X86::COND_AE;
546 case X86::JS: return X86::COND_S;
547 case X86::JNS: return X86::COND_NS;
548 case X86::JP: return X86::COND_P;
549 case X86::JNP: return X86::COND_NP;
550 case X86::JO: return X86::COND_O;
551 case X86::JNO: return X86::COND_NO;
555 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) {
557 default: assert(0 && "Illegal condition code!");
558 case X86::COND_E: return X86::JE;
559 case X86::COND_NE: return X86::JNE;
560 case X86::COND_L: return X86::JL;
561 case X86::COND_LE: return X86::JLE;
562 case X86::COND_G: return X86::JG;
563 case X86::COND_GE: return X86::JGE;
564 case X86::COND_B: return X86::JB;
565 case X86::COND_BE: return X86::JBE;
566 case X86::COND_A: return X86::JA;
567 case X86::COND_AE: return X86::JAE;
568 case X86::COND_S: return X86::JS;
569 case X86::COND_NS: return X86::JNS;
570 case X86::COND_P: return X86::JP;
571 case X86::COND_NP: return X86::JNP;
572 case X86::COND_O: return X86::JO;
573 case X86::COND_NO: return X86::JNO;
577 /// GetOppositeBranchCondition - Return the inverse of the specified condition,
578 /// e.g. turning COND_E to COND_NE.
579 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) {
581 default: assert(0 && "Illegal condition code!");
582 case X86::COND_E: return X86::COND_NE;
583 case X86::COND_NE: return X86::COND_E;
584 case X86::COND_L: return X86::COND_GE;
585 case X86::COND_LE: return X86::COND_G;
586 case X86::COND_G: return X86::COND_LE;
587 case X86::COND_GE: return X86::COND_L;
588 case X86::COND_B: return X86::COND_AE;
589 case X86::COND_BE: return X86::COND_A;
590 case X86::COND_A: return X86::COND_BE;
591 case X86::COND_AE: return X86::COND_B;
592 case X86::COND_S: return X86::COND_NS;
593 case X86::COND_NS: return X86::COND_S;
594 case X86::COND_P: return X86::COND_NP;
595 case X86::COND_NP: return X86::COND_P;
596 case X86::COND_O: return X86::COND_NO;
597 case X86::COND_NO: return X86::COND_O;
601 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
602 const TargetInstrDescriptor *TID = MI->getInstrDescriptor();
603 if (TID->Flags & M_TERMINATOR_FLAG) {
604 // Conditional branch is a special case.
605 if ((TID->Flags & M_BRANCH_FLAG) != 0 && (TID->Flags & M_BARRIER_FLAG) == 0)
607 if ((TID->Flags & M_PREDICABLE) == 0)
609 return !isPredicated(MI);
614 // For purposes of branch analysis do not count FP_REG_KILL as a terminator.
615 static bool isBrAnalysisUnpredicatedTerminator(const MachineInstr *MI,
616 const X86InstrInfo &TII) {
617 if (MI->getOpcode() == X86::FP_REG_KILL)
619 return TII.isUnpredicatedTerminator(MI);
622 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
623 MachineBasicBlock *&TBB,
624 MachineBasicBlock *&FBB,
625 std::vector<MachineOperand> &Cond) const {
626 // If the block has no terminators, it just falls into the block after it.
627 MachineBasicBlock::iterator I = MBB.end();
628 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this))
631 // Get the last instruction in the block.
632 MachineInstr *LastInst = I;
634 // If there is only one terminator instruction, process it.
635 if (I == MBB.begin() || !isBrAnalysisUnpredicatedTerminator(--I, *this)) {
636 if (!isBranch(LastInst->getOpcode()))
639 // If the block ends with a branch there are 3 possibilities:
640 // it's an unconditional, conditional, or indirect branch.
642 if (LastInst->getOpcode() == X86::JMP) {
643 TBB = LastInst->getOperand(0).getMBB();
646 X86::CondCode BranchCode = GetCondFromBranchOpc(LastInst->getOpcode());
647 if (BranchCode == X86::COND_INVALID)
648 return true; // Can't handle indirect branch.
650 // Otherwise, block ends with fall-through condbranch.
651 TBB = LastInst->getOperand(0).getMBB();
652 Cond.push_back(MachineOperand::CreateImm(BranchCode));
656 // Get the instruction before it if it's a terminator.
657 MachineInstr *SecondLastInst = I;
659 // If there are three terminators, we don't know what sort of block this is.
660 if (SecondLastInst && I != MBB.begin() &&
661 isBrAnalysisUnpredicatedTerminator(--I, *this))
664 // If the block ends with X86::JMP and a conditional branch, handle it.
665 X86::CondCode BranchCode = GetCondFromBranchOpc(SecondLastInst->getOpcode());
666 if (BranchCode != X86::COND_INVALID && LastInst->getOpcode() == X86::JMP) {
667 TBB = SecondLastInst->getOperand(0).getMBB();
668 Cond.push_back(MachineOperand::CreateImm(BranchCode));
669 FBB = LastInst->getOperand(0).getMBB();
673 // If the block ends with two X86::JMPs, handle it. The second one is not
674 // executed, so remove it.
675 if (SecondLastInst->getOpcode() == X86::JMP &&
676 LastInst->getOpcode() == X86::JMP) {
677 TBB = SecondLastInst->getOperand(0).getMBB();
679 I->eraseFromParent();
683 // Otherwise, can't handle this.
687 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
688 MachineBasicBlock::iterator I = MBB.end();
689 if (I == MBB.begin()) return 0;
691 if (I->getOpcode() != X86::JMP &&
692 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
695 // Remove the branch.
696 I->eraseFromParent();
700 if (I == MBB.begin()) return 1;
702 if (GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID)
705 // Remove the branch.
706 I->eraseFromParent();
710 static const MachineInstrBuilder &X86InstrAddOperand(MachineInstrBuilder &MIB,
711 MachineOperand &MO) {
713 MIB = MIB.addReg(MO.getReg(), MO.isDef(), MO.isImplicit(),
714 false, false, MO.getSubReg());
715 else if (MO.isImmediate())
716 MIB = MIB.addImm(MO.getImm());
717 else if (MO.isFrameIndex())
718 MIB = MIB.addFrameIndex(MO.getIndex());
719 else if (MO.isGlobalAddress())
720 MIB = MIB.addGlobalAddress(MO.getGlobal(), MO.getOffset());
721 else if (MO.isConstantPoolIndex())
722 MIB = MIB.addConstantPoolIndex(MO.getIndex(), MO.getOffset());
723 else if (MO.isJumpTableIndex())
724 MIB = MIB.addJumpTableIndex(MO.getIndex());
725 else if (MO.isExternalSymbol())
726 MIB = MIB.addExternalSymbol(MO.getSymbolName());
728 assert(0 && "Unknown operand for X86InstrAddOperand!");
734 X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
735 MachineBasicBlock *FBB,
736 const std::vector<MachineOperand> &Cond) const {
737 // Shouldn't be a fall through.
738 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
739 assert((Cond.size() == 1 || Cond.size() == 0) &&
740 "X86 branch conditions have one component!");
742 if (FBB == 0) { // One way branch.
744 // Unconditional branch?
745 BuildMI(&MBB, get(X86::JMP)).addMBB(TBB);
747 // Conditional branch.
748 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
749 BuildMI(&MBB, get(Opc)).addMBB(TBB);
754 // Two-way Conditional branch.
755 unsigned Opc = GetCondBranchFromCond((X86::CondCode)Cond[0].getImm());
756 BuildMI(&MBB, get(Opc)).addMBB(TBB);
757 BuildMI(&MBB, get(X86::JMP)).addMBB(FBB);
761 void X86InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
762 MachineBasicBlock::iterator MI,
763 unsigned DestReg, unsigned SrcReg,
764 const TargetRegisterClass *DestRC,
765 const TargetRegisterClass *SrcRC) const {
766 if (DestRC != SrcRC) {
767 // Moving EFLAGS to / from another register requires a push and a pop.
768 if (SrcRC == &X86::CCRRegClass) {
769 assert(SrcReg == X86::EFLAGS);
770 if (DestRC == &X86::GR64RegClass) {
771 BuildMI(MBB, MI, get(X86::PUSHFQ));
772 BuildMI(MBB, MI, get(X86::POP64r), DestReg);
774 } else if (DestRC == &X86::GR32RegClass) {
775 BuildMI(MBB, MI, get(X86::PUSHFD));
776 BuildMI(MBB, MI, get(X86::POP32r), DestReg);
779 } else if (DestRC == &X86::CCRRegClass) {
780 assert(DestReg == X86::EFLAGS);
781 if (SrcRC == &X86::GR64RegClass) {
782 BuildMI(MBB, MI, get(X86::PUSH64r)).addReg(SrcReg);
783 BuildMI(MBB, MI, get(X86::POPFQ));
785 } else if (SrcRC == &X86::GR32RegClass) {
786 BuildMI(MBB, MI, get(X86::PUSH32r)).addReg(SrcReg);
787 BuildMI(MBB, MI, get(X86::POPFD));
791 cerr << "Not yet supported!";
796 if (DestRC == &X86::GR64RegClass) {
798 } else if (DestRC == &X86::GR32RegClass) {
800 } else if (DestRC == &X86::GR16RegClass) {
802 } else if (DestRC == &X86::GR8RegClass) {
804 } else if (DestRC == &X86::GR32_RegClass) {
806 } else if (DestRC == &X86::GR16_RegClass) {
808 } else if (DestRC == &X86::RFP32RegClass) {
809 Opc = X86::MOV_Fp3232;
810 } else if (DestRC == &X86::RFP64RegClass || DestRC == &X86::RSTRegClass) {
811 Opc = X86::MOV_Fp6464;
812 } else if (DestRC == &X86::RFP80RegClass) {
813 Opc = X86::MOV_Fp8080;
814 } else if (DestRC == &X86::FR32RegClass) {
815 Opc = X86::FsMOVAPSrr;
816 } else if (DestRC == &X86::FR64RegClass) {
817 Opc = X86::FsMOVAPDrr;
818 } else if (DestRC == &X86::VR128RegClass) {
820 } else if (DestRC == &X86::VR64RegClass) {
821 Opc = X86::MMX_MOVQ64rr;
823 assert(0 && "Unknown regclass");
826 BuildMI(MBB, MI, get(Opc), DestReg).addReg(SrcReg);
829 static unsigned getStoreRegOpcode(const TargetRegisterClass *RC,
830 unsigned StackAlign) {
832 if (RC == &X86::GR64RegClass) {
834 } else if (RC == &X86::GR32RegClass) {
836 } else if (RC == &X86::GR16RegClass) {
838 } else if (RC == &X86::GR8RegClass) {
840 } else if (RC == &X86::GR32_RegClass) {
842 } else if (RC == &X86::GR16_RegClass) {
844 } else if (RC == &X86::RFP80RegClass) {
845 Opc = X86::ST_FpP80m; // pops
846 } else if (RC == &X86::RFP64RegClass) {
848 } else if (RC == &X86::RFP32RegClass) {
850 } else if (RC == &X86::FR32RegClass) {
852 } else if (RC == &X86::FR64RegClass) {
854 } else if (RC == &X86::VR128RegClass) {
855 // FIXME: Use movaps once we are capable of selectively
856 // aligning functions that spill SSE registers on 16-byte boundaries.
857 Opc = StackAlign >= 16 ? X86::MOVAPSmr : X86::MOVUPSmr;
858 } else if (RC == &X86::VR64RegClass) {
859 Opc = X86::MMX_MOVQ64mr;
861 assert(0 && "Unknown regclass");
868 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
869 MachineBasicBlock::iterator MI,
870 unsigned SrcReg, bool isKill, int FrameIdx,
871 const TargetRegisterClass *RC) const {
872 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
873 addFrameReference(BuildMI(MBB, MI, get(Opc)), FrameIdx)
874 .addReg(SrcReg, false, false, isKill);
877 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
879 SmallVectorImpl<MachineOperand> &Addr,
880 const TargetRegisterClass *RC,
881 SmallVectorImpl<MachineInstr*> &NewMIs) const {
882 unsigned Opc = getStoreRegOpcode(RC, RI.getStackAlignment());
883 MachineInstrBuilder MIB = BuildMI(get(Opc));
884 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
885 MIB = X86InstrAddOperand(MIB, Addr[i]);
886 MIB.addReg(SrcReg, false, false, isKill);
887 NewMIs.push_back(MIB);
890 static unsigned getLoadRegOpcode(const TargetRegisterClass *RC,
891 unsigned StackAlign) {
893 if (RC == &X86::GR64RegClass) {
895 } else if (RC == &X86::GR32RegClass) {
897 } else if (RC == &X86::GR16RegClass) {
899 } else if (RC == &X86::GR8RegClass) {
901 } else if (RC == &X86::GR32_RegClass) {
903 } else if (RC == &X86::GR16_RegClass) {
905 } else if (RC == &X86::RFP80RegClass) {
907 } else if (RC == &X86::RFP64RegClass) {
909 } else if (RC == &X86::RFP32RegClass) {
911 } else if (RC == &X86::FR32RegClass) {
913 } else if (RC == &X86::FR64RegClass) {
915 } else if (RC == &X86::VR128RegClass) {
916 // FIXME: Use movaps once we are capable of selectively
917 // aligning functions that spill SSE registers on 16-byte boundaries.
918 Opc = StackAlign >= 16 ? X86::MOVAPSrm : X86::MOVUPSrm;
919 } else if (RC == &X86::VR64RegClass) {
920 Opc = X86::MMX_MOVQ64rm;
922 assert(0 && "Unknown regclass");
929 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
930 MachineBasicBlock::iterator MI,
931 unsigned DestReg, int FrameIdx,
932 const TargetRegisterClass *RC) const{
933 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
934 addFrameReference(BuildMI(MBB, MI, get(Opc), DestReg), FrameIdx);
937 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
938 SmallVectorImpl<MachineOperand> &Addr,
939 const TargetRegisterClass *RC,
940 SmallVectorImpl<MachineInstr*> &NewMIs) const {
941 unsigned Opc = getLoadRegOpcode(RC, RI.getStackAlignment());
942 MachineInstrBuilder MIB = BuildMI(get(Opc), DestReg);
943 for (unsigned i = 0, e = Addr.size(); i != e; ++i)
944 MIB = X86InstrAddOperand(MIB, Addr[i]);
945 NewMIs.push_back(MIB);
948 bool X86InstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB,
949 MachineBasicBlock::iterator MI,
950 const std::vector<CalleeSavedInfo> &CSI) const {
954 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
955 unsigned SlotSize = is64Bit ? 8 : 4;
957 MachineFunction &MF = *MBB.getParent();
958 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
959 X86FI->setCalleeSavedFrameSize(CSI.size() * SlotSize);
961 unsigned Opc = is64Bit ? X86::PUSH64r : X86::PUSH32r;
962 for (unsigned i = CSI.size(); i != 0; --i) {
963 unsigned Reg = CSI[i-1].getReg();
964 // Add the callee-saved register as live-in. It's killed at the spill.
966 BuildMI(MBB, MI, get(Opc)).addReg(Reg);
971 bool X86InstrInfo::restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
972 MachineBasicBlock::iterator MI,
973 const std::vector<CalleeSavedInfo> &CSI) const {
977 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
979 unsigned Opc = is64Bit ? X86::POP64r : X86::POP32r;
980 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
981 unsigned Reg = CSI[i].getReg();
982 BuildMI(MBB, MI, get(Opc), Reg);
987 bool X86InstrInfo::BlockHasNoFallThrough(MachineBasicBlock &MBB) const {
988 if (MBB.empty()) return false;
990 switch (MBB.back().getOpcode()) {
991 case X86::TCRETURNri:
992 case X86::TCRETURNdi:
993 case X86::RET: // Return.
998 case X86::JMP: // Uncond branch.
999 case X86::JMP32r: // Indirect branch.
1000 case X86::JMP64r: // Indirect branch (64-bit).
1001 case X86::JMP32m: // Indirect branch through mem.
1002 case X86::JMP64m: // Indirect branch through mem (64-bit).
1004 default: return false;
1009 ReverseBranchCondition(std::vector<MachineOperand> &Cond) const {
1010 assert(Cond.size() == 1 && "Invalid X86 branch condition!");
1011 Cond[0].setImm(GetOppositeBranchCondition((X86::CondCode)Cond[0].getImm()));
1015 const TargetRegisterClass *X86InstrInfo::getPointerRegClass() const {
1016 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
1017 if (Subtarget->is64Bit())
1018 return &X86::GR64RegClass;
1020 return &X86::GR32RegClass;