X86InstrInfo.h revision 8d4a4225135f1f4647e675902359b64cc21af7ef
1//===- X86InstrInfo.h - X86 Instruction Information ------------*- C++ -*- ===// 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#ifndef X86INSTRUCTIONINFO_H 15#define X86INSTRUCTIONINFO_H 16 17#include "llvm/Target/TargetInstrInfo.h" 18#include "X86.h" 19#include "X86RegisterInfo.h" 20#include "llvm/ADT/DenseMap.h" 21 22namespace llvm { 23 class X86RegisterInfo; 24 class X86TargetMachine; 25 26namespace X86 { 27 // Enums for memory operand decoding. Each memory operand is represented with 28 // a 5 operand sequence in the form: 29 // [BaseReg, ScaleAmt, IndexReg, Disp, Segment] 30 // These enums help decode this. 31 enum { 32 AddrBaseReg = 0, 33 AddrScaleAmt = 1, 34 AddrIndexReg = 2, 35 AddrDisp = 3, 36 37 /// AddrSegmentReg - The operand # of the segment in the memory operand. 38 AddrSegmentReg = 4, 39 40 /// AddrNumOperands - Total number of operands in a memory reference. 41 AddrNumOperands = 5 42 }; 43 44 45 // X86 specific condition code. These correspond to X86_*_COND in 46 // X86InstrInfo.td. They must be kept in synch. 47 enum CondCode { 48 COND_A = 0, 49 COND_AE = 1, 50 COND_B = 2, 51 COND_BE = 3, 52 COND_E = 4, 53 COND_G = 5, 54 COND_GE = 6, 55 COND_L = 7, 56 COND_LE = 8, 57 COND_NE = 9, 58 COND_NO = 10, 59 COND_NP = 11, 60 COND_NS = 12, 61 COND_O = 13, 62 COND_P = 14, 63 COND_S = 15, 64 65 // Artificial condition codes. These are used by AnalyzeBranch 66 // to indicate a block terminated with two conditional branches to 67 // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE, 68 // which can't be represented on x86 with a single condition. These 69 // are never used in MachineInstrs. 70 COND_NE_OR_P, 71 COND_NP_OR_E, 72 73 COND_INVALID 74 }; 75 76 // Turn condition code into conditional branch opcode. 77 unsigned GetCondBranchFromCond(CondCode CC); 78 79 /// GetOppositeBranchCondition - Return the inverse of the specified cond, 80 /// e.g. turning COND_E to COND_NE. 81 CondCode GetOppositeBranchCondition(X86::CondCode CC); 82 83} 84 85/// X86II - This namespace holds all of the target specific flags that 86/// instruction info tracks. 87/// 88namespace X86II { 89 /// Target Operand Flag enum. 90 enum TOF { 91 //===------------------------------------------------------------------===// 92 // X86 Specific MachineOperand flags. 93 94 MO_NO_FLAG, 95 96 /// MO_GOT_ABSOLUTE_ADDRESS - On a symbol operand, this represents a 97 /// relocation of: 98 /// SYMBOL_LABEL + [. - PICBASELABEL] 99 MO_GOT_ABSOLUTE_ADDRESS, 100 101 /// MO_PIC_BASE_OFFSET - On a symbol operand this indicates that the 102 /// immediate should get the value of the symbol minus the PIC base label: 103 /// SYMBOL_LABEL - PICBASELABEL 104 MO_PIC_BASE_OFFSET, 105 106 /// MO_GOT - On a symbol operand this indicates that the immediate is the 107 /// offset to the GOT entry for the symbol name from the base of the GOT. 108 /// 109 /// See the X86-64 ELF ABI supplement for more details. 110 /// SYMBOL_LABEL @GOT 111 MO_GOT, 112 113 /// MO_GOTOFF - On a symbol operand this indicates that the immediate is 114 /// the offset to the location of the symbol name from the base of the GOT. 115 /// 116 /// See the X86-64 ELF ABI supplement for more details. 117 /// SYMBOL_LABEL @GOTOFF 118 MO_GOTOFF, 119 120 /// MO_GOTPCREL - On a symbol operand this indicates that the immediate is 121 /// offset to the GOT entry for the symbol name from the current code 122 /// location. 123 /// 124 /// See the X86-64 ELF ABI supplement for more details. 125 /// SYMBOL_LABEL @GOTPCREL 126 MO_GOTPCREL, 127 128 /// MO_PLT - On a symbol operand this indicates that the immediate is 129 /// offset to the PLT entry of symbol name from the current code location. 130 /// 131 /// See the X86-64 ELF ABI supplement for more details. 132 /// SYMBOL_LABEL @PLT 133 MO_PLT, 134 135 /// MO_TLSGD - On a symbol operand this indicates that the immediate is 136 /// some TLS offset. 137 /// 138 /// See 'ELF Handling for Thread-Local Storage' for more details. 139 /// SYMBOL_LABEL @TLSGD 140 MO_TLSGD, 141 142 /// MO_GOTTPOFF - On a symbol operand this indicates that the immediate is 143 /// some TLS offset. 144 /// 145 /// See 'ELF Handling for Thread-Local Storage' for more details. 146 /// SYMBOL_LABEL @GOTTPOFF 147 MO_GOTTPOFF, 148 149 /// MO_INDNTPOFF - On a symbol operand this indicates that the immediate is 150 /// some TLS offset. 151 /// 152 /// See 'ELF Handling for Thread-Local Storage' for more details. 153 /// SYMBOL_LABEL @INDNTPOFF 154 MO_INDNTPOFF, 155 156 /// MO_TPOFF - On a symbol operand this indicates that the immediate is 157 /// some TLS offset. 158 /// 159 /// See 'ELF Handling for Thread-Local Storage' for more details. 160 /// SYMBOL_LABEL @TPOFF 161 MO_TPOFF, 162 163 /// MO_NTPOFF - On a symbol operand this indicates that the immediate is 164 /// some TLS offset. 165 /// 166 /// See 'ELF Handling for Thread-Local Storage' for more details. 167 /// SYMBOL_LABEL @NTPOFF 168 MO_NTPOFF, 169 170 /// MO_DLLIMPORT - On a symbol operand "FOO", this indicates that the 171 /// reference is actually to the "__imp_FOO" symbol. This is used for 172 /// dllimport linkage on windows. 173 MO_DLLIMPORT, 174 175 /// MO_DARWIN_STUB - On a symbol operand "FOO", this indicates that the 176 /// reference is actually to the "FOO$stub" symbol. This is used for calls 177 /// and jumps to external functions on Tiger and earlier. 178 MO_DARWIN_STUB, 179 180 /// MO_DARWIN_NONLAZY - On a symbol operand "FOO", this indicates that the 181 /// reference is actually to the "FOO$non_lazy_ptr" symbol, which is a 182 /// non-PIC-base-relative reference to a non-hidden dyld lazy pointer stub. 183 MO_DARWIN_NONLAZY, 184 185 /// MO_DARWIN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this indicates 186 /// that the reference is actually to "FOO$non_lazy_ptr - PICBASE", which is 187 /// a PIC-base-relative reference to a non-hidden dyld lazy pointer stub. 188 MO_DARWIN_NONLAZY_PIC_BASE, 189 190 /// MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE - On a symbol operand "FOO", this 191 /// indicates that the reference is actually to "FOO$non_lazy_ptr -PICBASE", 192 /// which is a PIC-base-relative reference to a hidden dyld lazy pointer 193 /// stub. 194 MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE, 195 196 /// MO_TLVP - On a symbol operand this indicates that the immediate is 197 /// some TLS offset. 198 /// 199 /// This is the TLS offset for the Darwin TLS mechanism. 200 MO_TLVP, 201 202 /// MO_TLVP_PIC_BASE - On a symbol operand this indicates that the immediate 203 /// is some TLS offset from the picbase. 204 /// 205 /// This is the 32-bit TLS offset for Darwin TLS in PIC mode. 206 MO_TLVP_PIC_BASE 207 }; 208} 209 210/// isGlobalStubReference - Return true if the specified TargetFlag operand is 211/// a reference to a stub for a global, not the global itself. 212inline static bool isGlobalStubReference(unsigned char TargetFlag) { 213 switch (TargetFlag) { 214 case X86II::MO_DLLIMPORT: // dllimport stub. 215 case X86II::MO_GOTPCREL: // rip-relative GOT reference. 216 case X86II::MO_GOT: // normal GOT reference. 217 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: // Normal $non_lazy_ptr ref. 218 case X86II::MO_DARWIN_NONLAZY: // Normal $non_lazy_ptr ref. 219 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Hidden $non_lazy_ptr ref. 220 return true; 221 default: 222 return false; 223 } 224} 225 226/// isGlobalRelativeToPICBase - Return true if the specified global value 227/// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg). If this 228/// is true, the addressing mode has the PIC base register added in (e.g. EBX). 229inline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag) { 230 switch (TargetFlag) { 231 case X86II::MO_GOTOFF: // isPICStyleGOT: local global. 232 case X86II::MO_GOT: // isPICStyleGOT: other global. 233 case X86II::MO_PIC_BASE_OFFSET: // Darwin local global. 234 case X86II::MO_DARWIN_NONLAZY_PIC_BASE: // Darwin/32 external global. 235 case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Darwin/32 hidden global. 236 case X86II::MO_TLVP: // ??? Pretty sure.. 237 return true; 238 default: 239 return false; 240 } 241} 242 243/// X86II - This namespace holds all of the target specific flags that 244/// instruction info tracks. 245/// 246namespace X86II { 247 enum { 248 //===------------------------------------------------------------------===// 249 // Instruction encodings. These are the standard/most common forms for X86 250 // instructions. 251 // 252 253 // PseudoFrm - This represents an instruction that is a pseudo instruction 254 // or one that has not been implemented yet. It is illegal to code generate 255 // it, but tolerated for intermediate implementation stages. 256 Pseudo = 0, 257 258 /// Raw - This form is for instructions that don't have any operands, so 259 /// they are just a fixed opcode value, like 'leave'. 260 RawFrm = 1, 261 262 /// AddRegFrm - This form is used for instructions like 'push r32' that have 263 /// their one register operand added to their opcode. 264 AddRegFrm = 2, 265 266 /// MRMDestReg - This form is used for instructions that use the Mod/RM byte 267 /// to specify a destination, which in this case is a register. 268 /// 269 MRMDestReg = 3, 270 271 /// MRMDestMem - This form is used for instructions that use the Mod/RM byte 272 /// to specify a destination, which in this case is memory. 273 /// 274 MRMDestMem = 4, 275 276 /// MRMSrcReg - This form is used for instructions that use the Mod/RM byte 277 /// to specify a source, which in this case is a register. 278 /// 279 MRMSrcReg = 5, 280 281 /// MRMSrcMem - This form is used for instructions that use the Mod/RM byte 282 /// to specify a source, which in this case is memory. 283 /// 284 MRMSrcMem = 6, 285 286 /// MRM[0-7][rm] - These forms are used to represent instructions that use 287 /// a Mod/RM byte, and use the middle field to hold extended opcode 288 /// information. In the intel manual these are represented as /0, /1, ... 289 /// 290 291 // First, instructions that operate on a register r/m operand... 292 MRM0r = 16, MRM1r = 17, MRM2r = 18, MRM3r = 19, // Format /0 /1 /2 /3 293 MRM4r = 20, MRM5r = 21, MRM6r = 22, MRM7r = 23, // Format /4 /5 /6 /7 294 295 // Next, instructions that operate on a memory r/m operand... 296 MRM0m = 24, MRM1m = 25, MRM2m = 26, MRM3m = 27, // Format /0 /1 /2 /3 297 MRM4m = 28, MRM5m = 29, MRM6m = 30, MRM7m = 31, // Format /4 /5 /6 /7 298 299 // MRMInitReg - This form is used for instructions whose source and 300 // destinations are the same register. 301 MRMInitReg = 32, 302 303 //// MRM_C1 - A mod/rm byte of exactly 0xC1. 304 MRM_C1 = 33, 305 MRM_C2 = 34, 306 MRM_C3 = 35, 307 MRM_C4 = 36, 308 MRM_C8 = 37, 309 MRM_C9 = 38, 310 MRM_E8 = 39, 311 MRM_F0 = 40, 312 MRM_F8 = 41, 313 MRM_F9 = 42, 314 MRM_D0 = 45, 315 MRM_D1 = 46, 316 317 /// RawFrmImm8 - This is used for the ENTER instruction, which has two 318 /// immediates, the first of which is a 16-bit immediate (specified by 319 /// the imm encoding) and the second is a 8-bit fixed value. 320 RawFrmImm8 = 43, 321 322 /// RawFrmImm16 - This is used for CALL FAR instructions, which have two 323 /// immediates, the first of which is a 16 or 32-bit immediate (specified by 324 /// the imm encoding) and the second is a 16-bit fixed value. In the AMD 325 /// manual, this operand is described as pntr16:32 and pntr16:16 326 RawFrmImm16 = 44, 327 328 FormMask = 63, 329 330 //===------------------------------------------------------------------===// 331 // Actual flags... 332 333 // OpSize - Set if this instruction requires an operand size prefix (0x66), 334 // which most often indicates that the instruction operates on 16 bit data 335 // instead of 32 bit data. 336 OpSize = 1 << 6, 337 338 // AsSize - Set if this instruction requires an operand size prefix (0x67), 339 // which most often indicates that the instruction address 16 bit address 340 // instead of 32 bit address (or 32 bit address in 64 bit mode). 341 AdSize = 1 << 7, 342 343 //===------------------------------------------------------------------===// 344 // Op0Mask - There are several prefix bytes that are used to form two byte 345 // opcodes. These are currently 0x0F, 0xF3, and 0xD8-0xDF. This mask is 346 // used to obtain the setting of this field. If no bits in this field is 347 // set, there is no prefix byte for obtaining a multibyte opcode. 348 // 349 Op0Shift = 8, 350 Op0Mask = 0xF << Op0Shift, 351 352 // TB - TwoByte - Set if this instruction has a two byte opcode, which 353 // starts with a 0x0F byte before the real opcode. 354 TB = 1 << Op0Shift, 355 356 // REP - The 0xF3 prefix byte indicating repetition of the following 357 // instruction. 358 REP = 2 << Op0Shift, 359 360 // D8-DF - These escape opcodes are used by the floating point unit. These 361 // values must remain sequential. 362 D8 = 3 << Op0Shift, D9 = 4 << Op0Shift, 363 DA = 5 << Op0Shift, DB = 6 << Op0Shift, 364 DC = 7 << Op0Shift, DD = 8 << Op0Shift, 365 DE = 9 << Op0Shift, DF = 10 << Op0Shift, 366 367 // XS, XD - These prefix codes are for single and double precision scalar 368 // floating point operations performed in the SSE registers. 369 XD = 11 << Op0Shift, XS = 12 << Op0Shift, 370 371 // T8, TA - Prefix after the 0x0F prefix. 372 T8 = 13 << Op0Shift, TA = 14 << Op0Shift, 373 374 // TF - Prefix before and after 0x0F 375 TF = 15 << Op0Shift, 376 377 //===------------------------------------------------------------------===// 378 // REX_W - REX prefixes are instruction prefixes used in 64-bit mode. 379 // They are used to specify GPRs and SSE registers, 64-bit operand size, 380 // etc. We only cares about REX.W and REX.R bits and only the former is 381 // statically determined. 382 // 383 REXShift = 12, 384 REX_W = 1 << REXShift, 385 386 //===------------------------------------------------------------------===// 387 // This three-bit field describes the size of an immediate operand. Zero is 388 // unused so that we can tell if we forgot to set a value. 389 ImmShift = 13, 390 ImmMask = 7 << ImmShift, 391 Imm8 = 1 << ImmShift, 392 Imm8PCRel = 2 << ImmShift, 393 Imm16 = 3 << ImmShift, 394 Imm16PCRel = 4 << ImmShift, 395 Imm32 = 5 << ImmShift, 396 Imm32PCRel = 6 << ImmShift, 397 Imm64 = 7 << ImmShift, 398 399 //===------------------------------------------------------------------===// 400 // FP Instruction Classification... Zero is non-fp instruction. 401 402 // FPTypeMask - Mask for all of the FP types... 403 FPTypeShift = 16, 404 FPTypeMask = 7 << FPTypeShift, 405 406 // NotFP - The default, set for instructions that do not use FP registers. 407 NotFP = 0 << FPTypeShift, 408 409 // ZeroArgFP - 0 arg FP instruction which implicitly pushes ST(0), f.e. fld0 410 ZeroArgFP = 1 << FPTypeShift, 411 412 // OneArgFP - 1 arg FP instructions which implicitly read ST(0), such as fst 413 OneArgFP = 2 << FPTypeShift, 414 415 // OneArgFPRW - 1 arg FP instruction which implicitly read ST(0) and write a 416 // result back to ST(0). For example, fcos, fsqrt, etc. 417 // 418 OneArgFPRW = 3 << FPTypeShift, 419 420 // TwoArgFP - 2 arg FP instructions which implicitly read ST(0), and an 421 // explicit argument, storing the result to either ST(0) or the implicit 422 // argument. For example: fadd, fsub, fmul, etc... 423 TwoArgFP = 4 << FPTypeShift, 424 425 // CompareFP - 2 arg FP instructions which implicitly read ST(0) and an 426 // explicit argument, but have no destination. Example: fucom, fucomi, ... 427 CompareFP = 5 << FPTypeShift, 428 429 // CondMovFP - "2 operand" floating point conditional move instructions. 430 CondMovFP = 6 << FPTypeShift, 431 432 // SpecialFP - Special instruction forms. Dispatch by opcode explicitly. 433 SpecialFP = 7 << FPTypeShift, 434 435 // Lock prefix 436 LOCKShift = 19, 437 LOCK = 1 << LOCKShift, 438 439 // Segment override prefixes. Currently we just need ability to address 440 // stuff in gs and fs segments. 441 SegOvrShift = 20, 442 SegOvrMask = 3 << SegOvrShift, 443 FS = 1 << SegOvrShift, 444 GS = 2 << SegOvrShift, 445 446 // Execution domain for SSE instructions in bits 22, 23. 447 // 0 in bits 22-23 means normal, non-SSE instruction. 448 SSEDomainShift = 22, 449 450 OpcodeShift = 24, 451 OpcodeMask = 0xFF << OpcodeShift, 452 453 //===------------------------------------------------------------------===// 454 /// VEX - The opcode prefix used by AVX instructions 455 VEX = 1U << 0, 456 457 /// VEX_W - Has a opcode specific functionality, but is used in the same 458 /// way as REX_W is for regular SSE instructions. 459 VEX_W = 1U << 1, 460 461 /// VEX_4V - Used to specify an additional AVX/SSE register. Several 2 462 /// address instructions in SSE are represented as 3 address ones in AVX 463 /// and the additional register is encoded in VEX_VVVV prefix. 464 VEX_4V = 1U << 2, 465 466 /// VEX_I8IMM - Specifies that the last register used in a AVX instruction, 467 /// must be encoded in the i8 immediate field. This usually happens in 468 /// instructions with 4 operands. 469 VEX_I8IMM = 1U << 3, 470 471 /// VEX_L - Stands for a bit in the VEX opcode prefix meaning the current 472 /// instruction uses 256-bit wide registers. This is usually auto detected 473 /// if a VR256 register is used, but some AVX instructions also have this 474 /// field marked when using a f256 memory references. 475 VEX_L = 1U << 4, 476 477 /// Has3DNow0F0FOpcode - This flag indicates that the instruction uses the 478 /// wacky 0x0F 0x0F prefix for 3DNow! instructions. The manual documents 479 /// this as having a 0x0F prefix with a 0x0F opcode, and each instruction 480 /// storing a classifier in the imm8 field. To simplify our implementation, 481 /// we handle this by storeing the classifier in the opcode field and using 482 /// this flag to indicate that the encoder should do the wacky 3DNow! thing. 483 Has3DNow0F0FOpcode = 1U << 5 484 }; 485 486 // getBaseOpcodeFor - This function returns the "base" X86 opcode for the 487 // specified machine instruction. 488 // 489 static inline unsigned char getBaseOpcodeFor(uint64_t TSFlags) { 490 return TSFlags >> X86II::OpcodeShift; 491 } 492 493 static inline bool hasImm(uint64_t TSFlags) { 494 return (TSFlags & X86II::ImmMask) != 0; 495 } 496 497 /// getSizeOfImm - Decode the "size of immediate" field from the TSFlags field 498 /// of the specified instruction. 499 static inline unsigned getSizeOfImm(uint64_t TSFlags) { 500 switch (TSFlags & X86II::ImmMask) { 501 default: assert(0 && "Unknown immediate size"); 502 case X86II::Imm8: 503 case X86II::Imm8PCRel: return 1; 504 case X86II::Imm16: 505 case X86II::Imm16PCRel: return 2; 506 case X86II::Imm32: 507 case X86II::Imm32PCRel: return 4; 508 case X86II::Imm64: return 8; 509 } 510 } 511 512 /// isImmPCRel - Return true if the immediate of the specified instruction's 513 /// TSFlags indicates that it is pc relative. 514 static inline unsigned isImmPCRel(uint64_t TSFlags) { 515 switch (TSFlags & X86II::ImmMask) { 516 default: assert(0 && "Unknown immediate size"); 517 case X86II::Imm8PCRel: 518 case X86II::Imm16PCRel: 519 case X86II::Imm32PCRel: 520 return true; 521 case X86II::Imm8: 522 case X86II::Imm16: 523 case X86II::Imm32: 524 case X86II::Imm64: 525 return false; 526 } 527 } 528 529 /// getMemoryOperandNo - The function returns the MCInst operand # for the 530 /// first field of the memory operand. If the instruction doesn't have a 531 /// memory operand, this returns -1. 532 /// 533 /// Note that this ignores tied operands. If there is a tied register which 534 /// is duplicated in the MCInst (e.g. "EAX = addl EAX, [mem]") it is only 535 /// counted as one operand. 536 /// 537 static inline int getMemoryOperandNo(uint64_t TSFlags) { 538 switch (TSFlags & X86II::FormMask) { 539 case X86II::MRMInitReg: assert(0 && "FIXME: Remove this form"); 540 default: assert(0 && "Unknown FormMask value in getMemoryOperandNo!"); 541 case X86II::Pseudo: 542 case X86II::RawFrm: 543 case X86II::AddRegFrm: 544 case X86II::MRMDestReg: 545 case X86II::MRMSrcReg: 546 case X86II::RawFrmImm8: 547 case X86II::RawFrmImm16: 548 return -1; 549 case X86II::MRMDestMem: 550 return 0; 551 case X86II::MRMSrcMem: { 552 bool HasVEX_4V = (TSFlags >> 32) & X86II::VEX_4V; 553 unsigned FirstMemOp = 1; 554 if (HasVEX_4V) 555 ++FirstMemOp;// Skip the register source (which is encoded in VEX_VVVV). 556 557 // FIXME: Maybe lea should have its own form? This is a horrible hack. 558 //if (Opcode == X86::LEA64r || Opcode == X86::LEA64_32r || 559 // Opcode == X86::LEA16r || Opcode == X86::LEA32r) 560 return FirstMemOp; 561 } 562 case X86II::MRM0r: case X86II::MRM1r: 563 case X86II::MRM2r: case X86II::MRM3r: 564 case X86II::MRM4r: case X86II::MRM5r: 565 case X86II::MRM6r: case X86II::MRM7r: 566 return -1; 567 case X86II::MRM0m: case X86II::MRM1m: 568 case X86II::MRM2m: case X86II::MRM3m: 569 case X86II::MRM4m: case X86II::MRM5m: 570 case X86II::MRM6m: case X86II::MRM7m: 571 return 0; 572 case X86II::MRM_C1: 573 case X86II::MRM_C2: 574 case X86II::MRM_C3: 575 case X86II::MRM_C4: 576 case X86II::MRM_C8: 577 case X86II::MRM_C9: 578 case X86II::MRM_E8: 579 case X86II::MRM_F0: 580 case X86II::MRM_F8: 581 case X86II::MRM_F9: 582 case X86II::MRM_D0: 583 case X86II::MRM_D1: 584 return -1; 585 } 586 } 587} 588 589inline static bool isScale(const MachineOperand &MO) { 590 return MO.isImm() && 591 (MO.getImm() == 1 || MO.getImm() == 2 || 592 MO.getImm() == 4 || MO.getImm() == 8); 593} 594 595inline static bool isLeaMem(const MachineInstr *MI, unsigned Op) { 596 if (MI->getOperand(Op).isFI()) return true; 597 return Op+4 <= MI->getNumOperands() && 598 MI->getOperand(Op ).isReg() && isScale(MI->getOperand(Op+1)) && 599 MI->getOperand(Op+2).isReg() && 600 (MI->getOperand(Op+3).isImm() || 601 MI->getOperand(Op+3).isGlobal() || 602 MI->getOperand(Op+3).isCPI() || 603 MI->getOperand(Op+3).isJTI()); 604} 605 606inline static bool isMem(const MachineInstr *MI, unsigned Op) { 607 if (MI->getOperand(Op).isFI()) return true; 608 return Op+5 <= MI->getNumOperands() && 609 MI->getOperand(Op+4).isReg() && 610 isLeaMem(MI, Op); 611} 612 613class X86InstrInfo : public TargetInstrInfoImpl { 614 X86TargetMachine &TM; 615 const X86RegisterInfo RI; 616 617 /// RegOp2MemOpTable2Addr, RegOp2MemOpTable0, RegOp2MemOpTable1, 618 /// RegOp2MemOpTable2 - Load / store folding opcode maps. 619 /// 620 DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable2Addr; 621 DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable0; 622 DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable1; 623 DenseMap<unsigned, std::pair<unsigned,unsigned> > RegOp2MemOpTable2; 624 625 /// MemOp2RegOpTable - Load / store unfolding opcode map. 626 /// 627 DenseMap<unsigned, std::pair<unsigned, unsigned> > MemOp2RegOpTable; 628 629public: 630 explicit X86InstrInfo(X86TargetMachine &tm); 631 632 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As 633 /// such, whenever a client has an instance of instruction info, it should 634 /// always be able to get register info as well (through this method). 635 /// 636 virtual const X86RegisterInfo &getRegisterInfo() const { return RI; } 637 638 /// isCoalescableExtInstr - Return true if the instruction is a "coalescable" 639 /// extension instruction. That is, it's like a copy where it's legal for the 640 /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns 641 /// true, then it's expected the pre-extension value is available as a subreg 642 /// of the result register. This also returns the sub-register index in 643 /// SubIdx. 644 virtual bool isCoalescableExtInstr(const MachineInstr &MI, 645 unsigned &SrcReg, unsigned &DstReg, 646 unsigned &SubIdx) const; 647 648 unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const; 649 /// isLoadFromStackSlotPostFE - Check for post-frame ptr elimination 650 /// stack locations as well. This uses a heuristic so it isn't 651 /// reliable for correctness. 652 unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI, 653 int &FrameIndex) const; 654 655 /// hasLoadFromStackSlot - If the specified machine instruction has 656 /// a load from a stack slot, return true along with the FrameIndex 657 /// of the loaded stack slot and the machine mem operand containing 658 /// the reference. If not, return false. Unlike 659 /// isLoadFromStackSlot, this returns true for any instructions that 660 /// loads from the stack. This is a hint only and may not catch all 661 /// cases. 662 bool hasLoadFromStackSlot(const MachineInstr *MI, 663 const MachineMemOperand *&MMO, 664 int &FrameIndex) const; 665 666 unsigned isStoreToStackSlot(const MachineInstr *MI, int &FrameIndex) const; 667 /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination 668 /// stack locations as well. This uses a heuristic so it isn't 669 /// reliable for correctness. 670 unsigned isStoreToStackSlotPostFE(const MachineInstr *MI, 671 int &FrameIndex) const; 672 673 /// hasStoreToStackSlot - If the specified machine instruction has a 674 /// store to a stack slot, return true along with the FrameIndex of 675 /// the loaded stack slot and the machine mem operand containing the 676 /// reference. If not, return false. Unlike isStoreToStackSlot, 677 /// this returns true for any instructions that loads from the 678 /// stack. This is a hint only and may not catch all cases. 679 bool hasStoreToStackSlot(const MachineInstr *MI, 680 const MachineMemOperand *&MMO, 681 int &FrameIndex) const; 682 683 bool isReallyTriviallyReMaterializable(const MachineInstr *MI, 684 AliasAnalysis *AA) const; 685 void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, 686 unsigned DestReg, unsigned SubIdx, 687 const MachineInstr *Orig, 688 const TargetRegisterInfo &TRI) const; 689 690 /// convertToThreeAddress - This method must be implemented by targets that 691 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target 692 /// may be able to convert a two-address instruction into a true 693 /// three-address instruction on demand. This allows the X86 target (for 694 /// example) to convert ADD and SHL instructions into LEA instructions if they 695 /// would require register copies due to two-addressness. 696 /// 697 /// This method returns a null pointer if the transformation cannot be 698 /// performed, otherwise it returns the new instruction. 699 /// 700 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI, 701 MachineBasicBlock::iterator &MBBI, 702 LiveVariables *LV) const; 703 704 /// commuteInstruction - We have a few instructions that must be hacked on to 705 /// commute them. 706 /// 707 virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const; 708 709 // Branch analysis. 710 virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const; 711 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 712 MachineBasicBlock *&FBB, 713 SmallVectorImpl<MachineOperand> &Cond, 714 bool AllowModify) const; 715 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const; 716 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 717 MachineBasicBlock *FBB, 718 const SmallVectorImpl<MachineOperand> &Cond, 719 DebugLoc DL) const; 720 virtual void copyPhysReg(MachineBasicBlock &MBB, 721 MachineBasicBlock::iterator MI, DebugLoc DL, 722 unsigned DestReg, unsigned SrcReg, 723 bool KillSrc) const; 724 virtual void storeRegToStackSlot(MachineBasicBlock &MBB, 725 MachineBasicBlock::iterator MI, 726 unsigned SrcReg, bool isKill, int FrameIndex, 727 const TargetRegisterClass *RC, 728 const TargetRegisterInfo *TRI) const; 729 730 virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill, 731 SmallVectorImpl<MachineOperand> &Addr, 732 const TargetRegisterClass *RC, 733 MachineInstr::mmo_iterator MMOBegin, 734 MachineInstr::mmo_iterator MMOEnd, 735 SmallVectorImpl<MachineInstr*> &NewMIs) const; 736 737 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB, 738 MachineBasicBlock::iterator MI, 739 unsigned DestReg, int FrameIndex, 740 const TargetRegisterClass *RC, 741 const TargetRegisterInfo *TRI) const; 742 743 virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg, 744 SmallVectorImpl<MachineOperand> &Addr, 745 const TargetRegisterClass *RC, 746 MachineInstr::mmo_iterator MMOBegin, 747 MachineInstr::mmo_iterator MMOEnd, 748 SmallVectorImpl<MachineInstr*> &NewMIs) const; 749 virtual 750 MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF, 751 int FrameIx, uint64_t Offset, 752 const MDNode *MDPtr, 753 DebugLoc DL) const; 754 755 /// foldMemoryOperand - If this target supports it, fold a load or store of 756 /// the specified stack slot into the specified machine instruction for the 757 /// specified operand(s). If this is possible, the target should perform the 758 /// folding and return true, otherwise it should return false. If it folds 759 /// the instruction, it is likely that the MachineInstruction the iterator 760 /// references has been changed. 761 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF, 762 MachineInstr* MI, 763 const SmallVectorImpl<unsigned> &Ops, 764 int FrameIndex) const; 765 766 /// foldMemoryOperand - Same as the previous version except it allows folding 767 /// of any load and store from / to any address, not just from a specific 768 /// stack slot. 769 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF, 770 MachineInstr* MI, 771 const SmallVectorImpl<unsigned> &Ops, 772 MachineInstr* LoadMI) const; 773 774 /// canFoldMemoryOperand - Returns true if the specified load / store is 775 /// folding is possible. 776 virtual bool canFoldMemoryOperand(const MachineInstr*, 777 const SmallVectorImpl<unsigned> &) const; 778 779 /// unfoldMemoryOperand - Separate a single instruction which folded a load or 780 /// a store or a load and a store into two or more instruction. If this is 781 /// possible, returns true as well as the new instructions by reference. 782 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI, 783 unsigned Reg, bool UnfoldLoad, bool UnfoldStore, 784 SmallVectorImpl<MachineInstr*> &NewMIs) const; 785 786 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, 787 SmallVectorImpl<SDNode*> &NewNodes) const; 788 789 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new 790 /// instruction after load / store are unfolded from an instruction of the 791 /// specified opcode. It returns zero if the specified unfolding is not 792 /// possible. If LoadRegIndex is non-null, it is filled in with the operand 793 /// index of the operand which will hold the register holding the loaded 794 /// value. 795 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc, 796 bool UnfoldLoad, bool UnfoldStore, 797 unsigned *LoadRegIndex = 0) const; 798 799 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler 800 /// to determine if two loads are loading from the same base address. It 801 /// should only return true if the base pointers are the same and the 802 /// only differences between the two addresses are the offset. It also returns 803 /// the offsets by reference. 804 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 805 int64_t &Offset1, int64_t &Offset2) const; 806 807 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to 808 /// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should 809 /// be scheduled togther. On some targets if two loads are loading from 810 /// addresses in the same cache line, it's better if they are scheduled 811 /// together. This function takes two integers that represent the load offsets 812 /// from the common base address. It returns true if it decides it's desirable 813 /// to schedule the two loads together. "NumLoads" is the number of loads that 814 /// have already been scheduled after Load1. 815 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 816 int64_t Offset1, int64_t Offset2, 817 unsigned NumLoads) const; 818 819 virtual void getNoopForMachoTarget(MCInst &NopInst) const; 820 821 virtual 822 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const; 823 824 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine 825 /// instruction that defines the specified register class. 826 bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const; 827 828 static bool isX86_64NonExtLowByteReg(unsigned reg) { 829 return (reg == X86::SPL || reg == X86::BPL || 830 reg == X86::SIL || reg == X86::DIL); 831 } 832 833 static bool isX86_64ExtendedReg(const MachineOperand &MO) { 834 if (!MO.isReg()) return false; 835 return isX86_64ExtendedReg(MO.getReg()); 836 } 837 838 /// isX86_64ExtendedReg - Is the MachineOperand a x86-64 extended (r8 or 839 /// higher) register? e.g. r8, xmm8, xmm13, etc. 840 static bool isX86_64ExtendedReg(unsigned RegNo); 841 842 /// getGlobalBaseReg - Return a virtual register initialized with the 843 /// the global base register value. Output instructions required to 844 /// initialize the register in the function entry block, if necessary. 845 /// 846 unsigned getGlobalBaseReg(MachineFunction *MF) const; 847 848 /// GetSSEDomain - Return the SSE execution domain of MI as the first element, 849 /// and a bitmask of possible arguments to SetSSEDomain ase the second. 850 std::pair<uint16_t, uint16_t> GetSSEDomain(const MachineInstr *MI) const; 851 852 /// SetSSEDomain - Set the SSEDomain of MI. 853 void SetSSEDomain(MachineInstr *MI, unsigned Domain) const; 854 855 MachineInstr* foldMemoryOperandImpl(MachineFunction &MF, 856 MachineInstr* MI, 857 unsigned OpNum, 858 const SmallVectorImpl<MachineOperand> &MOs, 859 unsigned Size, unsigned Alignment) const; 860 861 bool hasHighOperandLatency(const InstrItineraryData *ItinData, 862 const MachineRegisterInfo *MRI, 863 const MachineInstr *DefMI, unsigned DefIdx, 864 const MachineInstr *UseMI, unsigned UseIdx) const; 865 866private: 867 MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc, 868 MachineFunction::iterator &MFI, 869 MachineBasicBlock::iterator &MBBI, 870 LiveVariables *LV) const; 871 872 /// isFrameOperand - Return true and the FrameIndex if the specified 873 /// operand and follow operands form a reference to the stack frame. 874 bool isFrameOperand(const MachineInstr *MI, unsigned int Op, 875 int &FrameIndex) const; 876}; 877 878} // End llvm namespace 879 880#endif 881