1 //===-- X86InstrInfo.cpp - X86 Instruction Information --------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the X86 implementation of the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "X86InstrInfo.h" 15 #include "X86.h" 16 #include "X86InstrBuilder.h" 17 #include "X86InstrFoldTables.h" 18 #include "X86MachineFunctionInfo.h" 19 #include "X86Subtarget.h" 20 #include "X86TargetMachine.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/Sequence.h" 23 #include "llvm/CodeGen/LivePhysRegs.h" 24 #include "llvm/CodeGen/LiveVariables.h" 25 #include "llvm/CodeGen/MachineConstantPool.h" 26 #include "llvm/CodeGen/MachineDominators.h" 27 #include "llvm/CodeGen/MachineFrameInfo.h" 28 #include "llvm/CodeGen/MachineInstrBuilder.h" 29 #include "llvm/CodeGen/MachineModuleInfo.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/StackMaps.h" 32 #include "llvm/IR/DerivedTypes.h" 33 #include "llvm/IR/Function.h" 34 #include "llvm/IR/LLVMContext.h" 35 #include "llvm/MC/MCAsmInfo.h" 36 #include "llvm/MC/MCExpr.h" 37 #include "llvm/MC/MCInst.h" 38 #include "llvm/Support/CommandLine.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/raw_ostream.h" 42 #include "llvm/Target/TargetOptions.h" 43 44 using namespace llvm; 45 46 #define DEBUG_TYPE "x86-instr-info" 47 48 #define GET_INSTRINFO_CTOR_DTOR 49 #include "X86GenInstrInfo.inc" 50 51 static cl::opt<bool> 52 NoFusing("disable-spill-fusing", 53 cl::desc("Disable fusing of spill code into instructions"), 54 cl::Hidden); 55 static cl::opt<bool> 56 PrintFailedFusing("print-failed-fuse-candidates", 57 cl::desc("Print instructions that the allocator wants to" 58 " fuse, but the X86 backend currently can't"), 59 cl::Hidden); 60 static cl::opt<bool> 61 ReMatPICStubLoad("remat-pic-stub-load", 62 cl::desc("Re-materialize load from stub in PIC mode"), 63 cl::init(false), cl::Hidden); 64 static cl::opt<unsigned> 65 PartialRegUpdateClearance("partial-reg-update-clearance", 66 cl::desc("Clearance between two register writes " 67 "for inserting XOR to avoid partial " 68 "register update"), 69 cl::init(64), cl::Hidden); 70 static cl::opt<unsigned> 71 UndefRegClearance("undef-reg-clearance", 72 cl::desc("How many idle instructions we would like before " 73 "certain undef register reads"), 74 cl::init(128), cl::Hidden); 75 76 77 // Pin the vtable to this file. 78 void X86InstrInfo::anchor() {} 79 80 X86InstrInfo::X86InstrInfo(X86Subtarget &STI) 81 : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64 82 : X86::ADJCALLSTACKDOWN32), 83 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64 84 : X86::ADJCALLSTACKUP32), 85 X86::CATCHRET, 86 (STI.is64Bit() ? X86::RETQ : X86::RETL)), 87 Subtarget(STI), RI(STI.getTargetTriple()) { 88 } 89 90 bool 91 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI, 92 unsigned &SrcReg, unsigned &DstReg, 93 unsigned &SubIdx) const { 94 switch (MI.getOpcode()) { 95 default: break; 96 case X86::MOVSX16rr8: 97 case X86::MOVZX16rr8: 98 case X86::MOVSX32rr8: 99 case X86::MOVZX32rr8: 100 case X86::MOVSX64rr8: 101 if (!Subtarget.is64Bit()) 102 // It's not always legal to reference the low 8-bit of the larger 103 // register in 32-bit mode. 104 return false; 105 LLVM_FALLTHROUGH; 106 case X86::MOVSX32rr16: 107 case X86::MOVZX32rr16: 108 case X86::MOVSX64rr16: 109 case X86::MOVSX64rr32: { 110 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg()) 111 // Be conservative. 112 return false; 113 SrcReg = MI.getOperand(1).getReg(); 114 DstReg = MI.getOperand(0).getReg(); 115 switch (MI.getOpcode()) { 116 default: llvm_unreachable("Unreachable!"); 117 case X86::MOVSX16rr8: 118 case X86::MOVZX16rr8: 119 case X86::MOVSX32rr8: 120 case X86::MOVZX32rr8: 121 case X86::MOVSX64rr8: 122 SubIdx = X86::sub_8bit; 123 break; 124 case X86::MOVSX32rr16: 125 case X86::MOVZX32rr16: 126 case X86::MOVSX64rr16: 127 SubIdx = X86::sub_16bit; 128 break; 129 case X86::MOVSX64rr32: 130 SubIdx = X86::sub_32bit; 131 break; 132 } 133 return true; 134 } 135 } 136 return false; 137 } 138 139 int X86InstrInfo::getSPAdjust(const MachineInstr &MI) const { 140 const MachineFunction *MF = MI.getParent()->getParent(); 141 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 142 143 if (isFrameInstr(MI)) { 144 unsigned StackAlign = TFI->getStackAlignment(); 145 int SPAdj = alignTo(getFrameSize(MI), StackAlign); 146 SPAdj -= getFrameAdjustment(MI); 147 if (!isFrameSetup(MI)) 148 SPAdj = -SPAdj; 149 return SPAdj; 150 } 151 152 // To know whether a call adjusts the stack, we need information 153 // that is bound to the following ADJCALLSTACKUP pseudo. 154 // Look for the next ADJCALLSTACKUP that follows the call. 155 if (MI.isCall()) { 156 const MachineBasicBlock *MBB = MI.getParent(); 157 auto I = ++MachineBasicBlock::const_iterator(MI); 158 for (auto E = MBB->end(); I != E; ++I) { 159 if (I->getOpcode() == getCallFrameDestroyOpcode() || 160 I->isCall()) 161 break; 162 } 163 164 // If we could not find a frame destroy opcode, then it has already 165 // been simplified, so we don't care. 166 if (I->getOpcode() != getCallFrameDestroyOpcode()) 167 return 0; 168 169 return -(I->getOperand(1).getImm()); 170 } 171 172 // Currently handle only PUSHes we can reasonably expect to see 173 // in call sequences 174 switch (MI.getOpcode()) { 175 default: 176 return 0; 177 case X86::PUSH32i8: 178 case X86::PUSH32r: 179 case X86::PUSH32rmm: 180 case X86::PUSH32rmr: 181 case X86::PUSHi32: 182 return 4; 183 case X86::PUSH64i8: 184 case X86::PUSH64r: 185 case X86::PUSH64rmm: 186 case X86::PUSH64rmr: 187 case X86::PUSH64i32: 188 return 8; 189 } 190 } 191 192 /// Return true and the FrameIndex if the specified 193 /// operand and follow operands form a reference to the stack frame. 194 bool X86InstrInfo::isFrameOperand(const MachineInstr &MI, unsigned int Op, 195 int &FrameIndex) const { 196 if (MI.getOperand(Op + X86::AddrBaseReg).isFI() && 197 MI.getOperand(Op + X86::AddrScaleAmt).isImm() && 198 MI.getOperand(Op + X86::AddrIndexReg).isReg() && 199 MI.getOperand(Op + X86::AddrDisp).isImm() && 200 MI.getOperand(Op + X86::AddrScaleAmt).getImm() == 1 && 201 MI.getOperand(Op + X86::AddrIndexReg).getReg() == 0 && 202 MI.getOperand(Op + X86::AddrDisp).getImm() == 0) { 203 FrameIndex = MI.getOperand(Op + X86::AddrBaseReg).getIndex(); 204 return true; 205 } 206 return false; 207 } 208 209 static bool isFrameLoadOpcode(int Opcode, unsigned &MemBytes) { 210 switch (Opcode) { 211 default: 212 return false; 213 case X86::MOV8rm: 214 case X86::KMOVBkm: 215 MemBytes = 1; 216 return true; 217 case X86::MOV16rm: 218 case X86::KMOVWkm: 219 MemBytes = 2; 220 return true; 221 case X86::MOV32rm: 222 case X86::MOVSSrm: 223 case X86::VMOVSSZrm: 224 case X86::VMOVSSrm: 225 case X86::KMOVDkm: 226 MemBytes = 4; 227 return true; 228 case X86::MOV64rm: 229 case X86::LD_Fp64m: 230 case X86::MOVSDrm: 231 case X86::VMOVSDrm: 232 case X86::VMOVSDZrm: 233 case X86::MMX_MOVD64rm: 234 case X86::MMX_MOVQ64rm: 235 case X86::KMOVQkm: 236 MemBytes = 8; 237 return true; 238 case X86::MOVAPSrm: 239 case X86::MOVUPSrm: 240 case X86::MOVAPDrm: 241 case X86::MOVUPDrm: 242 case X86::MOVDQArm: 243 case X86::MOVDQUrm: 244 case X86::VMOVAPSrm: 245 case X86::VMOVUPSrm: 246 case X86::VMOVAPDrm: 247 case X86::VMOVUPDrm: 248 case X86::VMOVDQArm: 249 case X86::VMOVDQUrm: 250 case X86::VMOVAPSZ128rm: 251 case X86::VMOVUPSZ128rm: 252 case X86::VMOVAPSZ128rm_NOVLX: 253 case X86::VMOVUPSZ128rm_NOVLX: 254 case X86::VMOVAPDZ128rm: 255 case X86::VMOVUPDZ128rm: 256 case X86::VMOVDQU8Z128rm: 257 case X86::VMOVDQU16Z128rm: 258 case X86::VMOVDQA32Z128rm: 259 case X86::VMOVDQU32Z128rm: 260 case X86::VMOVDQA64Z128rm: 261 case X86::VMOVDQU64Z128rm: 262 MemBytes = 16; 263 return true; 264 case X86::VMOVAPSYrm: 265 case X86::VMOVUPSYrm: 266 case X86::VMOVAPDYrm: 267 case X86::VMOVUPDYrm: 268 case X86::VMOVDQAYrm: 269 case X86::VMOVDQUYrm: 270 case X86::VMOVAPSZ256rm: 271 case X86::VMOVUPSZ256rm: 272 case X86::VMOVAPSZ256rm_NOVLX: 273 case X86::VMOVUPSZ256rm_NOVLX: 274 case X86::VMOVAPDZ256rm: 275 case X86::VMOVUPDZ256rm: 276 case X86::VMOVDQU8Z256rm: 277 case X86::VMOVDQU16Z256rm: 278 case X86::VMOVDQA32Z256rm: 279 case X86::VMOVDQU32Z256rm: 280 case X86::VMOVDQA64Z256rm: 281 case X86::VMOVDQU64Z256rm: 282 MemBytes = 32; 283 return true; 284 case X86::VMOVAPSZrm: 285 case X86::VMOVUPSZrm: 286 case X86::VMOVAPDZrm: 287 case X86::VMOVUPDZrm: 288 case X86::VMOVDQU8Zrm: 289 case X86::VMOVDQU16Zrm: 290 case X86::VMOVDQA32Zrm: 291 case X86::VMOVDQU32Zrm: 292 case X86::VMOVDQA64Zrm: 293 case X86::VMOVDQU64Zrm: 294 MemBytes = 64; 295 return true; 296 } 297 } 298 299 static bool isFrameStoreOpcode(int Opcode, unsigned &MemBytes) { 300 switch (Opcode) { 301 default: 302 return false; 303 case X86::MOV8mr: 304 case X86::KMOVBmk: 305 MemBytes = 1; 306 return true; 307 case X86::MOV16mr: 308 case X86::KMOVWmk: 309 MemBytes = 2; 310 return true; 311 case X86::MOV32mr: 312 case X86::MOVSSmr: 313 case X86::VMOVSSmr: 314 case X86::VMOVSSZmr: 315 case X86::KMOVDmk: 316 MemBytes = 4; 317 return true; 318 case X86::MOV64mr: 319 case X86::ST_FpP64m: 320 case X86::MOVSDmr: 321 case X86::VMOVSDmr: 322 case X86::VMOVSDZmr: 323 case X86::MMX_MOVD64mr: 324 case X86::MMX_MOVQ64mr: 325 case X86::MMX_MOVNTQmr: 326 case X86::KMOVQmk: 327 MemBytes = 8; 328 return true; 329 case X86::MOVAPSmr: 330 case X86::MOVUPSmr: 331 case X86::MOVAPDmr: 332 case X86::MOVUPDmr: 333 case X86::MOVDQAmr: 334 case X86::MOVDQUmr: 335 case X86::VMOVAPSmr: 336 case X86::VMOVUPSmr: 337 case X86::VMOVAPDmr: 338 case X86::VMOVUPDmr: 339 case X86::VMOVDQAmr: 340 case X86::VMOVDQUmr: 341 case X86::VMOVUPSZ128mr: 342 case X86::VMOVAPSZ128mr: 343 case X86::VMOVUPSZ128mr_NOVLX: 344 case X86::VMOVAPSZ128mr_NOVLX: 345 case X86::VMOVUPDZ128mr: 346 case X86::VMOVAPDZ128mr: 347 case X86::VMOVDQA32Z128mr: 348 case X86::VMOVDQU32Z128mr: 349 case X86::VMOVDQA64Z128mr: 350 case X86::VMOVDQU64Z128mr: 351 case X86::VMOVDQU8Z128mr: 352 case X86::VMOVDQU16Z128mr: 353 MemBytes = 16; 354 return true; 355 case X86::VMOVUPSYmr: 356 case X86::VMOVAPSYmr: 357 case X86::VMOVUPDYmr: 358 case X86::VMOVAPDYmr: 359 case X86::VMOVDQUYmr: 360 case X86::VMOVDQAYmr: 361 case X86::VMOVUPSZ256mr: 362 case X86::VMOVAPSZ256mr: 363 case X86::VMOVUPSZ256mr_NOVLX: 364 case X86::VMOVAPSZ256mr_NOVLX: 365 case X86::VMOVUPDZ256mr: 366 case X86::VMOVAPDZ256mr: 367 case X86::VMOVDQU8Z256mr: 368 case X86::VMOVDQU16Z256mr: 369 case X86::VMOVDQA32Z256mr: 370 case X86::VMOVDQU32Z256mr: 371 case X86::VMOVDQA64Z256mr: 372 case X86::VMOVDQU64Z256mr: 373 MemBytes = 32; 374 return true; 375 case X86::VMOVUPSZmr: 376 case X86::VMOVAPSZmr: 377 case X86::VMOVUPDZmr: 378 case X86::VMOVAPDZmr: 379 case X86::VMOVDQU8Zmr: 380 case X86::VMOVDQU16Zmr: 381 case X86::VMOVDQA32Zmr: 382 case X86::VMOVDQU32Zmr: 383 case X86::VMOVDQA64Zmr: 384 case X86::VMOVDQU64Zmr: 385 MemBytes = 64; 386 return true; 387 } 388 return false; 389 } 390 391 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 392 int &FrameIndex) const { 393 unsigned Dummy; 394 return X86InstrInfo::isLoadFromStackSlot(MI, FrameIndex, Dummy); 395 } 396 397 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 398 int &FrameIndex, 399 unsigned &MemBytes) const { 400 if (isFrameLoadOpcode(MI.getOpcode(), MemBytes)) 401 if (MI.getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex)) 402 return MI.getOperand(0).getReg(); 403 return 0; 404 } 405 406 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr &MI, 407 int &FrameIndex) const { 408 unsigned Dummy; 409 if (isFrameLoadOpcode(MI.getOpcode(), Dummy)) { 410 unsigned Reg; 411 if ((Reg = isLoadFromStackSlot(MI, FrameIndex))) 412 return Reg; 413 // Check for post-frame index elimination operations 414 SmallVector<const MachineMemOperand *, 1> Accesses; 415 if (hasLoadFromStackSlot(MI, Accesses)) { 416 FrameIndex = 417 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue()) 418 ->getFrameIndex(); 419 return 1; 420 } 421 } 422 return 0; 423 } 424 425 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI, 426 int &FrameIndex) const { 427 unsigned Dummy; 428 return X86InstrInfo::isStoreToStackSlot(MI, FrameIndex, Dummy); 429 } 430 431 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr &MI, 432 int &FrameIndex, 433 unsigned &MemBytes) const { 434 if (isFrameStoreOpcode(MI.getOpcode(), MemBytes)) 435 if (MI.getOperand(X86::AddrNumOperands).getSubReg() == 0 && 436 isFrameOperand(MI, 0, FrameIndex)) 437 return MI.getOperand(X86::AddrNumOperands).getReg(); 438 return 0; 439 } 440 441 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr &MI, 442 int &FrameIndex) const { 443 unsigned Dummy; 444 if (isFrameStoreOpcode(MI.getOpcode(), Dummy)) { 445 unsigned Reg; 446 if ((Reg = isStoreToStackSlot(MI, FrameIndex))) 447 return Reg; 448 // Check for post-frame index elimination operations 449 SmallVector<const MachineMemOperand *, 1> Accesses; 450 if (hasStoreToStackSlot(MI, Accesses)) { 451 FrameIndex = 452 cast<FixedStackPseudoSourceValue>(Accesses.front()->getPseudoValue()) 453 ->getFrameIndex(); 454 return 1; 455 } 456 } 457 return 0; 458 } 459 460 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r. 461 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) { 462 // Don't waste compile time scanning use-def chains of physregs. 463 if (!TargetRegisterInfo::isVirtualRegister(BaseReg)) 464 return false; 465 bool isPICBase = false; 466 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg), 467 E = MRI.def_instr_end(); I != E; ++I) { 468 MachineInstr *DefMI = &*I; 469 if (DefMI->getOpcode() != X86::MOVPC32r) 470 return false; 471 assert(!isPICBase && "More than one PIC base?"); 472 isPICBase = true; 473 } 474 return isPICBase; 475 } 476 477 bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI, 478 AliasAnalysis *AA) const { 479 switch (MI.getOpcode()) { 480 default: break; 481 case X86::MOV8rm: 482 case X86::MOV8rm_NOREX: 483 case X86::MOV16rm: 484 case X86::MOV32rm: 485 case X86::MOV64rm: 486 case X86::LD_Fp64m: 487 case X86::MOVSSrm: 488 case X86::MOVSDrm: 489 case X86::MOVAPSrm: 490 case X86::MOVUPSrm: 491 case X86::MOVAPDrm: 492 case X86::MOVUPDrm: 493 case X86::MOVDQArm: 494 case X86::MOVDQUrm: 495 case X86::VMOVSSrm: 496 case X86::VMOVSDrm: 497 case X86::VMOVAPSrm: 498 case X86::VMOVUPSrm: 499 case X86::VMOVAPDrm: 500 case X86::VMOVUPDrm: 501 case X86::VMOVDQArm: 502 case X86::VMOVDQUrm: 503 case X86::VMOVAPSYrm: 504 case X86::VMOVUPSYrm: 505 case X86::VMOVAPDYrm: 506 case X86::VMOVUPDYrm: 507 case X86::VMOVDQAYrm: 508 case X86::VMOVDQUYrm: 509 case X86::MMX_MOVD64rm: 510 case X86::MMX_MOVQ64rm: 511 // AVX-512 512 case X86::VMOVSSZrm: 513 case X86::VMOVSDZrm: 514 case X86::VMOVAPDZ128rm: 515 case X86::VMOVAPDZ256rm: 516 case X86::VMOVAPDZrm: 517 case X86::VMOVAPSZ128rm: 518 case X86::VMOVAPSZ256rm: 519 case X86::VMOVAPSZ128rm_NOVLX: 520 case X86::VMOVAPSZ256rm_NOVLX: 521 case X86::VMOVAPSZrm: 522 case X86::VMOVDQA32Z128rm: 523 case X86::VMOVDQA32Z256rm: 524 case X86::VMOVDQA32Zrm: 525 case X86::VMOVDQA64Z128rm: 526 case X86::VMOVDQA64Z256rm: 527 case X86::VMOVDQA64Zrm: 528 case X86::VMOVDQU16Z128rm: 529 case X86::VMOVDQU16Z256rm: 530 case X86::VMOVDQU16Zrm: 531 case X86::VMOVDQU32Z128rm: 532 case X86::VMOVDQU32Z256rm: 533 case X86::VMOVDQU32Zrm: 534 case X86::VMOVDQU64Z128rm: 535 case X86::VMOVDQU64Z256rm: 536 case X86::VMOVDQU64Zrm: 537 case X86::VMOVDQU8Z128rm: 538 case X86::VMOVDQU8Z256rm: 539 case X86::VMOVDQU8Zrm: 540 case X86::VMOVUPDZ128rm: 541 case X86::VMOVUPDZ256rm: 542 case X86::VMOVUPDZrm: 543 case X86::VMOVUPSZ128rm: 544 case X86::VMOVUPSZ256rm: 545 case X86::VMOVUPSZ128rm_NOVLX: 546 case X86::VMOVUPSZ256rm_NOVLX: 547 case X86::VMOVUPSZrm: { 548 // Loads from constant pools are trivially rematerializable. 549 if (MI.getOperand(1 + X86::AddrBaseReg).isReg() && 550 MI.getOperand(1 + X86::AddrScaleAmt).isImm() && 551 MI.getOperand(1 + X86::AddrIndexReg).isReg() && 552 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 && 553 MI.isDereferenceableInvariantLoad(AA)) { 554 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg(); 555 if (BaseReg == 0 || BaseReg == X86::RIP) 556 return true; 557 // Allow re-materialization of PIC load. 558 if (!ReMatPICStubLoad && MI.getOperand(1 + X86::AddrDisp).isGlobal()) 559 return false; 560 const MachineFunction &MF = *MI.getParent()->getParent(); 561 const MachineRegisterInfo &MRI = MF.getRegInfo(); 562 return regIsPICBase(BaseReg, MRI); 563 } 564 return false; 565 } 566 567 case X86::LEA32r: 568 case X86::LEA64r: { 569 if (MI.getOperand(1 + X86::AddrScaleAmt).isImm() && 570 MI.getOperand(1 + X86::AddrIndexReg).isReg() && 571 MI.getOperand(1 + X86::AddrIndexReg).getReg() == 0 && 572 !MI.getOperand(1 + X86::AddrDisp).isReg()) { 573 // lea fi#, lea GV, etc. are all rematerializable. 574 if (!MI.getOperand(1 + X86::AddrBaseReg).isReg()) 575 return true; 576 unsigned BaseReg = MI.getOperand(1 + X86::AddrBaseReg).getReg(); 577 if (BaseReg == 0) 578 return true; 579 // Allow re-materialization of lea PICBase + x. 580 const MachineFunction &MF = *MI.getParent()->getParent(); 581 const MachineRegisterInfo &MRI = MF.getRegInfo(); 582 return regIsPICBase(BaseReg, MRI); 583 } 584 return false; 585 } 586 } 587 588 // All other instructions marked M_REMATERIALIZABLE are always trivially 589 // rematerializable. 590 return true; 591 } 592 593 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB, 594 MachineBasicBlock::iterator I) const { 595 MachineBasicBlock::iterator E = MBB.end(); 596 597 // For compile time consideration, if we are not able to determine the 598 // safety after visiting 4 instructions in each direction, we will assume 599 // it's not safe. 600 MachineBasicBlock::iterator Iter = I; 601 for (unsigned i = 0; Iter != E && i < 4; ++i) { 602 bool SeenDef = false; 603 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 604 MachineOperand &MO = Iter->getOperand(j); 605 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 606 SeenDef = true; 607 if (!MO.isReg()) 608 continue; 609 if (MO.getReg() == X86::EFLAGS) { 610 if (MO.isUse()) 611 return false; 612 SeenDef = true; 613 } 614 } 615 616 if (SeenDef) 617 // This instruction defines EFLAGS, no need to look any further. 618 return true; 619 ++Iter; 620 // Skip over debug instructions. 621 while (Iter != E && Iter->isDebugInstr()) 622 ++Iter; 623 } 624 625 // It is safe to clobber EFLAGS at the end of a block of no successor has it 626 // live in. 627 if (Iter == E) { 628 for (MachineBasicBlock *S : MBB.successors()) 629 if (S->isLiveIn(X86::EFLAGS)) 630 return false; 631 return true; 632 } 633 634 MachineBasicBlock::iterator B = MBB.begin(); 635 Iter = I; 636 for (unsigned i = 0; i < 4; ++i) { 637 // If we make it to the beginning of the block, it's safe to clobber 638 // EFLAGS iff EFLAGS is not live-in. 639 if (Iter == B) 640 return !MBB.isLiveIn(X86::EFLAGS); 641 642 --Iter; 643 // Skip over debug instructions. 644 while (Iter != B && Iter->isDebugInstr()) 645 --Iter; 646 647 bool SawKill = false; 648 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 649 MachineOperand &MO = Iter->getOperand(j); 650 // A register mask may clobber EFLAGS, but we should still look for a 651 // live EFLAGS def. 652 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 653 SawKill = true; 654 if (MO.isReg() && MO.getReg() == X86::EFLAGS) { 655 if (MO.isDef()) return MO.isDead(); 656 if (MO.isKill()) SawKill = true; 657 } 658 } 659 660 if (SawKill) 661 // This instruction kills EFLAGS and doesn't redefine it, so 662 // there's no need to look further. 663 return true; 664 } 665 666 // Conservative answer. 667 return false; 668 } 669 670 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB, 671 MachineBasicBlock::iterator I, 672 unsigned DestReg, unsigned SubIdx, 673 const MachineInstr &Orig, 674 const TargetRegisterInfo &TRI) const { 675 bool ClobbersEFLAGS = false; 676 for (const MachineOperand &MO : Orig.operands()) { 677 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) { 678 ClobbersEFLAGS = true; 679 break; 680 } 681 } 682 683 if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) { 684 // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side 685 // effects. 686 unsigned NewOpc = X86::MOV32ri; 687 int Value; 688 switch (Orig.getOpcode()) { 689 case X86::MOV64r0: NewOpc = X86::MOV32ri64; Value = 0; break; 690 case X86::MOV32r0: Value = 0; break; 691 case X86::MOV32r1: Value = 1; break; 692 case X86::MOV32r_1: Value = -1; break; 693 default: 694 llvm_unreachable("Unexpected instruction!"); 695 } 696 697 const DebugLoc &DL = Orig.getDebugLoc(); 698 BuildMI(MBB, I, DL, get(NewOpc)) 699 .add(Orig.getOperand(0)) 700 .addImm(Value); 701 } else { 702 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig); 703 MBB.insert(I, MI); 704 } 705 706 MachineInstr &NewMI = *std::prev(I); 707 NewMI.substituteRegister(Orig.getOperand(0).getReg(), DestReg, SubIdx, TRI); 708 } 709 710 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead. 711 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr &MI) const { 712 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 713 MachineOperand &MO = MI.getOperand(i); 714 if (MO.isReg() && MO.isDef() && 715 MO.getReg() == X86::EFLAGS && !MO.isDead()) { 716 return true; 717 } 718 } 719 return false; 720 } 721 722 /// Check whether the shift count for a machine operand is non-zero. 723 inline static unsigned getTruncatedShiftCount(MachineInstr &MI, 724 unsigned ShiftAmtOperandIdx) { 725 // The shift count is six bits with the REX.W prefix and five bits without. 726 unsigned ShiftCountMask = (MI.getDesc().TSFlags & X86II::REX_W) ? 63 : 31; 727 unsigned Imm = MI.getOperand(ShiftAmtOperandIdx).getImm(); 728 return Imm & ShiftCountMask; 729 } 730 731 /// Check whether the given shift count is appropriate 732 /// can be represented by a LEA instruction. 733 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) { 734 // Left shift instructions can be transformed into load-effective-address 735 // instructions if we can encode them appropriately. 736 // A LEA instruction utilizes a SIB byte to encode its scale factor. 737 // The SIB.scale field is two bits wide which means that we can encode any 738 // shift amount less than 4. 739 return ShAmt < 4 && ShAmt > 0; 740 } 741 742 bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const MachineOperand &Src, 743 unsigned Opc, bool AllowSP, unsigned &NewSrc, 744 bool &isKill, bool &isUndef, 745 MachineOperand &ImplicitOp, 746 LiveVariables *LV) const { 747 MachineFunction &MF = *MI.getParent()->getParent(); 748 const TargetRegisterClass *RC; 749 if (AllowSP) { 750 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass; 751 } else { 752 RC = Opc != X86::LEA32r ? 753 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass; 754 } 755 unsigned SrcReg = Src.getReg(); 756 757 // For both LEA64 and LEA32 the register already has essentially the right 758 // type (32-bit or 64-bit) we may just need to forbid SP. 759 if (Opc != X86::LEA64_32r) { 760 NewSrc = SrcReg; 761 isKill = Src.isKill(); 762 isUndef = Src.isUndef(); 763 764 if (TargetRegisterInfo::isVirtualRegister(NewSrc) && 765 !MF.getRegInfo().constrainRegClass(NewSrc, RC)) 766 return false; 767 768 return true; 769 } 770 771 // This is for an LEA64_32r and incoming registers are 32-bit. One way or 772 // another we need to add 64-bit registers to the final MI. 773 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) { 774 ImplicitOp = Src; 775 ImplicitOp.setImplicit(); 776 777 NewSrc = getX86SubSuperRegister(Src.getReg(), 64); 778 isKill = Src.isKill(); 779 isUndef = Src.isUndef(); 780 } else { 781 // Virtual register of the wrong class, we have to create a temporary 64-bit 782 // vreg to feed into the LEA. 783 NewSrc = MF.getRegInfo().createVirtualRegister(RC); 784 MachineInstr *Copy = 785 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(TargetOpcode::COPY)) 786 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit) 787 .add(Src); 788 789 // Which is obviously going to be dead after we're done with it. 790 isKill = true; 791 isUndef = false; 792 793 if (LV) 794 LV->replaceKillInstruction(SrcReg, MI, *Copy); 795 } 796 797 // We've set all the parameters without issue. 798 return true; 799 } 800 801 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit 802 /// LEA to form 3-address code by promoting to a 32-bit superregister and then 803 /// truncating back down to a 16-bit subregister. 804 MachineInstr *X86InstrInfo::convertToThreeAddressWithLEA( 805 unsigned MIOpc, MachineFunction::iterator &MFI, MachineInstr &MI, 806 LiveVariables *LV) const { 807 MachineBasicBlock::iterator MBBI = MI.getIterator(); 808 unsigned Dest = MI.getOperand(0).getReg(); 809 unsigned Src = MI.getOperand(1).getReg(); 810 bool isDead = MI.getOperand(0).isDead(); 811 bool isKill = MI.getOperand(1).isKill(); 812 813 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo(); 814 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass); 815 unsigned Opc, leaInReg; 816 if (Subtarget.is64Bit()) { 817 Opc = X86::LEA64_32r; 818 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass); 819 } else { 820 Opc = X86::LEA32r; 821 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 822 } 823 824 // Build and insert into an implicit UNDEF value. This is OK because 825 // well be shifting and then extracting the lower 16-bits. 826 // This has the potential to cause partial register stall. e.g. 827 // movw (%rbp,%rcx,2), %dx 828 // leal -65(%rdx), %esi 829 // But testing has shown this *does* help performance in 64-bit mode (at 830 // least on modern x86 machines). 831 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg); 832 MachineInstr *InsMI = 833 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY)) 834 .addReg(leaInReg, RegState::Define, X86::sub_16bit) 835 .addReg(Src, getKillRegState(isKill)); 836 837 MachineInstrBuilder MIB = 838 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(Opc), leaOutReg); 839 switch (MIOpc) { 840 default: llvm_unreachable("Unreachable!"); 841 case X86::SHL16ri: { 842 unsigned ShAmt = MI.getOperand(2).getImm(); 843 MIB.addReg(0).addImm(1ULL << ShAmt) 844 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0); 845 break; 846 } 847 case X86::INC16r: 848 addRegOffset(MIB, leaInReg, true, 1); 849 break; 850 case X86::DEC16r: 851 addRegOffset(MIB, leaInReg, true, -1); 852 break; 853 case X86::ADD16ri: 854 case X86::ADD16ri8: 855 case X86::ADD16ri_DB: 856 case X86::ADD16ri8_DB: 857 addRegOffset(MIB, leaInReg, true, MI.getOperand(2).getImm()); 858 break; 859 case X86::ADD16rr: 860 case X86::ADD16rr_DB: { 861 unsigned Src2 = MI.getOperand(2).getReg(); 862 bool isKill2 = MI.getOperand(2).isKill(); 863 unsigned leaInReg2 = 0; 864 MachineInstr *InsMI2 = nullptr; 865 if (Src == Src2) { 866 // ADD16rr killed %reg1028, %reg1028 867 // just a single insert_subreg. 868 addRegReg(MIB, leaInReg, true, leaInReg, false); 869 } else { 870 if (Subtarget.is64Bit()) 871 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass); 872 else 873 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 874 // Build and insert into an implicit UNDEF value. This is OK because 875 // well be shifting and then extracting the lower 16-bits. 876 BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg2); 877 InsMI2 = BuildMI(*MFI, &*MIB, MI.getDebugLoc(), get(TargetOpcode::COPY)) 878 .addReg(leaInReg2, RegState::Define, X86::sub_16bit) 879 .addReg(Src2, getKillRegState(isKill2)); 880 addRegReg(MIB, leaInReg, true, leaInReg2, true); 881 } 882 if (LV && isKill2 && InsMI2) 883 LV->replaceKillInstruction(Src2, MI, *InsMI2); 884 break; 885 } 886 } 887 888 MachineInstr *NewMI = MIB; 889 MachineInstr *ExtMI = 890 BuildMI(*MFI, MBBI, MI.getDebugLoc(), get(TargetOpcode::COPY)) 891 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 892 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit); 893 894 if (LV) { 895 // Update live variables 896 LV->getVarInfo(leaInReg).Kills.push_back(NewMI); 897 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI); 898 if (isKill) 899 LV->replaceKillInstruction(Src, MI, *InsMI); 900 if (isDead) 901 LV->replaceKillInstruction(Dest, MI, *ExtMI); 902 } 903 904 return ExtMI; 905 } 906 907 /// This method must be implemented by targets that 908 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target 909 /// may be able to convert a two-address instruction into a true 910 /// three-address instruction on demand. This allows the X86 target (for 911 /// example) to convert ADD and SHL instructions into LEA instructions if they 912 /// would require register copies due to two-addressness. 913 /// 914 /// This method returns a null pointer if the transformation cannot be 915 /// performed, otherwise it returns the new instruction. 916 /// 917 MachineInstr * 918 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI, 919 MachineInstr &MI, LiveVariables *LV) const { 920 // The following opcodes also sets the condition code register(s). Only 921 // convert them to equivalent lea if the condition code register def's 922 // are dead! 923 if (hasLiveCondCodeDef(MI)) 924 return nullptr; 925 926 MachineFunction &MF = *MI.getParent()->getParent(); 927 // All instructions input are two-addr instructions. Get the known operands. 928 const MachineOperand &Dest = MI.getOperand(0); 929 const MachineOperand &Src = MI.getOperand(1); 930 931 MachineInstr *NewMI = nullptr; 932 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When 933 // we have better subtarget support, enable the 16-bit LEA generation here. 934 // 16-bit LEA is also slow on Core2. 935 bool DisableLEA16 = true; 936 bool is64Bit = Subtarget.is64Bit(); 937 938 unsigned MIOpc = MI.getOpcode(); 939 switch (MIOpc) { 940 default: return nullptr; 941 case X86::SHL64ri: { 942 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!"); 943 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 944 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 945 946 // LEA can't handle RSP. 947 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) && 948 !MF.getRegInfo().constrainRegClass(Src.getReg(), 949 &X86::GR64_NOSPRegClass)) 950 return nullptr; 951 952 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)) 953 .add(Dest) 954 .addReg(0) 955 .addImm(1ULL << ShAmt) 956 .add(Src) 957 .addImm(0) 958 .addReg(0); 959 break; 960 } 961 case X86::SHL32ri: { 962 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!"); 963 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 964 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 965 966 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 967 968 // LEA can't handle ESP. 969 bool isKill, isUndef; 970 unsigned SrcReg; 971 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 972 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 973 SrcReg, isKill, isUndef, ImplicitOp, LV)) 974 return nullptr; 975 976 MachineInstrBuilder MIB = 977 BuildMI(MF, MI.getDebugLoc(), get(Opc)) 978 .add(Dest) 979 .addReg(0) 980 .addImm(1ULL << ShAmt) 981 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef)) 982 .addImm(0) 983 .addReg(0); 984 if (ImplicitOp.getReg() != 0) 985 MIB.add(ImplicitOp); 986 NewMI = MIB; 987 988 break; 989 } 990 case X86::SHL16ri: { 991 assert(MI.getNumOperands() >= 3 && "Unknown shift instruction!"); 992 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 993 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 994 995 if (DisableLEA16) 996 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV) 997 : nullptr; 998 NewMI = BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)) 999 .add(Dest) 1000 .addReg(0) 1001 .addImm(1ULL << ShAmt) 1002 .add(Src) 1003 .addImm(0) 1004 .addReg(0); 1005 break; 1006 } 1007 case X86::INC64r: 1008 case X86::INC32r: { 1009 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!"); 1010 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r 1011 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 1012 bool isKill, isUndef; 1013 unsigned SrcReg; 1014 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 1015 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 1016 SrcReg, isKill, isUndef, ImplicitOp, LV)) 1017 return nullptr; 1018 1019 MachineInstrBuilder MIB = 1020 BuildMI(MF, MI.getDebugLoc(), get(Opc)) 1021 .add(Dest) 1022 .addReg(SrcReg, 1023 getKillRegState(isKill) | getUndefRegState(isUndef)); 1024 if (ImplicitOp.getReg() != 0) 1025 MIB.add(ImplicitOp); 1026 1027 NewMI = addOffset(MIB, 1); 1028 break; 1029 } 1030 case X86::INC16r: 1031 if (DisableLEA16) 1032 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV) 1033 : nullptr; 1034 assert(MI.getNumOperands() >= 2 && "Unknown inc instruction!"); 1035 NewMI = addOffset( 1036 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src), 1); 1037 break; 1038 case X86::DEC64r: 1039 case X86::DEC32r: { 1040 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!"); 1041 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r 1042 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 1043 1044 bool isKill, isUndef; 1045 unsigned SrcReg; 1046 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 1047 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 1048 SrcReg, isKill, isUndef, ImplicitOp, LV)) 1049 return nullptr; 1050 1051 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)) 1052 .add(Dest) 1053 .addReg(SrcReg, getUndefRegState(isUndef) | 1054 getKillRegState(isKill)); 1055 if (ImplicitOp.getReg() != 0) 1056 MIB.add(ImplicitOp); 1057 1058 NewMI = addOffset(MIB, -1); 1059 1060 break; 1061 } 1062 case X86::DEC16r: 1063 if (DisableLEA16) 1064 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV) 1065 : nullptr; 1066 assert(MI.getNumOperands() >= 2 && "Unknown dec instruction!"); 1067 NewMI = addOffset( 1068 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src), -1); 1069 break; 1070 case X86::ADD64rr: 1071 case X86::ADD64rr_DB: 1072 case X86::ADD32rr: 1073 case X86::ADD32rr_DB: { 1074 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!"); 1075 unsigned Opc; 1076 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) 1077 Opc = X86::LEA64r; 1078 else 1079 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 1080 1081 bool isKill, isUndef; 1082 unsigned SrcReg; 1083 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 1084 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true, 1085 SrcReg, isKill, isUndef, ImplicitOp, LV)) 1086 return nullptr; 1087 1088 const MachineOperand &Src2 = MI.getOperand(2); 1089 bool isKill2, isUndef2; 1090 unsigned SrcReg2; 1091 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false); 1092 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false, 1093 SrcReg2, isKill2, isUndef2, ImplicitOp2, LV)) 1094 return nullptr; 1095 1096 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)).add(Dest); 1097 if (ImplicitOp.getReg() != 0) 1098 MIB.add(ImplicitOp); 1099 if (ImplicitOp2.getReg() != 0) 1100 MIB.add(ImplicitOp2); 1101 1102 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2); 1103 1104 // Preserve undefness of the operands. 1105 NewMI->getOperand(1).setIsUndef(isUndef); 1106 NewMI->getOperand(3).setIsUndef(isUndef2); 1107 1108 if (LV && Src2.isKill()) 1109 LV->replaceKillInstruction(SrcReg2, MI, *NewMI); 1110 break; 1111 } 1112 case X86::ADD16rr: 1113 case X86::ADD16rr_DB: { 1114 if (DisableLEA16) 1115 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV) 1116 : nullptr; 1117 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!"); 1118 unsigned Src2 = MI.getOperand(2).getReg(); 1119 bool isKill2 = MI.getOperand(2).isKill(); 1120 NewMI = addRegReg(BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest), 1121 Src.getReg(), Src.isKill(), Src2, isKill2); 1122 1123 // Preserve undefness of the operands. 1124 bool isUndef = MI.getOperand(1).isUndef(); 1125 bool isUndef2 = MI.getOperand(2).isUndef(); 1126 NewMI->getOperand(1).setIsUndef(isUndef); 1127 NewMI->getOperand(3).setIsUndef(isUndef2); 1128 1129 if (LV && isKill2) 1130 LV->replaceKillInstruction(Src2, MI, *NewMI); 1131 break; 1132 } 1133 case X86::ADD64ri32: 1134 case X86::ADD64ri8: 1135 case X86::ADD64ri32_DB: 1136 case X86::ADD64ri8_DB: 1137 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!"); 1138 NewMI = addOffset( 1139 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA64r)).add(Dest).add(Src), 1140 MI.getOperand(2)); 1141 break; 1142 case X86::ADD32ri: 1143 case X86::ADD32ri8: 1144 case X86::ADD32ri_DB: 1145 case X86::ADD32ri8_DB: { 1146 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!"); 1147 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 1148 1149 bool isKill, isUndef; 1150 unsigned SrcReg; 1151 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 1152 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true, 1153 SrcReg, isKill, isUndef, ImplicitOp, LV)) 1154 return nullptr; 1155 1156 MachineInstrBuilder MIB = BuildMI(MF, MI.getDebugLoc(), get(Opc)) 1157 .add(Dest) 1158 .addReg(SrcReg, getUndefRegState(isUndef) | 1159 getKillRegState(isKill)); 1160 if (ImplicitOp.getReg() != 0) 1161 MIB.add(ImplicitOp); 1162 1163 NewMI = addOffset(MIB, MI.getOperand(2)); 1164 break; 1165 } 1166 case X86::ADD16ri: 1167 case X86::ADD16ri8: 1168 case X86::ADD16ri_DB: 1169 case X86::ADD16ri8_DB: 1170 if (DisableLEA16) 1171 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MI, LV) 1172 : nullptr; 1173 assert(MI.getNumOperands() >= 3 && "Unknown add instruction!"); 1174 NewMI = addOffset( 1175 BuildMI(MF, MI.getDebugLoc(), get(X86::LEA16r)).add(Dest).add(Src), 1176 MI.getOperand(2)); 1177 break; 1178 1179 case X86::VMOVDQU8Z128rmk: 1180 case X86::VMOVDQU8Z256rmk: 1181 case X86::VMOVDQU8Zrmk: 1182 case X86::VMOVDQU16Z128rmk: 1183 case X86::VMOVDQU16Z256rmk: 1184 case X86::VMOVDQU16Zrmk: 1185 case X86::VMOVDQU32Z128rmk: case X86::VMOVDQA32Z128rmk: 1186 case X86::VMOVDQU32Z256rmk: case X86::VMOVDQA32Z256rmk: 1187 case X86::VMOVDQU32Zrmk: case X86::VMOVDQA32Zrmk: 1188 case X86::VMOVDQU64Z128rmk: case X86::VMOVDQA64Z128rmk: 1189 case X86::VMOVDQU64Z256rmk: case X86::VMOVDQA64Z256rmk: 1190 case X86::VMOVDQU64Zrmk: case X86::VMOVDQA64Zrmk: 1191 case X86::VMOVUPDZ128rmk: case X86::VMOVAPDZ128rmk: 1192 case X86::VMOVUPDZ256rmk: case X86::VMOVAPDZ256rmk: 1193 case X86::VMOVUPDZrmk: case X86::VMOVAPDZrmk: 1194 case X86::VMOVUPSZ128rmk: case X86::VMOVAPSZ128rmk: 1195 case X86::VMOVUPSZ256rmk: case X86::VMOVAPSZ256rmk: 1196 case X86::VMOVUPSZrmk: case X86::VMOVAPSZrmk: { 1197 unsigned Opc; 1198 switch (MIOpc) { 1199 default: llvm_unreachable("Unreachable!"); 1200 case X86::VMOVDQU8Z128rmk: Opc = X86::VPBLENDMBZ128rmk; break; 1201 case X86::VMOVDQU8Z256rmk: Opc = X86::VPBLENDMBZ256rmk; break; 1202 case X86::VMOVDQU8Zrmk: Opc = X86::VPBLENDMBZrmk; break; 1203 case X86::VMOVDQU16Z128rmk: Opc = X86::VPBLENDMWZ128rmk; break; 1204 case X86::VMOVDQU16Z256rmk: Opc = X86::VPBLENDMWZ256rmk; break; 1205 case X86::VMOVDQU16Zrmk: Opc = X86::VPBLENDMWZrmk; break; 1206 case X86::VMOVDQU32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break; 1207 case X86::VMOVDQU32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break; 1208 case X86::VMOVDQU32Zrmk: Opc = X86::VPBLENDMDZrmk; break; 1209 case X86::VMOVDQU64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break; 1210 case X86::VMOVDQU64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break; 1211 case X86::VMOVDQU64Zrmk: Opc = X86::VPBLENDMQZrmk; break; 1212 case X86::VMOVUPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break; 1213 case X86::VMOVUPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break; 1214 case X86::VMOVUPDZrmk: Opc = X86::VBLENDMPDZrmk; break; 1215 case X86::VMOVUPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break; 1216 case X86::VMOVUPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break; 1217 case X86::VMOVUPSZrmk: Opc = X86::VBLENDMPSZrmk; break; 1218 case X86::VMOVDQA32Z128rmk: Opc = X86::VPBLENDMDZ128rmk; break; 1219 case X86::VMOVDQA32Z256rmk: Opc = X86::VPBLENDMDZ256rmk; break; 1220 case X86::VMOVDQA32Zrmk: Opc = X86::VPBLENDMDZrmk; break; 1221 case X86::VMOVDQA64Z128rmk: Opc = X86::VPBLENDMQZ128rmk; break; 1222 case X86::VMOVDQA64Z256rmk: Opc = X86::VPBLENDMQZ256rmk; break; 1223 case X86::VMOVDQA64Zrmk: Opc = X86::VPBLENDMQZrmk; break; 1224 case X86::VMOVAPDZ128rmk: Opc = X86::VBLENDMPDZ128rmk; break; 1225 case X86::VMOVAPDZ256rmk: Opc = X86::VBLENDMPDZ256rmk; break; 1226 case X86::VMOVAPDZrmk: Opc = X86::VBLENDMPDZrmk; break; 1227 case X86::VMOVAPSZ128rmk: Opc = X86::VBLENDMPSZ128rmk; break; 1228 case X86::VMOVAPSZ256rmk: Opc = X86::VBLENDMPSZ256rmk; break; 1229 case X86::VMOVAPSZrmk: Opc = X86::VBLENDMPSZrmk; break; 1230 } 1231 1232 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc)) 1233 .add(Dest) 1234 .add(MI.getOperand(2)) 1235 .add(Src) 1236 .add(MI.getOperand(3)) 1237 .add(MI.getOperand(4)) 1238 .add(MI.getOperand(5)) 1239 .add(MI.getOperand(6)) 1240 .add(MI.getOperand(7)); 1241 break; 1242 } 1243 case X86::VMOVDQU8Z128rrk: 1244 case X86::VMOVDQU8Z256rrk: 1245 case X86::VMOVDQU8Zrrk: 1246 case X86::VMOVDQU16Z128rrk: 1247 case X86::VMOVDQU16Z256rrk: 1248 case X86::VMOVDQU16Zrrk: 1249 case X86::VMOVDQU32Z128rrk: case X86::VMOVDQA32Z128rrk: 1250 case X86::VMOVDQU32Z256rrk: case X86::VMOVDQA32Z256rrk: 1251 case X86::VMOVDQU32Zrrk: case X86::VMOVDQA32Zrrk: 1252 case X86::VMOVDQU64Z128rrk: case X86::VMOVDQA64Z128rrk: 1253 case X86::VMOVDQU64Z256rrk: case X86::VMOVDQA64Z256rrk: 1254 case X86::VMOVDQU64Zrrk: case X86::VMOVDQA64Zrrk: 1255 case X86::VMOVUPDZ128rrk: case X86::VMOVAPDZ128rrk: 1256 case X86::VMOVUPDZ256rrk: case X86::VMOVAPDZ256rrk: 1257 case X86::VMOVUPDZrrk: case X86::VMOVAPDZrrk: 1258 case X86::VMOVUPSZ128rrk: case X86::VMOVAPSZ128rrk: 1259 case X86::VMOVUPSZ256rrk: case X86::VMOVAPSZ256rrk: 1260 case X86::VMOVUPSZrrk: case X86::VMOVAPSZrrk: { 1261 unsigned Opc; 1262 switch (MIOpc) { 1263 default: llvm_unreachable("Unreachable!"); 1264 case X86::VMOVDQU8Z128rrk: Opc = X86::VPBLENDMBZ128rrk; break; 1265 case X86::VMOVDQU8Z256rrk: Opc = X86::VPBLENDMBZ256rrk; break; 1266 case X86::VMOVDQU8Zrrk: Opc = X86::VPBLENDMBZrrk; break; 1267 case X86::VMOVDQU16Z128rrk: Opc = X86::VPBLENDMWZ128rrk; break; 1268 case X86::VMOVDQU16Z256rrk: Opc = X86::VPBLENDMWZ256rrk; break; 1269 case X86::VMOVDQU16Zrrk: Opc = X86::VPBLENDMWZrrk; break; 1270 case X86::VMOVDQU32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break; 1271 case X86::VMOVDQU32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break; 1272 case X86::VMOVDQU32Zrrk: Opc = X86::VPBLENDMDZrrk; break; 1273 case X86::VMOVDQU64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break; 1274 case X86::VMOVDQU64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break; 1275 case X86::VMOVDQU64Zrrk: Opc = X86::VPBLENDMQZrrk; break; 1276 case X86::VMOVUPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break; 1277 case X86::VMOVUPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break; 1278 case X86::VMOVUPDZrrk: Opc = X86::VBLENDMPDZrrk; break; 1279 case X86::VMOVUPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break; 1280 case X86::VMOVUPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break; 1281 case X86::VMOVUPSZrrk: Opc = X86::VBLENDMPSZrrk; break; 1282 case X86::VMOVDQA32Z128rrk: Opc = X86::VPBLENDMDZ128rrk; break; 1283 case X86::VMOVDQA32Z256rrk: Opc = X86::VPBLENDMDZ256rrk; break; 1284 case X86::VMOVDQA32Zrrk: Opc = X86::VPBLENDMDZrrk; break; 1285 case X86::VMOVDQA64Z128rrk: Opc = X86::VPBLENDMQZ128rrk; break; 1286 case X86::VMOVDQA64Z256rrk: Opc = X86::VPBLENDMQZ256rrk; break; 1287 case X86::VMOVDQA64Zrrk: Opc = X86::VPBLENDMQZrrk; break; 1288 case X86::VMOVAPDZ128rrk: Opc = X86::VBLENDMPDZ128rrk; break; 1289 case X86::VMOVAPDZ256rrk: Opc = X86::VBLENDMPDZ256rrk; break; 1290 case X86::VMOVAPDZrrk: Opc = X86::VBLENDMPDZrrk; break; 1291 case X86::VMOVAPSZ128rrk: Opc = X86::VBLENDMPSZ128rrk; break; 1292 case X86::VMOVAPSZ256rrk: Opc = X86::VBLENDMPSZ256rrk; break; 1293 case X86::VMOVAPSZrrk: Opc = X86::VBLENDMPSZrrk; break; 1294 } 1295 1296 NewMI = BuildMI(MF, MI.getDebugLoc(), get(Opc)) 1297 .add(Dest) 1298 .add(MI.getOperand(2)) 1299 .add(Src) 1300 .add(MI.getOperand(3)); 1301 break; 1302 } 1303 } 1304 1305 if (!NewMI) return nullptr; 1306 1307 if (LV) { // Update live variables 1308 if (Src.isKill()) 1309 LV->replaceKillInstruction(Src.getReg(), MI, *NewMI); 1310 if (Dest.isDead()) 1311 LV->replaceKillInstruction(Dest.getReg(), MI, *NewMI); 1312 } 1313 1314 MFI->insert(MI.getIterator(), NewMI); // Insert the new inst 1315 return NewMI; 1316 } 1317 1318 /// This determines which of three possible cases of a three source commute 1319 /// the source indexes correspond to taking into account any mask operands. 1320 /// All prevents commuting a passthru operand. Returns -1 if the commute isn't 1321 /// possible. 1322 /// Case 0 - Possible to commute the first and second operands. 1323 /// Case 1 - Possible to commute the first and third operands. 1324 /// Case 2 - Possible to commute the second and third operands. 1325 static unsigned getThreeSrcCommuteCase(uint64_t TSFlags, unsigned SrcOpIdx1, 1326 unsigned SrcOpIdx2) { 1327 // Put the lowest index to SrcOpIdx1 to simplify the checks below. 1328 if (SrcOpIdx1 > SrcOpIdx2) 1329 std::swap(SrcOpIdx1, SrcOpIdx2); 1330 1331 unsigned Op1 = 1, Op2 = 2, Op3 = 3; 1332 if (X86II::isKMasked(TSFlags)) { 1333 Op2++; 1334 Op3++; 1335 } 1336 1337 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op2) 1338 return 0; 1339 if (SrcOpIdx1 == Op1 && SrcOpIdx2 == Op3) 1340 return 1; 1341 if (SrcOpIdx1 == Op2 && SrcOpIdx2 == Op3) 1342 return 2; 1343 llvm_unreachable("Unknown three src commute case."); 1344 } 1345 1346 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands( 1347 const MachineInstr &MI, unsigned SrcOpIdx1, unsigned SrcOpIdx2, 1348 const X86InstrFMA3Group &FMA3Group) const { 1349 1350 unsigned Opc = MI.getOpcode(); 1351 1352 // TODO: Commuting the 1st operand of FMA*_Int requires some additional 1353 // analysis. The commute optimization is legal only if all users of FMA*_Int 1354 // use only the lowest element of the FMA*_Int instruction. Such analysis are 1355 // not implemented yet. So, just return 0 in that case. 1356 // When such analysis are available this place will be the right place for 1357 // calling it. 1358 assert(!(FMA3Group.isIntrinsic() && (SrcOpIdx1 == 1 || SrcOpIdx2 == 1)) && 1359 "Intrinsic instructions can't commute operand 1"); 1360 1361 // Determine which case this commute is or if it can't be done. 1362 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1, 1363 SrcOpIdx2); 1364 assert(Case < 3 && "Unexpected case number!"); 1365 1366 // Define the FMA forms mapping array that helps to map input FMA form 1367 // to output FMA form to preserve the operation semantics after 1368 // commuting the operands. 1369 const unsigned Form132Index = 0; 1370 const unsigned Form213Index = 1; 1371 const unsigned Form231Index = 2; 1372 static const unsigned FormMapping[][3] = { 1373 // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2; 1374 // FMA132 A, C, b; ==> FMA231 C, A, b; 1375 // FMA213 B, A, c; ==> FMA213 A, B, c; 1376 // FMA231 C, A, b; ==> FMA132 A, C, b; 1377 { Form231Index, Form213Index, Form132Index }, 1378 // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3; 1379 // FMA132 A, c, B; ==> FMA132 B, c, A; 1380 // FMA213 B, a, C; ==> FMA231 C, a, B; 1381 // FMA231 C, a, B; ==> FMA213 B, a, C; 1382 { Form132Index, Form231Index, Form213Index }, 1383 // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3; 1384 // FMA132 a, C, B; ==> FMA213 a, B, C; 1385 // FMA213 b, A, C; ==> FMA132 b, C, A; 1386 // FMA231 c, A, B; ==> FMA231 c, B, A; 1387 { Form213Index, Form132Index, Form231Index } 1388 }; 1389 1390 unsigned FMAForms[3]; 1391 FMAForms[0] = FMA3Group.get132Opcode(); 1392 FMAForms[1] = FMA3Group.get213Opcode(); 1393 FMAForms[2] = FMA3Group.get231Opcode(); 1394 unsigned FormIndex; 1395 for (FormIndex = 0; FormIndex < 3; FormIndex++) 1396 if (Opc == FMAForms[FormIndex]) 1397 break; 1398 1399 // Everything is ready, just adjust the FMA opcode and return it. 1400 FormIndex = FormMapping[Case][FormIndex]; 1401 return FMAForms[FormIndex]; 1402 } 1403 1404 static void commuteVPTERNLOG(MachineInstr &MI, unsigned SrcOpIdx1, 1405 unsigned SrcOpIdx2) { 1406 // Determine which case this commute is or if it can't be done. 1407 unsigned Case = getThreeSrcCommuteCase(MI.getDesc().TSFlags, SrcOpIdx1, 1408 SrcOpIdx2); 1409 assert(Case < 3 && "Unexpected case value!"); 1410 1411 // For each case we need to swap two pairs of bits in the final immediate. 1412 static const uint8_t SwapMasks[3][4] = { 1413 { 0x04, 0x10, 0x08, 0x20 }, // Swap bits 2/4 and 3/5. 1414 { 0x02, 0x10, 0x08, 0x40 }, // Swap bits 1/4 and 3/6. 1415 { 0x02, 0x04, 0x20, 0x40 }, // Swap bits 1/2 and 5/6. 1416 }; 1417 1418 uint8_t Imm = MI.getOperand(MI.getNumOperands()-1).getImm(); 1419 // Clear out the bits we are swapping. 1420 uint8_t NewImm = Imm & ~(SwapMasks[Case][0] | SwapMasks[Case][1] | 1421 SwapMasks[Case][2] | SwapMasks[Case][3]); 1422 // If the immediate had a bit of the pair set, then set the opposite bit. 1423 if (Imm & SwapMasks[Case][0]) NewImm |= SwapMasks[Case][1]; 1424 if (Imm & SwapMasks[Case][1]) NewImm |= SwapMasks[Case][0]; 1425 if (Imm & SwapMasks[Case][2]) NewImm |= SwapMasks[Case][3]; 1426 if (Imm & SwapMasks[Case][3]) NewImm |= SwapMasks[Case][2]; 1427 MI.getOperand(MI.getNumOperands()-1).setImm(NewImm); 1428 } 1429 1430 // Returns true if this is a VPERMI2 or VPERMT2 instruction that can be 1431 // commuted. 1432 static bool isCommutableVPERMV3Instruction(unsigned Opcode) { 1433 #define VPERM_CASES(Suffix) \ 1434 case X86::VPERMI2##Suffix##128rr: case X86::VPERMT2##Suffix##128rr: \ 1435 case X86::VPERMI2##Suffix##256rr: case X86::VPERMT2##Suffix##256rr: \ 1436 case X86::VPERMI2##Suffix##rr: case X86::VPERMT2##Suffix##rr: \ 1437 case X86::VPERMI2##Suffix##128rm: case X86::VPERMT2##Suffix##128rm: \ 1438 case X86::VPERMI2##Suffix##256rm: case X86::VPERMT2##Suffix##256rm: \ 1439 case X86::VPERMI2##Suffix##rm: case X86::VPERMT2##Suffix##rm: \ 1440 case X86::VPERMI2##Suffix##128rrkz: case X86::VPERMT2##Suffix##128rrkz: \ 1441 case X86::VPERMI2##Suffix##256rrkz: case X86::VPERMT2##Suffix##256rrkz: \ 1442 case X86::VPERMI2##Suffix##rrkz: case X86::VPERMT2##Suffix##rrkz: \ 1443 case X86::VPERMI2##Suffix##128rmkz: case X86::VPERMT2##Suffix##128rmkz: \ 1444 case X86::VPERMI2##Suffix##256rmkz: case X86::VPERMT2##Suffix##256rmkz: \ 1445 case X86::VPERMI2##Suffix##rmkz: case X86::VPERMT2##Suffix##rmkz: 1446 1447 #define VPERM_CASES_BROADCAST(Suffix) \ 1448 VPERM_CASES(Suffix) \ 1449 case X86::VPERMI2##Suffix##128rmb: case X86::VPERMT2##Suffix##128rmb: \ 1450 case X86::VPERMI2##Suffix##256rmb: case X86::VPERMT2##Suffix##256rmb: \ 1451 case X86::VPERMI2##Suffix##rmb: case X86::VPERMT2##Suffix##rmb: \ 1452 case X86::VPERMI2##Suffix##128rmbkz: case X86::VPERMT2##Suffix##128rmbkz: \ 1453 case X86::VPERMI2##Suffix##256rmbkz: case X86::VPERMT2##Suffix##256rmbkz: \ 1454 case X86::VPERMI2##Suffix##rmbkz: case X86::VPERMT2##Suffix##rmbkz: 1455 1456 switch (Opcode) { 1457 default: return false; 1458 VPERM_CASES(B) 1459 VPERM_CASES_BROADCAST(D) 1460 VPERM_CASES_BROADCAST(PD) 1461 VPERM_CASES_BROADCAST(PS) 1462 VPERM_CASES_BROADCAST(Q) 1463 VPERM_CASES(W) 1464 return true; 1465 } 1466 #undef VPERM_CASES_BROADCAST 1467 #undef VPERM_CASES 1468 } 1469 1470 // Returns commuted opcode for VPERMI2 and VPERMT2 instructions by switching 1471 // from the I opcode to the T opcode and vice versa. 1472 static unsigned getCommutedVPERMV3Opcode(unsigned Opcode) { 1473 #define VPERM_CASES(Orig, New) \ 1474 case X86::Orig##128rr: return X86::New##128rr; \ 1475 case X86::Orig##128rrkz: return X86::New##128rrkz; \ 1476 case X86::Orig##128rm: return X86::New##128rm; \ 1477 case X86::Orig##128rmkz: return X86::New##128rmkz; \ 1478 case X86::Orig##256rr: return X86::New##256rr; \ 1479 case X86::Orig##256rrkz: return X86::New##256rrkz; \ 1480 case X86::Orig##256rm: return X86::New##256rm; \ 1481 case X86::Orig##256rmkz: return X86::New##256rmkz; \ 1482 case X86::Orig##rr: return X86::New##rr; \ 1483 case X86::Orig##rrkz: return X86::New##rrkz; \ 1484 case X86::Orig##rm: return X86::New##rm; \ 1485 case X86::Orig##rmkz: return X86::New##rmkz; 1486 1487 #define VPERM_CASES_BROADCAST(Orig, New) \ 1488 VPERM_CASES(Orig, New) \ 1489 case X86::Orig##128rmb: return X86::New##128rmb; \ 1490 case X86::Orig##128rmbkz: return X86::New##128rmbkz; \ 1491 case X86::Orig##256rmb: return X86::New##256rmb; \ 1492 case X86::Orig##256rmbkz: return X86::New##256rmbkz; \ 1493 case X86::Orig##rmb: return X86::New##rmb; \ 1494 case X86::Orig##rmbkz: return X86::New##rmbkz; 1495 1496 switch (Opcode) { 1497 VPERM_CASES(VPERMI2B, VPERMT2B) 1498 VPERM_CASES_BROADCAST(VPERMI2D, VPERMT2D) 1499 VPERM_CASES_BROADCAST(VPERMI2PD, VPERMT2PD) 1500 VPERM_CASES_BROADCAST(VPERMI2PS, VPERMT2PS) 1501 VPERM_CASES_BROADCAST(VPERMI2Q, VPERMT2Q) 1502 VPERM_CASES(VPERMI2W, VPERMT2W) 1503 VPERM_CASES(VPERMT2B, VPERMI2B) 1504 VPERM_CASES_BROADCAST(VPERMT2D, VPERMI2D) 1505 VPERM_CASES_BROADCAST(VPERMT2PD, VPERMI2PD) 1506 VPERM_CASES_BROADCAST(VPERMT2PS, VPERMI2PS) 1507 VPERM_CASES_BROADCAST(VPERMT2Q, VPERMI2Q) 1508 VPERM_CASES(VPERMT2W, VPERMI2W) 1509 } 1510 1511 llvm_unreachable("Unreachable!"); 1512 #undef VPERM_CASES_BROADCAST 1513 #undef VPERM_CASES 1514 } 1515 1516 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI, 1517 unsigned OpIdx1, 1518 unsigned OpIdx2) const { 1519 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 1520 if (NewMI) 1521 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 1522 return MI; 1523 }; 1524 1525 switch (MI.getOpcode()) { 1526 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I) 1527 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I) 1528 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I) 1529 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I) 1530 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I) 1531 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I) 1532 unsigned Opc; 1533 unsigned Size; 1534 switch (MI.getOpcode()) { 1535 default: llvm_unreachable("Unreachable!"); 1536 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break; 1537 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break; 1538 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break; 1539 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break; 1540 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break; 1541 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break; 1542 } 1543 unsigned Amt = MI.getOperand(3).getImm(); 1544 auto &WorkingMI = cloneIfNew(MI); 1545 WorkingMI.setDesc(get(Opc)); 1546 WorkingMI.getOperand(3).setImm(Size - Amt); 1547 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1548 OpIdx1, OpIdx2); 1549 } 1550 case X86::PFSUBrr: 1551 case X86::PFSUBRrr: { 1552 // PFSUB x, y: x = x - y 1553 // PFSUBR x, y: x = y - x 1554 unsigned Opc = 1555 (X86::PFSUBRrr == MI.getOpcode() ? X86::PFSUBrr : X86::PFSUBRrr); 1556 auto &WorkingMI = cloneIfNew(MI); 1557 WorkingMI.setDesc(get(Opc)); 1558 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1559 OpIdx1, OpIdx2); 1560 } 1561 case X86::BLENDPDrri: 1562 case X86::BLENDPSrri: 1563 case X86::VBLENDPDrri: 1564 case X86::VBLENDPSrri: 1565 // If we're optimizing for size, try to use MOVSD/MOVSS. 1566 if (MI.getParent()->getParent()->getFunction().optForSize()) { 1567 unsigned Mask, Opc; 1568 switch (MI.getOpcode()) { 1569 default: llvm_unreachable("Unreachable!"); 1570 case X86::BLENDPDrri: Opc = X86::MOVSDrr; Mask = 0x03; break; 1571 case X86::BLENDPSrri: Opc = X86::MOVSSrr; Mask = 0x0F; break; 1572 case X86::VBLENDPDrri: Opc = X86::VMOVSDrr; Mask = 0x03; break; 1573 case X86::VBLENDPSrri: Opc = X86::VMOVSSrr; Mask = 0x0F; break; 1574 } 1575 if ((MI.getOperand(3).getImm() ^ Mask) == 1) { 1576 auto &WorkingMI = cloneIfNew(MI); 1577 WorkingMI.setDesc(get(Opc)); 1578 WorkingMI.RemoveOperand(3); 1579 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, 1580 /*NewMI=*/false, 1581 OpIdx1, OpIdx2); 1582 } 1583 } 1584 LLVM_FALLTHROUGH; 1585 case X86::PBLENDWrri: 1586 case X86::VBLENDPDYrri: 1587 case X86::VBLENDPSYrri: 1588 case X86::VPBLENDDrri: 1589 case X86::VPBLENDWrri: 1590 case X86::VPBLENDDYrri: 1591 case X86::VPBLENDWYrri:{ 1592 unsigned Mask; 1593 switch (MI.getOpcode()) { 1594 default: llvm_unreachable("Unreachable!"); 1595 case X86::BLENDPDrri: Mask = 0x03; break; 1596 case X86::BLENDPSrri: Mask = 0x0F; break; 1597 case X86::PBLENDWrri: Mask = 0xFF; break; 1598 case X86::VBLENDPDrri: Mask = 0x03; break; 1599 case X86::VBLENDPSrri: Mask = 0x0F; break; 1600 case X86::VBLENDPDYrri: Mask = 0x0F; break; 1601 case X86::VBLENDPSYrri: Mask = 0xFF; break; 1602 case X86::VPBLENDDrri: Mask = 0x0F; break; 1603 case X86::VPBLENDWrri: Mask = 0xFF; break; 1604 case X86::VPBLENDDYrri: Mask = 0xFF; break; 1605 case X86::VPBLENDWYrri: Mask = 0xFF; break; 1606 } 1607 // Only the least significant bits of Imm are used. 1608 unsigned Imm = MI.getOperand(3).getImm() & Mask; 1609 auto &WorkingMI = cloneIfNew(MI); 1610 WorkingMI.getOperand(3).setImm(Mask ^ Imm); 1611 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1612 OpIdx1, OpIdx2); 1613 } 1614 case X86::MOVSDrr: 1615 case X86::MOVSSrr: 1616 case X86::VMOVSDrr: 1617 case X86::VMOVSSrr:{ 1618 // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD. 1619 assert(Subtarget.hasSSE41() && "Commuting MOVSD/MOVSS requires SSE41!"); 1620 1621 unsigned Mask, Opc; 1622 switch (MI.getOpcode()) { 1623 default: llvm_unreachable("Unreachable!"); 1624 case X86::MOVSDrr: Opc = X86::BLENDPDrri; Mask = 0x02; break; 1625 case X86::MOVSSrr: Opc = X86::BLENDPSrri; Mask = 0x0E; break; 1626 case X86::VMOVSDrr: Opc = X86::VBLENDPDrri; Mask = 0x02; break; 1627 case X86::VMOVSSrr: Opc = X86::VBLENDPSrri; Mask = 0x0E; break; 1628 } 1629 1630 auto &WorkingMI = cloneIfNew(MI); 1631 WorkingMI.setDesc(get(Opc)); 1632 WorkingMI.addOperand(MachineOperand::CreateImm(Mask)); 1633 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1634 OpIdx1, OpIdx2); 1635 } 1636 case X86::PCLMULQDQrr: 1637 case X86::VPCLMULQDQrr: 1638 case X86::VPCLMULQDQYrr: 1639 case X86::VPCLMULQDQZrr: 1640 case X86::VPCLMULQDQZ128rr: 1641 case X86::VPCLMULQDQZ256rr: { 1642 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0] 1643 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0] 1644 unsigned Imm = MI.getOperand(3).getImm(); 1645 unsigned Src1Hi = Imm & 0x01; 1646 unsigned Src2Hi = Imm & 0x10; 1647 auto &WorkingMI = cloneIfNew(MI); 1648 WorkingMI.getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4)); 1649 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1650 OpIdx1, OpIdx2); 1651 } 1652 case X86::VPCMPBZ128rri: case X86::VPCMPUBZ128rri: 1653 case X86::VPCMPBZ256rri: case X86::VPCMPUBZ256rri: 1654 case X86::VPCMPBZrri: case X86::VPCMPUBZrri: 1655 case X86::VPCMPDZ128rri: case X86::VPCMPUDZ128rri: 1656 case X86::VPCMPDZ256rri: case X86::VPCMPUDZ256rri: 1657 case X86::VPCMPDZrri: case X86::VPCMPUDZrri: 1658 case X86::VPCMPQZ128rri: case X86::VPCMPUQZ128rri: 1659 case X86::VPCMPQZ256rri: case X86::VPCMPUQZ256rri: 1660 case X86::VPCMPQZrri: case X86::VPCMPUQZrri: 1661 case X86::VPCMPWZ128rri: case X86::VPCMPUWZ128rri: 1662 case X86::VPCMPWZ256rri: case X86::VPCMPUWZ256rri: 1663 case X86::VPCMPWZrri: case X86::VPCMPUWZrri: 1664 case X86::VPCMPBZ128rrik: case X86::VPCMPUBZ128rrik: 1665 case X86::VPCMPBZ256rrik: case X86::VPCMPUBZ256rrik: 1666 case X86::VPCMPBZrrik: case X86::VPCMPUBZrrik: 1667 case X86::VPCMPDZ128rrik: case X86::VPCMPUDZ128rrik: 1668 case X86::VPCMPDZ256rrik: case X86::VPCMPUDZ256rrik: 1669 case X86::VPCMPDZrrik: case X86::VPCMPUDZrrik: 1670 case X86::VPCMPQZ128rrik: case X86::VPCMPUQZ128rrik: 1671 case X86::VPCMPQZ256rrik: case X86::VPCMPUQZ256rrik: 1672 case X86::VPCMPQZrrik: case X86::VPCMPUQZrrik: 1673 case X86::VPCMPWZ128rrik: case X86::VPCMPUWZ128rrik: 1674 case X86::VPCMPWZ256rrik: case X86::VPCMPUWZ256rrik: 1675 case X86::VPCMPWZrrik: case X86::VPCMPUWZrrik: { 1676 // Flip comparison mode immediate (if necessary). 1677 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm() & 0x7; 1678 Imm = X86::getSwappedVPCMPImm(Imm); 1679 auto &WorkingMI = cloneIfNew(MI); 1680 WorkingMI.getOperand(MI.getNumOperands() - 1).setImm(Imm); 1681 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1682 OpIdx1, OpIdx2); 1683 } 1684 case X86::VPCOMBri: case X86::VPCOMUBri: 1685 case X86::VPCOMDri: case X86::VPCOMUDri: 1686 case X86::VPCOMQri: case X86::VPCOMUQri: 1687 case X86::VPCOMWri: case X86::VPCOMUWri: { 1688 // Flip comparison mode immediate (if necessary). 1689 unsigned Imm = MI.getOperand(3).getImm() & 0x7; 1690 Imm = X86::getSwappedVPCOMImm(Imm); 1691 auto &WorkingMI = cloneIfNew(MI); 1692 WorkingMI.getOperand(3).setImm(Imm); 1693 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1694 OpIdx1, OpIdx2); 1695 } 1696 case X86::VPERM2F128rr: 1697 case X86::VPERM2I128rr: { 1698 // Flip permute source immediate. 1699 // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi. 1700 // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi. 1701 unsigned Imm = MI.getOperand(3).getImm() & 0xFF; 1702 auto &WorkingMI = cloneIfNew(MI); 1703 WorkingMI.getOperand(3).setImm(Imm ^ 0x22); 1704 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1705 OpIdx1, OpIdx2); 1706 } 1707 case X86::MOVHLPSrr: 1708 case X86::UNPCKHPDrr: 1709 case X86::VMOVHLPSrr: 1710 case X86::VUNPCKHPDrr: 1711 case X86::VMOVHLPSZrr: 1712 case X86::VUNPCKHPDZ128rr: { 1713 assert(Subtarget.hasSSE2() && "Commuting MOVHLP/UNPCKHPD requires SSE2!"); 1714 1715 unsigned Opc = MI.getOpcode(); 1716 switch (Opc) { 1717 default: llvm_unreachable("Unreachable!"); 1718 case X86::MOVHLPSrr: Opc = X86::UNPCKHPDrr; break; 1719 case X86::UNPCKHPDrr: Opc = X86::MOVHLPSrr; break; 1720 case X86::VMOVHLPSrr: Opc = X86::VUNPCKHPDrr; break; 1721 case X86::VUNPCKHPDrr: Opc = X86::VMOVHLPSrr; break; 1722 case X86::VMOVHLPSZrr: Opc = X86::VUNPCKHPDZ128rr; break; 1723 case X86::VUNPCKHPDZ128rr: Opc = X86::VMOVHLPSZrr; break; 1724 } 1725 auto &WorkingMI = cloneIfNew(MI); 1726 WorkingMI.setDesc(get(Opc)); 1727 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1728 OpIdx1, OpIdx2); 1729 } 1730 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr: 1731 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr: 1732 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr: 1733 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr: 1734 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr: 1735 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr: 1736 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr: 1737 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr: 1738 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr: 1739 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr: 1740 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr: 1741 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr: 1742 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr: 1743 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr: 1744 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr: 1745 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: { 1746 unsigned Opc; 1747 switch (MI.getOpcode()) { 1748 default: llvm_unreachable("Unreachable!"); 1749 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break; 1750 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break; 1751 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break; 1752 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break; 1753 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break; 1754 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break; 1755 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break; 1756 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break; 1757 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break; 1758 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break; 1759 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break; 1760 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break; 1761 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break; 1762 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break; 1763 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break; 1764 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break; 1765 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break; 1766 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break; 1767 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break; 1768 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break; 1769 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break; 1770 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break; 1771 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break; 1772 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break; 1773 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break; 1774 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break; 1775 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break; 1776 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break; 1777 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break; 1778 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break; 1779 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break; 1780 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break; 1781 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break; 1782 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break; 1783 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break; 1784 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break; 1785 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break; 1786 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break; 1787 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break; 1788 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break; 1789 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break; 1790 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break; 1791 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break; 1792 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break; 1793 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break; 1794 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break; 1795 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break; 1796 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break; 1797 } 1798 auto &WorkingMI = cloneIfNew(MI); 1799 WorkingMI.setDesc(get(Opc)); 1800 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1801 OpIdx1, OpIdx2); 1802 } 1803 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi: 1804 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi: 1805 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi: 1806 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi: 1807 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi: 1808 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi: 1809 case X86::VPTERNLOGDZrrik: 1810 case X86::VPTERNLOGDZ128rrik: 1811 case X86::VPTERNLOGDZ256rrik: 1812 case X86::VPTERNLOGQZrrik: 1813 case X86::VPTERNLOGQZ128rrik: 1814 case X86::VPTERNLOGQZ256rrik: 1815 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz: 1816 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz: 1817 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz: 1818 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz: 1819 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz: 1820 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz: 1821 case X86::VPTERNLOGDZ128rmbi: 1822 case X86::VPTERNLOGDZ256rmbi: 1823 case X86::VPTERNLOGDZrmbi: 1824 case X86::VPTERNLOGQZ128rmbi: 1825 case X86::VPTERNLOGQZ256rmbi: 1826 case X86::VPTERNLOGQZrmbi: 1827 case X86::VPTERNLOGDZ128rmbikz: 1828 case X86::VPTERNLOGDZ256rmbikz: 1829 case X86::VPTERNLOGDZrmbikz: 1830 case X86::VPTERNLOGQZ128rmbikz: 1831 case X86::VPTERNLOGQZ256rmbikz: 1832 case X86::VPTERNLOGQZrmbikz: { 1833 auto &WorkingMI = cloneIfNew(MI); 1834 commuteVPTERNLOG(WorkingMI, OpIdx1, OpIdx2); 1835 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1836 OpIdx1, OpIdx2); 1837 } 1838 default: { 1839 if (isCommutableVPERMV3Instruction(MI.getOpcode())) { 1840 unsigned Opc = getCommutedVPERMV3Opcode(MI.getOpcode()); 1841 auto &WorkingMI = cloneIfNew(MI); 1842 WorkingMI.setDesc(get(Opc)); 1843 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1844 OpIdx1, OpIdx2); 1845 } 1846 1847 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(), 1848 MI.getDesc().TSFlags); 1849 if (FMA3Group) { 1850 unsigned Opc = 1851 getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2, *FMA3Group); 1852 auto &WorkingMI = cloneIfNew(MI); 1853 WorkingMI.setDesc(get(Opc)); 1854 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 1855 OpIdx1, OpIdx2); 1856 } 1857 1858 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 1859 } 1860 } 1861 } 1862 1863 bool 1864 X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI, 1865 unsigned &SrcOpIdx1, 1866 unsigned &SrcOpIdx2, 1867 bool IsIntrinsic) const { 1868 uint64_t TSFlags = MI.getDesc().TSFlags; 1869 1870 unsigned FirstCommutableVecOp = 1; 1871 unsigned LastCommutableVecOp = 3; 1872 unsigned KMaskOp = -1U; 1873 if (X86II::isKMasked(TSFlags)) { 1874 // For k-zero-masked operations it is Ok to commute the first vector 1875 // operand. 1876 // For regular k-masked operations a conservative choice is done as the 1877 // elements of the first vector operand, for which the corresponding bit 1878 // in the k-mask operand is set to 0, are copied to the result of the 1879 // instruction. 1880 // TODO/FIXME: The commute still may be legal if it is known that the 1881 // k-mask operand is set to either all ones or all zeroes. 1882 // It is also Ok to commute the 1st operand if all users of MI use only 1883 // the elements enabled by the k-mask operand. For example, 1884 // v4 = VFMADD213PSZrk v1, k, v2, v3; // v1[i] = k[i] ? v2[i]*v1[i]+v3[i] 1885 // : v1[i]; 1886 // VMOVAPSZmrk <mem_addr>, k, v4; // this is the ONLY user of v4 -> 1887 // // Ok, to commute v1 in FMADD213PSZrk. 1888 1889 // The k-mask operand has index = 2 for masked and zero-masked operations. 1890 KMaskOp = 2; 1891 1892 // The operand with index = 1 is used as a source for those elements for 1893 // which the corresponding bit in the k-mask is set to 0. 1894 if (X86II::isKMergeMasked(TSFlags)) 1895 FirstCommutableVecOp = 3; 1896 1897 LastCommutableVecOp++; 1898 } else if (IsIntrinsic) { 1899 // Commuting the first operand of an intrinsic instruction isn't possible 1900 // unless we can prove that only the lowest element of the result is used. 1901 FirstCommutableVecOp = 2; 1902 } 1903 1904 if (isMem(MI, LastCommutableVecOp)) 1905 LastCommutableVecOp--; 1906 1907 // Only the first RegOpsNum operands are commutable. 1908 // Also, the value 'CommuteAnyOperandIndex' is valid here as it means 1909 // that the operand is not specified/fixed. 1910 if (SrcOpIdx1 != CommuteAnyOperandIndex && 1911 (SrcOpIdx1 < FirstCommutableVecOp || SrcOpIdx1 > LastCommutableVecOp || 1912 SrcOpIdx1 == KMaskOp)) 1913 return false; 1914 if (SrcOpIdx2 != CommuteAnyOperandIndex && 1915 (SrcOpIdx2 < FirstCommutableVecOp || SrcOpIdx2 > LastCommutableVecOp || 1916 SrcOpIdx2 == KMaskOp)) 1917 return false; 1918 1919 // Look for two different register operands assumed to be commutable 1920 // regardless of the FMA opcode. The FMA opcode is adjusted later. 1921 if (SrcOpIdx1 == CommuteAnyOperandIndex || 1922 SrcOpIdx2 == CommuteAnyOperandIndex) { 1923 unsigned CommutableOpIdx1 = SrcOpIdx1; 1924 unsigned CommutableOpIdx2 = SrcOpIdx2; 1925 1926 // At least one of operands to be commuted is not specified and 1927 // this method is free to choose appropriate commutable operands. 1928 if (SrcOpIdx1 == SrcOpIdx2) 1929 // Both of operands are not fixed. By default set one of commutable 1930 // operands to the last register operand of the instruction. 1931 CommutableOpIdx2 = LastCommutableVecOp; 1932 else if (SrcOpIdx2 == CommuteAnyOperandIndex) 1933 // Only one of operands is not fixed. 1934 CommutableOpIdx2 = SrcOpIdx1; 1935 1936 // CommutableOpIdx2 is well defined now. Let's choose another commutable 1937 // operand and assign its index to CommutableOpIdx1. 1938 unsigned Op2Reg = MI.getOperand(CommutableOpIdx2).getReg(); 1939 for (CommutableOpIdx1 = LastCommutableVecOp; 1940 CommutableOpIdx1 >= FirstCommutableVecOp; CommutableOpIdx1--) { 1941 // Just ignore and skip the k-mask operand. 1942 if (CommutableOpIdx1 == KMaskOp) 1943 continue; 1944 1945 // The commuted operands must have different registers. 1946 // Otherwise, the commute transformation does not change anything and 1947 // is useless then. 1948 if (Op2Reg != MI.getOperand(CommutableOpIdx1).getReg()) 1949 break; 1950 } 1951 1952 // No appropriate commutable operands were found. 1953 if (CommutableOpIdx1 < FirstCommutableVecOp) 1954 return false; 1955 1956 // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2 1957 // to return those values. 1958 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1959 CommutableOpIdx1, CommutableOpIdx2)) 1960 return false; 1961 } 1962 1963 return true; 1964 } 1965 1966 bool X86InstrInfo::findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1, 1967 unsigned &SrcOpIdx2) const { 1968 const MCInstrDesc &Desc = MI.getDesc(); 1969 if (!Desc.isCommutable()) 1970 return false; 1971 1972 switch (MI.getOpcode()) { 1973 case X86::CMPSDrr: 1974 case X86::CMPSSrr: 1975 case X86::CMPPDrri: 1976 case X86::CMPPSrri: 1977 case X86::VCMPSDrr: 1978 case X86::VCMPSSrr: 1979 case X86::VCMPPDrri: 1980 case X86::VCMPPSrri: 1981 case X86::VCMPPDYrri: 1982 case X86::VCMPPSYrri: 1983 case X86::VCMPSDZrr: 1984 case X86::VCMPSSZrr: 1985 case X86::VCMPPDZrri: 1986 case X86::VCMPPSZrri: 1987 case X86::VCMPPDZ128rri: 1988 case X86::VCMPPSZ128rri: 1989 case X86::VCMPPDZ256rri: 1990 case X86::VCMPPSZ256rri: { 1991 // Float comparison can be safely commuted for 1992 // Ordered/Unordered/Equal/NotEqual tests 1993 unsigned Imm = MI.getOperand(3).getImm() & 0x7; 1994 switch (Imm) { 1995 case 0x00: // EQUAL 1996 case 0x03: // UNORDERED 1997 case 0x04: // NOT EQUAL 1998 case 0x07: // ORDERED 1999 // The indices of the commutable operands are 1 and 2. 2000 // Assign them to the returned operand indices here. 2001 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2); 2002 } 2003 return false; 2004 } 2005 case X86::MOVSDrr: 2006 case X86::MOVSSrr: 2007 case X86::VMOVSDrr: 2008 case X86::VMOVSSrr: 2009 if (Subtarget.hasSSE41()) 2010 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 2011 return false; 2012 case X86::MOVHLPSrr: 2013 case X86::UNPCKHPDrr: 2014 case X86::VMOVHLPSrr: 2015 case X86::VUNPCKHPDrr: 2016 case X86::VMOVHLPSZrr: 2017 case X86::VUNPCKHPDZ128rr: 2018 if (Subtarget.hasSSE2()) 2019 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 2020 return false; 2021 case X86::VPTERNLOGDZrri: case X86::VPTERNLOGDZrmi: 2022 case X86::VPTERNLOGDZ128rri: case X86::VPTERNLOGDZ128rmi: 2023 case X86::VPTERNLOGDZ256rri: case X86::VPTERNLOGDZ256rmi: 2024 case X86::VPTERNLOGQZrri: case X86::VPTERNLOGQZrmi: 2025 case X86::VPTERNLOGQZ128rri: case X86::VPTERNLOGQZ128rmi: 2026 case X86::VPTERNLOGQZ256rri: case X86::VPTERNLOGQZ256rmi: 2027 case X86::VPTERNLOGDZrrik: 2028 case X86::VPTERNLOGDZ128rrik: 2029 case X86::VPTERNLOGDZ256rrik: 2030 case X86::VPTERNLOGQZrrik: 2031 case X86::VPTERNLOGQZ128rrik: 2032 case X86::VPTERNLOGQZ256rrik: 2033 case X86::VPTERNLOGDZrrikz: case X86::VPTERNLOGDZrmikz: 2034 case X86::VPTERNLOGDZ128rrikz: case X86::VPTERNLOGDZ128rmikz: 2035 case X86::VPTERNLOGDZ256rrikz: case X86::VPTERNLOGDZ256rmikz: 2036 case X86::VPTERNLOGQZrrikz: case X86::VPTERNLOGQZrmikz: 2037 case X86::VPTERNLOGQZ128rrikz: case X86::VPTERNLOGQZ128rmikz: 2038 case X86::VPTERNLOGQZ256rrikz: case X86::VPTERNLOGQZ256rmikz: 2039 case X86::VPTERNLOGDZ128rmbi: 2040 case X86::VPTERNLOGDZ256rmbi: 2041 case X86::VPTERNLOGDZrmbi: 2042 case X86::VPTERNLOGQZ128rmbi: 2043 case X86::VPTERNLOGQZ256rmbi: 2044 case X86::VPTERNLOGQZrmbi: 2045 case X86::VPTERNLOGDZ128rmbikz: 2046 case X86::VPTERNLOGDZ256rmbikz: 2047 case X86::VPTERNLOGDZrmbikz: 2048 case X86::VPTERNLOGQZ128rmbikz: 2049 case X86::VPTERNLOGQZ256rmbikz: 2050 case X86::VPTERNLOGQZrmbikz: 2051 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 2052 case X86::VPMADD52HUQZ128r: 2053 case X86::VPMADD52HUQZ128rk: 2054 case X86::VPMADD52HUQZ128rkz: 2055 case X86::VPMADD52HUQZ256r: 2056 case X86::VPMADD52HUQZ256rk: 2057 case X86::VPMADD52HUQZ256rkz: 2058 case X86::VPMADD52HUQZr: 2059 case X86::VPMADD52HUQZrk: 2060 case X86::VPMADD52HUQZrkz: 2061 case X86::VPMADD52LUQZ128r: 2062 case X86::VPMADD52LUQZ128rk: 2063 case X86::VPMADD52LUQZ128rkz: 2064 case X86::VPMADD52LUQZ256r: 2065 case X86::VPMADD52LUQZ256rk: 2066 case X86::VPMADD52LUQZ256rkz: 2067 case X86::VPMADD52LUQZr: 2068 case X86::VPMADD52LUQZrk: 2069 case X86::VPMADD52LUQZrkz: { 2070 unsigned CommutableOpIdx1 = 2; 2071 unsigned CommutableOpIdx2 = 3; 2072 if (X86II::isKMasked(Desc.TSFlags)) { 2073 // Skip the mask register. 2074 ++CommutableOpIdx1; 2075 ++CommutableOpIdx2; 2076 } 2077 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2078 CommutableOpIdx1, CommutableOpIdx2)) 2079 return false; 2080 if (!MI.getOperand(SrcOpIdx1).isReg() || 2081 !MI.getOperand(SrcOpIdx2).isReg()) 2082 // No idea. 2083 return false; 2084 return true; 2085 } 2086 2087 default: 2088 const X86InstrFMA3Group *FMA3Group = getFMA3Group(MI.getOpcode(), 2089 MI.getDesc().TSFlags); 2090 if (FMA3Group) 2091 return findThreeSrcCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2, 2092 FMA3Group->isIntrinsic()); 2093 2094 // Handled masked instructions since we need to skip over the mask input 2095 // and the preserved input. 2096 if (X86II::isKMasked(Desc.TSFlags)) { 2097 // First assume that the first input is the mask operand and skip past it. 2098 unsigned CommutableOpIdx1 = Desc.getNumDefs() + 1; 2099 unsigned CommutableOpIdx2 = Desc.getNumDefs() + 2; 2100 // Check if the first input is tied. If there isn't one then we only 2101 // need to skip the mask operand which we did above. 2102 if ((MI.getDesc().getOperandConstraint(Desc.getNumDefs(), 2103 MCOI::TIED_TO) != -1)) { 2104 // If this is zero masking instruction with a tied operand, we need to 2105 // move the first index back to the first input since this must 2106 // be a 3 input instruction and we want the first two non-mask inputs. 2107 // Otherwise this is a 2 input instruction with a preserved input and 2108 // mask, so we need to move the indices to skip one more input. 2109 if (X86II::isKMergeMasked(Desc.TSFlags)) { 2110 ++CommutableOpIdx1; 2111 ++CommutableOpIdx2; 2112 } else { 2113 --CommutableOpIdx1; 2114 } 2115 } 2116 2117 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2118 CommutableOpIdx1, CommutableOpIdx2)) 2119 return false; 2120 2121 if (!MI.getOperand(SrcOpIdx1).isReg() || 2122 !MI.getOperand(SrcOpIdx2).isReg()) 2123 // No idea. 2124 return false; 2125 return true; 2126 } 2127 2128 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 2129 } 2130 return false; 2131 } 2132 2133 X86::CondCode X86::getCondFromBranchOpc(unsigned BrOpc) { 2134 switch (BrOpc) { 2135 default: return X86::COND_INVALID; 2136 case X86::JE_1: return X86::COND_E; 2137 case X86::JNE_1: return X86::COND_NE; 2138 case X86::JL_1: return X86::COND_L; 2139 case X86::JLE_1: return X86::COND_LE; 2140 case X86::JG_1: return X86::COND_G; 2141 case X86::JGE_1: return X86::COND_GE; 2142 case X86::JB_1: return X86::COND_B; 2143 case X86::JBE_1: return X86::COND_BE; 2144 case X86::JA_1: return X86::COND_A; 2145 case X86::JAE_1: return X86::COND_AE; 2146 case X86::JS_1: return X86::COND_S; 2147 case X86::JNS_1: return X86::COND_NS; 2148 case X86::JP_1: return X86::COND_P; 2149 case X86::JNP_1: return X86::COND_NP; 2150 case X86::JO_1: return X86::COND_O; 2151 case X86::JNO_1: return X86::COND_NO; 2152 } 2153 } 2154 2155 /// Return condition code of a SET opcode. 2156 X86::CondCode X86::getCondFromSETOpc(unsigned Opc) { 2157 switch (Opc) { 2158 default: return X86::COND_INVALID; 2159 case X86::SETAr: case X86::SETAm: return X86::COND_A; 2160 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE; 2161 case X86::SETBr: case X86::SETBm: return X86::COND_B; 2162 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE; 2163 case X86::SETEr: case X86::SETEm: return X86::COND_E; 2164 case X86::SETGr: case X86::SETGm: return X86::COND_G; 2165 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE; 2166 case X86::SETLr: case X86::SETLm: return X86::COND_L; 2167 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE; 2168 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE; 2169 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO; 2170 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP; 2171 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS; 2172 case X86::SETOr: case X86::SETOm: return X86::COND_O; 2173 case X86::SETPr: case X86::SETPm: return X86::COND_P; 2174 case X86::SETSr: case X86::SETSm: return X86::COND_S; 2175 } 2176 } 2177 2178 /// Return condition code of a CMov opcode. 2179 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) { 2180 switch (Opc) { 2181 default: return X86::COND_INVALID; 2182 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm: 2183 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr: 2184 return X86::COND_A; 2185 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm: 2186 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr: 2187 return X86::COND_AE; 2188 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm: 2189 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr: 2190 return X86::COND_B; 2191 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm: 2192 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr: 2193 return X86::COND_BE; 2194 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm: 2195 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr: 2196 return X86::COND_E; 2197 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm: 2198 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr: 2199 return X86::COND_G; 2200 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm: 2201 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr: 2202 return X86::COND_GE; 2203 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm: 2204 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr: 2205 return X86::COND_L; 2206 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm: 2207 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr: 2208 return X86::COND_LE; 2209 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm: 2210 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr: 2211 return X86::COND_NE; 2212 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm: 2213 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr: 2214 return X86::COND_NO; 2215 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm: 2216 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr: 2217 return X86::COND_NP; 2218 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm: 2219 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr: 2220 return X86::COND_NS; 2221 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm: 2222 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr: 2223 return X86::COND_O; 2224 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm: 2225 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr: 2226 return X86::COND_P; 2227 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm: 2228 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr: 2229 return X86::COND_S; 2230 } 2231 } 2232 2233 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) { 2234 switch (CC) { 2235 default: llvm_unreachable("Illegal condition code!"); 2236 case X86::COND_E: return X86::JE_1; 2237 case X86::COND_NE: return X86::JNE_1; 2238 case X86::COND_L: return X86::JL_1; 2239 case X86::COND_LE: return X86::JLE_1; 2240 case X86::COND_G: return X86::JG_1; 2241 case X86::COND_GE: return X86::JGE_1; 2242 case X86::COND_B: return X86::JB_1; 2243 case X86::COND_BE: return X86::JBE_1; 2244 case X86::COND_A: return X86::JA_1; 2245 case X86::COND_AE: return X86::JAE_1; 2246 case X86::COND_S: return X86::JS_1; 2247 case X86::COND_NS: return X86::JNS_1; 2248 case X86::COND_P: return X86::JP_1; 2249 case X86::COND_NP: return X86::JNP_1; 2250 case X86::COND_O: return X86::JO_1; 2251 case X86::COND_NO: return X86::JNO_1; 2252 } 2253 } 2254 2255 /// Return the inverse of the specified condition, 2256 /// e.g. turning COND_E to COND_NE. 2257 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) { 2258 switch (CC) { 2259 default: llvm_unreachable("Illegal condition code!"); 2260 case X86::COND_E: return X86::COND_NE; 2261 case X86::COND_NE: return X86::COND_E; 2262 case X86::COND_L: return X86::COND_GE; 2263 case X86::COND_LE: return X86::COND_G; 2264 case X86::COND_G: return X86::COND_LE; 2265 case X86::COND_GE: return X86::COND_L; 2266 case X86::COND_B: return X86::COND_AE; 2267 case X86::COND_BE: return X86::COND_A; 2268 case X86::COND_A: return X86::COND_BE; 2269 case X86::COND_AE: return X86::COND_B; 2270 case X86::COND_S: return X86::COND_NS; 2271 case X86::COND_NS: return X86::COND_S; 2272 case X86::COND_P: return X86::COND_NP; 2273 case X86::COND_NP: return X86::COND_P; 2274 case X86::COND_O: return X86::COND_NO; 2275 case X86::COND_NO: return X86::COND_O; 2276 case X86::COND_NE_OR_P: return X86::COND_E_AND_NP; 2277 case X86::COND_E_AND_NP: return X86::COND_NE_OR_P; 2278 } 2279 } 2280 2281 /// Assuming the flags are set by MI(a,b), return the condition code if we 2282 /// modify the instructions such that flags are set by MI(b,a). 2283 static X86::CondCode getSwappedCondition(X86::CondCode CC) { 2284 switch (CC) { 2285 default: return X86::COND_INVALID; 2286 case X86::COND_E: return X86::COND_E; 2287 case X86::COND_NE: return X86::COND_NE; 2288 case X86::COND_L: return X86::COND_G; 2289 case X86::COND_LE: return X86::COND_GE; 2290 case X86::COND_G: return X86::COND_L; 2291 case X86::COND_GE: return X86::COND_LE; 2292 case X86::COND_B: return X86::COND_A; 2293 case X86::COND_BE: return X86::COND_AE; 2294 case X86::COND_A: return X86::COND_B; 2295 case X86::COND_AE: return X86::COND_BE; 2296 } 2297 } 2298 2299 std::pair<X86::CondCode, bool> 2300 X86::getX86ConditionCode(CmpInst::Predicate Predicate) { 2301 X86::CondCode CC = X86::COND_INVALID; 2302 bool NeedSwap = false; 2303 switch (Predicate) { 2304 default: break; 2305 // Floating-point Predicates 2306 case CmpInst::FCMP_UEQ: CC = X86::COND_E; break; 2307 case CmpInst::FCMP_OLT: NeedSwap = true; LLVM_FALLTHROUGH; 2308 case CmpInst::FCMP_OGT: CC = X86::COND_A; break; 2309 case CmpInst::FCMP_OLE: NeedSwap = true; LLVM_FALLTHROUGH; 2310 case CmpInst::FCMP_OGE: CC = X86::COND_AE; break; 2311 case CmpInst::FCMP_UGT: NeedSwap = true; LLVM_FALLTHROUGH; 2312 case CmpInst::FCMP_ULT: CC = X86::COND_B; break; 2313 case CmpInst::FCMP_UGE: NeedSwap = true; LLVM_FALLTHROUGH; 2314 case CmpInst::FCMP_ULE: CC = X86::COND_BE; break; 2315 case CmpInst::FCMP_ONE: CC = X86::COND_NE; break; 2316 case CmpInst::FCMP_UNO: CC = X86::COND_P; break; 2317 case CmpInst::FCMP_ORD: CC = X86::COND_NP; break; 2318 case CmpInst::FCMP_OEQ: LLVM_FALLTHROUGH; 2319 case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break; 2320 2321 // Integer Predicates 2322 case CmpInst::ICMP_EQ: CC = X86::COND_E; break; 2323 case CmpInst::ICMP_NE: CC = X86::COND_NE; break; 2324 case CmpInst::ICMP_UGT: CC = X86::COND_A; break; 2325 case CmpInst::ICMP_UGE: CC = X86::COND_AE; break; 2326 case CmpInst::ICMP_ULT: CC = X86::COND_B; break; 2327 case CmpInst::ICMP_ULE: CC = X86::COND_BE; break; 2328 case CmpInst::ICMP_SGT: CC = X86::COND_G; break; 2329 case CmpInst::ICMP_SGE: CC = X86::COND_GE; break; 2330 case CmpInst::ICMP_SLT: CC = X86::COND_L; break; 2331 case CmpInst::ICMP_SLE: CC = X86::COND_LE; break; 2332 } 2333 2334 return std::make_pair(CC, NeedSwap); 2335 } 2336 2337 /// Return a set opcode for the given condition and 2338 /// whether it has memory operand. 2339 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) { 2340 static const uint16_t Opc[16][2] = { 2341 { X86::SETAr, X86::SETAm }, 2342 { X86::SETAEr, X86::SETAEm }, 2343 { X86::SETBr, X86::SETBm }, 2344 { X86::SETBEr, X86::SETBEm }, 2345 { X86::SETEr, X86::SETEm }, 2346 { X86::SETGr, X86::SETGm }, 2347 { X86::SETGEr, X86::SETGEm }, 2348 { X86::SETLr, X86::SETLm }, 2349 { X86::SETLEr, X86::SETLEm }, 2350 { X86::SETNEr, X86::SETNEm }, 2351 { X86::SETNOr, X86::SETNOm }, 2352 { X86::SETNPr, X86::SETNPm }, 2353 { X86::SETNSr, X86::SETNSm }, 2354 { X86::SETOr, X86::SETOm }, 2355 { X86::SETPr, X86::SETPm }, 2356 { X86::SETSr, X86::SETSm } 2357 }; 2358 2359 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes"); 2360 return Opc[CC][HasMemoryOperand ? 1 : 0]; 2361 } 2362 2363 /// Return a cmov opcode for the given condition, 2364 /// register size in bytes, and operand type. 2365 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes, 2366 bool HasMemoryOperand) { 2367 static const uint16_t Opc[32][3] = { 2368 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr }, 2369 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr }, 2370 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr }, 2371 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr }, 2372 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr }, 2373 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr }, 2374 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr }, 2375 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr }, 2376 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr }, 2377 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr }, 2378 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr }, 2379 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr }, 2380 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr }, 2381 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr }, 2382 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr }, 2383 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr }, 2384 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm }, 2385 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm }, 2386 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm }, 2387 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm }, 2388 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm }, 2389 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm }, 2390 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm }, 2391 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm }, 2392 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm }, 2393 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm }, 2394 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm }, 2395 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm }, 2396 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm }, 2397 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm }, 2398 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm }, 2399 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm } 2400 }; 2401 2402 assert(CC < 16 && "Can only handle standard cond codes"); 2403 unsigned Idx = HasMemoryOperand ? 16+CC : CC; 2404 switch(RegBytes) { 2405 default: llvm_unreachable("Illegal register size!"); 2406 case 2: return Opc[Idx][0]; 2407 case 4: return Opc[Idx][1]; 2408 case 8: return Opc[Idx][2]; 2409 } 2410 } 2411 2412 /// Get the VPCMP immediate for the given condition. 2413 unsigned X86::getVPCMPImmForCond(ISD::CondCode CC) { 2414 switch (CC) { 2415 default: llvm_unreachable("Unexpected SETCC condition"); 2416 case ISD::SETNE: return 4; 2417 case ISD::SETEQ: return 0; 2418 case ISD::SETULT: 2419 case ISD::SETLT: return 1; 2420 case ISD::SETUGT: 2421 case ISD::SETGT: return 6; 2422 case ISD::SETUGE: 2423 case ISD::SETGE: return 5; 2424 case ISD::SETULE: 2425 case ISD::SETLE: return 2; 2426 } 2427 } 2428 2429 /// Get the VPCMP immediate if the opcodes are swapped. 2430 unsigned X86::getSwappedVPCMPImm(unsigned Imm) { 2431 switch (Imm) { 2432 default: llvm_unreachable("Unreachable!"); 2433 case 0x01: Imm = 0x06; break; // LT -> NLE 2434 case 0x02: Imm = 0x05; break; // LE -> NLT 2435 case 0x05: Imm = 0x02; break; // NLT -> LE 2436 case 0x06: Imm = 0x01; break; // NLE -> LT 2437 case 0x00: // EQ 2438 case 0x03: // FALSE 2439 case 0x04: // NE 2440 case 0x07: // TRUE 2441 break; 2442 } 2443 2444 return Imm; 2445 } 2446 2447 /// Get the VPCOM immediate if the opcodes are swapped. 2448 unsigned X86::getSwappedVPCOMImm(unsigned Imm) { 2449 switch (Imm) { 2450 default: llvm_unreachable("Unreachable!"); 2451 case 0x00: Imm = 0x02; break; // LT -> GT 2452 case 0x01: Imm = 0x03; break; // LE -> GE 2453 case 0x02: Imm = 0x00; break; // GT -> LT 2454 case 0x03: Imm = 0x01; break; // GE -> LE 2455 case 0x04: // EQ 2456 case 0x05: // NE 2457 case 0x06: // FALSE 2458 case 0x07: // TRUE 2459 break; 2460 } 2461 2462 return Imm; 2463 } 2464 2465 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { 2466 if (!MI.isTerminator()) return false; 2467 2468 // Conditional branch is a special case. 2469 if (MI.isBranch() && !MI.isBarrier()) 2470 return true; 2471 if (!MI.isPredicable()) 2472 return true; 2473 return !isPredicated(MI); 2474 } 2475 2476 bool X86InstrInfo::isUnconditionalTailCall(const MachineInstr &MI) const { 2477 switch (MI.getOpcode()) { 2478 case X86::TCRETURNdi: 2479 case X86::TCRETURNri: 2480 case X86::TCRETURNmi: 2481 case X86::TCRETURNdi64: 2482 case X86::TCRETURNri64: 2483 case X86::TCRETURNmi64: 2484 return true; 2485 default: 2486 return false; 2487 } 2488 } 2489 2490 bool X86InstrInfo::canMakeTailCallConditional( 2491 SmallVectorImpl<MachineOperand> &BranchCond, 2492 const MachineInstr &TailCall) const { 2493 if (TailCall.getOpcode() != X86::TCRETURNdi && 2494 TailCall.getOpcode() != X86::TCRETURNdi64) { 2495 // Only direct calls can be done with a conditional branch. 2496 return false; 2497 } 2498 2499 const MachineFunction *MF = TailCall.getParent()->getParent(); 2500 if (Subtarget.isTargetWin64() && MF->hasWinCFI()) { 2501 // Conditional tail calls confuse the Win64 unwinder. 2502 return false; 2503 } 2504 2505 assert(BranchCond.size() == 1); 2506 if (BranchCond[0].getImm() > X86::LAST_VALID_COND) { 2507 // Can't make a conditional tail call with this condition. 2508 return false; 2509 } 2510 2511 const X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); 2512 if (X86FI->getTCReturnAddrDelta() != 0 || 2513 TailCall.getOperand(1).getImm() != 0) { 2514 // A conditional tail call cannot do any stack adjustment. 2515 return false; 2516 } 2517 2518 return true; 2519 } 2520 2521 void X86InstrInfo::replaceBranchWithTailCall( 2522 MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &BranchCond, 2523 const MachineInstr &TailCall) const { 2524 assert(canMakeTailCallConditional(BranchCond, TailCall)); 2525 2526 MachineBasicBlock::iterator I = MBB.end(); 2527 while (I != MBB.begin()) { 2528 --I; 2529 if (I->isDebugInstr()) 2530 continue; 2531 if (!I->isBranch()) 2532 assert(0 && "Can't find the branch to replace!"); 2533 2534 X86::CondCode CC = X86::getCondFromBranchOpc(I->getOpcode()); 2535 assert(BranchCond.size() == 1); 2536 if (CC != BranchCond[0].getImm()) 2537 continue; 2538 2539 break; 2540 } 2541 2542 unsigned Opc = TailCall.getOpcode() == X86::TCRETURNdi ? X86::TCRETURNdicc 2543 : X86::TCRETURNdi64cc; 2544 2545 auto MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opc)); 2546 MIB->addOperand(TailCall.getOperand(0)); // Destination. 2547 MIB.addImm(0); // Stack offset (not used). 2548 MIB->addOperand(BranchCond[0]); // Condition. 2549 MIB.copyImplicitOps(TailCall); // Regmask and (imp-used) parameters. 2550 2551 // Add implicit uses and defs of all live regs potentially clobbered by the 2552 // call. This way they still appear live across the call. 2553 LivePhysRegs LiveRegs(getRegisterInfo()); 2554 LiveRegs.addLiveOuts(MBB); 2555 SmallVector<std::pair<unsigned, const MachineOperand *>, 8> Clobbers; 2556 LiveRegs.stepForward(*MIB, Clobbers); 2557 for (const auto &C : Clobbers) { 2558 MIB.addReg(C.first, RegState::Implicit); 2559 MIB.addReg(C.first, RegState::Implicit | RegState::Define); 2560 } 2561 2562 I->eraseFromParent(); 2563 } 2564 2565 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may 2566 // not be a fallthrough MBB now due to layout changes). Return nullptr if the 2567 // fallthrough MBB cannot be identified. 2568 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB, 2569 MachineBasicBlock *TBB) { 2570 // Look for non-EHPad successors other than TBB. If we find exactly one, it 2571 // is the fallthrough MBB. If we find zero, then TBB is both the target MBB 2572 // and fallthrough MBB. If we find more than one, we cannot identify the 2573 // fallthrough MBB and should return nullptr. 2574 MachineBasicBlock *FallthroughBB = nullptr; 2575 for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) { 2576 if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB)) 2577 continue; 2578 // Return a nullptr if we found more than one fallthrough successor. 2579 if (FallthroughBB && FallthroughBB != TBB) 2580 return nullptr; 2581 FallthroughBB = *SI; 2582 } 2583 return FallthroughBB; 2584 } 2585 2586 bool X86InstrInfo::AnalyzeBranchImpl( 2587 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, 2588 SmallVectorImpl<MachineOperand> &Cond, 2589 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const { 2590 2591 // Start from the bottom of the block and work up, examining the 2592 // terminator instructions. 2593 MachineBasicBlock::iterator I = MBB.end(); 2594 MachineBasicBlock::iterator UnCondBrIter = MBB.end(); 2595 while (I != MBB.begin()) { 2596 --I; 2597 if (I->isDebugInstr()) 2598 continue; 2599 2600 // Working from the bottom, when we see a non-terminator instruction, we're 2601 // done. 2602 if (!isUnpredicatedTerminator(*I)) 2603 break; 2604 2605 // A terminator that isn't a branch can't easily be handled by this 2606 // analysis. 2607 if (!I->isBranch()) 2608 return true; 2609 2610 // Handle unconditional branches. 2611 if (I->getOpcode() == X86::JMP_1) { 2612 UnCondBrIter = I; 2613 2614 if (!AllowModify) { 2615 TBB = I->getOperand(0).getMBB(); 2616 continue; 2617 } 2618 2619 // If the block has any instructions after a JMP, delete them. 2620 while (std::next(I) != MBB.end()) 2621 std::next(I)->eraseFromParent(); 2622 2623 Cond.clear(); 2624 FBB = nullptr; 2625 2626 // Delete the JMP if it's equivalent to a fall-through. 2627 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 2628 TBB = nullptr; 2629 I->eraseFromParent(); 2630 I = MBB.end(); 2631 UnCondBrIter = MBB.end(); 2632 continue; 2633 } 2634 2635 // TBB is used to indicate the unconditional destination. 2636 TBB = I->getOperand(0).getMBB(); 2637 continue; 2638 } 2639 2640 // Handle conditional branches. 2641 X86::CondCode BranchCode = X86::getCondFromBranchOpc(I->getOpcode()); 2642 if (BranchCode == X86::COND_INVALID) 2643 return true; // Can't handle indirect branch. 2644 2645 // In practice we should never have an undef eflags operand, if we do 2646 // abort here as we are not prepared to preserve the flag. 2647 if (I->getOperand(1).isUndef()) 2648 return true; 2649 2650 // Working from the bottom, handle the first conditional branch. 2651 if (Cond.empty()) { 2652 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB(); 2653 if (AllowModify && UnCondBrIter != MBB.end() && 2654 MBB.isLayoutSuccessor(TargetBB)) { 2655 // If we can modify the code and it ends in something like: 2656 // 2657 // jCC L1 2658 // jmp L2 2659 // L1: 2660 // ... 2661 // L2: 2662 // 2663 // Then we can change this to: 2664 // 2665 // jnCC L2 2666 // L1: 2667 // ... 2668 // L2: 2669 // 2670 // Which is a bit more efficient. 2671 // We conditionally jump to the fall-through block. 2672 BranchCode = GetOppositeBranchCondition(BranchCode); 2673 unsigned JNCC = GetCondBranchFromCond(BranchCode); 2674 MachineBasicBlock::iterator OldInst = I; 2675 2676 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC)) 2677 .addMBB(UnCondBrIter->getOperand(0).getMBB()); 2678 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1)) 2679 .addMBB(TargetBB); 2680 2681 OldInst->eraseFromParent(); 2682 UnCondBrIter->eraseFromParent(); 2683 2684 // Restart the analysis. 2685 UnCondBrIter = MBB.end(); 2686 I = MBB.end(); 2687 continue; 2688 } 2689 2690 FBB = TBB; 2691 TBB = I->getOperand(0).getMBB(); 2692 Cond.push_back(MachineOperand::CreateImm(BranchCode)); 2693 CondBranches.push_back(&*I); 2694 continue; 2695 } 2696 2697 // Handle subsequent conditional branches. Only handle the case where all 2698 // conditional branches branch to the same destination and their condition 2699 // opcodes fit one of the special multi-branch idioms. 2700 assert(Cond.size() == 1); 2701 assert(TBB); 2702 2703 // If the conditions are the same, we can leave them alone. 2704 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm(); 2705 auto NewTBB = I->getOperand(0).getMBB(); 2706 if (OldBranchCode == BranchCode && TBB == NewTBB) 2707 continue; 2708 2709 // If they differ, see if they fit one of the known patterns. Theoretically, 2710 // we could handle more patterns here, but we shouldn't expect to see them 2711 // if instruction selection has done a reasonable job. 2712 if (TBB == NewTBB && 2713 ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) || 2714 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) { 2715 BranchCode = X86::COND_NE_OR_P; 2716 } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) || 2717 (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) { 2718 if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB))) 2719 return true; 2720 2721 // X86::COND_E_AND_NP usually has two different branch destinations. 2722 // 2723 // JP B1 2724 // JE B2 2725 // JMP B1 2726 // B1: 2727 // B2: 2728 // 2729 // Here this condition branches to B2 only if NP && E. It has another 2730 // equivalent form: 2731 // 2732 // JNE B1 2733 // JNP B2 2734 // JMP B1 2735 // B1: 2736 // B2: 2737 // 2738 // Similarly it branches to B2 only if E && NP. That is why this condition 2739 // is named with COND_E_AND_NP. 2740 BranchCode = X86::COND_E_AND_NP; 2741 } else 2742 return true; 2743 2744 // Update the MachineOperand. 2745 Cond[0].setImm(BranchCode); 2746 CondBranches.push_back(&*I); 2747 } 2748 2749 return false; 2750 } 2751 2752 bool X86InstrInfo::analyzeBranch(MachineBasicBlock &MBB, 2753 MachineBasicBlock *&TBB, 2754 MachineBasicBlock *&FBB, 2755 SmallVectorImpl<MachineOperand> &Cond, 2756 bool AllowModify) const { 2757 SmallVector<MachineInstr *, 4> CondBranches; 2758 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify); 2759 } 2760 2761 bool X86InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB, 2762 MachineBranchPredicate &MBP, 2763 bool AllowModify) const { 2764 using namespace std::placeholders; 2765 2766 SmallVector<MachineOperand, 4> Cond; 2767 SmallVector<MachineInstr *, 4> CondBranches; 2768 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches, 2769 AllowModify)) 2770 return true; 2771 2772 if (Cond.size() != 1) 2773 return true; 2774 2775 assert(MBP.TrueDest && "expected!"); 2776 2777 if (!MBP.FalseDest) 2778 MBP.FalseDest = MBB.getNextNode(); 2779 2780 const TargetRegisterInfo *TRI = &getRegisterInfo(); 2781 2782 MachineInstr *ConditionDef = nullptr; 2783 bool SingleUseCondition = true; 2784 2785 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) { 2786 if (I->modifiesRegister(X86::EFLAGS, TRI)) { 2787 ConditionDef = &*I; 2788 break; 2789 } 2790 2791 if (I->readsRegister(X86::EFLAGS, TRI)) 2792 SingleUseCondition = false; 2793 } 2794 2795 if (!ConditionDef) 2796 return true; 2797 2798 if (SingleUseCondition) { 2799 for (auto *Succ : MBB.successors()) 2800 if (Succ->isLiveIn(X86::EFLAGS)) 2801 SingleUseCondition = false; 2802 } 2803 2804 MBP.ConditionDef = ConditionDef; 2805 MBP.SingleUseCondition = SingleUseCondition; 2806 2807 // Currently we only recognize the simple pattern: 2808 // 2809 // test %reg, %reg 2810 // je %label 2811 // 2812 const unsigned TestOpcode = 2813 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr; 2814 2815 if (ConditionDef->getOpcode() == TestOpcode && 2816 ConditionDef->getNumOperands() == 3 && 2817 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) && 2818 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) { 2819 MBP.LHS = ConditionDef->getOperand(0); 2820 MBP.RHS = MachineOperand::CreateImm(0); 2821 MBP.Predicate = Cond[0].getImm() == X86::COND_NE 2822 ? MachineBranchPredicate::PRED_NE 2823 : MachineBranchPredicate::PRED_EQ; 2824 return false; 2825 } 2826 2827 return true; 2828 } 2829 2830 unsigned X86InstrInfo::removeBranch(MachineBasicBlock &MBB, 2831 int *BytesRemoved) const { 2832 assert(!BytesRemoved && "code size not handled"); 2833 2834 MachineBasicBlock::iterator I = MBB.end(); 2835 unsigned Count = 0; 2836 2837 while (I != MBB.begin()) { 2838 --I; 2839 if (I->isDebugInstr()) 2840 continue; 2841 if (I->getOpcode() != X86::JMP_1 && 2842 X86::getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID) 2843 break; 2844 // Remove the branch. 2845 I->eraseFromParent(); 2846 I = MBB.end(); 2847 ++Count; 2848 } 2849 2850 return Count; 2851 } 2852 2853 unsigned X86InstrInfo::insertBranch(MachineBasicBlock &MBB, 2854 MachineBasicBlock *TBB, 2855 MachineBasicBlock *FBB, 2856 ArrayRef<MachineOperand> Cond, 2857 const DebugLoc &DL, 2858 int *BytesAdded) const { 2859 // Shouldn't be a fall through. 2860 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 2861 assert((Cond.size() == 1 || Cond.size() == 0) && 2862 "X86 branch conditions have one component!"); 2863 assert(!BytesAdded && "code size not handled"); 2864 2865 if (Cond.empty()) { 2866 // Unconditional branch? 2867 assert(!FBB && "Unconditional branch with multiple successors!"); 2868 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB); 2869 return 1; 2870 } 2871 2872 // If FBB is null, it is implied to be a fall-through block. 2873 bool FallThru = FBB == nullptr; 2874 2875 // Conditional branch. 2876 unsigned Count = 0; 2877 X86::CondCode CC = (X86::CondCode)Cond[0].getImm(); 2878 switch (CC) { 2879 case X86::COND_NE_OR_P: 2880 // Synthesize NE_OR_P with two branches. 2881 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB); 2882 ++Count; 2883 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB); 2884 ++Count; 2885 break; 2886 case X86::COND_E_AND_NP: 2887 // Use the next block of MBB as FBB if it is null. 2888 if (FBB == nullptr) { 2889 FBB = getFallThroughMBB(&MBB, TBB); 2890 assert(FBB && "MBB cannot be the last block in function when the false " 2891 "body is a fall-through."); 2892 } 2893 // Synthesize COND_E_AND_NP with two branches. 2894 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB); 2895 ++Count; 2896 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB); 2897 ++Count; 2898 break; 2899 default: { 2900 unsigned Opc = GetCondBranchFromCond(CC); 2901 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB); 2902 ++Count; 2903 } 2904 } 2905 if (!FallThru) { 2906 // Two-way Conditional branch. Insert the second branch. 2907 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB); 2908 ++Count; 2909 } 2910 return Count; 2911 } 2912 2913 bool X86InstrInfo:: 2914 canInsertSelect(const MachineBasicBlock &MBB, 2915 ArrayRef<MachineOperand> Cond, 2916 unsigned TrueReg, unsigned FalseReg, 2917 int &CondCycles, int &TrueCycles, int &FalseCycles) const { 2918 // Not all subtargets have cmov instructions. 2919 if (!Subtarget.hasCMov()) 2920 return false; 2921 if (Cond.size() != 1) 2922 return false; 2923 // We cannot do the composite conditions, at least not in SSA form. 2924 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S) 2925 return false; 2926 2927 // Check register classes. 2928 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 2929 const TargetRegisterClass *RC = 2930 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 2931 if (!RC) 2932 return false; 2933 2934 // We have cmov instructions for 16, 32, and 64 bit general purpose registers. 2935 if (X86::GR16RegClass.hasSubClassEq(RC) || 2936 X86::GR32RegClass.hasSubClassEq(RC) || 2937 X86::GR64RegClass.hasSubClassEq(RC)) { 2938 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy 2939 // Bridge. Probably Ivy Bridge as well. 2940 CondCycles = 2; 2941 TrueCycles = 2; 2942 FalseCycles = 2; 2943 return true; 2944 } 2945 2946 // Can't do vectors. 2947 return false; 2948 } 2949 2950 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB, 2951 MachineBasicBlock::iterator I, 2952 const DebugLoc &DL, unsigned DstReg, 2953 ArrayRef<MachineOperand> Cond, unsigned TrueReg, 2954 unsigned FalseReg) const { 2955 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 2956 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo(); 2957 const TargetRegisterClass &RC = *MRI.getRegClass(DstReg); 2958 assert(Cond.size() == 1 && "Invalid Cond array"); 2959 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(), 2960 TRI.getRegSizeInBits(RC) / 8, 2961 false /*HasMemoryOperand*/); 2962 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg); 2963 } 2964 2965 /// Test if the given register is a physical h register. 2966 static bool isHReg(unsigned Reg) { 2967 return X86::GR8_ABCD_HRegClass.contains(Reg); 2968 } 2969 2970 // Try and copy between VR128/VR64 and GR64 registers. 2971 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg, 2972 const X86Subtarget &Subtarget) { 2973 bool HasAVX = Subtarget.hasAVX(); 2974 bool HasAVX512 = Subtarget.hasAVX512(); 2975 2976 // SrcReg(MaskReg) -> DestReg(GR64) 2977 // SrcReg(MaskReg) -> DestReg(GR32) 2978 2979 // All KMASK RegClasses hold the same k registers, can be tested against anyone. 2980 if (X86::VK16RegClass.contains(SrcReg)) { 2981 if (X86::GR64RegClass.contains(DestReg)) { 2982 assert(Subtarget.hasBWI()); 2983 return X86::KMOVQrk; 2984 } 2985 if (X86::GR32RegClass.contains(DestReg)) 2986 return Subtarget.hasBWI() ? X86::KMOVDrk : X86::KMOVWrk; 2987 } 2988 2989 // SrcReg(GR64) -> DestReg(MaskReg) 2990 // SrcReg(GR32) -> DestReg(MaskReg) 2991 2992 // All KMASK RegClasses hold the same k registers, can be tested against anyone. 2993 if (X86::VK16RegClass.contains(DestReg)) { 2994 if (X86::GR64RegClass.contains(SrcReg)) { 2995 assert(Subtarget.hasBWI()); 2996 return X86::KMOVQkr; 2997 } 2998 if (X86::GR32RegClass.contains(SrcReg)) 2999 return Subtarget.hasBWI() ? X86::KMOVDkr : X86::KMOVWkr; 3000 } 3001 3002 3003 // SrcReg(VR128) -> DestReg(GR64) 3004 // SrcReg(VR64) -> DestReg(GR64) 3005 // SrcReg(GR64) -> DestReg(VR128) 3006 // SrcReg(GR64) -> DestReg(VR64) 3007 3008 if (X86::GR64RegClass.contains(DestReg)) { 3009 if (X86::VR128XRegClass.contains(SrcReg)) 3010 // Copy from a VR128 register to a GR64 register. 3011 return HasAVX512 ? X86::VMOVPQIto64Zrr : 3012 HasAVX ? X86::VMOVPQIto64rr : 3013 X86::MOVPQIto64rr; 3014 if (X86::VR64RegClass.contains(SrcReg)) 3015 // Copy from a VR64 register to a GR64 register. 3016 return X86::MMX_MOVD64from64rr; 3017 } else if (X86::GR64RegClass.contains(SrcReg)) { 3018 // Copy from a GR64 register to a VR128 register. 3019 if (X86::VR128XRegClass.contains(DestReg)) 3020 return HasAVX512 ? X86::VMOV64toPQIZrr : 3021 HasAVX ? X86::VMOV64toPQIrr : 3022 X86::MOV64toPQIrr; 3023 // Copy from a GR64 register to a VR64 register. 3024 if (X86::VR64RegClass.contains(DestReg)) 3025 return X86::MMX_MOVD64to64rr; 3026 } 3027 3028 // SrcReg(FR32) -> DestReg(GR32) 3029 // SrcReg(GR32) -> DestReg(FR32) 3030 3031 if (X86::GR32RegClass.contains(DestReg) && 3032 X86::FR32XRegClass.contains(SrcReg)) 3033 // Copy from a FR32 register to a GR32 register. 3034 return HasAVX512 ? X86::VMOVSS2DIZrr : 3035 HasAVX ? X86::VMOVSS2DIrr : 3036 X86::MOVSS2DIrr; 3037 3038 if (X86::FR32XRegClass.contains(DestReg) && 3039 X86::GR32RegClass.contains(SrcReg)) 3040 // Copy from a GR32 register to a FR32 register. 3041 return HasAVX512 ? X86::VMOVDI2SSZrr : 3042 HasAVX ? X86::VMOVDI2SSrr : 3043 X86::MOVDI2SSrr; 3044 return 0; 3045 } 3046 3047 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB, 3048 MachineBasicBlock::iterator MI, 3049 const DebugLoc &DL, unsigned DestReg, 3050 unsigned SrcReg, bool KillSrc) const { 3051 // First deal with the normal symmetric copies. 3052 bool HasAVX = Subtarget.hasAVX(); 3053 bool HasVLX = Subtarget.hasVLX(); 3054 unsigned Opc = 0; 3055 if (X86::GR64RegClass.contains(DestReg, SrcReg)) 3056 Opc = X86::MOV64rr; 3057 else if (X86::GR32RegClass.contains(DestReg, SrcReg)) 3058 Opc = X86::MOV32rr; 3059 else if (X86::GR16RegClass.contains(DestReg, SrcReg)) 3060 Opc = X86::MOV16rr; 3061 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) { 3062 // Copying to or from a physical H register on x86-64 requires a NOREX 3063 // move. Otherwise use a normal move. 3064 if ((isHReg(DestReg) || isHReg(SrcReg)) && 3065 Subtarget.is64Bit()) { 3066 Opc = X86::MOV8rr_NOREX; 3067 // Both operands must be encodable without an REX prefix. 3068 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) && 3069 "8-bit H register can not be copied outside GR8_NOREX"); 3070 } else 3071 Opc = X86::MOV8rr; 3072 } 3073 else if (X86::VR64RegClass.contains(DestReg, SrcReg)) 3074 Opc = X86::MMX_MOVQ64rr; 3075 else if (X86::VR128XRegClass.contains(DestReg, SrcReg)) { 3076 if (HasVLX) 3077 Opc = X86::VMOVAPSZ128rr; 3078 else if (X86::VR128RegClass.contains(DestReg, SrcReg)) 3079 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr; 3080 else { 3081 // If this an extended register and we don't have VLX we need to use a 3082 // 512-bit move. 3083 Opc = X86::VMOVAPSZrr; 3084 const TargetRegisterInfo *TRI = &getRegisterInfo(); 3085 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_xmm, 3086 &X86::VR512RegClass); 3087 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, 3088 &X86::VR512RegClass); 3089 } 3090 } else if (X86::VR256XRegClass.contains(DestReg, SrcReg)) { 3091 if (HasVLX) 3092 Opc = X86::VMOVAPSZ256rr; 3093 else if (X86::VR256RegClass.contains(DestReg, SrcReg)) 3094 Opc = X86::VMOVAPSYrr; 3095 else { 3096 // If this an extended register and we don't have VLX we need to use a 3097 // 512-bit move. 3098 Opc = X86::VMOVAPSZrr; 3099 const TargetRegisterInfo *TRI = &getRegisterInfo(); 3100 DestReg = TRI->getMatchingSuperReg(DestReg, X86::sub_ymm, 3101 &X86::VR512RegClass); 3102 SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm, 3103 &X86::VR512RegClass); 3104 } 3105 } else if (X86::VR512RegClass.contains(DestReg, SrcReg)) 3106 Opc = X86::VMOVAPSZrr; 3107 // All KMASK RegClasses hold the same k registers, can be tested against anyone. 3108 else if (X86::VK16RegClass.contains(DestReg, SrcReg)) 3109 Opc = Subtarget.hasBWI() ? X86::KMOVQkk : X86::KMOVWkk; 3110 if (!Opc) 3111 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget); 3112 3113 if (Opc) { 3114 BuildMI(MBB, MI, DL, get(Opc), DestReg) 3115 .addReg(SrcReg, getKillRegState(KillSrc)); 3116 return; 3117 } 3118 3119 if (SrcReg == X86::EFLAGS || DestReg == X86::EFLAGS) { 3120 // FIXME: We use a fatal error here because historically LLVM has tried 3121 // lower some of these physreg copies and we want to ensure we get 3122 // reasonable bug reports if someone encounters a case no other testing 3123 // found. This path should be removed after the LLVM 7 release. 3124 report_fatal_error("Unable to copy EFLAGS physical register!"); 3125 } 3126 3127 LLVM_DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) << " to " 3128 << RI.getName(DestReg) << '\n'); 3129 report_fatal_error("Cannot emit physreg copy instruction"); 3130 } 3131 3132 bool X86InstrInfo::isCopyInstrImpl(const MachineInstr &MI, 3133 const MachineOperand *&Src, 3134 const MachineOperand *&Dest) const { 3135 if (MI.isMoveReg()) { 3136 Dest = &MI.getOperand(0); 3137 Src = &MI.getOperand(1); 3138 return true; 3139 } 3140 return false; 3141 } 3142 3143 static unsigned getLoadStoreRegOpcode(unsigned Reg, 3144 const TargetRegisterClass *RC, 3145 bool isStackAligned, 3146 const X86Subtarget &STI, 3147 bool load) { 3148 bool HasAVX = STI.hasAVX(); 3149 bool HasAVX512 = STI.hasAVX512(); 3150 bool HasVLX = STI.hasVLX(); 3151 3152 switch (STI.getRegisterInfo()->getSpillSize(*RC)) { 3153 default: 3154 llvm_unreachable("Unknown spill size"); 3155 case 1: 3156 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass"); 3157 if (STI.is64Bit()) 3158 // Copying to or from a physical H register on x86-64 requires a NOREX 3159 // move. Otherwise use a normal move. 3160 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC)) 3161 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX; 3162 return load ? X86::MOV8rm : X86::MOV8mr; 3163 case 2: 3164 if (X86::VK16RegClass.hasSubClassEq(RC)) 3165 return load ? X86::KMOVWkm : X86::KMOVWmk; 3166 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass"); 3167 return load ? X86::MOV16rm : X86::MOV16mr; 3168 case 4: 3169 if (X86::GR32RegClass.hasSubClassEq(RC)) 3170 return load ? X86::MOV32rm : X86::MOV32mr; 3171 if (X86::FR32XRegClass.hasSubClassEq(RC)) 3172 return load ? 3173 (HasAVX512 ? X86::VMOVSSZrm : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) : 3174 (HasAVX512 ? X86::VMOVSSZmr : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr); 3175 if (X86::RFP32RegClass.hasSubClassEq(RC)) 3176 return load ? X86::LD_Fp32m : X86::ST_Fp32m; 3177 if (X86::VK32RegClass.hasSubClassEq(RC)) { 3178 assert(STI.hasBWI() && "KMOVD requires BWI"); 3179 return load ? X86::KMOVDkm : X86::KMOVDmk; 3180 } 3181 llvm_unreachable("Unknown 4-byte regclass"); 3182 case 8: 3183 if (X86::GR64RegClass.hasSubClassEq(RC)) 3184 return load ? X86::MOV64rm : X86::MOV64mr; 3185 if (X86::FR64XRegClass.hasSubClassEq(RC)) 3186 return load ? 3187 (HasAVX512 ? X86::VMOVSDZrm : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) : 3188 (HasAVX512 ? X86::VMOVSDZmr : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr); 3189 if (X86::VR64RegClass.hasSubClassEq(RC)) 3190 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr; 3191 if (X86::RFP64RegClass.hasSubClassEq(RC)) 3192 return load ? X86::LD_Fp64m : X86::ST_Fp64m; 3193 if (X86::VK64RegClass.hasSubClassEq(RC)) { 3194 assert(STI.hasBWI() && "KMOVQ requires BWI"); 3195 return load ? X86::KMOVQkm : X86::KMOVQmk; 3196 } 3197 llvm_unreachable("Unknown 8-byte regclass"); 3198 case 10: 3199 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass"); 3200 return load ? X86::LD_Fp80m : X86::ST_FpP80m; 3201 case 16: { 3202 if (X86::VR128XRegClass.hasSubClassEq(RC)) { 3203 // If stack is realigned we can use aligned stores. 3204 if (isStackAligned) 3205 return load ? 3206 (HasVLX ? X86::VMOVAPSZ128rm : 3207 HasAVX512 ? X86::VMOVAPSZ128rm_NOVLX : 3208 HasAVX ? X86::VMOVAPSrm : 3209 X86::MOVAPSrm): 3210 (HasVLX ? X86::VMOVAPSZ128mr : 3211 HasAVX512 ? X86::VMOVAPSZ128mr_NOVLX : 3212 HasAVX ? X86::VMOVAPSmr : 3213 X86::MOVAPSmr); 3214 else 3215 return load ? 3216 (HasVLX ? X86::VMOVUPSZ128rm : 3217 HasAVX512 ? X86::VMOVUPSZ128rm_NOVLX : 3218 HasAVX ? X86::VMOVUPSrm : 3219 X86::MOVUPSrm): 3220 (HasVLX ? X86::VMOVUPSZ128mr : 3221 HasAVX512 ? X86::VMOVUPSZ128mr_NOVLX : 3222 HasAVX ? X86::VMOVUPSmr : 3223 X86::MOVUPSmr); 3224 } 3225 if (X86::BNDRRegClass.hasSubClassEq(RC)) { 3226 if (STI.is64Bit()) 3227 return load ? X86::BNDMOV64rm : X86::BNDMOV64mr; 3228 else 3229 return load ? X86::BNDMOV32rm : X86::BNDMOV32mr; 3230 } 3231 llvm_unreachable("Unknown 16-byte regclass"); 3232 } 3233 case 32: 3234 assert(X86::VR256XRegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass"); 3235 // If stack is realigned we can use aligned stores. 3236 if (isStackAligned) 3237 return load ? 3238 (HasVLX ? X86::VMOVAPSZ256rm : 3239 HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX : 3240 X86::VMOVAPSYrm) : 3241 (HasVLX ? X86::VMOVAPSZ256mr : 3242 HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX : 3243 X86::VMOVAPSYmr); 3244 else 3245 return load ? 3246 (HasVLX ? X86::VMOVUPSZ256rm : 3247 HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX : 3248 X86::VMOVUPSYrm) : 3249 (HasVLX ? X86::VMOVUPSZ256mr : 3250 HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX : 3251 X86::VMOVUPSYmr); 3252 case 64: 3253 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass"); 3254 assert(STI.hasAVX512() && "Using 512-bit register requires AVX512"); 3255 if (isStackAligned) 3256 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr; 3257 else 3258 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr; 3259 } 3260 } 3261 3262 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr &MemOp, unsigned &BaseReg, 3263 int64_t &Offset, 3264 const TargetRegisterInfo *TRI) const { 3265 const MCInstrDesc &Desc = MemOp.getDesc(); 3266 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags); 3267 if (MemRefBegin < 0) 3268 return false; 3269 3270 MemRefBegin += X86II::getOperandBias(Desc); 3271 3272 MachineOperand &BaseMO = MemOp.getOperand(MemRefBegin + X86::AddrBaseReg); 3273 if (!BaseMO.isReg()) // Can be an MO_FrameIndex 3274 return false; 3275 3276 BaseReg = BaseMO.getReg(); 3277 if (MemOp.getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1) 3278 return false; 3279 3280 if (MemOp.getOperand(MemRefBegin + X86::AddrIndexReg).getReg() != 3281 X86::NoRegister) 3282 return false; 3283 3284 const MachineOperand &DispMO = MemOp.getOperand(MemRefBegin + X86::AddrDisp); 3285 3286 // Displacement can be symbolic 3287 if (!DispMO.isImm()) 3288 return false; 3289 3290 Offset = DispMO.getImm(); 3291 3292 return true; 3293 } 3294 3295 static unsigned getStoreRegOpcode(unsigned SrcReg, 3296 const TargetRegisterClass *RC, 3297 bool isStackAligned, 3298 const X86Subtarget &STI) { 3299 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false); 3300 } 3301 3302 3303 static unsigned getLoadRegOpcode(unsigned DestReg, 3304 const TargetRegisterClass *RC, 3305 bool isStackAligned, 3306 const X86Subtarget &STI) { 3307 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true); 3308 } 3309 3310 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 3311 MachineBasicBlock::iterator MI, 3312 unsigned SrcReg, bool isKill, int FrameIdx, 3313 const TargetRegisterClass *RC, 3314 const TargetRegisterInfo *TRI) const { 3315 const MachineFunction &MF = *MBB.getParent(); 3316 assert(MF.getFrameInfo().getObjectSize(FrameIdx) >= TRI->getSpillSize(*RC) && 3317 "Stack slot too small for store"); 3318 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16); 3319 bool isAligned = 3320 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) || 3321 RI.canRealignStack(MF); 3322 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget); 3323 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc)), FrameIdx) 3324 .addReg(SrcReg, getKillRegState(isKill)); 3325 } 3326 3327 void X86InstrInfo::storeRegToAddr( 3328 MachineFunction &MF, unsigned SrcReg, bool isKill, 3329 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC, 3330 ArrayRef<MachineMemOperand *> MMOs, 3331 SmallVectorImpl<MachineInstr *> &NewMIs) const { 3332 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 3333 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16); 3334 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment; 3335 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget); 3336 DebugLoc DL; 3337 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc)); 3338 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 3339 MIB.add(Addr[i]); 3340 MIB.addReg(SrcReg, getKillRegState(isKill)); 3341 MIB.setMemRefs(MMOs); 3342 NewMIs.push_back(MIB); 3343 } 3344 3345 3346 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 3347 MachineBasicBlock::iterator MI, 3348 unsigned DestReg, int FrameIdx, 3349 const TargetRegisterClass *RC, 3350 const TargetRegisterInfo *TRI) const { 3351 const MachineFunction &MF = *MBB.getParent(); 3352 unsigned Alignment = std::max<uint32_t>(TRI->getSpillSize(*RC), 16); 3353 bool isAligned = 3354 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) || 3355 RI.canRealignStack(MF); 3356 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget); 3357 addFrameReference(BuildMI(MBB, MI, DebugLoc(), get(Opc), DestReg), FrameIdx); 3358 } 3359 3360 void X86InstrInfo::loadRegFromAddr( 3361 MachineFunction &MF, unsigned DestReg, 3362 SmallVectorImpl<MachineOperand> &Addr, const TargetRegisterClass *RC, 3363 ArrayRef<MachineMemOperand *> MMOs, 3364 SmallVectorImpl<MachineInstr *> &NewMIs) const { 3365 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 3366 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16); 3367 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment; 3368 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget); 3369 DebugLoc DL; 3370 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg); 3371 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 3372 MIB.add(Addr[i]); 3373 MIB.setMemRefs(MMOs); 3374 NewMIs.push_back(MIB); 3375 } 3376 3377 bool X86InstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 3378 unsigned &SrcReg2, int &CmpMask, 3379 int &CmpValue) const { 3380 switch (MI.getOpcode()) { 3381 default: break; 3382 case X86::CMP64ri32: 3383 case X86::CMP64ri8: 3384 case X86::CMP32ri: 3385 case X86::CMP32ri8: 3386 case X86::CMP16ri: 3387 case X86::CMP16ri8: 3388 case X86::CMP8ri: 3389 SrcReg = MI.getOperand(0).getReg(); 3390 SrcReg2 = 0; 3391 if (MI.getOperand(1).isImm()) { 3392 CmpMask = ~0; 3393 CmpValue = MI.getOperand(1).getImm(); 3394 } else { 3395 CmpMask = CmpValue = 0; 3396 } 3397 return true; 3398 // A SUB can be used to perform comparison. 3399 case X86::SUB64rm: 3400 case X86::SUB32rm: 3401 case X86::SUB16rm: 3402 case X86::SUB8rm: 3403 SrcReg = MI.getOperand(1).getReg(); 3404 SrcReg2 = 0; 3405 CmpMask = 0; 3406 CmpValue = 0; 3407 return true; 3408 case X86::SUB64rr: 3409 case X86::SUB32rr: 3410 case X86::SUB16rr: 3411 case X86::SUB8rr: 3412 SrcReg = MI.getOperand(1).getReg(); 3413 SrcReg2 = MI.getOperand(2).getReg(); 3414 CmpMask = 0; 3415 CmpValue = 0; 3416 return true; 3417 case X86::SUB64ri32: 3418 case X86::SUB64ri8: 3419 case X86::SUB32ri: 3420 case X86::SUB32ri8: 3421 case X86::SUB16ri: 3422 case X86::SUB16ri8: 3423 case X86::SUB8ri: 3424 SrcReg = MI.getOperand(1).getReg(); 3425 SrcReg2 = 0; 3426 if (MI.getOperand(2).isImm()) { 3427 CmpMask = ~0; 3428 CmpValue = MI.getOperand(2).getImm(); 3429 } else { 3430 CmpMask = CmpValue = 0; 3431 } 3432 return true; 3433 case X86::CMP64rr: 3434 case X86::CMP32rr: 3435 case X86::CMP16rr: 3436 case X86::CMP8rr: 3437 SrcReg = MI.getOperand(0).getReg(); 3438 SrcReg2 = MI.getOperand(1).getReg(); 3439 CmpMask = 0; 3440 CmpValue = 0; 3441 return true; 3442 case X86::TEST8rr: 3443 case X86::TEST16rr: 3444 case X86::TEST32rr: 3445 case X86::TEST64rr: 3446 SrcReg = MI.getOperand(0).getReg(); 3447 if (MI.getOperand(1).getReg() != SrcReg) 3448 return false; 3449 // Compare against zero. 3450 SrcReg2 = 0; 3451 CmpMask = ~0; 3452 CmpValue = 0; 3453 return true; 3454 } 3455 return false; 3456 } 3457 3458 /// Check whether the first instruction, whose only 3459 /// purpose is to update flags, can be made redundant. 3460 /// CMPrr can be made redundant by SUBrr if the operands are the same. 3461 /// This function can be extended later on. 3462 /// SrcReg, SrcRegs: register operands for FlagI. 3463 /// ImmValue: immediate for FlagI if it takes an immediate. 3464 inline static bool isRedundantFlagInstr(MachineInstr &FlagI, unsigned SrcReg, 3465 unsigned SrcReg2, int ImmMask, 3466 int ImmValue, MachineInstr &OI) { 3467 if (((FlagI.getOpcode() == X86::CMP64rr && OI.getOpcode() == X86::SUB64rr) || 3468 (FlagI.getOpcode() == X86::CMP32rr && OI.getOpcode() == X86::SUB32rr) || 3469 (FlagI.getOpcode() == X86::CMP16rr && OI.getOpcode() == X86::SUB16rr) || 3470 (FlagI.getOpcode() == X86::CMP8rr && OI.getOpcode() == X86::SUB8rr)) && 3471 ((OI.getOperand(1).getReg() == SrcReg && 3472 OI.getOperand(2).getReg() == SrcReg2) || 3473 (OI.getOperand(1).getReg() == SrcReg2 && 3474 OI.getOperand(2).getReg() == SrcReg))) 3475 return true; 3476 3477 if (ImmMask != 0 && 3478 ((FlagI.getOpcode() == X86::CMP64ri32 && 3479 OI.getOpcode() == X86::SUB64ri32) || 3480 (FlagI.getOpcode() == X86::CMP64ri8 && 3481 OI.getOpcode() == X86::SUB64ri8) || 3482 (FlagI.getOpcode() == X86::CMP32ri && OI.getOpcode() == X86::SUB32ri) || 3483 (FlagI.getOpcode() == X86::CMP32ri8 && 3484 OI.getOpcode() == X86::SUB32ri8) || 3485 (FlagI.getOpcode() == X86::CMP16ri && OI.getOpcode() == X86::SUB16ri) || 3486 (FlagI.getOpcode() == X86::CMP16ri8 && 3487 OI.getOpcode() == X86::SUB16ri8) || 3488 (FlagI.getOpcode() == X86::CMP8ri && OI.getOpcode() == X86::SUB8ri)) && 3489 OI.getOperand(1).getReg() == SrcReg && 3490 OI.getOperand(2).getImm() == ImmValue) 3491 return true; 3492 return false; 3493 } 3494 3495 /// Check whether the definition can be converted 3496 /// to remove a comparison against zero. 3497 inline static bool isDefConvertible(MachineInstr &MI) { 3498 switch (MI.getOpcode()) { 3499 default: return false; 3500 3501 // The shift instructions only modify ZF if their shift count is non-zero. 3502 // N.B.: The processor truncates the shift count depending on the encoding. 3503 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri: 3504 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri: 3505 return getTruncatedShiftCount(MI, 2) != 0; 3506 3507 // Some left shift instructions can be turned into LEA instructions but only 3508 // if their flags aren't used. Avoid transforming such instructions. 3509 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{ 3510 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 3511 if (isTruncatedShiftCountForLEA(ShAmt)) return false; 3512 return ShAmt != 0; 3513 } 3514 3515 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8: 3516 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8: 3517 return getTruncatedShiftCount(MI, 3) != 0; 3518 3519 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri: 3520 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8: 3521 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr: 3522 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm: 3523 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm: 3524 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r: 3525 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri: 3526 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8: 3527 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr: 3528 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm: 3529 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm: 3530 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r: 3531 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri: 3532 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8: 3533 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr: 3534 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm: 3535 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm: 3536 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri: 3537 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8: 3538 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr: 3539 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm: 3540 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm: 3541 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri: 3542 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8: 3543 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr: 3544 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm: 3545 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm: 3546 case X86::ADC64ri32: case X86::ADC64ri8: case X86::ADC32ri: 3547 case X86::ADC32ri8: case X86::ADC16ri: case X86::ADC16ri8: 3548 case X86::ADC8ri: case X86::ADC64rr: case X86::ADC32rr: 3549 case X86::ADC16rr: case X86::ADC8rr: case X86::ADC64rm: 3550 case X86::ADC32rm: case X86::ADC16rm: case X86::ADC8rm: 3551 case X86::SBB64ri32: case X86::SBB64ri8: case X86::SBB32ri: 3552 case X86::SBB32ri8: case X86::SBB16ri: case X86::SBB16ri8: 3553 case X86::SBB8ri: case X86::SBB64rr: case X86::SBB32rr: 3554 case X86::SBB16rr: case X86::SBB8rr: case X86::SBB64rm: 3555 case X86::SBB32rm: case X86::SBB16rm: case X86::SBB8rm: 3556 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r: 3557 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1: 3558 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1: 3559 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1: 3560 case X86::ANDN32rr: case X86::ANDN32rm: 3561 case X86::ANDN64rr: case X86::ANDN64rm: 3562 case X86::BEXTR32rr: case X86::BEXTR64rr: 3563 case X86::BEXTR32rm: case X86::BEXTR64rm: 3564 case X86::BLSI32rr: case X86::BLSI32rm: 3565 case X86::BLSI64rr: case X86::BLSI64rm: 3566 case X86::BLSMSK32rr:case X86::BLSMSK32rm: 3567 case X86::BLSMSK64rr:case X86::BLSMSK64rm: 3568 case X86::BLSR32rr: case X86::BLSR32rm: 3569 case X86::BLSR64rr: case X86::BLSR64rm: 3570 case X86::BZHI32rr: case X86::BZHI32rm: 3571 case X86::BZHI64rr: case X86::BZHI64rm: 3572 case X86::LZCNT16rr: case X86::LZCNT16rm: 3573 case X86::LZCNT32rr: case X86::LZCNT32rm: 3574 case X86::LZCNT64rr: case X86::LZCNT64rm: 3575 case X86::POPCNT16rr:case X86::POPCNT16rm: 3576 case X86::POPCNT32rr:case X86::POPCNT32rm: 3577 case X86::POPCNT64rr:case X86::POPCNT64rm: 3578 case X86::TZCNT16rr: case X86::TZCNT16rm: 3579 case X86::TZCNT32rr: case X86::TZCNT32rm: 3580 case X86::TZCNT64rr: case X86::TZCNT64rm: 3581 case X86::BEXTRI32ri: case X86::BEXTRI32mi: 3582 case X86::BEXTRI64ri: case X86::BEXTRI64mi: 3583 case X86::BLCFILL32rr: case X86::BLCFILL32rm: 3584 case X86::BLCFILL64rr: case X86::BLCFILL64rm: 3585 case X86::BLCI32rr: case X86::BLCI32rm: 3586 case X86::BLCI64rr: case X86::BLCI64rm: 3587 case X86::BLCIC32rr: case X86::BLCIC32rm: 3588 case X86::BLCIC64rr: case X86::BLCIC64rm: 3589 case X86::BLCMSK32rr: case X86::BLCMSK32rm: 3590 case X86::BLCMSK64rr: case X86::BLCMSK64rm: 3591 case X86::BLCS32rr: case X86::BLCS32rm: 3592 case X86::BLCS64rr: case X86::BLCS64rm: 3593 case X86::BLSFILL32rr: case X86::BLSFILL32rm: 3594 case X86::BLSFILL64rr: case X86::BLSFILL64rm: 3595 case X86::BLSIC32rr: case X86::BLSIC32rm: 3596 case X86::BLSIC64rr: case X86::BLSIC64rm: 3597 return true; 3598 } 3599 } 3600 3601 /// Check whether the use can be converted to remove a comparison against zero. 3602 static X86::CondCode isUseDefConvertible(MachineInstr &MI) { 3603 switch (MI.getOpcode()) { 3604 default: return X86::COND_INVALID; 3605 case X86::LZCNT16rr: case X86::LZCNT16rm: 3606 case X86::LZCNT32rr: case X86::LZCNT32rm: 3607 case X86::LZCNT64rr: case X86::LZCNT64rm: 3608 return X86::COND_B; 3609 case X86::POPCNT16rr:case X86::POPCNT16rm: 3610 case X86::POPCNT32rr:case X86::POPCNT32rm: 3611 case X86::POPCNT64rr:case X86::POPCNT64rm: 3612 return X86::COND_E; 3613 case X86::TZCNT16rr: case X86::TZCNT16rm: 3614 case X86::TZCNT32rr: case X86::TZCNT32rm: 3615 case X86::TZCNT64rr: case X86::TZCNT64rm: 3616 return X86::COND_B; 3617 case X86::BSF16rr: 3618 case X86::BSF16rm: 3619 case X86::BSF32rr: 3620 case X86::BSF32rm: 3621 case X86::BSF64rr: 3622 case X86::BSF64rm: 3623 return X86::COND_E; 3624 } 3625 } 3626 3627 /// Check if there exists an earlier instruction that 3628 /// operates on the same source operands and sets flags in the same way as 3629 /// Compare; remove Compare if possible. 3630 bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg, 3631 unsigned SrcReg2, int CmpMask, 3632 int CmpValue, 3633 const MachineRegisterInfo *MRI) const { 3634 // Check whether we can replace SUB with CMP. 3635 unsigned NewOpcode = 0; 3636 switch (CmpInstr.getOpcode()) { 3637 default: break; 3638 case X86::SUB64ri32: 3639 case X86::SUB64ri8: 3640 case X86::SUB32ri: 3641 case X86::SUB32ri8: 3642 case X86::SUB16ri: 3643 case X86::SUB16ri8: 3644 case X86::SUB8ri: 3645 case X86::SUB64rm: 3646 case X86::SUB32rm: 3647 case X86::SUB16rm: 3648 case X86::SUB8rm: 3649 case X86::SUB64rr: 3650 case X86::SUB32rr: 3651 case X86::SUB16rr: 3652 case X86::SUB8rr: { 3653 if (!MRI->use_nodbg_empty(CmpInstr.getOperand(0).getReg())) 3654 return false; 3655 // There is no use of the destination register, we can replace SUB with CMP. 3656 switch (CmpInstr.getOpcode()) { 3657 default: llvm_unreachable("Unreachable!"); 3658 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break; 3659 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break; 3660 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break; 3661 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break; 3662 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break; 3663 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break; 3664 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break; 3665 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break; 3666 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break; 3667 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break; 3668 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break; 3669 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break; 3670 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break; 3671 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break; 3672 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break; 3673 } 3674 CmpInstr.setDesc(get(NewOpcode)); 3675 CmpInstr.RemoveOperand(0); 3676 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri. 3677 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm || 3678 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm) 3679 return false; 3680 } 3681 } 3682 3683 // Get the unique definition of SrcReg. 3684 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 3685 if (!MI) return false; 3686 3687 // CmpInstr is the first instruction of the BB. 3688 MachineBasicBlock::iterator I = CmpInstr, Def = MI; 3689 3690 // If we are comparing against zero, check whether we can use MI to update 3691 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize. 3692 bool IsCmpZero = (CmpMask != 0 && CmpValue == 0); 3693 if (IsCmpZero && MI->getParent() != CmpInstr.getParent()) 3694 return false; 3695 3696 // If we have a use of the source register between the def and our compare 3697 // instruction we can eliminate the compare iff the use sets EFLAGS in the 3698 // right way. 3699 bool ShouldUpdateCC = false; 3700 X86::CondCode NewCC = X86::COND_INVALID; 3701 if (IsCmpZero && !isDefConvertible(*MI)) { 3702 // Scan forward from the use until we hit the use we're looking for or the 3703 // compare instruction. 3704 for (MachineBasicBlock::iterator J = MI;; ++J) { 3705 // Do we have a convertible instruction? 3706 NewCC = isUseDefConvertible(*J); 3707 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() && 3708 J->getOperand(1).getReg() == SrcReg) { 3709 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!"); 3710 ShouldUpdateCC = true; // Update CC later on. 3711 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going 3712 // with the new def. 3713 Def = J; 3714 MI = &*Def; 3715 break; 3716 } 3717 3718 if (J == I) 3719 return false; 3720 } 3721 } 3722 3723 // We are searching for an earlier instruction that can make CmpInstr 3724 // redundant and that instruction will be saved in Sub. 3725 MachineInstr *Sub = nullptr; 3726 const TargetRegisterInfo *TRI = &getRegisterInfo(); 3727 3728 // We iterate backward, starting from the instruction before CmpInstr and 3729 // stop when reaching the definition of a source register or done with the BB. 3730 // RI points to the instruction before CmpInstr. 3731 // If the definition is in this basic block, RE points to the definition; 3732 // otherwise, RE is the rend of the basic block. 3733 MachineBasicBlock::reverse_iterator 3734 RI = ++I.getReverse(), 3735 RE = CmpInstr.getParent() == MI->getParent() 3736 ? Def.getReverse() /* points to MI */ 3737 : CmpInstr.getParent()->rend(); 3738 MachineInstr *Movr0Inst = nullptr; 3739 for (; RI != RE; ++RI) { 3740 MachineInstr &Instr = *RI; 3741 // Check whether CmpInstr can be made redundant by the current instruction. 3742 if (!IsCmpZero && isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpMask, 3743 CmpValue, Instr)) { 3744 Sub = &Instr; 3745 break; 3746 } 3747 3748 if (Instr.modifiesRegister(X86::EFLAGS, TRI) || 3749 Instr.readsRegister(X86::EFLAGS, TRI)) { 3750 // This instruction modifies or uses EFLAGS. 3751 3752 // MOV32r0 etc. are implemented with xor which clobbers condition code. 3753 // They are safe to move up, if the definition to EFLAGS is dead and 3754 // earlier instructions do not read or write EFLAGS. 3755 if (!Movr0Inst && 3756 (Instr.getOpcode() == X86::MOV32r0 || 3757 Instr.getOpcode() == X86::MOV64r0) && 3758 Instr.registerDefIsDead(X86::EFLAGS, TRI)) { 3759 Movr0Inst = &Instr; 3760 continue; 3761 } 3762 3763 // We can't remove CmpInstr. 3764 return false; 3765 } 3766 } 3767 3768 // Return false if no candidates exist. 3769 if (!IsCmpZero && !Sub) 3770 return false; 3771 3772 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 3773 Sub->getOperand(2).getReg() == SrcReg); 3774 3775 // Scan forward from the instruction after CmpInstr for uses of EFLAGS. 3776 // It is safe to remove CmpInstr if EFLAGS is redefined or killed. 3777 // If we are done with the basic block, we need to check whether EFLAGS is 3778 // live-out. 3779 bool IsSafe = false; 3780 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate; 3781 MachineBasicBlock::iterator E = CmpInstr.getParent()->end(); 3782 for (++I; I != E; ++I) { 3783 const MachineInstr &Instr = *I; 3784 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI); 3785 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI); 3786 // We should check the usage if this instruction uses and updates EFLAGS. 3787 if (!UseEFLAGS && ModifyEFLAGS) { 3788 // It is safe to remove CmpInstr if EFLAGS is updated again. 3789 IsSafe = true; 3790 break; 3791 } 3792 if (!UseEFLAGS && !ModifyEFLAGS) 3793 continue; 3794 3795 // EFLAGS is used by this instruction. 3796 X86::CondCode OldCC = X86::COND_INVALID; 3797 bool OpcIsSET = false; 3798 if (IsCmpZero || IsSwapped) { 3799 // We decode the condition code from opcode. 3800 if (Instr.isBranch()) 3801 OldCC = X86::getCondFromBranchOpc(Instr.getOpcode()); 3802 else { 3803 OldCC = X86::getCondFromSETOpc(Instr.getOpcode()); 3804 if (OldCC != X86::COND_INVALID) 3805 OpcIsSET = true; 3806 else 3807 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode()); 3808 } 3809 if (OldCC == X86::COND_INVALID) return false; 3810 } 3811 X86::CondCode ReplacementCC = X86::COND_INVALID; 3812 if (IsCmpZero) { 3813 switch (OldCC) { 3814 default: break; 3815 case X86::COND_A: case X86::COND_AE: 3816 case X86::COND_B: case X86::COND_BE: 3817 case X86::COND_G: case X86::COND_GE: 3818 case X86::COND_L: case X86::COND_LE: 3819 case X86::COND_O: case X86::COND_NO: 3820 // CF and OF are used, we can't perform this optimization. 3821 return false; 3822 } 3823 3824 // If we're updating the condition code check if we have to reverse the 3825 // condition. 3826 if (ShouldUpdateCC) 3827 switch (OldCC) { 3828 default: 3829 return false; 3830 case X86::COND_E: 3831 ReplacementCC = NewCC; 3832 break; 3833 case X86::COND_NE: 3834 ReplacementCC = GetOppositeBranchCondition(NewCC); 3835 break; 3836 } 3837 } else if (IsSwapped) { 3838 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs 3839 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 3840 // We swap the condition code and synthesize the new opcode. 3841 ReplacementCC = getSwappedCondition(OldCC); 3842 if (ReplacementCC == X86::COND_INVALID) return false; 3843 } 3844 3845 if ((ShouldUpdateCC || IsSwapped) && ReplacementCC != OldCC) { 3846 // Synthesize the new opcode. 3847 bool HasMemoryOperand = Instr.hasOneMemOperand(); 3848 unsigned NewOpc; 3849 if (Instr.isBranch()) 3850 NewOpc = GetCondBranchFromCond(ReplacementCC); 3851 else if(OpcIsSET) 3852 NewOpc = getSETFromCond(ReplacementCC, HasMemoryOperand); 3853 else { 3854 unsigned DstReg = Instr.getOperand(0).getReg(); 3855 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg); 3856 NewOpc = getCMovFromCond(ReplacementCC, TRI->getRegSizeInBits(*DstRC)/8, 3857 HasMemoryOperand); 3858 } 3859 3860 // Push the MachineInstr to OpsToUpdate. 3861 // If it is safe to remove CmpInstr, the condition code of these 3862 // instructions will be modified. 3863 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc)); 3864 } 3865 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) { 3866 // It is safe to remove CmpInstr if EFLAGS is updated again or killed. 3867 IsSafe = true; 3868 break; 3869 } 3870 } 3871 3872 // If EFLAGS is not killed nor re-defined, we should check whether it is 3873 // live-out. If it is live-out, do not optimize. 3874 if ((IsCmpZero || IsSwapped) && !IsSafe) { 3875 MachineBasicBlock *MBB = CmpInstr.getParent(); 3876 for (MachineBasicBlock *Successor : MBB->successors()) 3877 if (Successor->isLiveIn(X86::EFLAGS)) 3878 return false; 3879 } 3880 3881 // The instruction to be updated is either Sub or MI. 3882 Sub = IsCmpZero ? MI : Sub; 3883 // Move Movr0Inst to the appropriate place before Sub. 3884 if (Movr0Inst) { 3885 // Look backwards until we find a def that doesn't use the current EFLAGS. 3886 Def = Sub; 3887 MachineBasicBlock::reverse_iterator InsertI = Def.getReverse(), 3888 InsertE = Sub->getParent()->rend(); 3889 for (; InsertI != InsertE; ++InsertI) { 3890 MachineInstr *Instr = &*InsertI; 3891 if (!Instr->readsRegister(X86::EFLAGS, TRI) && 3892 Instr->modifiesRegister(X86::EFLAGS, TRI)) { 3893 Sub->getParent()->remove(Movr0Inst); 3894 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr), 3895 Movr0Inst); 3896 break; 3897 } 3898 } 3899 if (InsertI == InsertE) 3900 return false; 3901 } 3902 3903 // Make sure Sub instruction defines EFLAGS and mark the def live. 3904 unsigned i = 0, e = Sub->getNumOperands(); 3905 for (; i != e; ++i) { 3906 MachineOperand &MO = Sub->getOperand(i); 3907 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) { 3908 MO.setIsDead(false); 3909 break; 3910 } 3911 } 3912 assert(i != e && "Unable to locate a def EFLAGS operand"); 3913 3914 CmpInstr.eraseFromParent(); 3915 3916 // Modify the condition code of instructions in OpsToUpdate. 3917 for (auto &Op : OpsToUpdate) 3918 Op.first->setDesc(get(Op.second)); 3919 return true; 3920 } 3921 3922 /// Try to remove the load by folding it to a register 3923 /// operand at the use. We fold the load instructions if load defines a virtual 3924 /// register, the virtual register is used once in the same BB, and the 3925 /// instructions in-between do not load or store, and have no side effects. 3926 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr &MI, 3927 const MachineRegisterInfo *MRI, 3928 unsigned &FoldAsLoadDefReg, 3929 MachineInstr *&DefMI) const { 3930 // Check whether we can move DefMI here. 3931 DefMI = MRI->getVRegDef(FoldAsLoadDefReg); 3932 assert(DefMI); 3933 bool SawStore = false; 3934 if (!DefMI->isSafeToMove(nullptr, SawStore)) 3935 return nullptr; 3936 3937 // Collect information about virtual register operands of MI. 3938 SmallVector<unsigned, 1> SrcOperandIds; 3939 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 3940 MachineOperand &MO = MI.getOperand(i); 3941 if (!MO.isReg()) 3942 continue; 3943 unsigned Reg = MO.getReg(); 3944 if (Reg != FoldAsLoadDefReg) 3945 continue; 3946 // Do not fold if we have a subreg use or a def. 3947 if (MO.getSubReg() || MO.isDef()) 3948 return nullptr; 3949 SrcOperandIds.push_back(i); 3950 } 3951 if (SrcOperandIds.empty()) 3952 return nullptr; 3953 3954 // Check whether we can fold the def into SrcOperandId. 3955 if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandIds, *DefMI)) { 3956 FoldAsLoadDefReg = 0; 3957 return FoldMI; 3958 } 3959 3960 return nullptr; 3961 } 3962 3963 /// Expand a single-def pseudo instruction to a two-addr 3964 /// instruction with two undef reads of the register being defined. 3965 /// This is used for mapping: 3966 /// %xmm4 = V_SET0 3967 /// to: 3968 /// %xmm4 = PXORrr undef %xmm4, undef %xmm4 3969 /// 3970 static bool Expand2AddrUndef(MachineInstrBuilder &MIB, 3971 const MCInstrDesc &Desc) { 3972 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction."); 3973 unsigned Reg = MIB->getOperand(0).getReg(); 3974 MIB->setDesc(Desc); 3975 3976 // MachineInstr::addOperand() will insert explicit operands before any 3977 // implicit operands. 3978 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 3979 // But we don't trust that. 3980 assert(MIB->getOperand(1).getReg() == Reg && 3981 MIB->getOperand(2).getReg() == Reg && "Misplaced operand"); 3982 return true; 3983 } 3984 3985 /// Expand a single-def pseudo instruction to a two-addr 3986 /// instruction with two %k0 reads. 3987 /// This is used for mapping: 3988 /// %k4 = K_SET1 3989 /// to: 3990 /// %k4 = KXNORrr %k0, %k0 3991 static bool Expand2AddrKreg(MachineInstrBuilder &MIB, 3992 const MCInstrDesc &Desc, unsigned Reg) { 3993 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction."); 3994 MIB->setDesc(Desc); 3995 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 3996 return true; 3997 } 3998 3999 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII, 4000 bool MinusOne) { 4001 MachineBasicBlock &MBB = *MIB->getParent(); 4002 DebugLoc DL = MIB->getDebugLoc(); 4003 unsigned Reg = MIB->getOperand(0).getReg(); 4004 4005 // Insert the XOR. 4006 BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg) 4007 .addReg(Reg, RegState::Undef) 4008 .addReg(Reg, RegState::Undef); 4009 4010 // Turn the pseudo into an INC or DEC. 4011 MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r)); 4012 MIB.addReg(Reg); 4013 4014 return true; 4015 } 4016 4017 static bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB, 4018 const TargetInstrInfo &TII, 4019 const X86Subtarget &Subtarget) { 4020 MachineBasicBlock &MBB = *MIB->getParent(); 4021 DebugLoc DL = MIB->getDebugLoc(); 4022 int64_t Imm = MIB->getOperand(1).getImm(); 4023 assert(Imm != 0 && "Using push/pop for 0 is not efficient."); 4024 MachineBasicBlock::iterator I = MIB.getInstr(); 4025 4026 int StackAdjustment; 4027 4028 if (Subtarget.is64Bit()) { 4029 assert(MIB->getOpcode() == X86::MOV64ImmSExti8 || 4030 MIB->getOpcode() == X86::MOV32ImmSExti8); 4031 4032 // Can't use push/pop lowering if the function might write to the red zone. 4033 X86MachineFunctionInfo *X86FI = 4034 MBB.getParent()->getInfo<X86MachineFunctionInfo>(); 4035 if (X86FI->getUsesRedZone()) { 4036 MIB->setDesc(TII.get(MIB->getOpcode() == 4037 X86::MOV32ImmSExti8 ? X86::MOV32ri : X86::MOV64ri)); 4038 return true; 4039 } 4040 4041 // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and 4042 // widen the register if necessary. 4043 StackAdjustment = 8; 4044 BuildMI(MBB, I, DL, TII.get(X86::PUSH64i8)).addImm(Imm); 4045 MIB->setDesc(TII.get(X86::POP64r)); 4046 MIB->getOperand(0) 4047 .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64)); 4048 } else { 4049 assert(MIB->getOpcode() == X86::MOV32ImmSExti8); 4050 StackAdjustment = 4; 4051 BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm); 4052 MIB->setDesc(TII.get(X86::POP32r)); 4053 } 4054 4055 // Build CFI if necessary. 4056 MachineFunction &MF = *MBB.getParent(); 4057 const X86FrameLowering *TFL = Subtarget.getFrameLowering(); 4058 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI(); 4059 bool NeedsDwarfCFI = 4060 !IsWin64Prologue && 4061 (MF.getMMI().hasDebugInfo() || MF.getFunction().needsUnwindTableEntry()); 4062 bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI; 4063 if (EmitCFI) { 4064 TFL->BuildCFI(MBB, I, DL, 4065 MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment)); 4066 TFL->BuildCFI(MBB, std::next(I), DL, 4067 MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment)); 4068 } 4069 4070 return true; 4071 } 4072 4073 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different 4074 // code sequence is needed for other targets. 4075 static void expandLoadStackGuard(MachineInstrBuilder &MIB, 4076 const TargetInstrInfo &TII) { 4077 MachineBasicBlock &MBB = *MIB->getParent(); 4078 DebugLoc DL = MIB->getDebugLoc(); 4079 unsigned Reg = MIB->getOperand(0).getReg(); 4080 const GlobalValue *GV = 4081 cast<GlobalValue>((*MIB->memoperands_begin())->getValue()); 4082 auto Flags = MachineMemOperand::MOLoad | 4083 MachineMemOperand::MODereferenceable | 4084 MachineMemOperand::MOInvariant; 4085 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand( 4086 MachinePointerInfo::getGOT(*MBB.getParent()), Flags, 8, 8); 4087 MachineBasicBlock::iterator I = MIB.getInstr(); 4088 4089 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1) 4090 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0) 4091 .addMemOperand(MMO); 4092 MIB->setDebugLoc(DL); 4093 MIB->setDesc(TII.get(X86::MOV64rm)); 4094 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0); 4095 } 4096 4097 static bool expandXorFP(MachineInstrBuilder &MIB, const TargetInstrInfo &TII) { 4098 MachineBasicBlock &MBB = *MIB->getParent(); 4099 MachineFunction &MF = *MBB.getParent(); 4100 const X86Subtarget &Subtarget = MF.getSubtarget<X86Subtarget>(); 4101 const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); 4102 unsigned XorOp = 4103 MIB->getOpcode() == X86::XOR64_FP ? X86::XOR64rr : X86::XOR32rr; 4104 MIB->setDesc(TII.get(XorOp)); 4105 MIB.addReg(TRI->getFrameRegister(MF), RegState::Undef); 4106 return true; 4107 } 4108 4109 // This is used to handle spills for 128/256-bit registers when we have AVX512, 4110 // but not VLX. If it uses an extended register we need to use an instruction 4111 // that loads the lower 128/256-bit, but is available with only AVX512F. 4112 static bool expandNOVLXLoad(MachineInstrBuilder &MIB, 4113 const TargetRegisterInfo *TRI, 4114 const MCInstrDesc &LoadDesc, 4115 const MCInstrDesc &BroadcastDesc, 4116 unsigned SubIdx) { 4117 unsigned DestReg = MIB->getOperand(0).getReg(); 4118 // Check if DestReg is XMM16-31 or YMM16-31. 4119 if (TRI->getEncodingValue(DestReg) < 16) { 4120 // We can use a normal VEX encoded load. 4121 MIB->setDesc(LoadDesc); 4122 } else { 4123 // Use a 128/256-bit VBROADCAST instruction. 4124 MIB->setDesc(BroadcastDesc); 4125 // Change the destination to a 512-bit register. 4126 DestReg = TRI->getMatchingSuperReg(DestReg, SubIdx, &X86::VR512RegClass); 4127 MIB->getOperand(0).setReg(DestReg); 4128 } 4129 return true; 4130 } 4131 4132 // This is used to handle spills for 128/256-bit registers when we have AVX512, 4133 // but not VLX. If it uses an extended register we need to use an instruction 4134 // that stores the lower 128/256-bit, but is available with only AVX512F. 4135 static bool expandNOVLXStore(MachineInstrBuilder &MIB, 4136 const TargetRegisterInfo *TRI, 4137 const MCInstrDesc &StoreDesc, 4138 const MCInstrDesc &ExtractDesc, 4139 unsigned SubIdx) { 4140 unsigned SrcReg = MIB->getOperand(X86::AddrNumOperands).getReg(); 4141 // Check if DestReg is XMM16-31 or YMM16-31. 4142 if (TRI->getEncodingValue(SrcReg) < 16) { 4143 // We can use a normal VEX encoded store. 4144 MIB->setDesc(StoreDesc); 4145 } else { 4146 // Use a VEXTRACTF instruction. 4147 MIB->setDesc(ExtractDesc); 4148 // Change the destination to a 512-bit register. 4149 SrcReg = TRI->getMatchingSuperReg(SrcReg, SubIdx, &X86::VR512RegClass); 4150 MIB->getOperand(X86::AddrNumOperands).setReg(SrcReg); 4151 MIB.addImm(0x0); // Append immediate to extract from the lower bits. 4152 } 4153 4154 return true; 4155 } 4156 bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 4157 bool HasAVX = Subtarget.hasAVX(); 4158 MachineInstrBuilder MIB(*MI.getParent()->getParent(), MI); 4159 switch (MI.getOpcode()) { 4160 case X86::MOV32r0: 4161 return Expand2AddrUndef(MIB, get(X86::XOR32rr)); 4162 case X86::MOV64r0: { 4163 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4164 unsigned Reg = MIB->getOperand(0).getReg(); 4165 unsigned Reg32 = TRI->getSubReg(Reg, X86::sub_32bit); 4166 MIB->getOperand(0).setReg(Reg32); 4167 Expand2AddrUndef(MIB, get(X86::XOR32rr)); 4168 MIB.addReg(Reg, RegState::ImplicitDefine); 4169 return true; 4170 } 4171 case X86::MOV32r1: 4172 return expandMOV32r1(MIB, *this, /*MinusOne=*/ false); 4173 case X86::MOV32r_1: 4174 return expandMOV32r1(MIB, *this, /*MinusOne=*/ true); 4175 case X86::MOV32ImmSExti8: 4176 case X86::MOV64ImmSExti8: 4177 return ExpandMOVImmSExti8(MIB, *this, Subtarget); 4178 case X86::SETB_C8r: 4179 return Expand2AddrUndef(MIB, get(X86::SBB8rr)); 4180 case X86::SETB_C16r: 4181 return Expand2AddrUndef(MIB, get(X86::SBB16rr)); 4182 case X86::SETB_C32r: 4183 return Expand2AddrUndef(MIB, get(X86::SBB32rr)); 4184 case X86::SETB_C64r: 4185 return Expand2AddrUndef(MIB, get(X86::SBB64rr)); 4186 case X86::MMX_SET0: 4187 return Expand2AddrUndef(MIB, get(X86::MMX_PXORirr)); 4188 case X86::V_SET0: 4189 case X86::FsFLD0SS: 4190 case X86::FsFLD0SD: 4191 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr)); 4192 case X86::AVX_SET0: { 4193 assert(HasAVX && "AVX not supported"); 4194 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4195 unsigned SrcReg = MIB->getOperand(0).getReg(); 4196 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm); 4197 MIB->getOperand(0).setReg(XReg); 4198 Expand2AddrUndef(MIB, get(X86::VXORPSrr)); 4199 MIB.addReg(SrcReg, RegState::ImplicitDefine); 4200 return true; 4201 } 4202 case X86::AVX512_128_SET0: 4203 case X86::AVX512_FsFLD0SS: 4204 case X86::AVX512_FsFLD0SD: { 4205 bool HasVLX = Subtarget.hasVLX(); 4206 unsigned SrcReg = MIB->getOperand(0).getReg(); 4207 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4208 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) 4209 return Expand2AddrUndef(MIB, 4210 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr)); 4211 // Extended register without VLX. Use a larger XOR. 4212 SrcReg = 4213 TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass); 4214 MIB->getOperand(0).setReg(SrcReg); 4215 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr)); 4216 } 4217 case X86::AVX512_256_SET0: 4218 case X86::AVX512_512_SET0: { 4219 bool HasVLX = Subtarget.hasVLX(); 4220 unsigned SrcReg = MIB->getOperand(0).getReg(); 4221 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4222 if (HasVLX || TRI->getEncodingValue(SrcReg) < 16) { 4223 unsigned XReg = TRI->getSubReg(SrcReg, X86::sub_xmm); 4224 MIB->getOperand(0).setReg(XReg); 4225 Expand2AddrUndef(MIB, 4226 get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr)); 4227 MIB.addReg(SrcReg, RegState::ImplicitDefine); 4228 return true; 4229 } 4230 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr)); 4231 } 4232 case X86::V_SETALLONES: 4233 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr)); 4234 case X86::AVX2_SETALLONES: 4235 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr)); 4236 case X86::AVX1_SETALLONES: { 4237 unsigned Reg = MIB->getOperand(0).getReg(); 4238 // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS. 4239 MIB->setDesc(get(X86::VCMPPSYrri)); 4240 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf); 4241 return true; 4242 } 4243 case X86::AVX512_512_SETALLONES: { 4244 unsigned Reg = MIB->getOperand(0).getReg(); 4245 MIB->setDesc(get(X86::VPTERNLOGDZrri)); 4246 // VPTERNLOGD needs 3 register inputs and an immediate. 4247 // 0xff will return 1s for any input. 4248 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef) 4249 .addReg(Reg, RegState::Undef).addImm(0xff); 4250 return true; 4251 } 4252 case X86::AVX512_512_SEXT_MASK_32: 4253 case X86::AVX512_512_SEXT_MASK_64: { 4254 unsigned Reg = MIB->getOperand(0).getReg(); 4255 unsigned MaskReg = MIB->getOperand(1).getReg(); 4256 unsigned MaskState = getRegState(MIB->getOperand(1)); 4257 unsigned Opc = (MI.getOpcode() == X86::AVX512_512_SEXT_MASK_64) ? 4258 X86::VPTERNLOGQZrrikz : X86::VPTERNLOGDZrrikz; 4259 MI.RemoveOperand(1); 4260 MIB->setDesc(get(Opc)); 4261 // VPTERNLOG needs 3 register inputs and an immediate. 4262 // 0xff will return 1s for any input. 4263 MIB.addReg(Reg, RegState::Undef).addReg(MaskReg, MaskState) 4264 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xff); 4265 return true; 4266 } 4267 case X86::VMOVAPSZ128rm_NOVLX: 4268 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSrm), 4269 get(X86::VBROADCASTF32X4rm), X86::sub_xmm); 4270 case X86::VMOVUPSZ128rm_NOVLX: 4271 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSrm), 4272 get(X86::VBROADCASTF32X4rm), X86::sub_xmm); 4273 case X86::VMOVAPSZ256rm_NOVLX: 4274 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVAPSYrm), 4275 get(X86::VBROADCASTF64X4rm), X86::sub_ymm); 4276 case X86::VMOVUPSZ256rm_NOVLX: 4277 return expandNOVLXLoad(MIB, &getRegisterInfo(), get(X86::VMOVUPSYrm), 4278 get(X86::VBROADCASTF64X4rm), X86::sub_ymm); 4279 case X86::VMOVAPSZ128mr_NOVLX: 4280 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSmr), 4281 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm); 4282 case X86::VMOVUPSZ128mr_NOVLX: 4283 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSmr), 4284 get(X86::VEXTRACTF32x4Zmr), X86::sub_xmm); 4285 case X86::VMOVAPSZ256mr_NOVLX: 4286 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVAPSYmr), 4287 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm); 4288 case X86::VMOVUPSZ256mr_NOVLX: 4289 return expandNOVLXStore(MIB, &getRegisterInfo(), get(X86::VMOVUPSYmr), 4290 get(X86::VEXTRACTF64x4Zmr), X86::sub_ymm); 4291 case X86::MOV32ri64: { 4292 unsigned Reg = MIB->getOperand(0).getReg(); 4293 unsigned Reg32 = RI.getSubReg(Reg, X86::sub_32bit); 4294 MI.setDesc(get(X86::MOV32ri)); 4295 MIB->getOperand(0).setReg(Reg32); 4296 MIB.addReg(Reg, RegState::ImplicitDefine); 4297 return true; 4298 } 4299 4300 // KNL does not recognize dependency-breaking idioms for mask registers, 4301 // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1. 4302 // Using %k0 as the undef input register is a performance heuristic based 4303 // on the assumption that %k0 is used less frequently than the other mask 4304 // registers, since it is not usable as a write mask. 4305 // FIXME: A more advanced approach would be to choose the best input mask 4306 // register based on context. 4307 case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0); 4308 case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0); 4309 case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0); 4310 case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0); 4311 case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0); 4312 case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0); 4313 case TargetOpcode::LOAD_STACK_GUARD: 4314 expandLoadStackGuard(MIB, *this); 4315 return true; 4316 case X86::XOR64_FP: 4317 case X86::XOR32_FP: 4318 return expandXorFP(MIB, *this); 4319 } 4320 return false; 4321 } 4322 4323 /// Return true for all instructions that only update 4324 /// the first 32 or 64-bits of the destination register and leave the rest 4325 /// unmodified. This can be used to avoid folding loads if the instructions 4326 /// only update part of the destination register, and the non-updated part is 4327 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these 4328 /// instructions breaks the partial register dependency and it can improve 4329 /// performance. e.g.: 4330 /// 4331 /// movss (%rdi), %xmm0 4332 /// cvtss2sd %xmm0, %xmm0 4333 /// 4334 /// Instead of 4335 /// cvtss2sd (%rdi), %xmm0 4336 /// 4337 /// FIXME: This should be turned into a TSFlags. 4338 /// 4339 static bool hasPartialRegUpdate(unsigned Opcode, 4340 const X86Subtarget &Subtarget) { 4341 switch (Opcode) { 4342 case X86::CVTSI2SSrr: 4343 case X86::CVTSI2SSrm: 4344 case X86::CVTSI642SSrr: 4345 case X86::CVTSI642SSrm: 4346 case X86::CVTSI2SDrr: 4347 case X86::CVTSI2SDrm: 4348 case X86::CVTSI642SDrr: 4349 case X86::CVTSI642SDrm: 4350 case X86::CVTSD2SSrr: 4351 case X86::CVTSD2SSrm: 4352 case X86::CVTSS2SDrr: 4353 case X86::CVTSS2SDrm: 4354 case X86::MOVHPDrm: 4355 case X86::MOVHPSrm: 4356 case X86::MOVLPDrm: 4357 case X86::MOVLPSrm: 4358 case X86::RCPSSr: 4359 case X86::RCPSSm: 4360 case X86::RCPSSr_Int: 4361 case X86::RCPSSm_Int: 4362 case X86::ROUNDSDr: 4363 case X86::ROUNDSDm: 4364 case X86::ROUNDSSr: 4365 case X86::ROUNDSSm: 4366 case X86::RSQRTSSr: 4367 case X86::RSQRTSSm: 4368 case X86::RSQRTSSr_Int: 4369 case X86::RSQRTSSm_Int: 4370 case X86::SQRTSSr: 4371 case X86::SQRTSSm: 4372 case X86::SQRTSSr_Int: 4373 case X86::SQRTSSm_Int: 4374 case X86::SQRTSDr: 4375 case X86::SQRTSDm: 4376 case X86::SQRTSDr_Int: 4377 case X86::SQRTSDm_Int: 4378 return true; 4379 // GPR 4380 case X86::POPCNT32rm: 4381 case X86::POPCNT32rr: 4382 case X86::POPCNT64rm: 4383 case X86::POPCNT64rr: 4384 return Subtarget.hasPOPCNTFalseDeps(); 4385 case X86::LZCNT32rm: 4386 case X86::LZCNT32rr: 4387 case X86::LZCNT64rm: 4388 case X86::LZCNT64rr: 4389 case X86::TZCNT32rm: 4390 case X86::TZCNT32rr: 4391 case X86::TZCNT64rm: 4392 case X86::TZCNT64rr: 4393 return Subtarget.hasLZCNTFalseDeps(); 4394 } 4395 4396 return false; 4397 } 4398 4399 /// Inform the BreakFalseDeps pass how many idle 4400 /// instructions we would like before a partial register update. 4401 unsigned X86InstrInfo::getPartialRegUpdateClearance( 4402 const MachineInstr &MI, unsigned OpNum, 4403 const TargetRegisterInfo *TRI) const { 4404 if (OpNum != 0 || !hasPartialRegUpdate(MI.getOpcode(), Subtarget)) 4405 return 0; 4406 4407 // If MI is marked as reading Reg, the partial register update is wanted. 4408 const MachineOperand &MO = MI.getOperand(0); 4409 unsigned Reg = MO.getReg(); 4410 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 4411 if (MO.readsReg() || MI.readsVirtualRegister(Reg)) 4412 return 0; 4413 } else { 4414 if (MI.readsRegister(Reg, TRI)) 4415 return 0; 4416 } 4417 4418 // If any instructions in the clearance range are reading Reg, insert a 4419 // dependency breaking instruction, which is inexpensive and is likely to 4420 // be hidden in other instruction's cycles. 4421 return PartialRegUpdateClearance; 4422 } 4423 4424 // Return true for any instruction the copies the high bits of the first source 4425 // operand into the unused high bits of the destination operand. 4426 static bool hasUndefRegUpdate(unsigned Opcode) { 4427 switch (Opcode) { 4428 case X86::VCVTSI2SSrr: 4429 case X86::VCVTSI2SSrm: 4430 case X86::VCVTSI2SSrr_Int: 4431 case X86::VCVTSI2SSrm_Int: 4432 case X86::VCVTSI642SSrr: 4433 case X86::VCVTSI642SSrm: 4434 case X86::VCVTSI642SSrr_Int: 4435 case X86::VCVTSI642SSrm_Int: 4436 case X86::VCVTSI2SDrr: 4437 case X86::VCVTSI2SDrm: 4438 case X86::VCVTSI2SDrr_Int: 4439 case X86::VCVTSI2SDrm_Int: 4440 case X86::VCVTSI642SDrr: 4441 case X86::VCVTSI642SDrm: 4442 case X86::VCVTSI642SDrr_Int: 4443 case X86::VCVTSI642SDrm_Int: 4444 case X86::VCVTSD2SSrr: 4445 case X86::VCVTSD2SSrm: 4446 case X86::VCVTSD2SSrr_Int: 4447 case X86::VCVTSD2SSrm_Int: 4448 case X86::VCVTSS2SDrr: 4449 case X86::VCVTSS2SDrm: 4450 case X86::VCVTSS2SDrr_Int: 4451 case X86::VCVTSS2SDrm_Int: 4452 case X86::VRCPSSr: 4453 case X86::VRCPSSr_Int: 4454 case X86::VRCPSSm: 4455 case X86::VRCPSSm_Int: 4456 case X86::VROUNDSDr: 4457 case X86::VROUNDSDm: 4458 case X86::VROUNDSDr_Int: 4459 case X86::VROUNDSDm_Int: 4460 case X86::VROUNDSSr: 4461 case X86::VROUNDSSm: 4462 case X86::VROUNDSSr_Int: 4463 case X86::VROUNDSSm_Int: 4464 case X86::VRSQRTSSr: 4465 case X86::VRSQRTSSr_Int: 4466 case X86::VRSQRTSSm: 4467 case X86::VRSQRTSSm_Int: 4468 case X86::VSQRTSSr: 4469 case X86::VSQRTSSr_Int: 4470 case X86::VSQRTSSm: 4471 case X86::VSQRTSSm_Int: 4472 case X86::VSQRTSDr: 4473 case X86::VSQRTSDr_Int: 4474 case X86::VSQRTSDm: 4475 case X86::VSQRTSDm_Int: 4476 // AVX-512 4477 case X86::VCVTSI2SSZrr: 4478 case X86::VCVTSI2SSZrm: 4479 case X86::VCVTSI2SSZrr_Int: 4480 case X86::VCVTSI2SSZrrb_Int: 4481 case X86::VCVTSI2SSZrm_Int: 4482 case X86::VCVTSI642SSZrr: 4483 case X86::VCVTSI642SSZrm: 4484 case X86::VCVTSI642SSZrr_Int: 4485 case X86::VCVTSI642SSZrrb_Int: 4486 case X86::VCVTSI642SSZrm_Int: 4487 case X86::VCVTSI2SDZrr: 4488 case X86::VCVTSI2SDZrm: 4489 case X86::VCVTSI2SDZrr_Int: 4490 case X86::VCVTSI2SDZrrb_Int: 4491 case X86::VCVTSI2SDZrm_Int: 4492 case X86::VCVTSI642SDZrr: 4493 case X86::VCVTSI642SDZrm: 4494 case X86::VCVTSI642SDZrr_Int: 4495 case X86::VCVTSI642SDZrrb_Int: 4496 case X86::VCVTSI642SDZrm_Int: 4497 case X86::VCVTUSI2SSZrr: 4498 case X86::VCVTUSI2SSZrm: 4499 case X86::VCVTUSI2SSZrr_Int: 4500 case X86::VCVTUSI2SSZrrb_Int: 4501 case X86::VCVTUSI2SSZrm_Int: 4502 case X86::VCVTUSI642SSZrr: 4503 case X86::VCVTUSI642SSZrm: 4504 case X86::VCVTUSI642SSZrr_Int: 4505 case X86::VCVTUSI642SSZrrb_Int: 4506 case X86::VCVTUSI642SSZrm_Int: 4507 case X86::VCVTUSI2SDZrr: 4508 case X86::VCVTUSI2SDZrm: 4509 case X86::VCVTUSI2SDZrr_Int: 4510 case X86::VCVTUSI2SDZrm_Int: 4511 case X86::VCVTUSI642SDZrr: 4512 case X86::VCVTUSI642SDZrm: 4513 case X86::VCVTUSI642SDZrr_Int: 4514 case X86::VCVTUSI642SDZrrb_Int: 4515 case X86::VCVTUSI642SDZrm_Int: 4516 case X86::VCVTSD2SSZrr: 4517 case X86::VCVTSD2SSZrr_Int: 4518 case X86::VCVTSD2SSZrrb_Int: 4519 case X86::VCVTSD2SSZrm: 4520 case X86::VCVTSD2SSZrm_Int: 4521 case X86::VCVTSS2SDZrr: 4522 case X86::VCVTSS2SDZrr_Int: 4523 case X86::VCVTSS2SDZrrb_Int: 4524 case X86::VCVTSS2SDZrm: 4525 case X86::VCVTSS2SDZrm_Int: 4526 case X86::VGETEXPSDZr: 4527 case X86::VGETEXPSDZrb: 4528 case X86::VGETEXPSDZm: 4529 case X86::VGETEXPSSZr: 4530 case X86::VGETEXPSSZrb: 4531 case X86::VGETEXPSSZm: 4532 case X86::VGETMANTSDZrri: 4533 case X86::VGETMANTSDZrrib: 4534 case X86::VGETMANTSDZrmi: 4535 case X86::VGETMANTSSZrri: 4536 case X86::VGETMANTSSZrrib: 4537 case X86::VGETMANTSSZrmi: 4538 case X86::VRNDSCALESDZr: 4539 case X86::VRNDSCALESDZr_Int: 4540 case X86::VRNDSCALESDZrb_Int: 4541 case X86::VRNDSCALESDZm: 4542 case X86::VRNDSCALESDZm_Int: 4543 case X86::VRNDSCALESSZr: 4544 case X86::VRNDSCALESSZr_Int: 4545 case X86::VRNDSCALESSZrb_Int: 4546 case X86::VRNDSCALESSZm: 4547 case X86::VRNDSCALESSZm_Int: 4548 case X86::VRCP14SDZrr: 4549 case X86::VRCP14SDZrm: 4550 case X86::VRCP14SSZrr: 4551 case X86::VRCP14SSZrm: 4552 case X86::VRCP28SDZr: 4553 case X86::VRCP28SDZrb: 4554 case X86::VRCP28SDZm: 4555 case X86::VRCP28SSZr: 4556 case X86::VRCP28SSZrb: 4557 case X86::VRCP28SSZm: 4558 case X86::VREDUCESSZrmi: 4559 case X86::VREDUCESSZrri: 4560 case X86::VREDUCESSZrrib: 4561 case X86::VRSQRT14SDZrr: 4562 case X86::VRSQRT14SDZrm: 4563 case X86::VRSQRT14SSZrr: 4564 case X86::VRSQRT14SSZrm: 4565 case X86::VRSQRT28SDZr: 4566 case X86::VRSQRT28SDZrb: 4567 case X86::VRSQRT28SDZm: 4568 case X86::VRSQRT28SSZr: 4569 case X86::VRSQRT28SSZrb: 4570 case X86::VRSQRT28SSZm: 4571 case X86::VSQRTSSZr: 4572 case X86::VSQRTSSZr_Int: 4573 case X86::VSQRTSSZrb_Int: 4574 case X86::VSQRTSSZm: 4575 case X86::VSQRTSSZm_Int: 4576 case X86::VSQRTSDZr: 4577 case X86::VSQRTSDZr_Int: 4578 case X86::VSQRTSDZrb_Int: 4579 case X86::VSQRTSDZm: 4580 case X86::VSQRTSDZm_Int: 4581 return true; 4582 } 4583 4584 return false; 4585 } 4586 4587 /// Inform the BreakFalseDeps pass how many idle instructions we would like 4588 /// before certain undef register reads. 4589 /// 4590 /// This catches the VCVTSI2SD family of instructions: 4591 /// 4592 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14 4593 /// 4594 /// We should to be careful *not* to catch VXOR idioms which are presumably 4595 /// handled specially in the pipeline: 4596 /// 4597 /// vxorps undef %xmm1, undef %xmm1, %xmm1 4598 /// 4599 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the 4600 /// high bits that are passed-through are not live. 4601 unsigned 4602 X86InstrInfo::getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum, 4603 const TargetRegisterInfo *TRI) const { 4604 if (!hasUndefRegUpdate(MI.getOpcode())) 4605 return 0; 4606 4607 // Set the OpNum parameter to the first source operand. 4608 OpNum = 1; 4609 4610 const MachineOperand &MO = MI.getOperand(OpNum); 4611 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) { 4612 return UndefRegClearance; 4613 } 4614 return 0; 4615 } 4616 4617 void X86InstrInfo::breakPartialRegDependency( 4618 MachineInstr &MI, unsigned OpNum, const TargetRegisterInfo *TRI) const { 4619 unsigned Reg = MI.getOperand(OpNum).getReg(); 4620 // If MI kills this register, the false dependence is already broken. 4621 if (MI.killsRegister(Reg, TRI)) 4622 return; 4623 4624 if (X86::VR128RegClass.contains(Reg)) { 4625 // These instructions are all floating point domain, so xorps is the best 4626 // choice. 4627 unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr; 4628 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(Opc), Reg) 4629 .addReg(Reg, RegState::Undef) 4630 .addReg(Reg, RegState::Undef); 4631 MI.addRegisterKilled(Reg, TRI, true); 4632 } else if (X86::VR256RegClass.contains(Reg)) { 4633 // Use vxorps to clear the full ymm register. 4634 // It wants to read and write the xmm sub-register. 4635 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm); 4636 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::VXORPSrr), XReg) 4637 .addReg(XReg, RegState::Undef) 4638 .addReg(XReg, RegState::Undef) 4639 .addReg(Reg, RegState::ImplicitDefine); 4640 MI.addRegisterKilled(Reg, TRI, true); 4641 } else if (X86::GR64RegClass.contains(Reg)) { 4642 // Using XOR32rr because it has shorter encoding and zeros up the upper bits 4643 // as well. 4644 unsigned XReg = TRI->getSubReg(Reg, X86::sub_32bit); 4645 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), XReg) 4646 .addReg(XReg, RegState::Undef) 4647 .addReg(XReg, RegState::Undef) 4648 .addReg(Reg, RegState::ImplicitDefine); 4649 MI.addRegisterKilled(Reg, TRI, true); 4650 } else if (X86::GR32RegClass.contains(Reg)) { 4651 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), get(X86::XOR32rr), Reg) 4652 .addReg(Reg, RegState::Undef) 4653 .addReg(Reg, RegState::Undef); 4654 MI.addRegisterKilled(Reg, TRI, true); 4655 } 4656 } 4657 4658 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs, 4659 int PtrOffset = 0) { 4660 unsigned NumAddrOps = MOs.size(); 4661 4662 if (NumAddrOps < 4) { 4663 // FrameIndex only - add an immediate offset (whether its zero or not). 4664 for (unsigned i = 0; i != NumAddrOps; ++i) 4665 MIB.add(MOs[i]); 4666 addOffset(MIB, PtrOffset); 4667 } else { 4668 // General Memory Addressing - we need to add any offset to an existing 4669 // offset. 4670 assert(MOs.size() == 5 && "Unexpected memory operand list length"); 4671 for (unsigned i = 0; i != NumAddrOps; ++i) { 4672 const MachineOperand &MO = MOs[i]; 4673 if (i == 3 && PtrOffset != 0) { 4674 MIB.addDisp(MO, PtrOffset); 4675 } else { 4676 MIB.add(MO); 4677 } 4678 } 4679 } 4680 } 4681 4682 static void updateOperandRegConstraints(MachineFunction &MF, 4683 MachineInstr &NewMI, 4684 const TargetInstrInfo &TII) { 4685 MachineRegisterInfo &MRI = MF.getRegInfo(); 4686 const TargetRegisterInfo &TRI = *MRI.getTargetRegisterInfo(); 4687 4688 for (int Idx : llvm::seq<int>(0, NewMI.getNumOperands())) { 4689 MachineOperand &MO = NewMI.getOperand(Idx); 4690 // We only need to update constraints on virtual register operands. 4691 if (!MO.isReg()) 4692 continue; 4693 unsigned Reg = MO.getReg(); 4694 if (!TRI.isVirtualRegister(Reg)) 4695 continue; 4696 4697 auto *NewRC = MRI.constrainRegClass( 4698 Reg, TII.getRegClass(NewMI.getDesc(), Idx, &TRI, MF)); 4699 if (!NewRC) { 4700 LLVM_DEBUG( 4701 dbgs() << "WARNING: Unable to update register constraint for operand " 4702 << Idx << " of instruction:\n"; 4703 NewMI.dump(); dbgs() << "\n"); 4704 } 4705 } 4706 } 4707 4708 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode, 4709 ArrayRef<MachineOperand> MOs, 4710 MachineBasicBlock::iterator InsertPt, 4711 MachineInstr &MI, 4712 const TargetInstrInfo &TII) { 4713 // Create the base instruction with the memory operand as the first part. 4714 // Omit the implicit operands, something BuildMI can't do. 4715 MachineInstr *NewMI = 4716 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true); 4717 MachineInstrBuilder MIB(MF, NewMI); 4718 addOperands(MIB, MOs); 4719 4720 // Loop over the rest of the ri operands, converting them over. 4721 unsigned NumOps = MI.getDesc().getNumOperands() - 2; 4722 for (unsigned i = 0; i != NumOps; ++i) { 4723 MachineOperand &MO = MI.getOperand(i + 2); 4724 MIB.add(MO); 4725 } 4726 for (unsigned i = NumOps + 2, e = MI.getNumOperands(); i != e; ++i) { 4727 MachineOperand &MO = MI.getOperand(i); 4728 MIB.add(MO); 4729 } 4730 4731 updateOperandRegConstraints(MF, *NewMI, TII); 4732 4733 MachineBasicBlock *MBB = InsertPt->getParent(); 4734 MBB->insert(InsertPt, NewMI); 4735 4736 return MIB; 4737 } 4738 4739 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode, 4740 unsigned OpNo, ArrayRef<MachineOperand> MOs, 4741 MachineBasicBlock::iterator InsertPt, 4742 MachineInstr &MI, const TargetInstrInfo &TII, 4743 int PtrOffset = 0) { 4744 // Omit the implicit operands, something BuildMI can't do. 4745 MachineInstr *NewMI = 4746 MF.CreateMachineInstr(TII.get(Opcode), MI.getDebugLoc(), true); 4747 MachineInstrBuilder MIB(MF, NewMI); 4748 4749 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 4750 MachineOperand &MO = MI.getOperand(i); 4751 if (i == OpNo) { 4752 assert(MO.isReg() && "Expected to fold into reg operand!"); 4753 addOperands(MIB, MOs, PtrOffset); 4754 } else { 4755 MIB.add(MO); 4756 } 4757 } 4758 4759 updateOperandRegConstraints(MF, *NewMI, TII); 4760 4761 MachineBasicBlock *MBB = InsertPt->getParent(); 4762 MBB->insert(InsertPt, NewMI); 4763 4764 return MIB; 4765 } 4766 4767 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode, 4768 ArrayRef<MachineOperand> MOs, 4769 MachineBasicBlock::iterator InsertPt, 4770 MachineInstr &MI) { 4771 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 4772 MI.getDebugLoc(), TII.get(Opcode)); 4773 addOperands(MIB, MOs); 4774 return MIB.addImm(0); 4775 } 4776 4777 MachineInstr *X86InstrInfo::foldMemoryOperandCustom( 4778 MachineFunction &MF, MachineInstr &MI, unsigned OpNum, 4779 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt, 4780 unsigned Size, unsigned Align) const { 4781 switch (MI.getOpcode()) { 4782 case X86::INSERTPSrr: 4783 case X86::VINSERTPSrr: 4784 case X86::VINSERTPSZrr: 4785 // Attempt to convert the load of inserted vector into a fold load 4786 // of a single float. 4787 if (OpNum == 2) { 4788 unsigned Imm = MI.getOperand(MI.getNumOperands() - 1).getImm(); 4789 unsigned ZMask = Imm & 15; 4790 unsigned DstIdx = (Imm >> 4) & 3; 4791 unsigned SrcIdx = (Imm >> 6) & 3; 4792 4793 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 4794 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF); 4795 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8; 4796 if (Size <= RCSize && 4 <= Align) { 4797 int PtrOffset = SrcIdx * 4; 4798 unsigned NewImm = (DstIdx << 4) | ZMask; 4799 unsigned NewOpCode = 4800 (MI.getOpcode() == X86::VINSERTPSZrr) ? X86::VINSERTPSZrm : 4801 (MI.getOpcode() == X86::VINSERTPSrr) ? X86::VINSERTPSrm : 4802 X86::INSERTPSrm; 4803 MachineInstr *NewMI = 4804 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset); 4805 NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm); 4806 return NewMI; 4807 } 4808 } 4809 break; 4810 case X86::MOVHLPSrr: 4811 case X86::VMOVHLPSrr: 4812 case X86::VMOVHLPSZrr: 4813 // Move the upper 64-bits of the second operand to the lower 64-bits. 4814 // To fold the load, adjust the pointer to the upper and use (V)MOVLPS. 4815 // TODO: In most cases AVX doesn't have a 8-byte alignment requirement. 4816 if (OpNum == 2) { 4817 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 4818 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, &RI, MF); 4819 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8; 4820 if (Size <= RCSize && 8 <= Align) { 4821 unsigned NewOpCode = 4822 (MI.getOpcode() == X86::VMOVHLPSZrr) ? X86::VMOVLPSZ128rm : 4823 (MI.getOpcode() == X86::VMOVHLPSrr) ? X86::VMOVLPSrm : 4824 X86::MOVLPSrm; 4825 MachineInstr *NewMI = 4826 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8); 4827 return NewMI; 4828 } 4829 } 4830 break; 4831 }; 4832 4833 return nullptr; 4834 } 4835 4836 static bool shouldPreventUndefRegUpdateMemFold(MachineFunction &MF, MachineInstr &MI) { 4837 if (MF.getFunction().optForSize() || !hasUndefRegUpdate(MI.getOpcode()) || 4838 !MI.getOperand(1).isReg()) 4839 return false; 4840 4841 // The are two cases we need to handle depending on where in the pipeline 4842 // the folding attempt is being made. 4843 // -Register has the undef flag set. 4844 // -Register is produced by the IMPLICIT_DEF instruction. 4845 4846 if (MI.getOperand(1).isUndef()) 4847 return true; 4848 4849 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 4850 MachineInstr *VRegDef = RegInfo.getUniqueVRegDef(MI.getOperand(1).getReg()); 4851 return VRegDef && VRegDef->isImplicitDef(); 4852 } 4853 4854 4855 MachineInstr *X86InstrInfo::foldMemoryOperandImpl( 4856 MachineFunction &MF, MachineInstr &MI, unsigned OpNum, 4857 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt, 4858 unsigned Size, unsigned Align, bool AllowCommute) const { 4859 bool isSlowTwoMemOps = Subtarget.slowTwoMemOps(); 4860 bool isTwoAddrFold = false; 4861 4862 // For CPUs that favor the register form of a call or push, 4863 // do not fold loads into calls or pushes, unless optimizing for size 4864 // aggressively. 4865 if (isSlowTwoMemOps && !MF.getFunction().optForMinSize() && 4866 (MI.getOpcode() == X86::CALL32r || MI.getOpcode() == X86::CALL64r || 4867 MI.getOpcode() == X86::PUSH16r || MI.getOpcode() == X86::PUSH32r || 4868 MI.getOpcode() == X86::PUSH64r)) 4869 return nullptr; 4870 4871 // Avoid partial and undef register update stalls unless optimizing for size. 4872 if (!MF.getFunction().optForSize() && 4873 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) || 4874 shouldPreventUndefRegUpdateMemFold(MF, MI))) 4875 return nullptr; 4876 4877 unsigned NumOps = MI.getDesc().getNumOperands(); 4878 bool isTwoAddr = 4879 NumOps > 1 && MI.getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1; 4880 4881 // FIXME: AsmPrinter doesn't know how to handle 4882 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding. 4883 if (MI.getOpcode() == X86::ADD32ri && 4884 MI.getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS) 4885 return nullptr; 4886 4887 // GOTTPOFF relocation loads can only be folded into add instructions. 4888 // FIXME: Need to exclude other relocations that only support specific 4889 // instructions. 4890 if (MOs.size() == X86::AddrNumOperands && 4891 MOs[X86::AddrDisp].getTargetFlags() == X86II::MO_GOTTPOFF && 4892 MI.getOpcode() != X86::ADD64rr) 4893 return nullptr; 4894 4895 MachineInstr *NewMI = nullptr; 4896 4897 // Attempt to fold any custom cases we have. 4898 if (MachineInstr *CustomMI = 4899 foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align)) 4900 return CustomMI; 4901 4902 const X86MemoryFoldTableEntry *I = nullptr; 4903 4904 // Folding a memory location into the two-address part of a two-address 4905 // instruction is different than folding it other places. It requires 4906 // replacing the *two* registers with the memory location. 4907 if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() && 4908 MI.getOperand(1).isReg() && 4909 MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) { 4910 I = lookupTwoAddrFoldTable(MI.getOpcode()); 4911 isTwoAddrFold = true; 4912 } else { 4913 if (OpNum == 0) { 4914 if (MI.getOpcode() == X86::MOV32r0 || MI.getOpcode() == X86::MOV64r0) { 4915 unsigned NewOpc = MI.getOpcode() == X86::MOV64r0 ? X86::MOV64mi32 4916 : X86::MOV32mi; 4917 NewMI = MakeM0Inst(*this, NewOpc, MOs, InsertPt, MI); 4918 if (NewMI) 4919 return NewMI; 4920 } 4921 } 4922 4923 I = lookupFoldTable(MI.getOpcode(), OpNum); 4924 } 4925 4926 if (I != nullptr) { 4927 unsigned Opcode = I->DstOp; 4928 unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT; 4929 if (Align < MinAlign) 4930 return nullptr; 4931 bool NarrowToMOV32rm = false; 4932 if (Size) { 4933 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 4934 const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum, 4935 &RI, MF); 4936 unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8; 4937 if (Size < RCSize) { 4938 // Check if it's safe to fold the load. If the size of the object is 4939 // narrower than the load width, then it's not. 4940 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4) 4941 return nullptr; 4942 // If this is a 64-bit load, but the spill slot is 32, then we can do 4943 // a 32-bit load which is implicitly zero-extended. This likely is 4944 // due to live interval analysis remat'ing a load from stack slot. 4945 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg()) 4946 return nullptr; 4947 Opcode = X86::MOV32rm; 4948 NarrowToMOV32rm = true; 4949 } 4950 } 4951 4952 if (isTwoAddrFold) 4953 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this); 4954 else 4955 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this); 4956 4957 if (NarrowToMOV32rm) { 4958 // If this is the special case where we use a MOV32rm to load a 32-bit 4959 // value and zero-extend the top bits. Change the destination register 4960 // to a 32-bit one. 4961 unsigned DstReg = NewMI->getOperand(0).getReg(); 4962 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) 4963 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit)); 4964 else 4965 NewMI->getOperand(0).setSubReg(X86::sub_32bit); 4966 } 4967 return NewMI; 4968 } 4969 4970 // If the instruction and target operand are commutable, commute the 4971 // instruction and try again. 4972 if (AllowCommute) { 4973 unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex; 4974 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) { 4975 bool HasDef = MI.getDesc().getNumDefs(); 4976 unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0; 4977 unsigned Reg1 = MI.getOperand(CommuteOpIdx1).getReg(); 4978 unsigned Reg2 = MI.getOperand(CommuteOpIdx2).getReg(); 4979 bool Tied1 = 4980 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO); 4981 bool Tied2 = 4982 0 == MI.getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO); 4983 4984 // If either of the commutable operands are tied to the destination 4985 // then we can not commute + fold. 4986 if ((HasDef && Reg0 == Reg1 && Tied1) || 4987 (HasDef && Reg0 == Reg2 && Tied2)) 4988 return nullptr; 4989 4990 MachineInstr *CommutedMI = 4991 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2); 4992 if (!CommutedMI) { 4993 // Unable to commute. 4994 return nullptr; 4995 } 4996 if (CommutedMI != &MI) { 4997 // New instruction. We can't fold from this. 4998 CommutedMI->eraseFromParent(); 4999 return nullptr; 5000 } 5001 5002 // Attempt to fold with the commuted version of the instruction. 5003 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt, 5004 Size, Align, /*AllowCommute=*/false); 5005 if (NewMI) 5006 return NewMI; 5007 5008 // Folding failed again - undo the commute before returning. 5009 MachineInstr *UncommutedMI = 5010 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2); 5011 if (!UncommutedMI) { 5012 // Unable to commute. 5013 return nullptr; 5014 } 5015 if (UncommutedMI != &MI) { 5016 // New instruction. It doesn't need to be kept. 5017 UncommutedMI->eraseFromParent(); 5018 return nullptr; 5019 } 5020 5021 // Return here to prevent duplicate fuse failure report. 5022 return nullptr; 5023 } 5024 } 5025 5026 // No fusion 5027 if (PrintFailedFusing && !MI.isCopy()) 5028 dbgs() << "We failed to fuse operand " << OpNum << " in " << MI; 5029 return nullptr; 5030 } 5031 5032 MachineInstr * 5033 X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI, 5034 ArrayRef<unsigned> Ops, 5035 MachineBasicBlock::iterator InsertPt, 5036 int FrameIndex, LiveIntervals *LIS) const { 5037 // Check switch flag 5038 if (NoFusing) 5039 return nullptr; 5040 5041 // Avoid partial and undef register update stalls unless optimizing for size. 5042 if (!MF.getFunction().optForSize() && 5043 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) || 5044 shouldPreventUndefRegUpdateMemFold(MF, MI))) 5045 return nullptr; 5046 5047 // Don't fold subreg spills, or reloads that use a high subreg. 5048 for (auto Op : Ops) { 5049 MachineOperand &MO = MI.getOperand(Op); 5050 auto SubReg = MO.getSubReg(); 5051 if (SubReg && (MO.isDef() || SubReg == X86::sub_8bit_hi)) 5052 return nullptr; 5053 } 5054 5055 const MachineFrameInfo &MFI = MF.getFrameInfo(); 5056 unsigned Size = MFI.getObjectSize(FrameIndex); 5057 unsigned Alignment = MFI.getObjectAlignment(FrameIndex); 5058 // If the function stack isn't realigned we don't want to fold instructions 5059 // that need increased alignment. 5060 if (!RI.needsStackRealignment(MF)) 5061 Alignment = 5062 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment()); 5063 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 5064 unsigned NewOpc = 0; 5065 unsigned RCSize = 0; 5066 switch (MI.getOpcode()) { 5067 default: return nullptr; 5068 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break; 5069 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break; 5070 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break; 5071 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break; 5072 } 5073 // Check if it's safe to fold the load. If the size of the object is 5074 // narrower than the load width, then it's not. 5075 if (Size < RCSize) 5076 return nullptr; 5077 // Change to CMPXXri r, 0 first. 5078 MI.setDesc(get(NewOpc)); 5079 MI.getOperand(1).ChangeToImmediate(0); 5080 } else if (Ops.size() != 1) 5081 return nullptr; 5082 5083 return foldMemoryOperandImpl(MF, MI, Ops[0], 5084 MachineOperand::CreateFI(FrameIndex), InsertPt, 5085 Size, Alignment, /*AllowCommute=*/true); 5086 } 5087 5088 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI 5089 /// because the latter uses contents that wouldn't be defined in the folded 5090 /// version. For instance, this transformation isn't legal: 5091 /// movss (%rdi), %xmm0 5092 /// addps %xmm0, %xmm0 5093 /// -> 5094 /// addps (%rdi), %xmm0 5095 /// 5096 /// But this one is: 5097 /// movss (%rdi), %xmm0 5098 /// addss %xmm0, %xmm0 5099 /// -> 5100 /// addss (%rdi), %xmm0 5101 /// 5102 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI, 5103 const MachineInstr &UserMI, 5104 const MachineFunction &MF) { 5105 unsigned Opc = LoadMI.getOpcode(); 5106 unsigned UserOpc = UserMI.getOpcode(); 5107 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 5108 const TargetRegisterClass *RC = 5109 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg()); 5110 unsigned RegSize = TRI.getRegSizeInBits(*RC); 5111 5112 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm || Opc == X86::VMOVSSZrm) && 5113 RegSize > 32) { 5114 // These instructions only load 32 bits, we can't fold them if the 5115 // destination register is wider than 32 bits (4 bytes), and its user 5116 // instruction isn't scalar (SS). 5117 switch (UserOpc) { 5118 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: case X86::VADDSSZrr_Int: 5119 case X86::CMPSSrr_Int: case X86::VCMPSSrr_Int: case X86::VCMPSSZrr_Int: 5120 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: case X86::VDIVSSZrr_Int: 5121 case X86::MAXSSrr_Int: case X86::VMAXSSrr_Int: case X86::VMAXSSZrr_Int: 5122 case X86::MINSSrr_Int: case X86::VMINSSrr_Int: case X86::VMINSSZrr_Int: 5123 case X86::MULSSrr_Int: case X86::VMULSSrr_Int: case X86::VMULSSZrr_Int: 5124 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: case X86::VSUBSSZrr_Int: 5125 case X86::VADDSSZrr_Intk: case X86::VADDSSZrr_Intkz: 5126 case X86::VDIVSSZrr_Intk: case X86::VDIVSSZrr_Intkz: 5127 case X86::VMAXSSZrr_Intk: case X86::VMAXSSZrr_Intkz: 5128 case X86::VMINSSZrr_Intk: case X86::VMINSSZrr_Intkz: 5129 case X86::VMULSSZrr_Intk: case X86::VMULSSZrr_Intkz: 5130 case X86::VSUBSSZrr_Intk: case X86::VSUBSSZrr_Intkz: 5131 case X86::VFMADDSS4rr_Int: case X86::VFNMADDSS4rr_Int: 5132 case X86::VFMSUBSS4rr_Int: case X86::VFNMSUBSS4rr_Int: 5133 case X86::VFMADD132SSr_Int: case X86::VFNMADD132SSr_Int: 5134 case X86::VFMADD213SSr_Int: case X86::VFNMADD213SSr_Int: 5135 case X86::VFMADD231SSr_Int: case X86::VFNMADD231SSr_Int: 5136 case X86::VFMSUB132SSr_Int: case X86::VFNMSUB132SSr_Int: 5137 case X86::VFMSUB213SSr_Int: case X86::VFNMSUB213SSr_Int: 5138 case X86::VFMSUB231SSr_Int: case X86::VFNMSUB231SSr_Int: 5139 case X86::VFMADD132SSZr_Int: case X86::VFNMADD132SSZr_Int: 5140 case X86::VFMADD213SSZr_Int: case X86::VFNMADD213SSZr_Int: 5141 case X86::VFMADD231SSZr_Int: case X86::VFNMADD231SSZr_Int: 5142 case X86::VFMSUB132SSZr_Int: case X86::VFNMSUB132SSZr_Int: 5143 case X86::VFMSUB213SSZr_Int: case X86::VFNMSUB213SSZr_Int: 5144 case X86::VFMSUB231SSZr_Int: case X86::VFNMSUB231SSZr_Int: 5145 case X86::VFMADD132SSZr_Intk: case X86::VFNMADD132SSZr_Intk: 5146 case X86::VFMADD213SSZr_Intk: case X86::VFNMADD213SSZr_Intk: 5147 case X86::VFMADD231SSZr_Intk: case X86::VFNMADD231SSZr_Intk: 5148 case X86::VFMSUB132SSZr_Intk: case X86::VFNMSUB132SSZr_Intk: 5149 case X86::VFMSUB213SSZr_Intk: case X86::VFNMSUB213SSZr_Intk: 5150 case X86::VFMSUB231SSZr_Intk: case X86::VFNMSUB231SSZr_Intk: 5151 case X86::VFMADD132SSZr_Intkz: case X86::VFNMADD132SSZr_Intkz: 5152 case X86::VFMADD213SSZr_Intkz: case X86::VFNMADD213SSZr_Intkz: 5153 case X86::VFMADD231SSZr_Intkz: case X86::VFNMADD231SSZr_Intkz: 5154 case X86::VFMSUB132SSZr_Intkz: case X86::VFNMSUB132SSZr_Intkz: 5155 case X86::VFMSUB213SSZr_Intkz: case X86::VFNMSUB213SSZr_Intkz: 5156 case X86::VFMSUB231SSZr_Intkz: case X86::VFNMSUB231SSZr_Intkz: 5157 return false; 5158 default: 5159 return true; 5160 } 5161 } 5162 5163 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm || Opc == X86::VMOVSDZrm) && 5164 RegSize > 64) { 5165 // These instructions only load 64 bits, we can't fold them if the 5166 // destination register is wider than 64 bits (8 bytes), and its user 5167 // instruction isn't scalar (SD). 5168 switch (UserOpc) { 5169 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: case X86::VADDSDZrr_Int: 5170 case X86::CMPSDrr_Int: case X86::VCMPSDrr_Int: case X86::VCMPSDZrr_Int: 5171 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: case X86::VDIVSDZrr_Int: 5172 case X86::MAXSDrr_Int: case X86::VMAXSDrr_Int: case X86::VMAXSDZrr_Int: 5173 case X86::MINSDrr_Int: case X86::VMINSDrr_Int: case X86::VMINSDZrr_Int: 5174 case X86::MULSDrr_Int: case X86::VMULSDrr_Int: case X86::VMULSDZrr_Int: 5175 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: case X86::VSUBSDZrr_Int: 5176 case X86::VADDSDZrr_Intk: case X86::VADDSDZrr_Intkz: 5177 case X86::VDIVSDZrr_Intk: case X86::VDIVSDZrr_Intkz: 5178 case X86::VMAXSDZrr_Intk: case X86::VMAXSDZrr_Intkz: 5179 case X86::VMINSDZrr_Intk: case X86::VMINSDZrr_Intkz: 5180 case X86::VMULSDZrr_Intk: case X86::VMULSDZrr_Intkz: 5181 case X86::VSUBSDZrr_Intk: case X86::VSUBSDZrr_Intkz: 5182 case X86::VFMADDSD4rr_Int: case X86::VFNMADDSD4rr_Int: 5183 case X86::VFMSUBSD4rr_Int: case X86::VFNMSUBSD4rr_Int: 5184 case X86::VFMADD132SDr_Int: case X86::VFNMADD132SDr_Int: 5185 case X86::VFMADD213SDr_Int: case X86::VFNMADD213SDr_Int: 5186 case X86::VFMADD231SDr_Int: case X86::VFNMADD231SDr_Int: 5187 case X86::VFMSUB132SDr_Int: case X86::VFNMSUB132SDr_Int: 5188 case X86::VFMSUB213SDr_Int: case X86::VFNMSUB213SDr_Int: 5189 case X86::VFMSUB231SDr_Int: case X86::VFNMSUB231SDr_Int: 5190 case X86::VFMADD132SDZr_Int: case X86::VFNMADD132SDZr_Int: 5191 case X86::VFMADD213SDZr_Int: case X86::VFNMADD213SDZr_Int: 5192 case X86::VFMADD231SDZr_Int: case X86::VFNMADD231SDZr_Int: 5193 case X86::VFMSUB132SDZr_Int: case X86::VFNMSUB132SDZr_Int: 5194 case X86::VFMSUB213SDZr_Int: case X86::VFNMSUB213SDZr_Int: 5195 case X86::VFMSUB231SDZr_Int: case X86::VFNMSUB231SDZr_Int: 5196 case X86::VFMADD132SDZr_Intk: case X86::VFNMADD132SDZr_Intk: 5197 case X86::VFMADD213SDZr_Intk: case X86::VFNMADD213SDZr_Intk: 5198 case X86::VFMADD231SDZr_Intk: case X86::VFNMADD231SDZr_Intk: 5199 case X86::VFMSUB132SDZr_Intk: case X86::VFNMSUB132SDZr_Intk: 5200 case X86::VFMSUB213SDZr_Intk: case X86::VFNMSUB213SDZr_Intk: 5201 case X86::VFMSUB231SDZr_Intk: case X86::VFNMSUB231SDZr_Intk: 5202 case X86::VFMADD132SDZr_Intkz: case X86::VFNMADD132SDZr_Intkz: 5203 case X86::VFMADD213SDZr_Intkz: case X86::VFNMADD213SDZr_Intkz: 5204 case X86::VFMADD231SDZr_Intkz: case X86::VFNMADD231SDZr_Intkz: 5205 case X86::VFMSUB132SDZr_Intkz: case X86::VFNMSUB132SDZr_Intkz: 5206 case X86::VFMSUB213SDZr_Intkz: case X86::VFNMSUB213SDZr_Intkz: 5207 case X86::VFMSUB231SDZr_Intkz: case X86::VFNMSUB231SDZr_Intkz: 5208 return false; 5209 default: 5210 return true; 5211 } 5212 } 5213 5214 return false; 5215 } 5216 5217 MachineInstr *X86InstrInfo::foldMemoryOperandImpl( 5218 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 5219 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 5220 LiveIntervals *LIS) const { 5221 5222 // TODO: Support the case where LoadMI loads a wide register, but MI 5223 // only uses a subreg. 5224 for (auto Op : Ops) { 5225 if (MI.getOperand(Op).getSubReg()) 5226 return nullptr; 5227 } 5228 5229 // If loading from a FrameIndex, fold directly from the FrameIndex. 5230 unsigned NumOps = LoadMI.getDesc().getNumOperands(); 5231 int FrameIndex; 5232 if (isLoadFromStackSlot(LoadMI, FrameIndex)) { 5233 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF)) 5234 return nullptr; 5235 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS); 5236 } 5237 5238 // Check switch flag 5239 if (NoFusing) return nullptr; 5240 5241 // Avoid partial and undef register update stalls unless optimizing for size. 5242 if (!MF.getFunction().optForSize() && 5243 (hasPartialRegUpdate(MI.getOpcode(), Subtarget) || 5244 shouldPreventUndefRegUpdateMemFold(MF, MI))) 5245 return nullptr; 5246 5247 // Determine the alignment of the load. 5248 unsigned Alignment = 0; 5249 if (LoadMI.hasOneMemOperand()) 5250 Alignment = (*LoadMI.memoperands_begin())->getAlignment(); 5251 else 5252 switch (LoadMI.getOpcode()) { 5253 case X86::AVX512_512_SET0: 5254 case X86::AVX512_512_SETALLONES: 5255 Alignment = 64; 5256 break; 5257 case X86::AVX2_SETALLONES: 5258 case X86::AVX1_SETALLONES: 5259 case X86::AVX_SET0: 5260 case X86::AVX512_256_SET0: 5261 Alignment = 32; 5262 break; 5263 case X86::V_SET0: 5264 case X86::V_SETALLONES: 5265 case X86::AVX512_128_SET0: 5266 Alignment = 16; 5267 break; 5268 case X86::MMX_SET0: 5269 case X86::FsFLD0SD: 5270 case X86::AVX512_FsFLD0SD: 5271 Alignment = 8; 5272 break; 5273 case X86::FsFLD0SS: 5274 case X86::AVX512_FsFLD0SS: 5275 Alignment = 4; 5276 break; 5277 default: 5278 return nullptr; 5279 } 5280 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 5281 unsigned NewOpc = 0; 5282 switch (MI.getOpcode()) { 5283 default: return nullptr; 5284 case X86::TEST8rr: NewOpc = X86::CMP8ri; break; 5285 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break; 5286 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break; 5287 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break; 5288 } 5289 // Change to CMPXXri r, 0 first. 5290 MI.setDesc(get(NewOpc)); 5291 MI.getOperand(1).ChangeToImmediate(0); 5292 } else if (Ops.size() != 1) 5293 return nullptr; 5294 5295 // Make sure the subregisters match. 5296 // Otherwise we risk changing the size of the load. 5297 if (LoadMI.getOperand(0).getSubReg() != MI.getOperand(Ops[0]).getSubReg()) 5298 return nullptr; 5299 5300 SmallVector<MachineOperand,X86::AddrNumOperands> MOs; 5301 switch (LoadMI.getOpcode()) { 5302 case X86::MMX_SET0: 5303 case X86::V_SET0: 5304 case X86::V_SETALLONES: 5305 case X86::AVX2_SETALLONES: 5306 case X86::AVX1_SETALLONES: 5307 case X86::AVX_SET0: 5308 case X86::AVX512_128_SET0: 5309 case X86::AVX512_256_SET0: 5310 case X86::AVX512_512_SET0: 5311 case X86::AVX512_512_SETALLONES: 5312 case X86::FsFLD0SD: 5313 case X86::AVX512_FsFLD0SD: 5314 case X86::FsFLD0SS: 5315 case X86::AVX512_FsFLD0SS: { 5316 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure. 5317 // Create a constant-pool entry and operands to load from it. 5318 5319 // Medium and large mode can't fold loads this way. 5320 if (MF.getTarget().getCodeModel() != CodeModel::Small && 5321 MF.getTarget().getCodeModel() != CodeModel::Kernel) 5322 return nullptr; 5323 5324 // x86-32 PIC requires a PIC base register for constant pools. 5325 unsigned PICBase = 0; 5326 if (MF.getTarget().isPositionIndependent()) { 5327 if (Subtarget.is64Bit()) 5328 PICBase = X86::RIP; 5329 else 5330 // FIXME: PICBase = getGlobalBaseReg(&MF); 5331 // This doesn't work for several reasons. 5332 // 1. GlobalBaseReg may have been spilled. 5333 // 2. It may not be live at MI. 5334 return nullptr; 5335 } 5336 5337 // Create a constant-pool entry. 5338 MachineConstantPool &MCP = *MF.getConstantPool(); 5339 Type *Ty; 5340 unsigned Opc = LoadMI.getOpcode(); 5341 if (Opc == X86::FsFLD0SS || Opc == X86::AVX512_FsFLD0SS) 5342 Ty = Type::getFloatTy(MF.getFunction().getContext()); 5343 else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD) 5344 Ty = Type::getDoubleTy(MF.getFunction().getContext()); 5345 else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES) 5346 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16); 5347 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 || 5348 Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES) 5349 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8); 5350 else if (Opc == X86::MMX_SET0) 5351 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2); 5352 else 5353 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 4); 5354 5355 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES || 5356 Opc == X86::AVX512_512_SETALLONES || 5357 Opc == X86::AVX1_SETALLONES); 5358 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) : 5359 Constant::getNullValue(Ty); 5360 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment); 5361 5362 // Create operands to load from the constant pool entry. 5363 MOs.push_back(MachineOperand::CreateReg(PICBase, false)); 5364 MOs.push_back(MachineOperand::CreateImm(1)); 5365 MOs.push_back(MachineOperand::CreateReg(0, false)); 5366 MOs.push_back(MachineOperand::CreateCPI(CPI, 0)); 5367 MOs.push_back(MachineOperand::CreateReg(0, false)); 5368 break; 5369 } 5370 default: { 5371 if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF)) 5372 return nullptr; 5373 5374 // Folding a normal load. Just copy the load's address operands. 5375 MOs.append(LoadMI.operands_begin() + NumOps - X86::AddrNumOperands, 5376 LoadMI.operands_begin() + NumOps); 5377 break; 5378 } 5379 } 5380 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt, 5381 /*Size=*/0, Alignment, /*AllowCommute=*/true); 5382 } 5383 5384 static SmallVector<MachineMemOperand *, 2> 5385 extractLoadMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) { 5386 SmallVector<MachineMemOperand *, 2> LoadMMOs; 5387 5388 for (MachineMemOperand *MMO : MMOs) { 5389 if (!MMO->isLoad()) 5390 continue; 5391 5392 if (!MMO->isStore()) { 5393 // Reuse the MMO. 5394 LoadMMOs.push_back(MMO); 5395 } else { 5396 // Clone the MMO and unset the store flag. 5397 LoadMMOs.push_back(MF.getMachineMemOperand( 5398 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOStore, 5399 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr, 5400 MMO->getSyncScopeID(), MMO->getOrdering(), 5401 MMO->getFailureOrdering())); 5402 } 5403 } 5404 5405 return LoadMMOs; 5406 } 5407 5408 static SmallVector<MachineMemOperand *, 2> 5409 extractStoreMMOs(ArrayRef<MachineMemOperand *> MMOs, MachineFunction &MF) { 5410 SmallVector<MachineMemOperand *, 2> StoreMMOs; 5411 5412 for (MachineMemOperand *MMO : MMOs) { 5413 if (!MMO->isStore()) 5414 continue; 5415 5416 if (!MMO->isLoad()) { 5417 // Reuse the MMO. 5418 StoreMMOs.push_back(MMO); 5419 } else { 5420 // Clone the MMO and unset the load flag. 5421 StoreMMOs.push_back(MF.getMachineMemOperand( 5422 MMO->getPointerInfo(), MMO->getFlags() & ~MachineMemOperand::MOLoad, 5423 MMO->getSize(), MMO->getBaseAlignment(), MMO->getAAInfo(), nullptr, 5424 MMO->getSyncScopeID(), MMO->getOrdering(), 5425 MMO->getFailureOrdering())); 5426 } 5427 } 5428 5429 return StoreMMOs; 5430 } 5431 5432 bool X86InstrInfo::unfoldMemoryOperand( 5433 MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad, 5434 bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const { 5435 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode()); 5436 if (I == nullptr) 5437 return false; 5438 unsigned Opc = I->DstOp; 5439 unsigned Index = I->Flags & TB_INDEX_MASK; 5440 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD; 5441 bool FoldedStore = I->Flags & TB_FOLDED_STORE; 5442 if (UnfoldLoad && !FoldedLoad) 5443 return false; 5444 UnfoldLoad &= FoldedLoad; 5445 if (UnfoldStore && !FoldedStore) 5446 return false; 5447 UnfoldStore &= FoldedStore; 5448 5449 const MCInstrDesc &MCID = get(Opc); 5450 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 5451 // TODO: Check if 32-byte or greater accesses are slow too? 5452 if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass && 5453 Subtarget.isUnalignedMem16Slow()) 5454 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will 5455 // conservatively assume the address is unaligned. That's bad for 5456 // performance. 5457 return false; 5458 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps; 5459 SmallVector<MachineOperand,2> BeforeOps; 5460 SmallVector<MachineOperand,2> AfterOps; 5461 SmallVector<MachineOperand,4> ImpOps; 5462 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 5463 MachineOperand &Op = MI.getOperand(i); 5464 if (i >= Index && i < Index + X86::AddrNumOperands) 5465 AddrOps.push_back(Op); 5466 else if (Op.isReg() && Op.isImplicit()) 5467 ImpOps.push_back(Op); 5468 else if (i < Index) 5469 BeforeOps.push_back(Op); 5470 else if (i > Index) 5471 AfterOps.push_back(Op); 5472 } 5473 5474 // Emit the load instruction. 5475 if (UnfoldLoad) { 5476 auto MMOs = extractLoadMMOs(MI.memoperands(), MF); 5477 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs, NewMIs); 5478 if (UnfoldStore) { 5479 // Address operands cannot be marked isKill. 5480 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) { 5481 MachineOperand &MO = NewMIs[0]->getOperand(i); 5482 if (MO.isReg()) 5483 MO.setIsKill(false); 5484 } 5485 } 5486 } 5487 5488 // Emit the data processing instruction. 5489 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI.getDebugLoc(), true); 5490 MachineInstrBuilder MIB(MF, DataMI); 5491 5492 if (FoldedStore) 5493 MIB.addReg(Reg, RegState::Define); 5494 for (MachineOperand &BeforeOp : BeforeOps) 5495 MIB.add(BeforeOp); 5496 if (FoldedLoad) 5497 MIB.addReg(Reg); 5498 for (MachineOperand &AfterOp : AfterOps) 5499 MIB.add(AfterOp); 5500 for (MachineOperand &ImpOp : ImpOps) { 5501 MIB.addReg(ImpOp.getReg(), 5502 getDefRegState(ImpOp.isDef()) | 5503 RegState::Implicit | 5504 getKillRegState(ImpOp.isKill()) | 5505 getDeadRegState(ImpOp.isDead()) | 5506 getUndefRegState(ImpOp.isUndef())); 5507 } 5508 // Change CMP32ri r, 0 back to TEST32rr r, r, etc. 5509 switch (DataMI->getOpcode()) { 5510 default: break; 5511 case X86::CMP64ri32: 5512 case X86::CMP64ri8: 5513 case X86::CMP32ri: 5514 case X86::CMP32ri8: 5515 case X86::CMP16ri: 5516 case X86::CMP16ri8: 5517 case X86::CMP8ri: { 5518 MachineOperand &MO0 = DataMI->getOperand(0); 5519 MachineOperand &MO1 = DataMI->getOperand(1); 5520 if (MO1.getImm() == 0) { 5521 unsigned NewOpc; 5522 switch (DataMI->getOpcode()) { 5523 default: llvm_unreachable("Unreachable!"); 5524 case X86::CMP64ri8: 5525 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break; 5526 case X86::CMP32ri8: 5527 case X86::CMP32ri: NewOpc = X86::TEST32rr; break; 5528 case X86::CMP16ri8: 5529 case X86::CMP16ri: NewOpc = X86::TEST16rr; break; 5530 case X86::CMP8ri: NewOpc = X86::TEST8rr; break; 5531 } 5532 DataMI->setDesc(get(NewOpc)); 5533 MO1.ChangeToRegister(MO0.getReg(), false); 5534 } 5535 } 5536 } 5537 NewMIs.push_back(DataMI); 5538 5539 // Emit the store instruction. 5540 if (UnfoldStore) { 5541 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF); 5542 auto MMOs = extractStoreMMOs(MI.memoperands(), MF); 5543 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs, NewMIs); 5544 } 5545 5546 return true; 5547 } 5548 5549 bool 5550 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, 5551 SmallVectorImpl<SDNode*> &NewNodes) const { 5552 if (!N->isMachineOpcode()) 5553 return false; 5554 5555 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode()); 5556 if (I == nullptr) 5557 return false; 5558 unsigned Opc = I->DstOp; 5559 unsigned Index = I->Flags & TB_INDEX_MASK; 5560 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD; 5561 bool FoldedStore = I->Flags & TB_FOLDED_STORE; 5562 const MCInstrDesc &MCID = get(Opc); 5563 MachineFunction &MF = DAG.getMachineFunction(); 5564 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); 5565 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 5566 unsigned NumDefs = MCID.NumDefs; 5567 std::vector<SDValue> AddrOps; 5568 std::vector<SDValue> BeforeOps; 5569 std::vector<SDValue> AfterOps; 5570 SDLoc dl(N); 5571 unsigned NumOps = N->getNumOperands(); 5572 for (unsigned i = 0; i != NumOps-1; ++i) { 5573 SDValue Op = N->getOperand(i); 5574 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands) 5575 AddrOps.push_back(Op); 5576 else if (i < Index-NumDefs) 5577 BeforeOps.push_back(Op); 5578 else if (i > Index-NumDefs) 5579 AfterOps.push_back(Op); 5580 } 5581 SDValue Chain = N->getOperand(NumOps-1); 5582 AddrOps.push_back(Chain); 5583 5584 // Emit the load instruction. 5585 SDNode *Load = nullptr; 5586 if (FoldedLoad) { 5587 EVT VT = *TRI.legalclasstypes_begin(*RC); 5588 auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF); 5589 if (MMOs.empty() && RC == &X86::VR128RegClass && 5590 Subtarget.isUnalignedMem16Slow()) 5591 // Do not introduce a slow unaligned load. 5592 return false; 5593 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte 5594 // memory access is slow above. 5595 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16); 5596 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment; 5597 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl, 5598 VT, MVT::Other, AddrOps); 5599 NewNodes.push_back(Load); 5600 5601 // Preserve memory reference information. 5602 DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs); 5603 } 5604 5605 // Emit the data processing instruction. 5606 std::vector<EVT> VTs; 5607 const TargetRegisterClass *DstRC = nullptr; 5608 if (MCID.getNumDefs() > 0) { 5609 DstRC = getRegClass(MCID, 0, &RI, MF); 5610 VTs.push_back(*TRI.legalclasstypes_begin(*DstRC)); 5611 } 5612 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) { 5613 EVT VT = N->getValueType(i); 5614 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs()) 5615 VTs.push_back(VT); 5616 } 5617 if (Load) 5618 BeforeOps.push_back(SDValue(Load, 0)); 5619 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end()); 5620 // Change CMP32ri r, 0 back to TEST32rr r, r, etc. 5621 switch (Opc) { 5622 default: break; 5623 case X86::CMP64ri32: 5624 case X86::CMP64ri8: 5625 case X86::CMP32ri: 5626 case X86::CMP32ri8: 5627 case X86::CMP16ri: 5628 case X86::CMP16ri8: 5629 case X86::CMP8ri: 5630 if (isNullConstant(BeforeOps[1])) { 5631 switch (Opc) { 5632 default: llvm_unreachable("Unreachable!"); 5633 case X86::CMP64ri8: 5634 case X86::CMP64ri32: Opc = X86::TEST64rr; break; 5635 case X86::CMP32ri8: 5636 case X86::CMP32ri: Opc = X86::TEST32rr; break; 5637 case X86::CMP16ri8: 5638 case X86::CMP16ri: Opc = X86::TEST16rr; break; 5639 case X86::CMP8ri: Opc = X86::TEST8rr; break; 5640 } 5641 BeforeOps[1] = BeforeOps[0]; 5642 } 5643 } 5644 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps); 5645 NewNodes.push_back(NewNode); 5646 5647 // Emit the store instruction. 5648 if (FoldedStore) { 5649 AddrOps.pop_back(); 5650 AddrOps.push_back(SDValue(NewNode, 0)); 5651 AddrOps.push_back(Chain); 5652 auto MMOs = extractStoreMMOs(cast<MachineSDNode>(N)->memoperands(), MF); 5653 if (MMOs.empty() && RC == &X86::VR128RegClass && 5654 Subtarget.isUnalignedMem16Slow()) 5655 // Do not introduce a slow unaligned store. 5656 return false; 5657 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte 5658 // memory access is slow above. 5659 unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16); 5660 bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment; 5661 SDNode *Store = 5662 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget), 5663 dl, MVT::Other, AddrOps); 5664 NewNodes.push_back(Store); 5665 5666 // Preserve memory reference information. 5667 DAG.setNodeMemRefs(cast<MachineSDNode>(Store), MMOs); 5668 } 5669 5670 return true; 5671 } 5672 5673 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc, 5674 bool UnfoldLoad, bool UnfoldStore, 5675 unsigned *LoadRegIndex) const { 5676 const X86MemoryFoldTableEntry *I = lookupUnfoldTable(Opc); 5677 if (I == nullptr) 5678 return 0; 5679 bool FoldedLoad = I->Flags & TB_FOLDED_LOAD; 5680 bool FoldedStore = I->Flags & TB_FOLDED_STORE; 5681 if (UnfoldLoad && !FoldedLoad) 5682 return 0; 5683 if (UnfoldStore && !FoldedStore) 5684 return 0; 5685 if (LoadRegIndex) 5686 *LoadRegIndex = I->Flags & TB_INDEX_MASK; 5687 return I->DstOp; 5688 } 5689 5690 bool 5691 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 5692 int64_t &Offset1, int64_t &Offset2) const { 5693 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode()) 5694 return false; 5695 unsigned Opc1 = Load1->getMachineOpcode(); 5696 unsigned Opc2 = Load2->getMachineOpcode(); 5697 switch (Opc1) { 5698 default: return false; 5699 case X86::MOV8rm: 5700 case X86::MOV16rm: 5701 case X86::MOV32rm: 5702 case X86::MOV64rm: 5703 case X86::LD_Fp32m: 5704 case X86::LD_Fp64m: 5705 case X86::LD_Fp80m: 5706 case X86::MOVSSrm: 5707 case X86::MOVSDrm: 5708 case X86::MMX_MOVD64rm: 5709 case X86::MMX_MOVQ64rm: 5710 case X86::MOVAPSrm: 5711 case X86::MOVUPSrm: 5712 case X86::MOVAPDrm: 5713 case X86::MOVUPDrm: 5714 case X86::MOVDQArm: 5715 case X86::MOVDQUrm: 5716 // AVX load instructions 5717 case X86::VMOVSSrm: 5718 case X86::VMOVSDrm: 5719 case X86::VMOVAPSrm: 5720 case X86::VMOVUPSrm: 5721 case X86::VMOVAPDrm: 5722 case X86::VMOVUPDrm: 5723 case X86::VMOVDQArm: 5724 case X86::VMOVDQUrm: 5725 case X86::VMOVAPSYrm: 5726 case X86::VMOVUPSYrm: 5727 case X86::VMOVAPDYrm: 5728 case X86::VMOVUPDYrm: 5729 case X86::VMOVDQAYrm: 5730 case X86::VMOVDQUYrm: 5731 // AVX512 load instructions 5732 case X86::VMOVSSZrm: 5733 case X86::VMOVSDZrm: 5734 case X86::VMOVAPSZ128rm: 5735 case X86::VMOVUPSZ128rm: 5736 case X86::VMOVAPSZ128rm_NOVLX: 5737 case X86::VMOVUPSZ128rm_NOVLX: 5738 case X86::VMOVAPDZ128rm: 5739 case X86::VMOVUPDZ128rm: 5740 case X86::VMOVDQU8Z128rm: 5741 case X86::VMOVDQU16Z128rm: 5742 case X86::VMOVDQA32Z128rm: 5743 case X86::VMOVDQU32Z128rm: 5744 case X86::VMOVDQA64Z128rm: 5745 case X86::VMOVDQU64Z128rm: 5746 case X86::VMOVAPSZ256rm: 5747 case X86::VMOVUPSZ256rm: 5748 case X86::VMOVAPSZ256rm_NOVLX: 5749 case X86::VMOVUPSZ256rm_NOVLX: 5750 case X86::VMOVAPDZ256rm: 5751 case X86::VMOVUPDZ256rm: 5752 case X86::VMOVDQU8Z256rm: 5753 case X86::VMOVDQU16Z256rm: 5754 case X86::VMOVDQA32Z256rm: 5755 case X86::VMOVDQU32Z256rm: 5756 case X86::VMOVDQA64Z256rm: 5757 case X86::VMOVDQU64Z256rm: 5758 case X86::VMOVAPSZrm: 5759 case X86::VMOVUPSZrm: 5760 case X86::VMOVAPDZrm: 5761 case X86::VMOVUPDZrm: 5762 case X86::VMOVDQU8Zrm: 5763 case X86::VMOVDQU16Zrm: 5764 case X86::VMOVDQA32Zrm: 5765 case X86::VMOVDQU32Zrm: 5766 case X86::VMOVDQA64Zrm: 5767 case X86::VMOVDQU64Zrm: 5768 case X86::KMOVBkm: 5769 case X86::KMOVWkm: 5770 case X86::KMOVDkm: 5771 case X86::KMOVQkm: 5772 break; 5773 } 5774 switch (Opc2) { 5775 default: return false; 5776 case X86::MOV8rm: 5777 case X86::MOV16rm: 5778 case X86::MOV32rm: 5779 case X86::MOV64rm: 5780 case X86::LD_Fp32m: 5781 case X86::LD_Fp64m: 5782 case X86::LD_Fp80m: 5783 case X86::MOVSSrm: 5784 case X86::MOVSDrm: 5785 case X86::MMX_MOVD64rm: 5786 case X86::MMX_MOVQ64rm: 5787 case X86::MOVAPSrm: 5788 case X86::MOVUPSrm: 5789 case X86::MOVAPDrm: 5790 case X86::MOVUPDrm: 5791 case X86::MOVDQArm: 5792 case X86::MOVDQUrm: 5793 // AVX load instructions 5794 case X86::VMOVSSrm: 5795 case X86::VMOVSDrm: 5796 case X86::VMOVAPSrm: 5797 case X86::VMOVUPSrm: 5798 case X86::VMOVAPDrm: 5799 case X86::VMOVUPDrm: 5800 case X86::VMOVDQArm: 5801 case X86::VMOVDQUrm: 5802 case X86::VMOVAPSYrm: 5803 case X86::VMOVUPSYrm: 5804 case X86::VMOVAPDYrm: 5805 case X86::VMOVUPDYrm: 5806 case X86::VMOVDQAYrm: 5807 case X86::VMOVDQUYrm: 5808 // AVX512 load instructions 5809 case X86::VMOVSSZrm: 5810 case X86::VMOVSDZrm: 5811 case X86::VMOVAPSZ128rm: 5812 case X86::VMOVUPSZ128rm: 5813 case X86::VMOVAPSZ128rm_NOVLX: 5814 case X86::VMOVUPSZ128rm_NOVLX: 5815 case X86::VMOVAPDZ128rm: 5816 case X86::VMOVUPDZ128rm: 5817 case X86::VMOVDQU8Z128rm: 5818 case X86::VMOVDQU16Z128rm: 5819 case X86::VMOVDQA32Z128rm: 5820 case X86::VMOVDQU32Z128rm: 5821 case X86::VMOVDQA64Z128rm: 5822 case X86::VMOVDQU64Z128rm: 5823 case X86::VMOVAPSZ256rm: 5824 case X86::VMOVUPSZ256rm: 5825 case X86::VMOVAPSZ256rm_NOVLX: 5826 case X86::VMOVUPSZ256rm_NOVLX: 5827 case X86::VMOVAPDZ256rm: 5828 case X86::VMOVUPDZ256rm: 5829 case X86::VMOVDQU8Z256rm: 5830 case X86::VMOVDQU16Z256rm: 5831 case X86::VMOVDQA32Z256rm: 5832 case X86::VMOVDQU32Z256rm: 5833 case X86::VMOVDQA64Z256rm: 5834 case X86::VMOVDQU64Z256rm: 5835 case X86::VMOVAPSZrm: 5836 case X86::VMOVUPSZrm: 5837 case X86::VMOVAPDZrm: 5838 case X86::VMOVUPDZrm: 5839 case X86::VMOVDQU8Zrm: 5840 case X86::VMOVDQU16Zrm: 5841 case X86::VMOVDQA32Zrm: 5842 case X86::VMOVDQU32Zrm: 5843 case X86::VMOVDQA64Zrm: 5844 case X86::VMOVDQU64Zrm: 5845 case X86::KMOVBkm: 5846 case X86::KMOVWkm: 5847 case X86::KMOVDkm: 5848 case X86::KMOVQkm: 5849 break; 5850 } 5851 5852 // Lambda to check if both the loads have the same value for an operand index. 5853 auto HasSameOp = [&](int I) { 5854 return Load1->getOperand(I) == Load2->getOperand(I); 5855 }; 5856 5857 // All operands except the displacement should match. 5858 if (!HasSameOp(X86::AddrBaseReg) || !HasSameOp(X86::AddrScaleAmt) || 5859 !HasSameOp(X86::AddrIndexReg) || !HasSameOp(X86::AddrSegmentReg)) 5860 return false; 5861 5862 // Chain Operand must be the same. 5863 if (!HasSameOp(5)) 5864 return false; 5865 5866 // Now let's examine if the displacements are constants. 5867 auto Disp1 = dyn_cast<ConstantSDNode>(Load1->getOperand(X86::AddrDisp)); 5868 auto Disp2 = dyn_cast<ConstantSDNode>(Load2->getOperand(X86::AddrDisp)); 5869 if (!Disp1 || !Disp2) 5870 return false; 5871 5872 Offset1 = Disp1->getSExtValue(); 5873 Offset2 = Disp2->getSExtValue(); 5874 return true; 5875 } 5876 5877 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 5878 int64_t Offset1, int64_t Offset2, 5879 unsigned NumLoads) const { 5880 assert(Offset2 > Offset1); 5881 if ((Offset2 - Offset1) / 8 > 64) 5882 return false; 5883 5884 unsigned Opc1 = Load1->getMachineOpcode(); 5885 unsigned Opc2 = Load2->getMachineOpcode(); 5886 if (Opc1 != Opc2) 5887 return false; // FIXME: overly conservative? 5888 5889 switch (Opc1) { 5890 default: break; 5891 case X86::LD_Fp32m: 5892 case X86::LD_Fp64m: 5893 case X86::LD_Fp80m: 5894 case X86::MMX_MOVD64rm: 5895 case X86::MMX_MOVQ64rm: 5896 return false; 5897 } 5898 5899 EVT VT = Load1->getValueType(0); 5900 switch (VT.getSimpleVT().SimpleTy) { 5901 default: 5902 // XMM registers. In 64-bit mode we can be a bit more aggressive since we 5903 // have 16 of them to play with. 5904 if (Subtarget.is64Bit()) { 5905 if (NumLoads >= 3) 5906 return false; 5907 } else if (NumLoads) { 5908 return false; 5909 } 5910 break; 5911 case MVT::i8: 5912 case MVT::i16: 5913 case MVT::i32: 5914 case MVT::i64: 5915 case MVT::f32: 5916 case MVT::f64: 5917 if (NumLoads) 5918 return false; 5919 break; 5920 } 5921 5922 return true; 5923 } 5924 5925 bool X86InstrInfo:: 5926 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 5927 assert(Cond.size() == 1 && "Invalid X86 branch condition!"); 5928 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm()); 5929 Cond[0].setImm(GetOppositeBranchCondition(CC)); 5930 return false; 5931 } 5932 5933 bool X86InstrInfo:: 5934 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { 5935 // FIXME: Return false for x87 stack register classes for now. We can't 5936 // allow any loads of these registers before FpGet_ST0_80. 5937 return !(RC == &X86::CCRRegClass || RC == &X86::DFCCRRegClass || 5938 RC == &X86::RFP32RegClass || RC == &X86::RFP64RegClass || 5939 RC == &X86::RFP80RegClass); 5940 } 5941 5942 /// Return a virtual register initialized with the 5943 /// the global base register value. Output instructions required to 5944 /// initialize the register in the function entry block, if necessary. 5945 /// 5946 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo. 5947 /// 5948 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const { 5949 assert((!Subtarget.is64Bit() || 5950 MF->getTarget().getCodeModel() == CodeModel::Medium || 5951 MF->getTarget().getCodeModel() == CodeModel::Large) && 5952 "X86-64 PIC uses RIP relative addressing"); 5953 5954 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); 5955 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 5956 if (GlobalBaseReg != 0) 5957 return GlobalBaseReg; 5958 5959 // Create the register. The code to initialize it is inserted 5960 // later, by the CGBR pass (below). 5961 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 5962 GlobalBaseReg = RegInfo.createVirtualRegister( 5963 Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass); 5964 X86FI->setGlobalBaseReg(GlobalBaseReg); 5965 return GlobalBaseReg; 5966 } 5967 5968 // These are the replaceable SSE instructions. Some of these have Int variants 5969 // that we don't include here. We don't want to replace instructions selected 5970 // by intrinsics. 5971 static const uint16_t ReplaceableInstrs[][3] = { 5972 //PackedSingle PackedDouble PackedInt 5973 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr }, 5974 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm }, 5975 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr }, 5976 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr }, 5977 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm }, 5978 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr }, 5979 { X86::MOVSDmr, X86::MOVSDmr, X86::MOVPQI2QImr }, 5980 { X86::MOVSSmr, X86::MOVSSmr, X86::MOVPDI2DImr }, 5981 { X86::MOVSDrm, X86::MOVSDrm, X86::MOVQI2PQIrm }, 5982 { X86::MOVSSrm, X86::MOVSSrm, X86::MOVDI2PDIrm }, 5983 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr }, 5984 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm }, 5985 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr }, 5986 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm }, 5987 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr }, 5988 { X86::ORPSrm, X86::ORPDrm, X86::PORrm }, 5989 { X86::ORPSrr, X86::ORPDrr, X86::PORrr }, 5990 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm }, 5991 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr }, 5992 { X86::UNPCKLPDrm, X86::UNPCKLPDrm, X86::PUNPCKLQDQrm }, 5993 { X86::MOVLHPSrr, X86::UNPCKLPDrr, X86::PUNPCKLQDQrr }, 5994 { X86::UNPCKHPDrm, X86::UNPCKHPDrm, X86::PUNPCKHQDQrm }, 5995 { X86::UNPCKHPDrr, X86::UNPCKHPDrr, X86::PUNPCKHQDQrr }, 5996 { X86::UNPCKLPSrm, X86::UNPCKLPSrm, X86::PUNPCKLDQrm }, 5997 { X86::UNPCKLPSrr, X86::UNPCKLPSrr, X86::PUNPCKLDQrr }, 5998 { X86::UNPCKHPSrm, X86::UNPCKHPSrm, X86::PUNPCKHDQrm }, 5999 { X86::UNPCKHPSrr, X86::UNPCKHPSrr, X86::PUNPCKHDQrr }, 6000 { X86::EXTRACTPSmr, X86::EXTRACTPSmr, X86::PEXTRDmr }, 6001 { X86::EXTRACTPSrr, X86::EXTRACTPSrr, X86::PEXTRDrr }, 6002 // AVX 128-bit support 6003 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr }, 6004 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm }, 6005 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr }, 6006 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr }, 6007 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm }, 6008 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr }, 6009 { X86::VMOVSDmr, X86::VMOVSDmr, X86::VMOVPQI2QImr }, 6010 { X86::VMOVSSmr, X86::VMOVSSmr, X86::VMOVPDI2DImr }, 6011 { X86::VMOVSDrm, X86::VMOVSDrm, X86::VMOVQI2PQIrm }, 6012 { X86::VMOVSSrm, X86::VMOVSSrm, X86::VMOVDI2PDIrm }, 6013 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr }, 6014 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm }, 6015 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr }, 6016 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm }, 6017 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr }, 6018 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm }, 6019 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr }, 6020 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm }, 6021 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr }, 6022 { X86::VUNPCKLPDrm, X86::VUNPCKLPDrm, X86::VPUNPCKLQDQrm }, 6023 { X86::VMOVLHPSrr, X86::VUNPCKLPDrr, X86::VPUNPCKLQDQrr }, 6024 { X86::VUNPCKHPDrm, X86::VUNPCKHPDrm, X86::VPUNPCKHQDQrm }, 6025 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrr, X86::VPUNPCKHQDQrr }, 6026 { X86::VUNPCKLPSrm, X86::VUNPCKLPSrm, X86::VPUNPCKLDQrm }, 6027 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrr, X86::VPUNPCKLDQrr }, 6028 { X86::VUNPCKHPSrm, X86::VUNPCKHPSrm, X86::VPUNPCKHDQrm }, 6029 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrr, X86::VPUNPCKHDQrr }, 6030 { X86::VEXTRACTPSmr, X86::VEXTRACTPSmr, X86::VPEXTRDmr }, 6031 { X86::VEXTRACTPSrr, X86::VEXTRACTPSrr, X86::VPEXTRDrr }, 6032 // AVX 256-bit support 6033 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr }, 6034 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm }, 6035 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr }, 6036 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr }, 6037 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm }, 6038 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr }, 6039 { X86::VPERMPSYrm, X86::VPERMPSYrm, X86::VPERMDYrm }, 6040 { X86::VPERMPSYrr, X86::VPERMPSYrr, X86::VPERMDYrr }, 6041 { X86::VPERMPDYmi, X86::VPERMPDYmi, X86::VPERMQYmi }, 6042 { X86::VPERMPDYri, X86::VPERMPDYri, X86::VPERMQYri }, 6043 // AVX512 support 6044 { X86::VMOVLPSZ128mr, X86::VMOVLPDZ128mr, X86::VMOVPQI2QIZmr }, 6045 { X86::VMOVNTPSZ128mr, X86::VMOVNTPDZ128mr, X86::VMOVNTDQZ128mr }, 6046 { X86::VMOVNTPSZ256mr, X86::VMOVNTPDZ256mr, X86::VMOVNTDQZ256mr }, 6047 { X86::VMOVNTPSZmr, X86::VMOVNTPDZmr, X86::VMOVNTDQZmr }, 6048 { X86::VMOVSDZmr, X86::VMOVSDZmr, X86::VMOVPQI2QIZmr }, 6049 { X86::VMOVSSZmr, X86::VMOVSSZmr, X86::VMOVPDI2DIZmr }, 6050 { X86::VMOVSDZrm, X86::VMOVSDZrm, X86::VMOVQI2PQIZrm }, 6051 { X86::VMOVSSZrm, X86::VMOVSSZrm, X86::VMOVDI2PDIZrm }, 6052 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128r, X86::VPBROADCASTDZ128r }, 6053 { X86::VBROADCASTSSZ128m, X86::VBROADCASTSSZ128m, X86::VPBROADCASTDZ128m }, 6054 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256r, X86::VPBROADCASTDZ256r }, 6055 { X86::VBROADCASTSSZ256m, X86::VBROADCASTSSZ256m, X86::VPBROADCASTDZ256m }, 6056 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZr, X86::VPBROADCASTDZr }, 6057 { X86::VBROADCASTSSZm, X86::VBROADCASTSSZm, X86::VPBROADCASTDZm }, 6058 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256r, X86::VPBROADCASTQZ256r }, 6059 { X86::VBROADCASTSDZ256m, X86::VBROADCASTSDZ256m, X86::VPBROADCASTQZ256m }, 6060 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZr, X86::VPBROADCASTQZr }, 6061 { X86::VBROADCASTSDZm, X86::VBROADCASTSDZm, X86::VPBROADCASTQZm }, 6062 { X86::VINSERTF32x4Zrr, X86::VINSERTF32x4Zrr, X86::VINSERTI32x4Zrr }, 6063 { X86::VINSERTF32x4Zrm, X86::VINSERTF32x4Zrm, X86::VINSERTI32x4Zrm }, 6064 { X86::VINSERTF32x8Zrr, X86::VINSERTF32x8Zrr, X86::VINSERTI32x8Zrr }, 6065 { X86::VINSERTF32x8Zrm, X86::VINSERTF32x8Zrm, X86::VINSERTI32x8Zrm }, 6066 { X86::VINSERTF64x2Zrr, X86::VINSERTF64x2Zrr, X86::VINSERTI64x2Zrr }, 6067 { X86::VINSERTF64x2Zrm, X86::VINSERTF64x2Zrm, X86::VINSERTI64x2Zrm }, 6068 { X86::VINSERTF64x4Zrr, X86::VINSERTF64x4Zrr, X86::VINSERTI64x4Zrr }, 6069 { X86::VINSERTF64x4Zrm, X86::VINSERTF64x4Zrm, X86::VINSERTI64x4Zrm }, 6070 { X86::VINSERTF32x4Z256rr,X86::VINSERTF32x4Z256rr,X86::VINSERTI32x4Z256rr }, 6071 { X86::VINSERTF32x4Z256rm,X86::VINSERTF32x4Z256rm,X86::VINSERTI32x4Z256rm }, 6072 { X86::VINSERTF64x2Z256rr,X86::VINSERTF64x2Z256rr,X86::VINSERTI64x2Z256rr }, 6073 { X86::VINSERTF64x2Z256rm,X86::VINSERTF64x2Z256rm,X86::VINSERTI64x2Z256rm }, 6074 { X86::VEXTRACTF32x4Zrr, X86::VEXTRACTF32x4Zrr, X86::VEXTRACTI32x4Zrr }, 6075 { X86::VEXTRACTF32x4Zmr, X86::VEXTRACTF32x4Zmr, X86::VEXTRACTI32x4Zmr }, 6076 { X86::VEXTRACTF32x8Zrr, X86::VEXTRACTF32x8Zrr, X86::VEXTRACTI32x8Zrr }, 6077 { X86::VEXTRACTF32x8Zmr, X86::VEXTRACTF32x8Zmr, X86::VEXTRACTI32x8Zmr }, 6078 { X86::VEXTRACTF64x2Zrr, X86::VEXTRACTF64x2Zrr, X86::VEXTRACTI64x2Zrr }, 6079 { X86::VEXTRACTF64x2Zmr, X86::VEXTRACTF64x2Zmr, X86::VEXTRACTI64x2Zmr }, 6080 { X86::VEXTRACTF64x4Zrr, X86::VEXTRACTF64x4Zrr, X86::VEXTRACTI64x4Zrr }, 6081 { X86::VEXTRACTF64x4Zmr, X86::VEXTRACTF64x4Zmr, X86::VEXTRACTI64x4Zmr }, 6082 { X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTF32x4Z256rr,X86::VEXTRACTI32x4Z256rr }, 6083 { X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTF32x4Z256mr,X86::VEXTRACTI32x4Z256mr }, 6084 { X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTF64x2Z256rr,X86::VEXTRACTI64x2Z256rr }, 6085 { X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTF64x2Z256mr,X86::VEXTRACTI64x2Z256mr }, 6086 { X86::VPERMILPSmi, X86::VPERMILPSmi, X86::VPSHUFDmi }, 6087 { X86::VPERMILPSri, X86::VPERMILPSri, X86::VPSHUFDri }, 6088 { X86::VPERMILPSZ128mi, X86::VPERMILPSZ128mi, X86::VPSHUFDZ128mi }, 6089 { X86::VPERMILPSZ128ri, X86::VPERMILPSZ128ri, X86::VPSHUFDZ128ri }, 6090 { X86::VPERMILPSZ256mi, X86::VPERMILPSZ256mi, X86::VPSHUFDZ256mi }, 6091 { X86::VPERMILPSZ256ri, X86::VPERMILPSZ256ri, X86::VPSHUFDZ256ri }, 6092 { X86::VPERMILPSZmi, X86::VPERMILPSZmi, X86::VPSHUFDZmi }, 6093 { X86::VPERMILPSZri, X86::VPERMILPSZri, X86::VPSHUFDZri }, 6094 { X86::VPERMPSZ256rm, X86::VPERMPSZ256rm, X86::VPERMDZ256rm }, 6095 { X86::VPERMPSZ256rr, X86::VPERMPSZ256rr, X86::VPERMDZ256rr }, 6096 { X86::VPERMPDZ256mi, X86::VPERMPDZ256mi, X86::VPERMQZ256mi }, 6097 { X86::VPERMPDZ256ri, X86::VPERMPDZ256ri, X86::VPERMQZ256ri }, 6098 { X86::VPERMPDZ256rm, X86::VPERMPDZ256rm, X86::VPERMQZ256rm }, 6099 { X86::VPERMPDZ256rr, X86::VPERMPDZ256rr, X86::VPERMQZ256rr }, 6100 { X86::VPERMPSZrm, X86::VPERMPSZrm, X86::VPERMDZrm }, 6101 { X86::VPERMPSZrr, X86::VPERMPSZrr, X86::VPERMDZrr }, 6102 { X86::VPERMPDZmi, X86::VPERMPDZmi, X86::VPERMQZmi }, 6103 { X86::VPERMPDZri, X86::VPERMPDZri, X86::VPERMQZri }, 6104 { X86::VPERMPDZrm, X86::VPERMPDZrm, X86::VPERMQZrm }, 6105 { X86::VPERMPDZrr, X86::VPERMPDZrr, X86::VPERMQZrr }, 6106 { X86::VUNPCKLPDZ256rm, X86::VUNPCKLPDZ256rm, X86::VPUNPCKLQDQZ256rm }, 6107 { X86::VUNPCKLPDZ256rr, X86::VUNPCKLPDZ256rr, X86::VPUNPCKLQDQZ256rr }, 6108 { X86::VUNPCKHPDZ256rm, X86::VUNPCKHPDZ256rm, X86::VPUNPCKHQDQZ256rm }, 6109 { X86::VUNPCKHPDZ256rr, X86::VUNPCKHPDZ256rr, X86::VPUNPCKHQDQZ256rr }, 6110 { X86::VUNPCKLPSZ256rm, X86::VUNPCKLPSZ256rm, X86::VPUNPCKLDQZ256rm }, 6111 { X86::VUNPCKLPSZ256rr, X86::VUNPCKLPSZ256rr, X86::VPUNPCKLDQZ256rr }, 6112 { X86::VUNPCKHPSZ256rm, X86::VUNPCKHPSZ256rm, X86::VPUNPCKHDQZ256rm }, 6113 { X86::VUNPCKHPSZ256rr, X86::VUNPCKHPSZ256rr, X86::VPUNPCKHDQZ256rr }, 6114 { X86::VUNPCKLPDZ128rm, X86::VUNPCKLPDZ128rm, X86::VPUNPCKLQDQZ128rm }, 6115 { X86::VMOVLHPSZrr, X86::VUNPCKLPDZ128rr, X86::VPUNPCKLQDQZ128rr }, 6116 { X86::VUNPCKHPDZ128rm, X86::VUNPCKHPDZ128rm, X86::VPUNPCKHQDQZ128rm }, 6117 { X86::VUNPCKHPDZ128rr, X86::VUNPCKHPDZ128rr, X86::VPUNPCKHQDQZ128rr }, 6118 { X86::VUNPCKLPSZ128rm, X86::VUNPCKLPSZ128rm, X86::VPUNPCKLDQZ128rm }, 6119 { X86::VUNPCKLPSZ128rr, X86::VUNPCKLPSZ128rr, X86::VPUNPCKLDQZ128rr }, 6120 { X86::VUNPCKHPSZ128rm, X86::VUNPCKHPSZ128rm, X86::VPUNPCKHDQZ128rm }, 6121 { X86::VUNPCKHPSZ128rr, X86::VUNPCKHPSZ128rr, X86::VPUNPCKHDQZ128rr }, 6122 { X86::VUNPCKLPDZrm, X86::VUNPCKLPDZrm, X86::VPUNPCKLQDQZrm }, 6123 { X86::VUNPCKLPDZrr, X86::VUNPCKLPDZrr, X86::VPUNPCKLQDQZrr }, 6124 { X86::VUNPCKHPDZrm, X86::VUNPCKHPDZrm, X86::VPUNPCKHQDQZrm }, 6125 { X86::VUNPCKHPDZrr, X86::VUNPCKHPDZrr, X86::VPUNPCKHQDQZrr }, 6126 { X86::VUNPCKLPSZrm, X86::VUNPCKLPSZrm, X86::VPUNPCKLDQZrm }, 6127 { X86::VUNPCKLPSZrr, X86::VUNPCKLPSZrr, X86::VPUNPCKLDQZrr }, 6128 { X86::VUNPCKHPSZrm, X86::VUNPCKHPSZrm, X86::VPUNPCKHDQZrm }, 6129 { X86::VUNPCKHPSZrr, X86::VUNPCKHPSZrr, X86::VPUNPCKHDQZrr }, 6130 { X86::VEXTRACTPSZmr, X86::VEXTRACTPSZmr, X86::VPEXTRDZmr }, 6131 { X86::VEXTRACTPSZrr, X86::VEXTRACTPSZrr, X86::VPEXTRDZrr }, 6132 }; 6133 6134 static const uint16_t ReplaceableInstrsAVX2[][3] = { 6135 //PackedSingle PackedDouble PackedInt 6136 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm }, 6137 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr }, 6138 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm }, 6139 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr }, 6140 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm }, 6141 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr }, 6142 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm }, 6143 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr }, 6144 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm }, 6145 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }, 6146 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm}, 6147 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr}, 6148 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr}, 6149 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm}, 6150 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr}, 6151 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}, 6152 { X86::VBROADCASTF128, X86::VBROADCASTF128, X86::VBROADCASTI128 }, 6153 { X86::VBLENDPSYrri, X86::VBLENDPSYrri, X86::VPBLENDDYrri }, 6154 { X86::VBLENDPSYrmi, X86::VBLENDPSYrmi, X86::VPBLENDDYrmi }, 6155 { X86::VPERMILPSYmi, X86::VPERMILPSYmi, X86::VPSHUFDYmi }, 6156 { X86::VPERMILPSYri, X86::VPERMILPSYri, X86::VPSHUFDYri }, 6157 { X86::VUNPCKLPDYrm, X86::VUNPCKLPDYrm, X86::VPUNPCKLQDQYrm }, 6158 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrr, X86::VPUNPCKLQDQYrr }, 6159 { X86::VUNPCKHPDYrm, X86::VUNPCKHPDYrm, X86::VPUNPCKHQDQYrm }, 6160 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrr, X86::VPUNPCKHQDQYrr }, 6161 { X86::VUNPCKLPSYrm, X86::VUNPCKLPSYrm, X86::VPUNPCKLDQYrm }, 6162 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrr, X86::VPUNPCKLDQYrr }, 6163 { X86::VUNPCKHPSYrm, X86::VUNPCKHPSYrm, X86::VPUNPCKHDQYrm }, 6164 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrr, X86::VPUNPCKHDQYrr }, 6165 }; 6166 6167 static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = { 6168 //PackedSingle PackedDouble PackedInt 6169 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr }, 6170 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr }, 6171 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm }, 6172 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr }, 6173 }; 6174 6175 static const uint16_t ReplaceableInstrsAVX512[][4] = { 6176 // Two integer columns for 64-bit and 32-bit elements. 6177 //PackedSingle PackedDouble PackedInt PackedInt 6178 { X86::VMOVAPSZ128mr, X86::VMOVAPDZ128mr, X86::VMOVDQA64Z128mr, X86::VMOVDQA32Z128mr }, 6179 { X86::VMOVAPSZ128rm, X86::VMOVAPDZ128rm, X86::VMOVDQA64Z128rm, X86::VMOVDQA32Z128rm }, 6180 { X86::VMOVAPSZ128rr, X86::VMOVAPDZ128rr, X86::VMOVDQA64Z128rr, X86::VMOVDQA32Z128rr }, 6181 { X86::VMOVUPSZ128mr, X86::VMOVUPDZ128mr, X86::VMOVDQU64Z128mr, X86::VMOVDQU32Z128mr }, 6182 { X86::VMOVUPSZ128rm, X86::VMOVUPDZ128rm, X86::VMOVDQU64Z128rm, X86::VMOVDQU32Z128rm }, 6183 { X86::VMOVAPSZ256mr, X86::VMOVAPDZ256mr, X86::VMOVDQA64Z256mr, X86::VMOVDQA32Z256mr }, 6184 { X86::VMOVAPSZ256rm, X86::VMOVAPDZ256rm, X86::VMOVDQA64Z256rm, X86::VMOVDQA32Z256rm }, 6185 { X86::VMOVAPSZ256rr, X86::VMOVAPDZ256rr, X86::VMOVDQA64Z256rr, X86::VMOVDQA32Z256rr }, 6186 { X86::VMOVUPSZ256mr, X86::VMOVUPDZ256mr, X86::VMOVDQU64Z256mr, X86::VMOVDQU32Z256mr }, 6187 { X86::VMOVUPSZ256rm, X86::VMOVUPDZ256rm, X86::VMOVDQU64Z256rm, X86::VMOVDQU32Z256rm }, 6188 { X86::VMOVAPSZmr, X86::VMOVAPDZmr, X86::VMOVDQA64Zmr, X86::VMOVDQA32Zmr }, 6189 { X86::VMOVAPSZrm, X86::VMOVAPDZrm, X86::VMOVDQA64Zrm, X86::VMOVDQA32Zrm }, 6190 { X86::VMOVAPSZrr, X86::VMOVAPDZrr, X86::VMOVDQA64Zrr, X86::VMOVDQA32Zrr }, 6191 { X86::VMOVUPSZmr, X86::VMOVUPDZmr, X86::VMOVDQU64Zmr, X86::VMOVDQU32Zmr }, 6192 { X86::VMOVUPSZrm, X86::VMOVUPDZrm, X86::VMOVDQU64Zrm, X86::VMOVDQU32Zrm }, 6193 }; 6194 6195 static const uint16_t ReplaceableInstrsAVX512DQ[][4] = { 6196 // Two integer columns for 64-bit and 32-bit elements. 6197 //PackedSingle PackedDouble PackedInt PackedInt 6198 { X86::VANDNPSZ128rm, X86::VANDNPDZ128rm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm }, 6199 { X86::VANDNPSZ128rr, X86::VANDNPDZ128rr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr }, 6200 { X86::VANDPSZ128rm, X86::VANDPDZ128rm, X86::VPANDQZ128rm, X86::VPANDDZ128rm }, 6201 { X86::VANDPSZ128rr, X86::VANDPDZ128rr, X86::VPANDQZ128rr, X86::VPANDDZ128rr }, 6202 { X86::VORPSZ128rm, X86::VORPDZ128rm, X86::VPORQZ128rm, X86::VPORDZ128rm }, 6203 { X86::VORPSZ128rr, X86::VORPDZ128rr, X86::VPORQZ128rr, X86::VPORDZ128rr }, 6204 { X86::VXORPSZ128rm, X86::VXORPDZ128rm, X86::VPXORQZ128rm, X86::VPXORDZ128rm }, 6205 { X86::VXORPSZ128rr, X86::VXORPDZ128rr, X86::VPXORQZ128rr, X86::VPXORDZ128rr }, 6206 { X86::VANDNPSZ256rm, X86::VANDNPDZ256rm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm }, 6207 { X86::VANDNPSZ256rr, X86::VANDNPDZ256rr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr }, 6208 { X86::VANDPSZ256rm, X86::VANDPDZ256rm, X86::VPANDQZ256rm, X86::VPANDDZ256rm }, 6209 { X86::VANDPSZ256rr, X86::VANDPDZ256rr, X86::VPANDQZ256rr, X86::VPANDDZ256rr }, 6210 { X86::VORPSZ256rm, X86::VORPDZ256rm, X86::VPORQZ256rm, X86::VPORDZ256rm }, 6211 { X86::VORPSZ256rr, X86::VORPDZ256rr, X86::VPORQZ256rr, X86::VPORDZ256rr }, 6212 { X86::VXORPSZ256rm, X86::VXORPDZ256rm, X86::VPXORQZ256rm, X86::VPXORDZ256rm }, 6213 { X86::VXORPSZ256rr, X86::VXORPDZ256rr, X86::VPXORQZ256rr, X86::VPXORDZ256rr }, 6214 { X86::VANDNPSZrm, X86::VANDNPDZrm, X86::VPANDNQZrm, X86::VPANDNDZrm }, 6215 { X86::VANDNPSZrr, X86::VANDNPDZrr, X86::VPANDNQZrr, X86::VPANDNDZrr }, 6216 { X86::VANDPSZrm, X86::VANDPDZrm, X86::VPANDQZrm, X86::VPANDDZrm }, 6217 { X86::VANDPSZrr, X86::VANDPDZrr, X86::VPANDQZrr, X86::VPANDDZrr }, 6218 { X86::VORPSZrm, X86::VORPDZrm, X86::VPORQZrm, X86::VPORDZrm }, 6219 { X86::VORPSZrr, X86::VORPDZrr, X86::VPORQZrr, X86::VPORDZrr }, 6220 { X86::VXORPSZrm, X86::VXORPDZrm, X86::VPXORQZrm, X86::VPXORDZrm }, 6221 { X86::VXORPSZrr, X86::VXORPDZrr, X86::VPXORQZrr, X86::VPXORDZrr }, 6222 }; 6223 6224 static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = { 6225 // Two integer columns for 64-bit and 32-bit elements. 6226 //PackedSingle PackedDouble 6227 //PackedInt PackedInt 6228 { X86::VANDNPSZ128rmk, X86::VANDNPDZ128rmk, 6229 X86::VPANDNQZ128rmk, X86::VPANDNDZ128rmk }, 6230 { X86::VANDNPSZ128rmkz, X86::VANDNPDZ128rmkz, 6231 X86::VPANDNQZ128rmkz, X86::VPANDNDZ128rmkz }, 6232 { X86::VANDNPSZ128rrk, X86::VANDNPDZ128rrk, 6233 X86::VPANDNQZ128rrk, X86::VPANDNDZ128rrk }, 6234 { X86::VANDNPSZ128rrkz, X86::VANDNPDZ128rrkz, 6235 X86::VPANDNQZ128rrkz, X86::VPANDNDZ128rrkz }, 6236 { X86::VANDPSZ128rmk, X86::VANDPDZ128rmk, 6237 X86::VPANDQZ128rmk, X86::VPANDDZ128rmk }, 6238 { X86::VANDPSZ128rmkz, X86::VANDPDZ128rmkz, 6239 X86::VPANDQZ128rmkz, X86::VPANDDZ128rmkz }, 6240 { X86::VANDPSZ128rrk, X86::VANDPDZ128rrk, 6241 X86::VPANDQZ128rrk, X86::VPANDDZ128rrk }, 6242 { X86::VANDPSZ128rrkz, X86::VANDPDZ128rrkz, 6243 X86::VPANDQZ128rrkz, X86::VPANDDZ128rrkz }, 6244 { X86::VORPSZ128rmk, X86::VORPDZ128rmk, 6245 X86::VPORQZ128rmk, X86::VPORDZ128rmk }, 6246 { X86::VORPSZ128rmkz, X86::VORPDZ128rmkz, 6247 X86::VPORQZ128rmkz, X86::VPORDZ128rmkz }, 6248 { X86::VORPSZ128rrk, X86::VORPDZ128rrk, 6249 X86::VPORQZ128rrk, X86::VPORDZ128rrk }, 6250 { X86::VORPSZ128rrkz, X86::VORPDZ128rrkz, 6251 X86::VPORQZ128rrkz, X86::VPORDZ128rrkz }, 6252 { X86::VXORPSZ128rmk, X86::VXORPDZ128rmk, 6253 X86::VPXORQZ128rmk, X86::VPXORDZ128rmk }, 6254 { X86::VXORPSZ128rmkz, X86::VXORPDZ128rmkz, 6255 X86::VPXORQZ128rmkz, X86::VPXORDZ128rmkz }, 6256 { X86::VXORPSZ128rrk, X86::VXORPDZ128rrk, 6257 X86::VPXORQZ128rrk, X86::VPXORDZ128rrk }, 6258 { X86::VXORPSZ128rrkz, X86::VXORPDZ128rrkz, 6259 X86::VPXORQZ128rrkz, X86::VPXORDZ128rrkz }, 6260 { X86::VANDNPSZ256rmk, X86::VANDNPDZ256rmk, 6261 X86::VPANDNQZ256rmk, X86::VPANDNDZ256rmk }, 6262 { X86::VANDNPSZ256rmkz, X86::VANDNPDZ256rmkz, 6263 X86::VPANDNQZ256rmkz, X86::VPANDNDZ256rmkz }, 6264 { X86::VANDNPSZ256rrk, X86::VANDNPDZ256rrk, 6265 X86::VPANDNQZ256rrk, X86::VPANDNDZ256rrk }, 6266 { X86::VANDNPSZ256rrkz, X86::VANDNPDZ256rrkz, 6267 X86::VPANDNQZ256rrkz, X86::VPANDNDZ256rrkz }, 6268 { X86::VANDPSZ256rmk, X86::VANDPDZ256rmk, 6269 X86::VPANDQZ256rmk, X86::VPANDDZ256rmk }, 6270 { X86::VANDPSZ256rmkz, X86::VANDPDZ256rmkz, 6271 X86::VPANDQZ256rmkz, X86::VPANDDZ256rmkz }, 6272 { X86::VANDPSZ256rrk, X86::VANDPDZ256rrk, 6273 X86::VPANDQZ256rrk, X86::VPANDDZ256rrk }, 6274 { X86::VANDPSZ256rrkz, X86::VANDPDZ256rrkz, 6275 X86::VPANDQZ256rrkz, X86::VPANDDZ256rrkz }, 6276 { X86::VORPSZ256rmk, X86::VORPDZ256rmk, 6277 X86::VPORQZ256rmk, X86::VPORDZ256rmk }, 6278 { X86::VORPSZ256rmkz, X86::VORPDZ256rmkz, 6279 X86::VPORQZ256rmkz, X86::VPORDZ256rmkz }, 6280 { X86::VORPSZ256rrk, X86::VORPDZ256rrk, 6281 X86::VPORQZ256rrk, X86::VPORDZ256rrk }, 6282 { X86::VORPSZ256rrkz, X86::VORPDZ256rrkz, 6283 X86::VPORQZ256rrkz, X86::VPORDZ256rrkz }, 6284 { X86::VXORPSZ256rmk, X86::VXORPDZ256rmk, 6285 X86::VPXORQZ256rmk, X86::VPXORDZ256rmk }, 6286 { X86::VXORPSZ256rmkz, X86::VXORPDZ256rmkz, 6287 X86::VPXORQZ256rmkz, X86::VPXORDZ256rmkz }, 6288 { X86::VXORPSZ256rrk, X86::VXORPDZ256rrk, 6289 X86::VPXORQZ256rrk, X86::VPXORDZ256rrk }, 6290 { X86::VXORPSZ256rrkz, X86::VXORPDZ256rrkz, 6291 X86::VPXORQZ256rrkz, X86::VPXORDZ256rrkz }, 6292 { X86::VANDNPSZrmk, X86::VANDNPDZrmk, 6293 X86::VPANDNQZrmk, X86::VPANDNDZrmk }, 6294 { X86::VANDNPSZrmkz, X86::VANDNPDZrmkz, 6295 X86::VPANDNQZrmkz, X86::VPANDNDZrmkz }, 6296 { X86::VANDNPSZrrk, X86::VANDNPDZrrk, 6297 X86::VPANDNQZrrk, X86::VPANDNDZrrk }, 6298 { X86::VANDNPSZrrkz, X86::VANDNPDZrrkz, 6299 X86::VPANDNQZrrkz, X86::VPANDNDZrrkz }, 6300 { X86::VANDPSZrmk, X86::VANDPDZrmk, 6301 X86::VPANDQZrmk, X86::VPANDDZrmk }, 6302 { X86::VANDPSZrmkz, X86::VANDPDZrmkz, 6303 X86::VPANDQZrmkz, X86::VPANDDZrmkz }, 6304 { X86::VANDPSZrrk, X86::VANDPDZrrk, 6305 X86::VPANDQZrrk, X86::VPANDDZrrk }, 6306 { X86::VANDPSZrrkz, X86::VANDPDZrrkz, 6307 X86::VPANDQZrrkz, X86::VPANDDZrrkz }, 6308 { X86::VORPSZrmk, X86::VORPDZrmk, 6309 X86::VPORQZrmk, X86::VPORDZrmk }, 6310 { X86::VORPSZrmkz, X86::VORPDZrmkz, 6311 X86::VPORQZrmkz, X86::VPORDZrmkz }, 6312 { X86::VORPSZrrk, X86::VORPDZrrk, 6313 X86::VPORQZrrk, X86::VPORDZrrk }, 6314 { X86::VORPSZrrkz, X86::VORPDZrrkz, 6315 X86::VPORQZrrkz, X86::VPORDZrrkz }, 6316 { X86::VXORPSZrmk, X86::VXORPDZrmk, 6317 X86::VPXORQZrmk, X86::VPXORDZrmk }, 6318 { X86::VXORPSZrmkz, X86::VXORPDZrmkz, 6319 X86::VPXORQZrmkz, X86::VPXORDZrmkz }, 6320 { X86::VXORPSZrrk, X86::VXORPDZrrk, 6321 X86::VPXORQZrrk, X86::VPXORDZrrk }, 6322 { X86::VXORPSZrrkz, X86::VXORPDZrrkz, 6323 X86::VPXORQZrrkz, X86::VPXORDZrrkz }, 6324 // Broadcast loads can be handled the same as masked operations to avoid 6325 // changing element size. 6326 { X86::VANDNPSZ128rmb, X86::VANDNPDZ128rmb, 6327 X86::VPANDNQZ128rmb, X86::VPANDNDZ128rmb }, 6328 { X86::VANDPSZ128rmb, X86::VANDPDZ128rmb, 6329 X86::VPANDQZ128rmb, X86::VPANDDZ128rmb }, 6330 { X86::VORPSZ128rmb, X86::VORPDZ128rmb, 6331 X86::VPORQZ128rmb, X86::VPORDZ128rmb }, 6332 { X86::VXORPSZ128rmb, X86::VXORPDZ128rmb, 6333 X86::VPXORQZ128rmb, X86::VPXORDZ128rmb }, 6334 { X86::VANDNPSZ256rmb, X86::VANDNPDZ256rmb, 6335 X86::VPANDNQZ256rmb, X86::VPANDNDZ256rmb }, 6336 { X86::VANDPSZ256rmb, X86::VANDPDZ256rmb, 6337 X86::VPANDQZ256rmb, X86::VPANDDZ256rmb }, 6338 { X86::VORPSZ256rmb, X86::VORPDZ256rmb, 6339 X86::VPORQZ256rmb, X86::VPORDZ256rmb }, 6340 { X86::VXORPSZ256rmb, X86::VXORPDZ256rmb, 6341 X86::VPXORQZ256rmb, X86::VPXORDZ256rmb }, 6342 { X86::VANDNPSZrmb, X86::VANDNPDZrmb, 6343 X86::VPANDNQZrmb, X86::VPANDNDZrmb }, 6344 { X86::VANDPSZrmb, X86::VANDPDZrmb, 6345 X86::VPANDQZrmb, X86::VPANDDZrmb }, 6346 { X86::VANDPSZrmb, X86::VANDPDZrmb, 6347 X86::VPANDQZrmb, X86::VPANDDZrmb }, 6348 { X86::VORPSZrmb, X86::VORPDZrmb, 6349 X86::VPORQZrmb, X86::VPORDZrmb }, 6350 { X86::VXORPSZrmb, X86::VXORPDZrmb, 6351 X86::VPXORQZrmb, X86::VPXORDZrmb }, 6352 { X86::VANDNPSZ128rmbk, X86::VANDNPDZ128rmbk, 6353 X86::VPANDNQZ128rmbk, X86::VPANDNDZ128rmbk }, 6354 { X86::VANDPSZ128rmbk, X86::VANDPDZ128rmbk, 6355 X86::VPANDQZ128rmbk, X86::VPANDDZ128rmbk }, 6356 { X86::VORPSZ128rmbk, X86::VORPDZ128rmbk, 6357 X86::VPORQZ128rmbk, X86::VPORDZ128rmbk }, 6358 { X86::VXORPSZ128rmbk, X86::VXORPDZ128rmbk, 6359 X86::VPXORQZ128rmbk, X86::VPXORDZ128rmbk }, 6360 { X86::VANDNPSZ256rmbk, X86::VANDNPDZ256rmbk, 6361 X86::VPANDNQZ256rmbk, X86::VPANDNDZ256rmbk }, 6362 { X86::VANDPSZ256rmbk, X86::VANDPDZ256rmbk, 6363 X86::VPANDQZ256rmbk, X86::VPANDDZ256rmbk }, 6364 { X86::VORPSZ256rmbk, X86::VORPDZ256rmbk, 6365 X86::VPORQZ256rmbk, X86::VPORDZ256rmbk }, 6366 { X86::VXORPSZ256rmbk, X86::VXORPDZ256rmbk, 6367 X86::VPXORQZ256rmbk, X86::VPXORDZ256rmbk }, 6368 { X86::VANDNPSZrmbk, X86::VANDNPDZrmbk, 6369 X86::VPANDNQZrmbk, X86::VPANDNDZrmbk }, 6370 { X86::VANDPSZrmbk, X86::VANDPDZrmbk, 6371 X86::VPANDQZrmbk, X86::VPANDDZrmbk }, 6372 { X86::VANDPSZrmbk, X86::VANDPDZrmbk, 6373 X86::VPANDQZrmbk, X86::VPANDDZrmbk }, 6374 { X86::VORPSZrmbk, X86::VORPDZrmbk, 6375 X86::VPORQZrmbk, X86::VPORDZrmbk }, 6376 { X86::VXORPSZrmbk, X86::VXORPDZrmbk, 6377 X86::VPXORQZrmbk, X86::VPXORDZrmbk }, 6378 { X86::VANDNPSZ128rmbkz,X86::VANDNPDZ128rmbkz, 6379 X86::VPANDNQZ128rmbkz,X86::VPANDNDZ128rmbkz}, 6380 { X86::VANDPSZ128rmbkz, X86::VANDPDZ128rmbkz, 6381 X86::VPANDQZ128rmbkz, X86::VPANDDZ128rmbkz }, 6382 { X86::VORPSZ128rmbkz, X86::VORPDZ128rmbkz, 6383 X86::VPORQZ128rmbkz, X86::VPORDZ128rmbkz }, 6384 { X86::VXORPSZ128rmbkz, X86::VXORPDZ128rmbkz, 6385 X86::VPXORQZ128rmbkz, X86::VPXORDZ128rmbkz }, 6386 { X86::VANDNPSZ256rmbkz,X86::VANDNPDZ256rmbkz, 6387 X86::VPANDNQZ256rmbkz,X86::VPANDNDZ256rmbkz}, 6388 { X86::VANDPSZ256rmbkz, X86::VANDPDZ256rmbkz, 6389 X86::VPANDQZ256rmbkz, X86::VPANDDZ256rmbkz }, 6390 { X86::VORPSZ256rmbkz, X86::VORPDZ256rmbkz, 6391 X86::VPORQZ256rmbkz, X86::VPORDZ256rmbkz }, 6392 { X86::VXORPSZ256rmbkz, X86::VXORPDZ256rmbkz, 6393 X86::VPXORQZ256rmbkz, X86::VPXORDZ256rmbkz }, 6394 { X86::VANDNPSZrmbkz, X86::VANDNPDZrmbkz, 6395 X86::VPANDNQZrmbkz, X86::VPANDNDZrmbkz }, 6396 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz, 6397 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz }, 6398 { X86::VANDPSZrmbkz, X86::VANDPDZrmbkz, 6399 X86::VPANDQZrmbkz, X86::VPANDDZrmbkz }, 6400 { X86::VORPSZrmbkz, X86::VORPDZrmbkz, 6401 X86::VPORQZrmbkz, X86::VPORDZrmbkz }, 6402 { X86::VXORPSZrmbkz, X86::VXORPDZrmbkz, 6403 X86::VPXORQZrmbkz, X86::VPXORDZrmbkz }, 6404 }; 6405 6406 // NOTE: These should only be used by the custom domain methods. 6407 static const uint16_t ReplaceableCustomInstrs[][3] = { 6408 //PackedSingle PackedDouble PackedInt 6409 { X86::BLENDPSrmi, X86::BLENDPDrmi, X86::PBLENDWrmi }, 6410 { X86::BLENDPSrri, X86::BLENDPDrri, X86::PBLENDWrri }, 6411 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDWrmi }, 6412 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDWrri }, 6413 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDWYrmi }, 6414 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDWYrri }, 6415 }; 6416 static const uint16_t ReplaceableCustomAVX2Instrs[][3] = { 6417 //PackedSingle PackedDouble PackedInt 6418 { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDDrmi }, 6419 { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDDrri }, 6420 { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDDYrmi }, 6421 { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDDYrri }, 6422 }; 6423 6424 // Special table for changing EVEX logic instructions to VEX. 6425 // TODO: Should we run EVEX->VEX earlier? 6426 static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = { 6427 // Two integer columns for 64-bit and 32-bit elements. 6428 //PackedSingle PackedDouble PackedInt PackedInt 6429 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNQZ128rm, X86::VPANDNDZ128rm }, 6430 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNQZ128rr, X86::VPANDNDZ128rr }, 6431 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDQZ128rm, X86::VPANDDZ128rm }, 6432 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDQZ128rr, X86::VPANDDZ128rr }, 6433 { X86::VORPSrm, X86::VORPDrm, X86::VPORQZ128rm, X86::VPORDZ128rm }, 6434 { X86::VORPSrr, X86::VORPDrr, X86::VPORQZ128rr, X86::VPORDZ128rr }, 6435 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORQZ128rm, X86::VPXORDZ128rm }, 6436 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORQZ128rr, X86::VPXORDZ128rr }, 6437 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNQZ256rm, X86::VPANDNDZ256rm }, 6438 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNQZ256rr, X86::VPANDNDZ256rr }, 6439 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDQZ256rm, X86::VPANDDZ256rm }, 6440 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDQZ256rr, X86::VPANDDZ256rr }, 6441 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORQZ256rm, X86::VPORDZ256rm }, 6442 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORQZ256rr, X86::VPORDZ256rr }, 6443 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORQZ256rm, X86::VPXORDZ256rm }, 6444 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORQZ256rr, X86::VPXORDZ256rr }, 6445 }; 6446 6447 // FIXME: Some shuffle and unpack instructions have equivalents in different 6448 // domains, but they require a bit more work than just switching opcodes. 6449 6450 static const uint16_t *lookup(unsigned opcode, unsigned domain, 6451 ArrayRef<uint16_t[3]> Table) { 6452 for (const uint16_t (&Row)[3] : Table) 6453 if (Row[domain-1] == opcode) 6454 return Row; 6455 return nullptr; 6456 } 6457 6458 static const uint16_t *lookupAVX512(unsigned opcode, unsigned domain, 6459 ArrayRef<uint16_t[4]> Table) { 6460 // If this is the integer domain make sure to check both integer columns. 6461 for (const uint16_t (&Row)[4] : Table) 6462 if (Row[domain-1] == opcode || (domain == 3 && Row[3] == opcode)) 6463 return Row; 6464 return nullptr; 6465 } 6466 6467 // Helper to attempt to widen/narrow blend masks. 6468 static bool AdjustBlendMask(unsigned OldMask, unsigned OldWidth, 6469 unsigned NewWidth, unsigned *pNewMask = nullptr) { 6470 assert(((OldWidth % NewWidth) == 0 || (NewWidth % OldWidth) == 0) && 6471 "Illegal blend mask scale"); 6472 unsigned NewMask = 0; 6473 6474 if ((OldWidth % NewWidth) == 0) { 6475 unsigned Scale = OldWidth / NewWidth; 6476 unsigned SubMask = (1u << Scale) - 1; 6477 for (unsigned i = 0; i != NewWidth; ++i) { 6478 unsigned Sub = (OldMask >> (i * Scale)) & SubMask; 6479 if (Sub == SubMask) 6480 NewMask |= (1u << i); 6481 else if (Sub != 0x0) 6482 return false; 6483 } 6484 } else { 6485 unsigned Scale = NewWidth / OldWidth; 6486 unsigned SubMask = (1u << Scale) - 1; 6487 for (unsigned i = 0; i != OldWidth; ++i) { 6488 if (OldMask & (1 << i)) { 6489 NewMask |= (SubMask << (i * Scale)); 6490 } 6491 } 6492 } 6493 6494 if (pNewMask) 6495 *pNewMask = NewMask; 6496 return true; 6497 } 6498 6499 uint16_t X86InstrInfo::getExecutionDomainCustom(const MachineInstr &MI) const { 6500 unsigned Opcode = MI.getOpcode(); 6501 unsigned NumOperands = MI.getDesc().getNumOperands(); 6502 6503 auto GetBlendDomains = [&](unsigned ImmWidth, bool Is256) { 6504 uint16_t validDomains = 0; 6505 if (MI.getOperand(NumOperands - 1).isImm()) { 6506 unsigned Imm = MI.getOperand(NumOperands - 1).getImm(); 6507 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4)) 6508 validDomains |= 0x2; // PackedSingle 6509 if (AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2)) 6510 validDomains |= 0x4; // PackedDouble 6511 if (!Is256 || Subtarget.hasAVX2()) 6512 validDomains |= 0x8; // PackedInt 6513 } 6514 return validDomains; 6515 }; 6516 6517 switch (Opcode) { 6518 case X86::BLENDPDrmi: 6519 case X86::BLENDPDrri: 6520 case X86::VBLENDPDrmi: 6521 case X86::VBLENDPDrri: 6522 return GetBlendDomains(2, false); 6523 case X86::VBLENDPDYrmi: 6524 case X86::VBLENDPDYrri: 6525 return GetBlendDomains(4, true); 6526 case X86::BLENDPSrmi: 6527 case X86::BLENDPSrri: 6528 case X86::VBLENDPSrmi: 6529 case X86::VBLENDPSrri: 6530 case X86::VPBLENDDrmi: 6531 case X86::VPBLENDDrri: 6532 return GetBlendDomains(4, false); 6533 case X86::VBLENDPSYrmi: 6534 case X86::VBLENDPSYrri: 6535 case X86::VPBLENDDYrmi: 6536 case X86::VPBLENDDYrri: 6537 return GetBlendDomains(8, true); 6538 case X86::PBLENDWrmi: 6539 case X86::PBLENDWrri: 6540 case X86::VPBLENDWrmi: 6541 case X86::VPBLENDWrri: 6542 // Treat VPBLENDWY as a 128-bit vector as it repeats the lo/hi masks. 6543 case X86::VPBLENDWYrmi: 6544 case X86::VPBLENDWYrri: 6545 return GetBlendDomains(8, false); 6546 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm: 6547 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm: 6548 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm: 6549 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm: 6550 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm: 6551 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm: 6552 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm: 6553 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm: 6554 case X86::VPORDZ128rr: case X86::VPORDZ128rm: 6555 case X86::VPORDZ256rr: case X86::VPORDZ256rm: 6556 case X86::VPORQZ128rr: case X86::VPORQZ128rm: 6557 case X86::VPORQZ256rr: case X86::VPORQZ256rm: 6558 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm: 6559 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm: 6560 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm: 6561 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm: 6562 // If we don't have DQI see if we can still switch from an EVEX integer 6563 // instruction to a VEX floating point instruction. 6564 if (Subtarget.hasDQI()) 6565 return 0; 6566 6567 if (RI.getEncodingValue(MI.getOperand(0).getReg()) >= 16) 6568 return 0; 6569 if (RI.getEncodingValue(MI.getOperand(1).getReg()) >= 16) 6570 return 0; 6571 // Register forms will have 3 operands. Memory form will have more. 6572 if (NumOperands == 3 && 6573 RI.getEncodingValue(MI.getOperand(2).getReg()) >= 16) 6574 return 0; 6575 6576 // All domains are valid. 6577 return 0xe; 6578 case X86::MOVHLPSrr: 6579 // We can swap domains when both inputs are the same register. 6580 // FIXME: This doesn't catch all the cases we would like. If the input 6581 // register isn't KILLed by the instruction, the two address instruction 6582 // pass puts a COPY on one input. The other input uses the original 6583 // register. This prevents the same physical register from being used by 6584 // both inputs. 6585 if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg() && 6586 MI.getOperand(0).getSubReg() == 0 && 6587 MI.getOperand(1).getSubReg() == 0 && 6588 MI.getOperand(2).getSubReg() == 0) 6589 return 0x6; 6590 return 0; 6591 } 6592 return 0; 6593 } 6594 6595 bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI, 6596 unsigned Domain) const { 6597 assert(Domain > 0 && Domain < 4 && "Invalid execution domain"); 6598 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 6599 assert(dom && "Not an SSE instruction"); 6600 6601 unsigned Opcode = MI.getOpcode(); 6602 unsigned NumOperands = MI.getDesc().getNumOperands(); 6603 6604 auto SetBlendDomain = [&](unsigned ImmWidth, bool Is256) { 6605 if (MI.getOperand(NumOperands - 1).isImm()) { 6606 unsigned Imm = MI.getOperand(NumOperands - 1).getImm() & 255; 6607 Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm); 6608 unsigned NewImm = Imm; 6609 6610 const uint16_t *table = lookup(Opcode, dom, ReplaceableCustomInstrs); 6611 if (!table) 6612 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs); 6613 6614 if (Domain == 1) { // PackedSingle 6615 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm); 6616 } else if (Domain == 2) { // PackedDouble 6617 AdjustBlendMask(Imm, ImmWidth, Is256 ? 4 : 2, &NewImm); 6618 } else if (Domain == 3) { // PackedInt 6619 if (Subtarget.hasAVX2()) { 6620 // If we are already VPBLENDW use that, else use VPBLENDD. 6621 if ((ImmWidth / (Is256 ? 2 : 1)) != 8) { 6622 table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs); 6623 AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm); 6624 } 6625 } else { 6626 assert(!Is256 && "128-bit vector expected"); 6627 AdjustBlendMask(Imm, ImmWidth, 8, &NewImm); 6628 } 6629 } 6630 6631 assert(table && table[Domain - 1] && "Unknown domain op"); 6632 MI.setDesc(get(table[Domain - 1])); 6633 MI.getOperand(NumOperands - 1).setImm(NewImm & 255); 6634 } 6635 return true; 6636 }; 6637 6638 switch (Opcode) { 6639 case X86::BLENDPDrmi: 6640 case X86::BLENDPDrri: 6641 case X86::VBLENDPDrmi: 6642 case X86::VBLENDPDrri: 6643 return SetBlendDomain(2, false); 6644 case X86::VBLENDPDYrmi: 6645 case X86::VBLENDPDYrri: 6646 return SetBlendDomain(4, true); 6647 case X86::BLENDPSrmi: 6648 case X86::BLENDPSrri: 6649 case X86::VBLENDPSrmi: 6650 case X86::VBLENDPSrri: 6651 case X86::VPBLENDDrmi: 6652 case X86::VPBLENDDrri: 6653 return SetBlendDomain(4, false); 6654 case X86::VBLENDPSYrmi: 6655 case X86::VBLENDPSYrri: 6656 case X86::VPBLENDDYrmi: 6657 case X86::VPBLENDDYrri: 6658 return SetBlendDomain(8, true); 6659 case X86::PBLENDWrmi: 6660 case X86::PBLENDWrri: 6661 case X86::VPBLENDWrmi: 6662 case X86::VPBLENDWrri: 6663 return SetBlendDomain(8, false); 6664 case X86::VPBLENDWYrmi: 6665 case X86::VPBLENDWYrri: 6666 return SetBlendDomain(16, true); 6667 case X86::VPANDDZ128rr: case X86::VPANDDZ128rm: 6668 case X86::VPANDDZ256rr: case X86::VPANDDZ256rm: 6669 case X86::VPANDQZ128rr: case X86::VPANDQZ128rm: 6670 case X86::VPANDQZ256rr: case X86::VPANDQZ256rm: 6671 case X86::VPANDNDZ128rr: case X86::VPANDNDZ128rm: 6672 case X86::VPANDNDZ256rr: case X86::VPANDNDZ256rm: 6673 case X86::VPANDNQZ128rr: case X86::VPANDNQZ128rm: 6674 case X86::VPANDNQZ256rr: case X86::VPANDNQZ256rm: 6675 case X86::VPORDZ128rr: case X86::VPORDZ128rm: 6676 case X86::VPORDZ256rr: case X86::VPORDZ256rm: 6677 case X86::VPORQZ128rr: case X86::VPORQZ128rm: 6678 case X86::VPORQZ256rr: case X86::VPORQZ256rm: 6679 case X86::VPXORDZ128rr: case X86::VPXORDZ128rm: 6680 case X86::VPXORDZ256rr: case X86::VPXORDZ256rm: 6681 case X86::VPXORQZ128rr: case X86::VPXORQZ128rm: 6682 case X86::VPXORQZ256rr: case X86::VPXORQZ256rm: { 6683 // Without DQI, convert EVEX instructions to VEX instructions. 6684 if (Subtarget.hasDQI()) 6685 return false; 6686 6687 const uint16_t *table = lookupAVX512(MI.getOpcode(), dom, 6688 ReplaceableCustomAVX512LogicInstrs); 6689 assert(table && "Instruction not found in table?"); 6690 // Don't change integer Q instructions to D instructions and 6691 // use D intructions if we started with a PS instruction. 6692 if (Domain == 3 && (dom == 1 || table[3] == MI.getOpcode())) 6693 Domain = 4; 6694 MI.setDesc(get(table[Domain - 1])); 6695 return true; 6696 } 6697 case X86::UNPCKHPDrr: 6698 case X86::MOVHLPSrr: 6699 // We just need to commute the instruction which will switch the domains. 6700 if (Domain != dom && Domain != 3 && 6701 MI.getOperand(1).getReg() == MI.getOperand(2).getReg() && 6702 MI.getOperand(0).getSubReg() == 0 && 6703 MI.getOperand(1).getSubReg() == 0 && 6704 MI.getOperand(2).getSubReg() == 0) { 6705 commuteInstruction(MI, false); 6706 return true; 6707 } 6708 // We must always return true for MOVHLPSrr. 6709 if (Opcode == X86::MOVHLPSrr) 6710 return true; 6711 } 6712 return false; 6713 } 6714 6715 std::pair<uint16_t, uint16_t> 6716 X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const { 6717 uint16_t domain = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 6718 unsigned opcode = MI.getOpcode(); 6719 uint16_t validDomains = 0; 6720 if (domain) { 6721 // Attempt to match for custom instructions. 6722 validDomains = getExecutionDomainCustom(MI); 6723 if (validDomains) 6724 return std::make_pair(domain, validDomains); 6725 6726 if (lookup(opcode, domain, ReplaceableInstrs)) { 6727 validDomains = 0xe; 6728 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) { 6729 validDomains = Subtarget.hasAVX2() ? 0xe : 0x6; 6730 } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) { 6731 // Insert/extract instructions should only effect domain if AVX2 6732 // is enabled. 6733 if (!Subtarget.hasAVX2()) 6734 return std::make_pair(0, 0); 6735 validDomains = 0xe; 6736 } else if (lookupAVX512(opcode, domain, ReplaceableInstrsAVX512)) { 6737 validDomains = 0xe; 6738 } else if (Subtarget.hasDQI() && lookupAVX512(opcode, domain, 6739 ReplaceableInstrsAVX512DQ)) { 6740 validDomains = 0xe; 6741 } else if (Subtarget.hasDQI()) { 6742 if (const uint16_t *table = lookupAVX512(opcode, domain, 6743 ReplaceableInstrsAVX512DQMasked)) { 6744 if (domain == 1 || (domain == 3 && table[3] == opcode)) 6745 validDomains = 0xa; 6746 else 6747 validDomains = 0xc; 6748 } 6749 } 6750 } 6751 return std::make_pair(domain, validDomains); 6752 } 6753 6754 void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const { 6755 assert(Domain>0 && Domain<4 && "Invalid execution domain"); 6756 uint16_t dom = (MI.getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 6757 assert(dom && "Not an SSE instruction"); 6758 6759 // Attempt to match for custom instructions. 6760 if (setExecutionDomainCustom(MI, Domain)) 6761 return; 6762 6763 const uint16_t *table = lookup(MI.getOpcode(), dom, ReplaceableInstrs); 6764 if (!table) { // try the other table 6765 assert((Subtarget.hasAVX2() || Domain < 3) && 6766 "256-bit vector operations only available in AVX2"); 6767 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2); 6768 } 6769 if (!table) { // try the other table 6770 assert(Subtarget.hasAVX2() && 6771 "256-bit insert/extract only available in AVX2"); 6772 table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2InsertExtract); 6773 } 6774 if (!table) { // try the AVX512 table 6775 assert(Subtarget.hasAVX512() && "Requires AVX-512"); 6776 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512); 6777 // Don't change integer Q instructions to D instructions. 6778 if (table && Domain == 3 && table[3] == MI.getOpcode()) 6779 Domain = 4; 6780 } 6781 if (!table) { // try the AVX512DQ table 6782 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ"); 6783 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQ); 6784 // Don't change integer Q instructions to D instructions and 6785 // use D intructions if we started with a PS instruction. 6786 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode())) 6787 Domain = 4; 6788 } 6789 if (!table) { // try the AVX512DQMasked table 6790 assert((Subtarget.hasDQI() || Domain >= 3) && "Requires AVX-512DQ"); 6791 table = lookupAVX512(MI.getOpcode(), dom, ReplaceableInstrsAVX512DQMasked); 6792 if (table && Domain == 3 && (dom == 1 || table[3] == MI.getOpcode())) 6793 Domain = 4; 6794 } 6795 assert(table && "Cannot change domain"); 6796 MI.setDesc(get(table[Domain - 1])); 6797 } 6798 6799 /// Return the noop instruction to use for a noop. 6800 void X86InstrInfo::getNoop(MCInst &NopInst) const { 6801 NopInst.setOpcode(X86::NOOP); 6802 } 6803 6804 bool X86InstrInfo::isHighLatencyDef(int opc) const { 6805 switch (opc) { 6806 default: return false; 6807 case X86::DIVPDrm: 6808 case X86::DIVPDrr: 6809 case X86::DIVPSrm: 6810 case X86::DIVPSrr: 6811 case X86::DIVSDrm: 6812 case X86::DIVSDrm_Int: 6813 case X86::DIVSDrr: 6814 case X86::DIVSDrr_Int: 6815 case X86::DIVSSrm: 6816 case X86::DIVSSrm_Int: 6817 case X86::DIVSSrr: 6818 case X86::DIVSSrr_Int: 6819 case X86::SQRTPDm: 6820 case X86::SQRTPDr: 6821 case X86::SQRTPSm: 6822 case X86::SQRTPSr: 6823 case X86::SQRTSDm: 6824 case X86::SQRTSDm_Int: 6825 case X86::SQRTSDr: 6826 case X86::SQRTSDr_Int: 6827 case X86::SQRTSSm: 6828 case X86::SQRTSSm_Int: 6829 case X86::SQRTSSr: 6830 case X86::SQRTSSr_Int: 6831 // AVX instructions with high latency 6832 case X86::VDIVPDrm: 6833 case X86::VDIVPDrr: 6834 case X86::VDIVPDYrm: 6835 case X86::VDIVPDYrr: 6836 case X86::VDIVPSrm: 6837 case X86::VDIVPSrr: 6838 case X86::VDIVPSYrm: 6839 case X86::VDIVPSYrr: 6840 case X86::VDIVSDrm: 6841 case X86::VDIVSDrm_Int: 6842 case X86::VDIVSDrr: 6843 case X86::VDIVSDrr_Int: 6844 case X86::VDIVSSrm: 6845 case X86::VDIVSSrm_Int: 6846 case X86::VDIVSSrr: 6847 case X86::VDIVSSrr_Int: 6848 case X86::VSQRTPDm: 6849 case X86::VSQRTPDr: 6850 case X86::VSQRTPDYm: 6851 case X86::VSQRTPDYr: 6852 case X86::VSQRTPSm: 6853 case X86::VSQRTPSr: 6854 case X86::VSQRTPSYm: 6855 case X86::VSQRTPSYr: 6856 case X86::VSQRTSDm: 6857 case X86::VSQRTSDm_Int: 6858 case X86::VSQRTSDr: 6859 case X86::VSQRTSDr_Int: 6860 case X86::VSQRTSSm: 6861 case X86::VSQRTSSm_Int: 6862 case X86::VSQRTSSr: 6863 case X86::VSQRTSSr_Int: 6864 // AVX512 instructions with high latency 6865 case X86::VDIVPDZ128rm: 6866 case X86::VDIVPDZ128rmb: 6867 case X86::VDIVPDZ128rmbk: 6868 case X86::VDIVPDZ128rmbkz: 6869 case X86::VDIVPDZ128rmk: 6870 case X86::VDIVPDZ128rmkz: 6871 case X86::VDIVPDZ128rr: 6872 case X86::VDIVPDZ128rrk: 6873 case X86::VDIVPDZ128rrkz: 6874 case X86::VDIVPDZ256rm: 6875 case X86::VDIVPDZ256rmb: 6876 case X86::VDIVPDZ256rmbk: 6877 case X86::VDIVPDZ256rmbkz: 6878 case X86::VDIVPDZ256rmk: 6879 case X86::VDIVPDZ256rmkz: 6880 case X86::VDIVPDZ256rr: 6881 case X86::VDIVPDZ256rrk: 6882 case X86::VDIVPDZ256rrkz: 6883 case X86::VDIVPDZrrb: 6884 case X86::VDIVPDZrrbk: 6885 case X86::VDIVPDZrrbkz: 6886 case X86::VDIVPDZrm: 6887 case X86::VDIVPDZrmb: 6888 case X86::VDIVPDZrmbk: 6889 case X86::VDIVPDZrmbkz: 6890 case X86::VDIVPDZrmk: 6891 case X86::VDIVPDZrmkz: 6892 case X86::VDIVPDZrr: 6893 case X86::VDIVPDZrrk: 6894 case X86::VDIVPDZrrkz: 6895 case X86::VDIVPSZ128rm: 6896 case X86::VDIVPSZ128rmb: 6897 case X86::VDIVPSZ128rmbk: 6898 case X86::VDIVPSZ128rmbkz: 6899 case X86::VDIVPSZ128rmk: 6900 case X86::VDIVPSZ128rmkz: 6901 case X86::VDIVPSZ128rr: 6902 case X86::VDIVPSZ128rrk: 6903 case X86::VDIVPSZ128rrkz: 6904 case X86::VDIVPSZ256rm: 6905 case X86::VDIVPSZ256rmb: 6906 case X86::VDIVPSZ256rmbk: 6907 case X86::VDIVPSZ256rmbkz: 6908 case X86::VDIVPSZ256rmk: 6909 case X86::VDIVPSZ256rmkz: 6910 case X86::VDIVPSZ256rr: 6911 case X86::VDIVPSZ256rrk: 6912 case X86::VDIVPSZ256rrkz: 6913 case X86::VDIVPSZrrb: 6914 case X86::VDIVPSZrrbk: 6915 case X86::VDIVPSZrrbkz: 6916 case X86::VDIVPSZrm: 6917 case X86::VDIVPSZrmb: 6918 case X86::VDIVPSZrmbk: 6919 case X86::VDIVPSZrmbkz: 6920 case X86::VDIVPSZrmk: 6921 case X86::VDIVPSZrmkz: 6922 case X86::VDIVPSZrr: 6923 case X86::VDIVPSZrrk: 6924 case X86::VDIVPSZrrkz: 6925 case X86::VDIVSDZrm: 6926 case X86::VDIVSDZrr: 6927 case X86::VDIVSDZrm_Int: 6928 case X86::VDIVSDZrm_Intk: 6929 case X86::VDIVSDZrm_Intkz: 6930 case X86::VDIVSDZrr_Int: 6931 case X86::VDIVSDZrr_Intk: 6932 case X86::VDIVSDZrr_Intkz: 6933 case X86::VDIVSDZrrb_Int: 6934 case X86::VDIVSDZrrb_Intk: 6935 case X86::VDIVSDZrrb_Intkz: 6936 case X86::VDIVSSZrm: 6937 case X86::VDIVSSZrr: 6938 case X86::VDIVSSZrm_Int: 6939 case X86::VDIVSSZrm_Intk: 6940 case X86::VDIVSSZrm_Intkz: 6941 case X86::VDIVSSZrr_Int: 6942 case X86::VDIVSSZrr_Intk: 6943 case X86::VDIVSSZrr_Intkz: 6944 case X86::VDIVSSZrrb_Int: 6945 case X86::VDIVSSZrrb_Intk: 6946 case X86::VDIVSSZrrb_Intkz: 6947 case X86::VSQRTPDZ128m: 6948 case X86::VSQRTPDZ128mb: 6949 case X86::VSQRTPDZ128mbk: 6950 case X86::VSQRTPDZ128mbkz: 6951 case X86::VSQRTPDZ128mk: 6952 case X86::VSQRTPDZ128mkz: 6953 case X86::VSQRTPDZ128r: 6954 case X86::VSQRTPDZ128rk: 6955 case X86::VSQRTPDZ128rkz: 6956 case X86::VSQRTPDZ256m: 6957 case X86::VSQRTPDZ256mb: 6958 case X86::VSQRTPDZ256mbk: 6959 case X86::VSQRTPDZ256mbkz: 6960 case X86::VSQRTPDZ256mk: 6961 case X86::VSQRTPDZ256mkz: 6962 case X86::VSQRTPDZ256r: 6963 case X86::VSQRTPDZ256rk: 6964 case X86::VSQRTPDZ256rkz: 6965 case X86::VSQRTPDZm: 6966 case X86::VSQRTPDZmb: 6967 case X86::VSQRTPDZmbk: 6968 case X86::VSQRTPDZmbkz: 6969 case X86::VSQRTPDZmk: 6970 case X86::VSQRTPDZmkz: 6971 case X86::VSQRTPDZr: 6972 case X86::VSQRTPDZrb: 6973 case X86::VSQRTPDZrbk: 6974 case X86::VSQRTPDZrbkz: 6975 case X86::VSQRTPDZrk: 6976 case X86::VSQRTPDZrkz: 6977 case X86::VSQRTPSZ128m: 6978 case X86::VSQRTPSZ128mb: 6979 case X86::VSQRTPSZ128mbk: 6980 case X86::VSQRTPSZ128mbkz: 6981 case X86::VSQRTPSZ128mk: 6982 case X86::VSQRTPSZ128mkz: 6983 case X86::VSQRTPSZ128r: 6984 case X86::VSQRTPSZ128rk: 6985 case X86::VSQRTPSZ128rkz: 6986 case X86::VSQRTPSZ256m: 6987 case X86::VSQRTPSZ256mb: 6988 case X86::VSQRTPSZ256mbk: 6989 case X86::VSQRTPSZ256mbkz: 6990 case X86::VSQRTPSZ256mk: 6991 case X86::VSQRTPSZ256mkz: 6992 case X86::VSQRTPSZ256r: 6993 case X86::VSQRTPSZ256rk: 6994 case X86::VSQRTPSZ256rkz: 6995 case X86::VSQRTPSZm: 6996 case X86::VSQRTPSZmb: 6997 case X86::VSQRTPSZmbk: 6998 case X86::VSQRTPSZmbkz: 6999 case X86::VSQRTPSZmk: 7000 case X86::VSQRTPSZmkz: 7001 case X86::VSQRTPSZr: 7002 case X86::VSQRTPSZrb: 7003 case X86::VSQRTPSZrbk: 7004 case X86::VSQRTPSZrbkz: 7005 case X86::VSQRTPSZrk: 7006 case X86::VSQRTPSZrkz: 7007 case X86::VSQRTSDZm: 7008 case X86::VSQRTSDZm_Int: 7009 case X86::VSQRTSDZm_Intk: 7010 case X86::VSQRTSDZm_Intkz: 7011 case X86::VSQRTSDZr: 7012 case X86::VSQRTSDZr_Int: 7013 case X86::VSQRTSDZr_Intk: 7014 case X86::VSQRTSDZr_Intkz: 7015 case X86::VSQRTSDZrb_Int: 7016 case X86::VSQRTSDZrb_Intk: 7017 case X86::VSQRTSDZrb_Intkz: 7018 case X86::VSQRTSSZm: 7019 case X86::VSQRTSSZm_Int: 7020 case X86::VSQRTSSZm_Intk: 7021 case X86::VSQRTSSZm_Intkz: 7022 case X86::VSQRTSSZr: 7023 case X86::VSQRTSSZr_Int: 7024 case X86::VSQRTSSZr_Intk: 7025 case X86::VSQRTSSZr_Intkz: 7026 case X86::VSQRTSSZrb_Int: 7027 case X86::VSQRTSSZrb_Intk: 7028 case X86::VSQRTSSZrb_Intkz: 7029 7030 case X86::VGATHERDPDYrm: 7031 case X86::VGATHERDPDZ128rm: 7032 case X86::VGATHERDPDZ256rm: 7033 case X86::VGATHERDPDZrm: 7034 case X86::VGATHERDPDrm: 7035 case X86::VGATHERDPSYrm: 7036 case X86::VGATHERDPSZ128rm: 7037 case X86::VGATHERDPSZ256rm: 7038 case X86::VGATHERDPSZrm: 7039 case X86::VGATHERDPSrm: 7040 case X86::VGATHERPF0DPDm: 7041 case X86::VGATHERPF0DPSm: 7042 case X86::VGATHERPF0QPDm: 7043 case X86::VGATHERPF0QPSm: 7044 case X86::VGATHERPF1DPDm: 7045 case X86::VGATHERPF1DPSm: 7046 case X86::VGATHERPF1QPDm: 7047 case X86::VGATHERPF1QPSm: 7048 case X86::VGATHERQPDYrm: 7049 case X86::VGATHERQPDZ128rm: 7050 case X86::VGATHERQPDZ256rm: 7051 case X86::VGATHERQPDZrm: 7052 case X86::VGATHERQPDrm: 7053 case X86::VGATHERQPSYrm: 7054 case X86::VGATHERQPSZ128rm: 7055 case X86::VGATHERQPSZ256rm: 7056 case X86::VGATHERQPSZrm: 7057 case X86::VGATHERQPSrm: 7058 case X86::VPGATHERDDYrm: 7059 case X86::VPGATHERDDZ128rm: 7060 case X86::VPGATHERDDZ256rm: 7061 case X86::VPGATHERDDZrm: 7062 case X86::VPGATHERDDrm: 7063 case X86::VPGATHERDQYrm: 7064 case X86::VPGATHERDQZ128rm: 7065 case X86::VPGATHERDQZ256rm: 7066 case X86::VPGATHERDQZrm: 7067 case X86::VPGATHERDQrm: 7068 case X86::VPGATHERQDYrm: 7069 case X86::VPGATHERQDZ128rm: 7070 case X86::VPGATHERQDZ256rm: 7071 case X86::VPGATHERQDZrm: 7072 case X86::VPGATHERQDrm: 7073 case X86::VPGATHERQQYrm: 7074 case X86::VPGATHERQQZ128rm: 7075 case X86::VPGATHERQQZ256rm: 7076 case X86::VPGATHERQQZrm: 7077 case X86::VPGATHERQQrm: 7078 case X86::VSCATTERDPDZ128mr: 7079 case X86::VSCATTERDPDZ256mr: 7080 case X86::VSCATTERDPDZmr: 7081 case X86::VSCATTERDPSZ128mr: 7082 case X86::VSCATTERDPSZ256mr: 7083 case X86::VSCATTERDPSZmr: 7084 case X86::VSCATTERPF0DPDm: 7085 case X86::VSCATTERPF0DPSm: 7086 case X86::VSCATTERPF0QPDm: 7087 case X86::VSCATTERPF0QPSm: 7088 case X86::VSCATTERPF1DPDm: 7089 case X86::VSCATTERPF1DPSm: 7090 case X86::VSCATTERPF1QPDm: 7091 case X86::VSCATTERPF1QPSm: 7092 case X86::VSCATTERQPDZ128mr: 7093 case X86::VSCATTERQPDZ256mr: 7094 case X86::VSCATTERQPDZmr: 7095 case X86::VSCATTERQPSZ128mr: 7096 case X86::VSCATTERQPSZ256mr: 7097 case X86::VSCATTERQPSZmr: 7098 case X86::VPSCATTERDDZ128mr: 7099 case X86::VPSCATTERDDZ256mr: 7100 case X86::VPSCATTERDDZmr: 7101 case X86::VPSCATTERDQZ128mr: 7102 case X86::VPSCATTERDQZ256mr: 7103 case X86::VPSCATTERDQZmr: 7104 case X86::VPSCATTERQDZ128mr: 7105 case X86::VPSCATTERQDZ256mr: 7106 case X86::VPSCATTERQDZmr: 7107 case X86::VPSCATTERQQZ128mr: 7108 case X86::VPSCATTERQQZ256mr: 7109 case X86::VPSCATTERQQZmr: 7110 return true; 7111 } 7112 } 7113 7114 bool X86InstrInfo::hasHighOperandLatency(const TargetSchedModel &SchedModel, 7115 const MachineRegisterInfo *MRI, 7116 const MachineInstr &DefMI, 7117 unsigned DefIdx, 7118 const MachineInstr &UseMI, 7119 unsigned UseIdx) const { 7120 return isHighLatencyDef(DefMI.getOpcode()); 7121 } 7122 7123 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst, 7124 const MachineBasicBlock *MBB) const { 7125 assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) && 7126 "Reassociation needs binary operators"); 7127 7128 // Integer binary math/logic instructions have a third source operand: 7129 // the EFLAGS register. That operand must be both defined here and never 7130 // used; ie, it must be dead. If the EFLAGS operand is live, then we can 7131 // not change anything because rearranging the operands could affect other 7132 // instructions that depend on the exact status flags (zero, sign, etc.) 7133 // that are set by using these particular operands with this operation. 7134 if (Inst.getNumOperands() == 4) { 7135 assert(Inst.getOperand(3).isReg() && 7136 Inst.getOperand(3).getReg() == X86::EFLAGS && 7137 "Unexpected operand in reassociable instruction"); 7138 if (!Inst.getOperand(3).isDead()) 7139 return false; 7140 } 7141 7142 return TargetInstrInfo::hasReassociableOperands(Inst, MBB); 7143 } 7144 7145 // TODO: There are many more machine instruction opcodes to match: 7146 // 1. Other data types (integer, vectors) 7147 // 2. Other math / logic operations (xor, or) 7148 // 3. Other forms of the same operation (intrinsics and other variants) 7149 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const { 7150 switch (Inst.getOpcode()) { 7151 case X86::AND8rr: 7152 case X86::AND16rr: 7153 case X86::AND32rr: 7154 case X86::AND64rr: 7155 case X86::OR8rr: 7156 case X86::OR16rr: 7157 case X86::OR32rr: 7158 case X86::OR64rr: 7159 case X86::XOR8rr: 7160 case X86::XOR16rr: 7161 case X86::XOR32rr: 7162 case X86::XOR64rr: 7163 case X86::IMUL16rr: 7164 case X86::IMUL32rr: 7165 case X86::IMUL64rr: 7166 case X86::PANDrr: 7167 case X86::PORrr: 7168 case X86::PXORrr: 7169 case X86::ANDPDrr: 7170 case X86::ANDPSrr: 7171 case X86::ORPDrr: 7172 case X86::ORPSrr: 7173 case X86::XORPDrr: 7174 case X86::XORPSrr: 7175 case X86::PADDBrr: 7176 case X86::PADDWrr: 7177 case X86::PADDDrr: 7178 case X86::PADDQrr: 7179 case X86::VPANDrr: 7180 case X86::VPANDYrr: 7181 case X86::VPANDDZ128rr: 7182 case X86::VPANDDZ256rr: 7183 case X86::VPANDDZrr: 7184 case X86::VPANDQZ128rr: 7185 case X86::VPANDQZ256rr: 7186 case X86::VPANDQZrr: 7187 case X86::VPORrr: 7188 case X86::VPORYrr: 7189 case X86::VPORDZ128rr: 7190 case X86::VPORDZ256rr: 7191 case X86::VPORDZrr: 7192 case X86::VPORQZ128rr: 7193 case X86::VPORQZ256rr: 7194 case X86::VPORQZrr: 7195 case X86::VPXORrr: 7196 case X86::VPXORYrr: 7197 case X86::VPXORDZ128rr: 7198 case X86::VPXORDZ256rr: 7199 case X86::VPXORDZrr: 7200 case X86::VPXORQZ128rr: 7201 case X86::VPXORQZ256rr: 7202 case X86::VPXORQZrr: 7203 case X86::VANDPDrr: 7204 case X86::VANDPSrr: 7205 case X86::VANDPDYrr: 7206 case X86::VANDPSYrr: 7207 case X86::VANDPDZ128rr: 7208 case X86::VANDPSZ128rr: 7209 case X86::VANDPDZ256rr: 7210 case X86::VANDPSZ256rr: 7211 case X86::VANDPDZrr: 7212 case X86::VANDPSZrr: 7213 case X86::VORPDrr: 7214 case X86::VORPSrr: 7215 case X86::VORPDYrr: 7216 case X86::VORPSYrr: 7217 case X86::VORPDZ128rr: 7218 case X86::VORPSZ128rr: 7219 case X86::VORPDZ256rr: 7220 case X86::VORPSZ256rr: 7221 case X86::VORPDZrr: 7222 case X86::VORPSZrr: 7223 case X86::VXORPDrr: 7224 case X86::VXORPSrr: 7225 case X86::VXORPDYrr: 7226 case X86::VXORPSYrr: 7227 case X86::VXORPDZ128rr: 7228 case X86::VXORPSZ128rr: 7229 case X86::VXORPDZ256rr: 7230 case X86::VXORPSZ256rr: 7231 case X86::VXORPDZrr: 7232 case X86::VXORPSZrr: 7233 case X86::KADDBrr: 7234 case X86::KADDWrr: 7235 case X86::KADDDrr: 7236 case X86::KADDQrr: 7237 case X86::KANDBrr: 7238 case X86::KANDWrr: 7239 case X86::KANDDrr: 7240 case X86::KANDQrr: 7241 case X86::KORBrr: 7242 case X86::KORWrr: 7243 case X86::KORDrr: 7244 case X86::KORQrr: 7245 case X86::KXORBrr: 7246 case X86::KXORWrr: 7247 case X86::KXORDrr: 7248 case X86::KXORQrr: 7249 case X86::VPADDBrr: 7250 case X86::VPADDWrr: 7251 case X86::VPADDDrr: 7252 case X86::VPADDQrr: 7253 case X86::VPADDBYrr: 7254 case X86::VPADDWYrr: 7255 case X86::VPADDDYrr: 7256 case X86::VPADDQYrr: 7257 case X86::VPADDBZ128rr: 7258 case X86::VPADDWZ128rr: 7259 case X86::VPADDDZ128rr: 7260 case X86::VPADDQZ128rr: 7261 case X86::VPADDBZ256rr: 7262 case X86::VPADDWZ256rr: 7263 case X86::VPADDDZ256rr: 7264 case X86::VPADDQZ256rr: 7265 case X86::VPADDBZrr: 7266 case X86::VPADDWZrr: 7267 case X86::VPADDDZrr: 7268 case X86::VPADDQZrr: 7269 case X86::VPMULLWrr: 7270 case X86::VPMULLWYrr: 7271 case X86::VPMULLWZ128rr: 7272 case X86::VPMULLWZ256rr: 7273 case X86::VPMULLWZrr: 7274 case X86::VPMULLDrr: 7275 case X86::VPMULLDYrr: 7276 case X86::VPMULLDZ128rr: 7277 case X86::VPMULLDZ256rr: 7278 case X86::VPMULLDZrr: 7279 case X86::VPMULLQZ128rr: 7280 case X86::VPMULLQZ256rr: 7281 case X86::VPMULLQZrr: 7282 // Normal min/max instructions are not commutative because of NaN and signed 7283 // zero semantics, but these are. Thus, there's no need to check for global 7284 // relaxed math; the instructions themselves have the properties we need. 7285 case X86::MAXCPDrr: 7286 case X86::MAXCPSrr: 7287 case X86::MAXCSDrr: 7288 case X86::MAXCSSrr: 7289 case X86::MINCPDrr: 7290 case X86::MINCPSrr: 7291 case X86::MINCSDrr: 7292 case X86::MINCSSrr: 7293 case X86::VMAXCPDrr: 7294 case X86::VMAXCPSrr: 7295 case X86::VMAXCPDYrr: 7296 case X86::VMAXCPSYrr: 7297 case X86::VMAXCPDZ128rr: 7298 case X86::VMAXCPSZ128rr: 7299 case X86::VMAXCPDZ256rr: 7300 case X86::VMAXCPSZ256rr: 7301 case X86::VMAXCPDZrr: 7302 case X86::VMAXCPSZrr: 7303 case X86::VMAXCSDrr: 7304 case X86::VMAXCSSrr: 7305 case X86::VMAXCSDZrr: 7306 case X86::VMAXCSSZrr: 7307 case X86::VMINCPDrr: 7308 case X86::VMINCPSrr: 7309 case X86::VMINCPDYrr: 7310 case X86::VMINCPSYrr: 7311 case X86::VMINCPDZ128rr: 7312 case X86::VMINCPSZ128rr: 7313 case X86::VMINCPDZ256rr: 7314 case X86::VMINCPSZ256rr: 7315 case X86::VMINCPDZrr: 7316 case X86::VMINCPSZrr: 7317 case X86::VMINCSDrr: 7318 case X86::VMINCSSrr: 7319 case X86::VMINCSDZrr: 7320 case X86::VMINCSSZrr: 7321 return true; 7322 case X86::ADDPDrr: 7323 case X86::ADDPSrr: 7324 case X86::ADDSDrr: 7325 case X86::ADDSSrr: 7326 case X86::MULPDrr: 7327 case X86::MULPSrr: 7328 case X86::MULSDrr: 7329 case X86::MULSSrr: 7330 case X86::VADDPDrr: 7331 case X86::VADDPSrr: 7332 case X86::VADDPDYrr: 7333 case X86::VADDPSYrr: 7334 case X86::VADDPDZ128rr: 7335 case X86::VADDPSZ128rr: 7336 case X86::VADDPDZ256rr: 7337 case X86::VADDPSZ256rr: 7338 case X86::VADDPDZrr: 7339 case X86::VADDPSZrr: 7340 case X86::VADDSDrr: 7341 case X86::VADDSSrr: 7342 case X86::VADDSDZrr: 7343 case X86::VADDSSZrr: 7344 case X86::VMULPDrr: 7345 case X86::VMULPSrr: 7346 case X86::VMULPDYrr: 7347 case X86::VMULPSYrr: 7348 case X86::VMULPDZ128rr: 7349 case X86::VMULPSZ128rr: 7350 case X86::VMULPDZ256rr: 7351 case X86::VMULPSZ256rr: 7352 case X86::VMULPDZrr: 7353 case X86::VMULPSZrr: 7354 case X86::VMULSDrr: 7355 case X86::VMULSSrr: 7356 case X86::VMULSDZrr: 7357 case X86::VMULSSZrr: 7358 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath; 7359 default: 7360 return false; 7361 } 7362 } 7363 7364 /// This is an architecture-specific helper function of reassociateOps. 7365 /// Set special operand attributes for new instructions after reassociation. 7366 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1, 7367 MachineInstr &OldMI2, 7368 MachineInstr &NewMI1, 7369 MachineInstr &NewMI2) const { 7370 // Integer instructions define an implicit EFLAGS source register operand as 7371 // the third source (fourth total) operand. 7372 if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4) 7373 return; 7374 7375 assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 && 7376 "Unexpected instruction type for reassociation"); 7377 7378 MachineOperand &OldOp1 = OldMI1.getOperand(3); 7379 MachineOperand &OldOp2 = OldMI2.getOperand(3); 7380 MachineOperand &NewOp1 = NewMI1.getOperand(3); 7381 MachineOperand &NewOp2 = NewMI2.getOperand(3); 7382 7383 assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() && 7384 "Must have dead EFLAGS operand in reassociable instruction"); 7385 assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() && 7386 "Must have dead EFLAGS operand in reassociable instruction"); 7387 7388 (void)OldOp1; 7389 (void)OldOp2; 7390 7391 assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS && 7392 "Unexpected operand in reassociable instruction"); 7393 assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS && 7394 "Unexpected operand in reassociable instruction"); 7395 7396 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations 7397 // of this pass or other passes. The EFLAGS operands must be dead in these new 7398 // instructions because the EFLAGS operands in the original instructions must 7399 // be dead in order for reassociation to occur. 7400 NewOp1.setIsDead(); 7401 NewOp2.setIsDead(); 7402 } 7403 7404 std::pair<unsigned, unsigned> 7405 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 7406 return std::make_pair(TF, 0u); 7407 } 7408 7409 ArrayRef<std::pair<unsigned, const char *>> 7410 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 7411 using namespace X86II; 7412 static const std::pair<unsigned, const char *> TargetFlags[] = { 7413 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"}, 7414 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"}, 7415 {MO_GOT, "x86-got"}, 7416 {MO_GOTOFF, "x86-gotoff"}, 7417 {MO_GOTPCREL, "x86-gotpcrel"}, 7418 {MO_PLT, "x86-plt"}, 7419 {MO_TLSGD, "x86-tlsgd"}, 7420 {MO_TLSLD, "x86-tlsld"}, 7421 {MO_TLSLDM, "x86-tlsldm"}, 7422 {MO_GOTTPOFF, "x86-gottpoff"}, 7423 {MO_INDNTPOFF, "x86-indntpoff"}, 7424 {MO_TPOFF, "x86-tpoff"}, 7425 {MO_DTPOFF, "x86-dtpoff"}, 7426 {MO_NTPOFF, "x86-ntpoff"}, 7427 {MO_GOTNTPOFF, "x86-gotntpoff"}, 7428 {MO_DLLIMPORT, "x86-dllimport"}, 7429 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"}, 7430 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"}, 7431 {MO_TLVP, "x86-tlvp"}, 7432 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"}, 7433 {MO_SECREL, "x86-secrel"}, 7434 {MO_COFFSTUB, "x86-coffstub"}}; 7435 return makeArrayRef(TargetFlags); 7436 } 7437 7438 namespace { 7439 /// Create Global Base Reg pass. This initializes the PIC 7440 /// global base register for x86-32. 7441 struct CGBR : public MachineFunctionPass { 7442 static char ID; 7443 CGBR() : MachineFunctionPass(ID) {} 7444 7445 bool runOnMachineFunction(MachineFunction &MF) override { 7446 const X86TargetMachine *TM = 7447 static_cast<const X86TargetMachine *>(&MF.getTarget()); 7448 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>(); 7449 7450 // Don't do anything in the 64-bit small and kernel code models. They use 7451 // RIP-relative addressing for everything. 7452 if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small || 7453 TM->getCodeModel() == CodeModel::Kernel)) 7454 return false; 7455 7456 // Only emit a global base reg in PIC mode. 7457 if (!TM->isPositionIndependent()) 7458 return false; 7459 7460 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 7461 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 7462 7463 // If we didn't need a GlobalBaseReg, don't insert code. 7464 if (GlobalBaseReg == 0) 7465 return false; 7466 7467 // Insert the set of GlobalBaseReg into the first MBB of the function 7468 MachineBasicBlock &FirstMBB = MF.front(); 7469 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 7470 DebugLoc DL = FirstMBB.findDebugLoc(MBBI); 7471 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 7472 const X86InstrInfo *TII = STI.getInstrInfo(); 7473 7474 unsigned PC; 7475 if (STI.isPICStyleGOT()) 7476 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass); 7477 else 7478 PC = GlobalBaseReg; 7479 7480 if (STI.is64Bit()) { 7481 if (TM->getCodeModel() == CodeModel::Medium) { 7482 // In the medium code model, use a RIP-relative LEA to materialize the 7483 // GOT. 7484 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC) 7485 .addReg(X86::RIP) 7486 .addImm(0) 7487 .addReg(0) 7488 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_") 7489 .addReg(0); 7490 } else if (TM->getCodeModel() == CodeModel::Large) { 7491 // In the large code model, we are aiming for this code, though the 7492 // register allocation may vary: 7493 // leaq .LN$pb(%rip), %rax 7494 // movq $_GLOBAL_OFFSET_TABLE_ - .LN$pb, %rcx 7495 // addq %rcx, %rax 7496 // RAX now holds address of _GLOBAL_OFFSET_TABLE_. 7497 unsigned PBReg = RegInfo.createVirtualRegister(&X86::GR64RegClass); 7498 unsigned GOTReg = 7499 RegInfo.createVirtualRegister(&X86::GR64RegClass); 7500 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PBReg) 7501 .addReg(X86::RIP) 7502 .addImm(0) 7503 .addReg(0) 7504 .addSym(MF.getPICBaseSymbol()) 7505 .addReg(0); 7506 std::prev(MBBI)->setPreInstrSymbol(MF, MF.getPICBaseSymbol()); 7507 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOV64ri), GOTReg) 7508 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_", 7509 X86II::MO_PIC_BASE_OFFSET); 7510 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD64rr), PC) 7511 .addReg(PBReg, RegState::Kill) 7512 .addReg(GOTReg, RegState::Kill); 7513 } else { 7514 llvm_unreachable("unexpected code model"); 7515 } 7516 } else { 7517 // Operand of MovePCtoStack is completely ignored by asm printer. It's 7518 // only used in JIT code emission as displacement to pc. 7519 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0); 7520 7521 // If we're using vanilla 'GOT' PIC style, we should use relative 7522 // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external. 7523 if (STI.isPICStyleGOT()) { 7524 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], 7525 // %some_register 7526 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg) 7527 .addReg(PC) 7528 .addExternalSymbol("_GLOBAL_OFFSET_TABLE_", 7529 X86II::MO_GOT_ABSOLUTE_ADDRESS); 7530 } 7531 } 7532 7533 return true; 7534 } 7535 7536 StringRef getPassName() const override { 7537 return "X86 PIC Global Base Reg Initialization"; 7538 } 7539 7540 void getAnalysisUsage(AnalysisUsage &AU) const override { 7541 AU.setPreservesCFG(); 7542 MachineFunctionPass::getAnalysisUsage(AU); 7543 } 7544 }; 7545 } 7546 7547 char CGBR::ID = 0; 7548 FunctionPass* 7549 llvm::createX86GlobalBaseRegPass() { return new CGBR(); } 7550 7551 namespace { 7552 struct LDTLSCleanup : public MachineFunctionPass { 7553 static char ID; 7554 LDTLSCleanup() : MachineFunctionPass(ID) {} 7555 7556 bool runOnMachineFunction(MachineFunction &MF) override { 7557 if (skipFunction(MF.getFunction())) 7558 return false; 7559 7560 X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>(); 7561 if (MFI->getNumLocalDynamicTLSAccesses() < 2) { 7562 // No point folding accesses if there isn't at least two. 7563 return false; 7564 } 7565 7566 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>(); 7567 return VisitNode(DT->getRootNode(), 0); 7568 } 7569 7570 // Visit the dominator subtree rooted at Node in pre-order. 7571 // If TLSBaseAddrReg is non-null, then use that to replace any 7572 // TLS_base_addr instructions. Otherwise, create the register 7573 // when the first such instruction is seen, and then use it 7574 // as we encounter more instructions. 7575 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) { 7576 MachineBasicBlock *BB = Node->getBlock(); 7577 bool Changed = false; 7578 7579 // Traverse the current block. 7580 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; 7581 ++I) { 7582 switch (I->getOpcode()) { 7583 case X86::TLS_base_addr32: 7584 case X86::TLS_base_addr64: 7585 if (TLSBaseAddrReg) 7586 I = ReplaceTLSBaseAddrCall(*I, TLSBaseAddrReg); 7587 else 7588 I = SetRegister(*I, &TLSBaseAddrReg); 7589 Changed = true; 7590 break; 7591 default: 7592 break; 7593 } 7594 } 7595 7596 // Visit the children of this block in the dominator tree. 7597 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end(); 7598 I != E; ++I) { 7599 Changed |= VisitNode(*I, TLSBaseAddrReg); 7600 } 7601 7602 return Changed; 7603 } 7604 7605 // Replace the TLS_base_addr instruction I with a copy from 7606 // TLSBaseAddrReg, returning the new instruction. 7607 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr &I, 7608 unsigned TLSBaseAddrReg) { 7609 MachineFunction *MF = I.getParent()->getParent(); 7610 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>(); 7611 const bool is64Bit = STI.is64Bit(); 7612 const X86InstrInfo *TII = STI.getInstrInfo(); 7613 7614 // Insert a Copy from TLSBaseAddrReg to RAX/EAX. 7615 MachineInstr *Copy = 7616 BuildMI(*I.getParent(), I, I.getDebugLoc(), 7617 TII->get(TargetOpcode::COPY), is64Bit ? X86::RAX : X86::EAX) 7618 .addReg(TLSBaseAddrReg); 7619 7620 // Erase the TLS_base_addr instruction. 7621 I.eraseFromParent(); 7622 7623 return Copy; 7624 } 7625 7626 // Create a virtual register in *TLSBaseAddrReg, and populate it by 7627 // inserting a copy instruction after I. Returns the new instruction. 7628 MachineInstr *SetRegister(MachineInstr &I, unsigned *TLSBaseAddrReg) { 7629 MachineFunction *MF = I.getParent()->getParent(); 7630 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>(); 7631 const bool is64Bit = STI.is64Bit(); 7632 const X86InstrInfo *TII = STI.getInstrInfo(); 7633 7634 // Create a virtual register for the TLS base address. 7635 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 7636 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit 7637 ? &X86::GR64RegClass 7638 : &X86::GR32RegClass); 7639 7640 // Insert a copy from RAX/EAX to TLSBaseAddrReg. 7641 MachineInstr *Next = I.getNextNode(); 7642 MachineInstr *Copy = 7643 BuildMI(*I.getParent(), Next, I.getDebugLoc(), 7644 TII->get(TargetOpcode::COPY), *TLSBaseAddrReg) 7645 .addReg(is64Bit ? X86::RAX : X86::EAX); 7646 7647 return Copy; 7648 } 7649 7650 StringRef getPassName() const override { 7651 return "Local Dynamic TLS Access Clean-up"; 7652 } 7653 7654 void getAnalysisUsage(AnalysisUsage &AU) const override { 7655 AU.setPreservesCFG(); 7656 AU.addRequired<MachineDominatorTree>(); 7657 MachineFunctionPass::getAnalysisUsage(AU); 7658 } 7659 }; 7660 } 7661 7662 char LDTLSCleanup::ID = 0; 7663 FunctionPass* 7664 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); } 7665 7666 /// Constants defining how certain sequences should be outlined. 7667 /// 7668 /// \p MachineOutlinerDefault implies that the function is called with a call 7669 /// instruction, and a return must be emitted for the outlined function frame. 7670 /// 7671 /// That is, 7672 /// 7673 /// I1 OUTLINED_FUNCTION: 7674 /// I2 --> call OUTLINED_FUNCTION I1 7675 /// I3 I2 7676 /// I3 7677 /// ret 7678 /// 7679 /// * Call construction overhead: 1 (call instruction) 7680 /// * Frame construction overhead: 1 (return instruction) 7681 /// 7682 /// \p MachineOutlinerTailCall implies that the function is being tail called. 7683 /// A jump is emitted instead of a call, and the return is already present in 7684 /// the outlined sequence. That is, 7685 /// 7686 /// I1 OUTLINED_FUNCTION: 7687 /// I2 --> jmp OUTLINED_FUNCTION I1 7688 /// ret I2 7689 /// ret 7690 /// 7691 /// * Call construction overhead: 1 (jump instruction) 7692 /// * Frame construction overhead: 0 (don't need to return) 7693 /// 7694 enum MachineOutlinerClass { 7695 MachineOutlinerDefault, 7696 MachineOutlinerTailCall 7697 }; 7698 7699 outliner::OutlinedFunction X86InstrInfo::getOutliningCandidateInfo( 7700 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const { 7701 unsigned SequenceSize = 7702 std::accumulate(RepeatedSequenceLocs[0].front(), 7703 std::next(RepeatedSequenceLocs[0].back()), 0, 7704 [](unsigned Sum, const MachineInstr &MI) { 7705 // FIXME: x86 doesn't implement getInstSizeInBytes, so 7706 // we can't tell the cost. Just assume each instruction 7707 // is one byte. 7708 if (MI.isDebugInstr() || MI.isKill()) 7709 return Sum; 7710 return Sum + 1; 7711 }); 7712 7713 // FIXME: Use real size in bytes for call and ret instructions. 7714 if (RepeatedSequenceLocs[0].back()->isTerminator()) { 7715 for (outliner::Candidate &C : RepeatedSequenceLocs) 7716 C.setCallInfo(MachineOutlinerTailCall, 1); 7717 7718 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 7719 0, // Number of bytes to emit frame. 7720 MachineOutlinerTailCall // Type of frame. 7721 ); 7722 } 7723 7724 for (outliner::Candidate &C : RepeatedSequenceLocs) 7725 C.setCallInfo(MachineOutlinerDefault, 1); 7726 7727 return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, 1, 7728 MachineOutlinerDefault); 7729 } 7730 7731 bool X86InstrInfo::isFunctionSafeToOutlineFrom(MachineFunction &MF, 7732 bool OutlineFromLinkOnceODRs) const { 7733 const Function &F = MF.getFunction(); 7734 7735 // Does the function use a red zone? If it does, then we can't risk messing 7736 // with the stack. 7737 if (!F.hasFnAttribute(Attribute::NoRedZone)) { 7738 // It could have a red zone. If it does, then we don't want to touch it. 7739 const X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 7740 if (!X86FI || X86FI->getUsesRedZone()) 7741 return false; 7742 } 7743 7744 // If we *don't* want to outline from things that could potentially be deduped 7745 // then return false. 7746 if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage()) 7747 return false; 7748 7749 // This function is viable for outlining, so return true. 7750 return true; 7751 } 7752 7753 outliner::InstrType 7754 X86InstrInfo::getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const { 7755 MachineInstr &MI = *MIT; 7756 // Don't allow debug values to impact outlining type. 7757 if (MI.isDebugInstr() || MI.isIndirectDebugValue()) 7758 return outliner::InstrType::Invisible; 7759 7760 // At this point, KILL instructions don't really tell us much so we can go 7761 // ahead and skip over them. 7762 if (MI.isKill()) 7763 return outliner::InstrType::Invisible; 7764 7765 // Is this a tail call? If yes, we can outline as a tail call. 7766 if (isTailCall(MI)) 7767 return outliner::InstrType::Legal; 7768 7769 // Is this the terminator of a basic block? 7770 if (MI.isTerminator() || MI.isReturn()) { 7771 7772 // Does its parent have any successors in its MachineFunction? 7773 if (MI.getParent()->succ_empty()) 7774 return outliner::InstrType::Legal; 7775 7776 // It does, so we can't tail call it. 7777 return outliner::InstrType::Illegal; 7778 } 7779 7780 // Don't outline anything that modifies or reads from the stack pointer. 7781 // 7782 // FIXME: There are instructions which are being manually built without 7783 // explicit uses/defs so we also have to check the MCInstrDesc. We should be 7784 // able to remove the extra checks once those are fixed up. For example, 7785 // sometimes we might get something like %rax = POP64r 1. This won't be 7786 // caught by modifiesRegister or readsRegister even though the instruction 7787 // really ought to be formed so that modifiesRegister/readsRegister would 7788 // catch it. 7789 if (MI.modifiesRegister(X86::RSP, &RI) || MI.readsRegister(X86::RSP, &RI) || 7790 MI.getDesc().hasImplicitUseOfPhysReg(X86::RSP) || 7791 MI.getDesc().hasImplicitDefOfPhysReg(X86::RSP)) 7792 return outliner::InstrType::Illegal; 7793 7794 // Outlined calls change the instruction pointer, so don't read from it. 7795 if (MI.readsRegister(X86::RIP, &RI) || 7796 MI.getDesc().hasImplicitUseOfPhysReg(X86::RIP) || 7797 MI.getDesc().hasImplicitDefOfPhysReg(X86::RIP)) 7798 return outliner::InstrType::Illegal; 7799 7800 // Positions can't safely be outlined. 7801 if (MI.isPosition()) 7802 return outliner::InstrType::Illegal; 7803 7804 // Make sure none of the operands of this instruction do anything tricky. 7805 for (const MachineOperand &MOP : MI.operands()) 7806 if (MOP.isCPI() || MOP.isJTI() || MOP.isCFIIndex() || MOP.isFI() || 7807 MOP.isTargetIndex()) 7808 return outliner::InstrType::Illegal; 7809 7810 return outliner::InstrType::Legal; 7811 } 7812 7813 void X86InstrInfo::buildOutlinedFrame(MachineBasicBlock &MBB, 7814 MachineFunction &MF, 7815 const outliner::OutlinedFunction &OF) 7816 const { 7817 // If we're a tail call, we already have a return, so don't do anything. 7818 if (OF.FrameConstructionID == MachineOutlinerTailCall) 7819 return; 7820 7821 // We're a normal call, so our sequence doesn't have a return instruction. 7822 // Add it in. 7823 MachineInstr *retq = BuildMI(MF, DebugLoc(), get(X86::RETQ)); 7824 MBB.insert(MBB.end(), retq); 7825 } 7826 7827 MachineBasicBlock::iterator 7828 X86InstrInfo::insertOutlinedCall(Module &M, MachineBasicBlock &MBB, 7829 MachineBasicBlock::iterator &It, 7830 MachineFunction &MF, 7831 const outliner::Candidate &C) const { 7832 // Is it a tail call? 7833 if (C.CallConstructionID == MachineOutlinerTailCall) { 7834 // Yes, just insert a JMP. 7835 It = MBB.insert(It, 7836 BuildMI(MF, DebugLoc(), get(X86::TAILJMPd64)) 7837 .addGlobalAddress(M.getNamedValue(MF.getName()))); 7838 } else { 7839 // No, insert a call. 7840 It = MBB.insert(It, 7841 BuildMI(MF, DebugLoc(), get(X86::CALL64pcrel32)) 7842 .addGlobalAddress(M.getNamedValue(MF.getName()))); 7843 } 7844 7845 return It; 7846 } 7847