X86InstrInfo.cpp revision e7f702fc2d496aff1e5c1153519931e203b1ca76
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 "X86MachineFunctionInfo.h" 18#include "X86Subtarget.h" 19#include "X86TargetMachine.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/LLVMContext.h" 22#include "llvm/ADT/STLExtras.h" 23#include "llvm/CodeGen/MachineConstantPool.h" 24#include "llvm/CodeGen/MachineDominators.h" 25#include "llvm/CodeGen/MachineFrameInfo.h" 26#include "llvm/CodeGen/MachineInstrBuilder.h" 27#include "llvm/CodeGen/MachineRegisterInfo.h" 28#include "llvm/CodeGen/LiveVariables.h" 29#include "llvm/MC/MCAsmInfo.h" 30#include "llvm/MC/MCInst.h" 31#include "llvm/Support/CommandLine.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Support/ErrorHandling.h" 34#include "llvm/Support/raw_ostream.h" 35#include "llvm/Target/TargetOptions.h" 36#include <limits> 37 38#define GET_INSTRINFO_CTOR 39#include "X86GenInstrInfo.inc" 40 41using namespace llvm; 42 43static cl::opt<bool> 44NoFusing("disable-spill-fusing", 45 cl::desc("Disable fusing of spill code into instructions")); 46static cl::opt<bool> 47PrintFailedFusing("print-failed-fuse-candidates", 48 cl::desc("Print instructions that the allocator wants to" 49 " fuse, but the X86 backend currently can't"), 50 cl::Hidden); 51static cl::opt<bool> 52ReMatPICStubLoad("remat-pic-stub-load", 53 cl::desc("Re-materialize load from stub in PIC mode"), 54 cl::init(false), cl::Hidden); 55 56enum { 57 // Select which memory operand is being unfolded. 58 // (stored in bits 0 - 3) 59 TB_INDEX_0 = 0, 60 TB_INDEX_1 = 1, 61 TB_INDEX_2 = 2, 62 TB_INDEX_3 = 3, 63 TB_INDEX_MASK = 0xf, 64 65 // Do not insert the reverse map (MemOp -> RegOp) into the table. 66 // This may be needed because there is a many -> one mapping. 67 TB_NO_REVERSE = 1 << 4, 68 69 // Do not insert the forward map (RegOp -> MemOp) into the table. 70 // This is needed for Native Client, which prohibits branch 71 // instructions from using a memory operand. 72 TB_NO_FORWARD = 1 << 5, 73 74 TB_FOLDED_LOAD = 1 << 6, 75 TB_FOLDED_STORE = 1 << 7, 76 77 // Minimum alignment required for load/store. 78 // Used for RegOp->MemOp conversion. 79 // (stored in bits 8 - 15) 80 TB_ALIGN_SHIFT = 8, 81 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT, 82 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT, 83 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT, 84 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT 85}; 86 87struct X86OpTblEntry { 88 uint16_t RegOp; 89 uint16_t MemOp; 90 uint16_t Flags; 91}; 92 93X86InstrInfo::X86InstrInfo(X86TargetMachine &tm) 94 : X86GenInstrInfo((tm.getSubtarget<X86Subtarget>().is64Bit() 95 ? X86::ADJCALLSTACKDOWN64 96 : X86::ADJCALLSTACKDOWN32), 97 (tm.getSubtarget<X86Subtarget>().is64Bit() 98 ? X86::ADJCALLSTACKUP64 99 : X86::ADJCALLSTACKUP32)), 100 TM(tm), RI(tm, *this) { 101 102 static const X86OpTblEntry OpTbl2Addr[] = { 103 { X86::ADC32ri, X86::ADC32mi, 0 }, 104 { X86::ADC32ri8, X86::ADC32mi8, 0 }, 105 { X86::ADC32rr, X86::ADC32mr, 0 }, 106 { X86::ADC64ri32, X86::ADC64mi32, 0 }, 107 { X86::ADC64ri8, X86::ADC64mi8, 0 }, 108 { X86::ADC64rr, X86::ADC64mr, 0 }, 109 { X86::ADD16ri, X86::ADD16mi, 0 }, 110 { X86::ADD16ri8, X86::ADD16mi8, 0 }, 111 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE }, 112 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE }, 113 { X86::ADD16rr, X86::ADD16mr, 0 }, 114 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE }, 115 { X86::ADD32ri, X86::ADD32mi, 0 }, 116 { X86::ADD32ri8, X86::ADD32mi8, 0 }, 117 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE }, 118 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE }, 119 { X86::ADD32rr, X86::ADD32mr, 0 }, 120 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE }, 121 { X86::ADD64ri32, X86::ADD64mi32, 0 }, 122 { X86::ADD64ri8, X86::ADD64mi8, 0 }, 123 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE }, 124 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE }, 125 { X86::ADD64rr, X86::ADD64mr, 0 }, 126 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE }, 127 { X86::ADD8ri, X86::ADD8mi, 0 }, 128 { X86::ADD8rr, X86::ADD8mr, 0 }, 129 { X86::AND16ri, X86::AND16mi, 0 }, 130 { X86::AND16ri8, X86::AND16mi8, 0 }, 131 { X86::AND16rr, X86::AND16mr, 0 }, 132 { X86::AND32ri, X86::AND32mi, 0 }, 133 { X86::AND32ri8, X86::AND32mi8, 0 }, 134 { X86::AND32rr, X86::AND32mr, 0 }, 135 { X86::AND64ri32, X86::AND64mi32, 0 }, 136 { X86::AND64ri8, X86::AND64mi8, 0 }, 137 { X86::AND64rr, X86::AND64mr, 0 }, 138 { X86::AND8ri, X86::AND8mi, 0 }, 139 { X86::AND8rr, X86::AND8mr, 0 }, 140 { X86::DEC16r, X86::DEC16m, 0 }, 141 { X86::DEC32r, X86::DEC32m, 0 }, 142 { X86::DEC64_16r, X86::DEC64_16m, 0 }, 143 { X86::DEC64_32r, X86::DEC64_32m, 0 }, 144 { X86::DEC64r, X86::DEC64m, 0 }, 145 { X86::DEC8r, X86::DEC8m, 0 }, 146 { X86::INC16r, X86::INC16m, 0 }, 147 { X86::INC32r, X86::INC32m, 0 }, 148 { X86::INC64_16r, X86::INC64_16m, 0 }, 149 { X86::INC64_32r, X86::INC64_32m, 0 }, 150 { X86::INC64r, X86::INC64m, 0 }, 151 { X86::INC8r, X86::INC8m, 0 }, 152 { X86::NEG16r, X86::NEG16m, 0 }, 153 { X86::NEG32r, X86::NEG32m, 0 }, 154 { X86::NEG64r, X86::NEG64m, 0 }, 155 { X86::NEG8r, X86::NEG8m, 0 }, 156 { X86::NOT16r, X86::NOT16m, 0 }, 157 { X86::NOT32r, X86::NOT32m, 0 }, 158 { X86::NOT64r, X86::NOT64m, 0 }, 159 { X86::NOT8r, X86::NOT8m, 0 }, 160 { X86::OR16ri, X86::OR16mi, 0 }, 161 { X86::OR16ri8, X86::OR16mi8, 0 }, 162 { X86::OR16rr, X86::OR16mr, 0 }, 163 { X86::OR32ri, X86::OR32mi, 0 }, 164 { X86::OR32ri8, X86::OR32mi8, 0 }, 165 { X86::OR32rr, X86::OR32mr, 0 }, 166 { X86::OR64ri32, X86::OR64mi32, 0 }, 167 { X86::OR64ri8, X86::OR64mi8, 0 }, 168 { X86::OR64rr, X86::OR64mr, 0 }, 169 { X86::OR8ri, X86::OR8mi, 0 }, 170 { X86::OR8rr, X86::OR8mr, 0 }, 171 { X86::ROL16r1, X86::ROL16m1, 0 }, 172 { X86::ROL16rCL, X86::ROL16mCL, 0 }, 173 { X86::ROL16ri, X86::ROL16mi, 0 }, 174 { X86::ROL32r1, X86::ROL32m1, 0 }, 175 { X86::ROL32rCL, X86::ROL32mCL, 0 }, 176 { X86::ROL32ri, X86::ROL32mi, 0 }, 177 { X86::ROL64r1, X86::ROL64m1, 0 }, 178 { X86::ROL64rCL, X86::ROL64mCL, 0 }, 179 { X86::ROL64ri, X86::ROL64mi, 0 }, 180 { X86::ROL8r1, X86::ROL8m1, 0 }, 181 { X86::ROL8rCL, X86::ROL8mCL, 0 }, 182 { X86::ROL8ri, X86::ROL8mi, 0 }, 183 { X86::ROR16r1, X86::ROR16m1, 0 }, 184 { X86::ROR16rCL, X86::ROR16mCL, 0 }, 185 { X86::ROR16ri, X86::ROR16mi, 0 }, 186 { X86::ROR32r1, X86::ROR32m1, 0 }, 187 { X86::ROR32rCL, X86::ROR32mCL, 0 }, 188 { X86::ROR32ri, X86::ROR32mi, 0 }, 189 { X86::ROR64r1, X86::ROR64m1, 0 }, 190 { X86::ROR64rCL, X86::ROR64mCL, 0 }, 191 { X86::ROR64ri, X86::ROR64mi, 0 }, 192 { X86::ROR8r1, X86::ROR8m1, 0 }, 193 { X86::ROR8rCL, X86::ROR8mCL, 0 }, 194 { X86::ROR8ri, X86::ROR8mi, 0 }, 195 { X86::SAR16r1, X86::SAR16m1, 0 }, 196 { X86::SAR16rCL, X86::SAR16mCL, 0 }, 197 { X86::SAR16ri, X86::SAR16mi, 0 }, 198 { X86::SAR32r1, X86::SAR32m1, 0 }, 199 { X86::SAR32rCL, X86::SAR32mCL, 0 }, 200 { X86::SAR32ri, X86::SAR32mi, 0 }, 201 { X86::SAR64r1, X86::SAR64m1, 0 }, 202 { X86::SAR64rCL, X86::SAR64mCL, 0 }, 203 { X86::SAR64ri, X86::SAR64mi, 0 }, 204 { X86::SAR8r1, X86::SAR8m1, 0 }, 205 { X86::SAR8rCL, X86::SAR8mCL, 0 }, 206 { X86::SAR8ri, X86::SAR8mi, 0 }, 207 { X86::SBB32ri, X86::SBB32mi, 0 }, 208 { X86::SBB32ri8, X86::SBB32mi8, 0 }, 209 { X86::SBB32rr, X86::SBB32mr, 0 }, 210 { X86::SBB64ri32, X86::SBB64mi32, 0 }, 211 { X86::SBB64ri8, X86::SBB64mi8, 0 }, 212 { X86::SBB64rr, X86::SBB64mr, 0 }, 213 { X86::SHL16rCL, X86::SHL16mCL, 0 }, 214 { X86::SHL16ri, X86::SHL16mi, 0 }, 215 { X86::SHL32rCL, X86::SHL32mCL, 0 }, 216 { X86::SHL32ri, X86::SHL32mi, 0 }, 217 { X86::SHL64rCL, X86::SHL64mCL, 0 }, 218 { X86::SHL64ri, X86::SHL64mi, 0 }, 219 { X86::SHL8rCL, X86::SHL8mCL, 0 }, 220 { X86::SHL8ri, X86::SHL8mi, 0 }, 221 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 }, 222 { X86::SHLD16rri8, X86::SHLD16mri8, 0 }, 223 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 }, 224 { X86::SHLD32rri8, X86::SHLD32mri8, 0 }, 225 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 }, 226 { X86::SHLD64rri8, X86::SHLD64mri8, 0 }, 227 { X86::SHR16r1, X86::SHR16m1, 0 }, 228 { X86::SHR16rCL, X86::SHR16mCL, 0 }, 229 { X86::SHR16ri, X86::SHR16mi, 0 }, 230 { X86::SHR32r1, X86::SHR32m1, 0 }, 231 { X86::SHR32rCL, X86::SHR32mCL, 0 }, 232 { X86::SHR32ri, X86::SHR32mi, 0 }, 233 { X86::SHR64r1, X86::SHR64m1, 0 }, 234 { X86::SHR64rCL, X86::SHR64mCL, 0 }, 235 { X86::SHR64ri, X86::SHR64mi, 0 }, 236 { X86::SHR8r1, X86::SHR8m1, 0 }, 237 { X86::SHR8rCL, X86::SHR8mCL, 0 }, 238 { X86::SHR8ri, X86::SHR8mi, 0 }, 239 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 }, 240 { X86::SHRD16rri8, X86::SHRD16mri8, 0 }, 241 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 }, 242 { X86::SHRD32rri8, X86::SHRD32mri8, 0 }, 243 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 }, 244 { X86::SHRD64rri8, X86::SHRD64mri8, 0 }, 245 { X86::SUB16ri, X86::SUB16mi, 0 }, 246 { X86::SUB16ri8, X86::SUB16mi8, 0 }, 247 { X86::SUB16rr, X86::SUB16mr, 0 }, 248 { X86::SUB32ri, X86::SUB32mi, 0 }, 249 { X86::SUB32ri8, X86::SUB32mi8, 0 }, 250 { X86::SUB32rr, X86::SUB32mr, 0 }, 251 { X86::SUB64ri32, X86::SUB64mi32, 0 }, 252 { X86::SUB64ri8, X86::SUB64mi8, 0 }, 253 { X86::SUB64rr, X86::SUB64mr, 0 }, 254 { X86::SUB8ri, X86::SUB8mi, 0 }, 255 { X86::SUB8rr, X86::SUB8mr, 0 }, 256 { X86::XOR16ri, X86::XOR16mi, 0 }, 257 { X86::XOR16ri8, X86::XOR16mi8, 0 }, 258 { X86::XOR16rr, X86::XOR16mr, 0 }, 259 { X86::XOR32ri, X86::XOR32mi, 0 }, 260 { X86::XOR32ri8, X86::XOR32mi8, 0 }, 261 { X86::XOR32rr, X86::XOR32mr, 0 }, 262 { X86::XOR64ri32, X86::XOR64mi32, 0 }, 263 { X86::XOR64ri8, X86::XOR64mi8, 0 }, 264 { X86::XOR64rr, X86::XOR64mr, 0 }, 265 { X86::XOR8ri, X86::XOR8mi, 0 }, 266 { X86::XOR8rr, X86::XOR8mr, 0 } 267 }; 268 269 for (unsigned i = 0, e = array_lengthof(OpTbl2Addr); i != e; ++i) { 270 unsigned RegOp = OpTbl2Addr[i].RegOp; 271 unsigned MemOp = OpTbl2Addr[i].MemOp; 272 unsigned Flags = OpTbl2Addr[i].Flags; 273 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable, 274 RegOp, MemOp, 275 // Index 0, folded load and store, no alignment requirement. 276 Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE); 277 } 278 279 static const X86OpTblEntry OpTbl0[] = { 280 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD }, 281 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD }, 282 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD }, 283 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD }, 284 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD }, 285 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD }, 286 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD }, 287 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD }, 288 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD }, 289 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD }, 290 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD }, 291 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD }, 292 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD }, 293 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD }, 294 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD }, 295 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD }, 296 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD }, 297 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD }, 298 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD }, 299 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD }, 300 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 301 { X86::FsMOVAPDrr, X86::MOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE }, 302 { X86::FsMOVAPSrr, X86::MOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE }, 303 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD }, 304 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD }, 305 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD }, 306 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD }, 307 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD }, 308 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD }, 309 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD }, 310 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD }, 311 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD }, 312 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD }, 313 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE }, 314 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE }, 315 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE }, 316 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE }, 317 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE }, 318 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE }, 319 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE }, 320 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE }, 321 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE }, 322 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 323 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 324 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 325 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE }, 326 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE }, 327 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE }, 328 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE }, 329 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE }, 330 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE }, 331 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD }, 332 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD }, 333 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD }, 334 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD }, 335 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE }, 336 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE }, 337 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE }, 338 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE }, 339 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE }, 340 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE }, 341 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE }, 342 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE }, 343 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE }, 344 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE }, 345 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE }, 346 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE }, 347 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE }, 348 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE }, 349 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE }, 350 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE }, 351 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD }, 352 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD }, 353 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD }, 354 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD }, 355 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD }, 356 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD }, 357 // AVX 128-bit versions of foldable instructions 358 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 359 { X86::FsVMOVAPDrr, X86::VMOVSDmr, TB_FOLDED_STORE | TB_NO_REVERSE }, 360 { X86::FsVMOVAPSrr, X86::VMOVSSmr, TB_FOLDED_STORE | TB_NO_REVERSE }, 361 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 362 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 363 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 364 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 365 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE }, 366 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE }, 367 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE }, 368 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE }, 369 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE }, 370 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE }, 371 // AVX 256-bit foldable instructions 372 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 373 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 374 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 375 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 376 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE }, 377 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE } 378 }; 379 380 for (unsigned i = 0, e = array_lengthof(OpTbl0); i != e; ++i) { 381 unsigned RegOp = OpTbl0[i].RegOp; 382 unsigned MemOp = OpTbl0[i].MemOp; 383 unsigned Flags = OpTbl0[i].Flags; 384 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable, 385 RegOp, MemOp, TB_INDEX_0 | Flags); 386 } 387 388 static const X86OpTblEntry OpTbl1[] = { 389 { X86::CMP16rr, X86::CMP16rm, 0 }, 390 { X86::CMP32rr, X86::CMP32rm, 0 }, 391 { X86::CMP64rr, X86::CMP64rm, 0 }, 392 { X86::CMP8rr, X86::CMP8rm, 0 }, 393 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 }, 394 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 }, 395 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 }, 396 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 }, 397 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 }, 398 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 }, 399 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 }, 400 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 }, 401 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 }, 402 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 }, 403 { X86::FsMOVAPDrr, X86::MOVSDrm, TB_NO_REVERSE }, 404 { X86::FsMOVAPSrr, X86::MOVSSrm, TB_NO_REVERSE }, 405 { X86::IMUL16rri, X86::IMUL16rmi, 0 }, 406 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 }, 407 { X86::IMUL32rri, X86::IMUL32rmi, 0 }, 408 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 }, 409 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 }, 410 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 }, 411 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 }, 412 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 }, 413 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 }, 414 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 }, 415 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 }, 416 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 }, 417 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 }, 418 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 }, 419 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 }, 420 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 }, 421 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 }, 422 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 }, 423 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 }, 424 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 }, 425 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 }, 426 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 }, 427 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 }, 428 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 }, 429 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 }, 430 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 }, 431 { X86::MOV16rr, X86::MOV16rm, 0 }, 432 { X86::MOV32rr, X86::MOV32rm, 0 }, 433 { X86::MOV64rr, X86::MOV64rm, 0 }, 434 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 }, 435 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 }, 436 { X86::MOV8rr, X86::MOV8rm, 0 }, 437 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 }, 438 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 }, 439 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 }, 440 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 }, 441 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 }, 442 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 }, 443 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 }, 444 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 }, 445 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 }, 446 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 }, 447 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 }, 448 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 }, 449 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 }, 450 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 }, 451 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 }, 452 { X86::MOVUPSrr, X86::MOVUPSrm, 0 }, 453 { X86::MOVZDI2PDIrr, X86::MOVZDI2PDIrm, 0 }, 454 { X86::MOVZQI2PQIrr, X86::MOVZQI2PQIrm, 0 }, 455 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 }, 456 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 }, 457 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 }, 458 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 }, 459 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 }, 460 { X86::MOVZX64rr16, X86::MOVZX64rm16, 0 }, 461 { X86::MOVZX64rr32, X86::MOVZX64rm32, 0 }, 462 { X86::MOVZX64rr8, X86::MOVZX64rm8, 0 }, 463 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 }, 464 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 }, 465 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 }, 466 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 }, 467 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 }, 468 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 }, 469 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 }, 470 { X86::RCPPSr_Int, X86::RCPPSm_Int, TB_ALIGN_16 }, 471 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 }, 472 { X86::RSQRTPSr_Int, X86::RSQRTPSm_Int, TB_ALIGN_16 }, 473 { X86::RSQRTSSr, X86::RSQRTSSm, 0 }, 474 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 }, 475 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 }, 476 { X86::SQRTPDr_Int, X86::SQRTPDm_Int, TB_ALIGN_16 }, 477 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 }, 478 { X86::SQRTPSr_Int, X86::SQRTPSm_Int, TB_ALIGN_16 }, 479 { X86::SQRTSDr, X86::SQRTSDm, 0 }, 480 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 }, 481 { X86::SQRTSSr, X86::SQRTSSm, 0 }, 482 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 }, 483 { X86::TEST16rr, X86::TEST16rm, 0 }, 484 { X86::TEST32rr, X86::TEST32rm, 0 }, 485 { X86::TEST64rr, X86::TEST64rm, 0 }, 486 { X86::TEST8rr, X86::TEST8rm, 0 }, 487 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0 488 { X86::UCOMISDrr, X86::UCOMISDrm, 0 }, 489 { X86::UCOMISSrr, X86::UCOMISSrm, 0 }, 490 // AVX 128-bit versions of foldable instructions 491 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 }, 492 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 }, 493 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 }, 494 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 }, 495 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 }, 496 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 }, 497 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 }, 498 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 }, 499 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 }, 500 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 }, 501 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 }, 502 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 }, 503 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 }, 504 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 }, 505 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 }, 506 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 }, 507 { X86::FsVMOVAPDrr, X86::VMOVSDrm, TB_NO_REVERSE }, 508 { X86::FsVMOVAPSrr, X86::VMOVSSrm, TB_NO_REVERSE }, 509 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 }, 510 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 }, 511 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 }, 512 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 }, 513 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 }, 514 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 }, 515 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 }, 516 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 }, 517 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, TB_ALIGN_16 }, 518 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, TB_ALIGN_16 }, 519 { X86::VMOVUPDrr, X86::VMOVUPDrm, TB_ALIGN_16 }, 520 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 }, 521 { X86::VMOVZDI2PDIrr, X86::VMOVZDI2PDIrm, 0 }, 522 { X86::VMOVZQI2PQIrr, X86::VMOVZQI2PQIrm, 0 }, 523 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 }, 524 { X86::VPABSBrr128, X86::VPABSBrm128, TB_ALIGN_16 }, 525 { X86::VPABSDrr128, X86::VPABSDrm128, TB_ALIGN_16 }, 526 { X86::VPABSWrr128, X86::VPABSWrm128, TB_ALIGN_16 }, 527 { X86::VPERMILPDri, X86::VPERMILPDmi, TB_ALIGN_16 }, 528 { X86::VPERMILPSri, X86::VPERMILPSmi, TB_ALIGN_16 }, 529 { X86::VPSHUFDri, X86::VPSHUFDmi, TB_ALIGN_16 }, 530 { X86::VPSHUFHWri, X86::VPSHUFHWmi, TB_ALIGN_16 }, 531 { X86::VPSHUFLWri, X86::VPSHUFLWmi, TB_ALIGN_16 }, 532 { X86::VRCPPSr, X86::VRCPPSm, TB_ALIGN_16 }, 533 { X86::VRCPPSr_Int, X86::VRCPPSm_Int, TB_ALIGN_16 }, 534 { X86::VRSQRTPSr, X86::VRSQRTPSm, TB_ALIGN_16 }, 535 { X86::VRSQRTPSr_Int, X86::VRSQRTPSm_Int, TB_ALIGN_16 }, 536 { X86::VSQRTPDr, X86::VSQRTPDm, TB_ALIGN_16 }, 537 { X86::VSQRTPDr_Int, X86::VSQRTPDm_Int, TB_ALIGN_16 }, 538 { X86::VSQRTPSr, X86::VSQRTPSm, TB_ALIGN_16 }, 539 { X86::VSQRTPSr_Int, X86::VSQRTPSm_Int, TB_ALIGN_16 }, 540 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 }, 541 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 }, 542 // AVX 256-bit foldable instructions 543 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 }, 544 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 }, 545 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 }, 546 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 }, 547 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 }, 548 { X86::VPERMILPDYri, X86::VPERMILPDYmi, TB_ALIGN_32 }, 549 { X86::VPERMILPSYri, X86::VPERMILPSYmi, TB_ALIGN_32 }, 550 // AVX2 foldable instructions 551 { X86::VPABSBrr256, X86::VPABSBrm256, TB_ALIGN_32 }, 552 { X86::VPABSDrr256, X86::VPABSDrm256, TB_ALIGN_32 }, 553 { X86::VPABSWrr256, X86::VPABSWrm256, TB_ALIGN_32 }, 554 { X86::VPSHUFDYri, X86::VPSHUFDYmi, TB_ALIGN_32 }, 555 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, TB_ALIGN_32 }, 556 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, TB_ALIGN_32 }, 557 { X86::VRCPPSYr, X86::VRCPPSYm, TB_ALIGN_32 }, 558 { X86::VRCPPSYr_Int, X86::VRCPPSYm_Int, TB_ALIGN_32 }, 559 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, TB_ALIGN_32 }, 560 { X86::VRSQRTPSYr_Int, X86::VRSQRTPSYm_Int, TB_ALIGN_32 }, 561 { X86::VSQRTPDYr, X86::VSQRTPDYm, TB_ALIGN_32 }, 562 { X86::VSQRTPDYr_Int, X86::VSQRTPDYm_Int, TB_ALIGN_32 }, 563 { X86::VSQRTPSYr, X86::VSQRTPSYm, TB_ALIGN_32 }, 564 { X86::VSQRTPSYr_Int, X86::VSQRTPSYm_Int, TB_ALIGN_32 }, 565 }; 566 567 for (unsigned i = 0, e = array_lengthof(OpTbl1); i != e; ++i) { 568 unsigned RegOp = OpTbl1[i].RegOp; 569 unsigned MemOp = OpTbl1[i].MemOp; 570 unsigned Flags = OpTbl1[i].Flags; 571 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable, 572 RegOp, MemOp, 573 // Index 1, folded load 574 Flags | TB_INDEX_1 | TB_FOLDED_LOAD); 575 } 576 577 static const X86OpTblEntry OpTbl2[] = { 578 { X86::ADC32rr, X86::ADC32rm, 0 }, 579 { X86::ADC64rr, X86::ADC64rm, 0 }, 580 { X86::ADD16rr, X86::ADD16rm, 0 }, 581 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE }, 582 { X86::ADD32rr, X86::ADD32rm, 0 }, 583 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE }, 584 { X86::ADD64rr, X86::ADD64rm, 0 }, 585 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE }, 586 { X86::ADD8rr, X86::ADD8rm, 0 }, 587 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 }, 588 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 }, 589 { X86::ADDSDrr, X86::ADDSDrm, 0 }, 590 { X86::ADDSSrr, X86::ADDSSrm, 0 }, 591 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 }, 592 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 }, 593 { X86::AND16rr, X86::AND16rm, 0 }, 594 { X86::AND32rr, X86::AND32rm, 0 }, 595 { X86::AND64rr, X86::AND64rm, 0 }, 596 { X86::AND8rr, X86::AND8rm, 0 }, 597 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 }, 598 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 }, 599 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 }, 600 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 }, 601 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 }, 602 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 }, 603 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 }, 604 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 }, 605 { X86::CMOVA16rr, X86::CMOVA16rm, 0 }, 606 { X86::CMOVA32rr, X86::CMOVA32rm, 0 }, 607 { X86::CMOVA64rr, X86::CMOVA64rm, 0 }, 608 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 }, 609 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 }, 610 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 }, 611 { X86::CMOVB16rr, X86::CMOVB16rm, 0 }, 612 { X86::CMOVB32rr, X86::CMOVB32rm, 0 }, 613 { X86::CMOVB64rr, X86::CMOVB64rm, 0 }, 614 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 }, 615 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 }, 616 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 }, 617 { X86::CMOVE16rr, X86::CMOVE16rm, 0 }, 618 { X86::CMOVE32rr, X86::CMOVE32rm, 0 }, 619 { X86::CMOVE64rr, X86::CMOVE64rm, 0 }, 620 { X86::CMOVG16rr, X86::CMOVG16rm, 0 }, 621 { X86::CMOVG32rr, X86::CMOVG32rm, 0 }, 622 { X86::CMOVG64rr, X86::CMOVG64rm, 0 }, 623 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 }, 624 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 }, 625 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 }, 626 { X86::CMOVL16rr, X86::CMOVL16rm, 0 }, 627 { X86::CMOVL32rr, X86::CMOVL32rm, 0 }, 628 { X86::CMOVL64rr, X86::CMOVL64rm, 0 }, 629 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 }, 630 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 }, 631 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 }, 632 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 }, 633 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 }, 634 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 }, 635 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 }, 636 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 }, 637 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 }, 638 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 }, 639 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 }, 640 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 }, 641 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 }, 642 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 }, 643 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 }, 644 { X86::CMOVO16rr, X86::CMOVO16rm, 0 }, 645 { X86::CMOVO32rr, X86::CMOVO32rm, 0 }, 646 { X86::CMOVO64rr, X86::CMOVO64rm, 0 }, 647 { X86::CMOVP16rr, X86::CMOVP16rm, 0 }, 648 { X86::CMOVP32rr, X86::CMOVP32rm, 0 }, 649 { X86::CMOVP64rr, X86::CMOVP64rm, 0 }, 650 { X86::CMOVS16rr, X86::CMOVS16rm, 0 }, 651 { X86::CMOVS32rr, X86::CMOVS32rm, 0 }, 652 { X86::CMOVS64rr, X86::CMOVS64rm, 0 }, 653 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 }, 654 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 }, 655 { X86::CMPSDrr, X86::CMPSDrm, 0 }, 656 { X86::CMPSSrr, X86::CMPSSrm, 0 }, 657 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 }, 658 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 }, 659 { X86::DIVSDrr, X86::DIVSDrm, 0 }, 660 { X86::DIVSSrr, X86::DIVSSrm, 0 }, 661 { X86::FsANDNPDrr, X86::FsANDNPDrm, TB_ALIGN_16 }, 662 { X86::FsANDNPSrr, X86::FsANDNPSrm, TB_ALIGN_16 }, 663 { X86::FsANDPDrr, X86::FsANDPDrm, TB_ALIGN_16 }, 664 { X86::FsANDPSrr, X86::FsANDPSrm, TB_ALIGN_16 }, 665 { X86::FsORPDrr, X86::FsORPDrm, TB_ALIGN_16 }, 666 { X86::FsORPSrr, X86::FsORPSrm, TB_ALIGN_16 }, 667 { X86::FsXORPDrr, X86::FsXORPDrm, TB_ALIGN_16 }, 668 { X86::FsXORPSrr, X86::FsXORPSrm, TB_ALIGN_16 }, 669 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 }, 670 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 }, 671 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 }, 672 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 }, 673 { X86::IMUL16rr, X86::IMUL16rm, 0 }, 674 { X86::IMUL32rr, X86::IMUL32rm, 0 }, 675 { X86::IMUL64rr, X86::IMUL64rm, 0 }, 676 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 }, 677 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 }, 678 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 }, 679 { X86::MAXPDrr_Int, X86::MAXPDrm_Int, TB_ALIGN_16 }, 680 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 }, 681 { X86::MAXPSrr_Int, X86::MAXPSrm_Int, TB_ALIGN_16 }, 682 { X86::MAXSDrr, X86::MAXSDrm, 0 }, 683 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 }, 684 { X86::MAXSSrr, X86::MAXSSrm, 0 }, 685 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 }, 686 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 }, 687 { X86::MINPDrr_Int, X86::MINPDrm_Int, TB_ALIGN_16 }, 688 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 }, 689 { X86::MINPSrr_Int, X86::MINPSrm_Int, TB_ALIGN_16 }, 690 { X86::MINSDrr, X86::MINSDrm, 0 }, 691 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 }, 692 { X86::MINSSrr, X86::MINSSrm, 0 }, 693 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 }, 694 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 }, 695 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 }, 696 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 }, 697 { X86::MULSDrr, X86::MULSDrm, 0 }, 698 { X86::MULSSrr, X86::MULSSrm, 0 }, 699 { X86::OR16rr, X86::OR16rm, 0 }, 700 { X86::OR32rr, X86::OR32rm, 0 }, 701 { X86::OR64rr, X86::OR64rm, 0 }, 702 { X86::OR8rr, X86::OR8rm, 0 }, 703 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 }, 704 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 }, 705 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 }, 706 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 }, 707 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 }, 708 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 }, 709 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 }, 710 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 }, 711 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 }, 712 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 }, 713 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 }, 714 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 }, 715 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 }, 716 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 }, 717 { X86::PALIGNR128rr, X86::PALIGNR128rm, TB_ALIGN_16 }, 718 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 }, 719 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 }, 720 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 }, 721 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 }, 722 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 }, 723 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 }, 724 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 }, 725 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 }, 726 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 }, 727 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 }, 728 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 }, 729 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 }, 730 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 }, 731 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 }, 732 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 }, 733 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 }, 734 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 }, 735 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 }, 736 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 }, 737 { X86::PINSRWrri, X86::PINSRWrmi, TB_ALIGN_16 }, 738 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 }, 739 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 }, 740 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 }, 741 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 }, 742 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 }, 743 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 }, 744 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 }, 745 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 }, 746 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 }, 747 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 }, 748 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 }, 749 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 }, 750 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 }, 751 { X86::PORrr, X86::PORrm, TB_ALIGN_16 }, 752 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 }, 753 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 }, 754 { X86::PSIGNBrr, X86::PSIGNBrm, TB_ALIGN_16 }, 755 { X86::PSIGNWrr, X86::PSIGNWrm, TB_ALIGN_16 }, 756 { X86::PSIGNDrr, X86::PSIGNDrm, TB_ALIGN_16 }, 757 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 }, 758 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 }, 759 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 }, 760 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 }, 761 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 }, 762 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 }, 763 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 }, 764 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 }, 765 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 }, 766 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 }, 767 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 }, 768 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 }, 769 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 }, 770 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 }, 771 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 }, 772 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 }, 773 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 }, 774 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 }, 775 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 }, 776 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 }, 777 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 }, 778 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 }, 779 { X86::SBB32rr, X86::SBB32rm, 0 }, 780 { X86::SBB64rr, X86::SBB64rm, 0 }, 781 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 }, 782 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 }, 783 { X86::SUB16rr, X86::SUB16rm, 0 }, 784 { X86::SUB32rr, X86::SUB32rm, 0 }, 785 { X86::SUB64rr, X86::SUB64rm, 0 }, 786 { X86::SUB8rr, X86::SUB8rm, 0 }, 787 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 }, 788 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 }, 789 { X86::SUBSDrr, X86::SUBSDrm, 0 }, 790 { X86::SUBSSrr, X86::SUBSSrm, 0 }, 791 // FIXME: TEST*rr -> swapped operand of TEST*mr. 792 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 }, 793 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 }, 794 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 }, 795 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 }, 796 { X86::XOR16rr, X86::XOR16rm, 0 }, 797 { X86::XOR32rr, X86::XOR32rm, 0 }, 798 { X86::XOR64rr, X86::XOR64rm, 0 }, 799 { X86::XOR8rr, X86::XOR8rm, 0 }, 800 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 }, 801 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 }, 802 // AVX 128-bit versions of foldable instructions 803 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 }, 804 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 }, 805 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 }, 806 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 }, 807 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 }, 808 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 }, 809 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 }, 810 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 }, 811 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 }, 812 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 }, 813 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 }, 814 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 }, 815 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQrm, TB_ALIGN_16 }, 816 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, TB_ALIGN_16 }, 817 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 }, 818 { X86::VSQRTSDr, X86::VSQRTSDm, 0 }, 819 { X86::VSQRTSSr, X86::VSQRTSSm, 0 }, 820 { X86::VADDPDrr, X86::VADDPDrm, TB_ALIGN_16 }, 821 { X86::VADDPSrr, X86::VADDPSrm, TB_ALIGN_16 }, 822 { X86::VADDSDrr, X86::VADDSDrm, 0 }, 823 { X86::VADDSSrr, X86::VADDSSrm, 0 }, 824 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, TB_ALIGN_16 }, 825 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, TB_ALIGN_16 }, 826 { X86::VANDNPDrr, X86::VANDNPDrm, TB_ALIGN_16 }, 827 { X86::VANDNPSrr, X86::VANDNPSrm, TB_ALIGN_16 }, 828 { X86::VANDPDrr, X86::VANDPDrm, TB_ALIGN_16 }, 829 { X86::VANDPSrr, X86::VANDPSrm, TB_ALIGN_16 }, 830 { X86::VBLENDPDrri, X86::VBLENDPDrmi, TB_ALIGN_16 }, 831 { X86::VBLENDPSrri, X86::VBLENDPSrmi, TB_ALIGN_16 }, 832 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, TB_ALIGN_16 }, 833 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, TB_ALIGN_16 }, 834 { X86::VCMPPDrri, X86::VCMPPDrmi, TB_ALIGN_16 }, 835 { X86::VCMPPSrri, X86::VCMPPSrmi, TB_ALIGN_16 }, 836 { X86::VCMPSDrr, X86::VCMPSDrm, 0 }, 837 { X86::VCMPSSrr, X86::VCMPSSrm, 0 }, 838 { X86::VDIVPDrr, X86::VDIVPDrm, TB_ALIGN_16 }, 839 { X86::VDIVPSrr, X86::VDIVPSrm, TB_ALIGN_16 }, 840 { X86::VDIVSDrr, X86::VDIVSDrm, 0 }, 841 { X86::VDIVSSrr, X86::VDIVSSrm, 0 }, 842 { X86::VFsANDNPDrr, X86::VFsANDNPDrm, TB_ALIGN_16 }, 843 { X86::VFsANDNPSrr, X86::VFsANDNPSrm, TB_ALIGN_16 }, 844 { X86::VFsANDPDrr, X86::VFsANDPDrm, TB_ALIGN_16 }, 845 { X86::VFsANDPSrr, X86::VFsANDPSrm, TB_ALIGN_16 }, 846 { X86::VFsORPDrr, X86::VFsORPDrm, TB_ALIGN_16 }, 847 { X86::VFsORPSrr, X86::VFsORPSrm, TB_ALIGN_16 }, 848 { X86::VFsXORPDrr, X86::VFsXORPDrm, TB_ALIGN_16 }, 849 { X86::VFsXORPSrr, X86::VFsXORPSrm, TB_ALIGN_16 }, 850 { X86::VHADDPDrr, X86::VHADDPDrm, TB_ALIGN_16 }, 851 { X86::VHADDPSrr, X86::VHADDPSrm, TB_ALIGN_16 }, 852 { X86::VHSUBPDrr, X86::VHSUBPDrm, TB_ALIGN_16 }, 853 { X86::VHSUBPSrr, X86::VHSUBPSrm, TB_ALIGN_16 }, 854 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 }, 855 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 }, 856 { X86::VMAXPDrr, X86::VMAXPDrm, TB_ALIGN_16 }, 857 { X86::VMAXPDrr_Int, X86::VMAXPDrm_Int, TB_ALIGN_16 }, 858 { X86::VMAXPSrr, X86::VMAXPSrm, TB_ALIGN_16 }, 859 { X86::VMAXPSrr_Int, X86::VMAXPSrm_Int, TB_ALIGN_16 }, 860 { X86::VMAXSDrr, X86::VMAXSDrm, 0 }, 861 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 }, 862 { X86::VMAXSSrr, X86::VMAXSSrm, 0 }, 863 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 }, 864 { X86::VMINPDrr, X86::VMINPDrm, TB_ALIGN_16 }, 865 { X86::VMINPDrr_Int, X86::VMINPDrm_Int, TB_ALIGN_16 }, 866 { X86::VMINPSrr, X86::VMINPSrm, TB_ALIGN_16 }, 867 { X86::VMINPSrr_Int, X86::VMINPSrm_Int, TB_ALIGN_16 }, 868 { X86::VMINSDrr, X86::VMINSDrm, 0 }, 869 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 }, 870 { X86::VMINSSrr, X86::VMINSSrm, 0 }, 871 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 }, 872 { X86::VMPSADBWrri, X86::VMPSADBWrmi, TB_ALIGN_16 }, 873 { X86::VMULPDrr, X86::VMULPDrm, TB_ALIGN_16 }, 874 { X86::VMULPSrr, X86::VMULPSrm, TB_ALIGN_16 }, 875 { X86::VMULSDrr, X86::VMULSDrm, 0 }, 876 { X86::VMULSSrr, X86::VMULSSrm, 0 }, 877 { X86::VORPDrr, X86::VORPDrm, TB_ALIGN_16 }, 878 { X86::VORPSrr, X86::VORPSrm, TB_ALIGN_16 }, 879 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, TB_ALIGN_16 }, 880 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, TB_ALIGN_16 }, 881 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, TB_ALIGN_16 }, 882 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, TB_ALIGN_16 }, 883 { X86::VPADDBrr, X86::VPADDBrm, TB_ALIGN_16 }, 884 { X86::VPADDDrr, X86::VPADDDrm, TB_ALIGN_16 }, 885 { X86::VPADDQrr, X86::VPADDQrm, TB_ALIGN_16 }, 886 { X86::VPADDSBrr, X86::VPADDSBrm, TB_ALIGN_16 }, 887 { X86::VPADDSWrr, X86::VPADDSWrm, TB_ALIGN_16 }, 888 { X86::VPADDUSBrr, X86::VPADDUSBrm, TB_ALIGN_16 }, 889 { X86::VPADDUSWrr, X86::VPADDUSWrm, TB_ALIGN_16 }, 890 { X86::VPADDWrr, X86::VPADDWrm, TB_ALIGN_16 }, 891 { X86::VPALIGNR128rr, X86::VPALIGNR128rm, TB_ALIGN_16 }, 892 { X86::VPANDNrr, X86::VPANDNrm, TB_ALIGN_16 }, 893 { X86::VPANDrr, X86::VPANDrm, TB_ALIGN_16 }, 894 { X86::VPAVGBrr, X86::VPAVGBrm, TB_ALIGN_16 }, 895 { X86::VPAVGWrr, X86::VPAVGWrm, TB_ALIGN_16 }, 896 { X86::VPBLENDWrri, X86::VPBLENDWrmi, TB_ALIGN_16 }, 897 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, TB_ALIGN_16 }, 898 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, TB_ALIGN_16 }, 899 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, TB_ALIGN_16 }, 900 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, TB_ALIGN_16 }, 901 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, TB_ALIGN_16 }, 902 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, TB_ALIGN_16 }, 903 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, TB_ALIGN_16 }, 904 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, TB_ALIGN_16 }, 905 { X86::VPHADDDrr, X86::VPHADDDrm, TB_ALIGN_16 }, 906 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, TB_ALIGN_16 }, 907 { X86::VPHADDWrr, X86::VPHADDWrm, TB_ALIGN_16 }, 908 { X86::VPHSUBDrr, X86::VPHSUBDrm, TB_ALIGN_16 }, 909 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, TB_ALIGN_16 }, 910 { X86::VPHSUBWrr, X86::VPHSUBWrm, TB_ALIGN_16 }, 911 { X86::VPERMILPDrr, X86::VPERMILPDrm, TB_ALIGN_16 }, 912 { X86::VPERMILPSrr, X86::VPERMILPSrm, TB_ALIGN_16 }, 913 { X86::VPINSRWrri, X86::VPINSRWrmi, TB_ALIGN_16 }, 914 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, TB_ALIGN_16 }, 915 { X86::VPMADDWDrr, X86::VPMADDWDrm, TB_ALIGN_16 }, 916 { X86::VPMAXSWrr, X86::VPMAXSWrm, TB_ALIGN_16 }, 917 { X86::VPMAXUBrr, X86::VPMAXUBrm, TB_ALIGN_16 }, 918 { X86::VPMINSWrr, X86::VPMINSWrm, TB_ALIGN_16 }, 919 { X86::VPMINUBrr, X86::VPMINUBrm, TB_ALIGN_16 }, 920 { X86::VPMULDQrr, X86::VPMULDQrm, TB_ALIGN_16 }, 921 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, TB_ALIGN_16 }, 922 { X86::VPMULHUWrr, X86::VPMULHUWrm, TB_ALIGN_16 }, 923 { X86::VPMULHWrr, X86::VPMULHWrm, TB_ALIGN_16 }, 924 { X86::VPMULLDrr, X86::VPMULLDrm, TB_ALIGN_16 }, 925 { X86::VPMULLWrr, X86::VPMULLWrm, TB_ALIGN_16 }, 926 { X86::VPMULUDQrr, X86::VPMULUDQrm, TB_ALIGN_16 }, 927 { X86::VPORrr, X86::VPORrm, TB_ALIGN_16 }, 928 { X86::VPSADBWrr, X86::VPSADBWrm, TB_ALIGN_16 }, 929 { X86::VPSHUFBrr, X86::VPSHUFBrm, TB_ALIGN_16 }, 930 { X86::VPSIGNBrr, X86::VPSIGNBrm, TB_ALIGN_16 }, 931 { X86::VPSIGNWrr, X86::VPSIGNWrm, TB_ALIGN_16 }, 932 { X86::VPSIGNDrr, X86::VPSIGNDrm, TB_ALIGN_16 }, 933 { X86::VPSLLDrr, X86::VPSLLDrm, TB_ALIGN_16 }, 934 { X86::VPSLLQrr, X86::VPSLLQrm, TB_ALIGN_16 }, 935 { X86::VPSLLWrr, X86::VPSLLWrm, TB_ALIGN_16 }, 936 { X86::VPSRADrr, X86::VPSRADrm, TB_ALIGN_16 }, 937 { X86::VPSRAWrr, X86::VPSRAWrm, TB_ALIGN_16 }, 938 { X86::VPSRLDrr, X86::VPSRLDrm, TB_ALIGN_16 }, 939 { X86::VPSRLQrr, X86::VPSRLQrm, TB_ALIGN_16 }, 940 { X86::VPSRLWrr, X86::VPSRLWrm, TB_ALIGN_16 }, 941 { X86::VPSUBBrr, X86::VPSUBBrm, TB_ALIGN_16 }, 942 { X86::VPSUBDrr, X86::VPSUBDrm, TB_ALIGN_16 }, 943 { X86::VPSUBSBrr, X86::VPSUBSBrm, TB_ALIGN_16 }, 944 { X86::VPSUBSWrr, X86::VPSUBSWrm, TB_ALIGN_16 }, 945 { X86::VPSUBWrr, X86::VPSUBWrm, TB_ALIGN_16 }, 946 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, TB_ALIGN_16 }, 947 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, TB_ALIGN_16 }, 948 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, TB_ALIGN_16 }, 949 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, TB_ALIGN_16 }, 950 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, TB_ALIGN_16 }, 951 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, TB_ALIGN_16 }, 952 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, TB_ALIGN_16 }, 953 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, TB_ALIGN_16 }, 954 { X86::VPXORrr, X86::VPXORrm, TB_ALIGN_16 }, 955 { X86::VSHUFPDrri, X86::VSHUFPDrmi, TB_ALIGN_16 }, 956 { X86::VSHUFPSrri, X86::VSHUFPSrmi, TB_ALIGN_16 }, 957 { X86::VSUBPDrr, X86::VSUBPDrm, TB_ALIGN_16 }, 958 { X86::VSUBPSrr, X86::VSUBPSrm, TB_ALIGN_16 }, 959 { X86::VSUBSDrr, X86::VSUBSDrm, 0 }, 960 { X86::VSUBSSrr, X86::VSUBSSrm, 0 }, 961 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, TB_ALIGN_16 }, 962 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, TB_ALIGN_16 }, 963 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, TB_ALIGN_16 }, 964 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, TB_ALIGN_16 }, 965 { X86::VXORPDrr, X86::VXORPDrm, TB_ALIGN_16 }, 966 { X86::VXORPSrr, X86::VXORPSrm, TB_ALIGN_16 }, 967 // AVX 256-bit foldable instructions 968 { X86::VADDPDYrr, X86::VADDPDYrm, TB_ALIGN_32 }, 969 { X86::VADDPSYrr, X86::VADDPSYrm, TB_ALIGN_32 }, 970 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, TB_ALIGN_32 }, 971 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, TB_ALIGN_32 }, 972 { X86::VANDNPDYrr, X86::VANDNPDYrm, TB_ALIGN_32 }, 973 { X86::VANDNPSYrr, X86::VANDNPSYrm, TB_ALIGN_32 }, 974 { X86::VANDPDYrr, X86::VANDPDYrm, TB_ALIGN_32 }, 975 { X86::VANDPSYrr, X86::VANDPSYrm, TB_ALIGN_32 }, 976 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, TB_ALIGN_32 }, 977 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, TB_ALIGN_32 }, 978 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, TB_ALIGN_32 }, 979 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, TB_ALIGN_32 }, 980 { X86::VCMPPDYrri, X86::VCMPPDYrmi, TB_ALIGN_32 }, 981 { X86::VCMPPSYrri, X86::VCMPPSYrmi, TB_ALIGN_32 }, 982 { X86::VDIVPDYrr, X86::VDIVPDYrm, TB_ALIGN_32 }, 983 { X86::VDIVPSYrr, X86::VDIVPSYrm, TB_ALIGN_32 }, 984 { X86::VHADDPDYrr, X86::VHADDPDYrm, TB_ALIGN_32 }, 985 { X86::VHADDPSYrr, X86::VHADDPSYrm, TB_ALIGN_32 }, 986 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, TB_ALIGN_32 }, 987 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, TB_ALIGN_32 }, 988 { X86::VINSERTF128rr, X86::VINSERTF128rm, TB_ALIGN_32 }, 989 { X86::VMAXPDYrr, X86::VMAXPDYrm, TB_ALIGN_32 }, 990 { X86::VMAXPDYrr_Int, X86::VMAXPDYrm_Int, TB_ALIGN_32 }, 991 { X86::VMAXPSYrr, X86::VMAXPSYrm, TB_ALIGN_32 }, 992 { X86::VMAXPSYrr_Int, X86::VMAXPSYrm_Int, TB_ALIGN_32 }, 993 { X86::VMINPDYrr, X86::VMINPDYrm, TB_ALIGN_32 }, 994 { X86::VMINPDYrr_Int, X86::VMINPDYrm_Int, TB_ALIGN_32 }, 995 { X86::VMINPSYrr, X86::VMINPSYrm, TB_ALIGN_32 }, 996 { X86::VMINPSYrr_Int, X86::VMINPSYrm_Int, TB_ALIGN_32 }, 997 { X86::VMULPDYrr, X86::VMULPDYrm, TB_ALIGN_32 }, 998 { X86::VMULPSYrr, X86::VMULPSYrm, TB_ALIGN_32 }, 999 { X86::VORPDYrr, X86::VORPDYrm, TB_ALIGN_32 }, 1000 { X86::VORPSYrr, X86::VORPSYrm, TB_ALIGN_32 }, 1001 { X86::VPERM2F128rr, X86::VPERM2F128rm, TB_ALIGN_32 }, 1002 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, TB_ALIGN_32 }, 1003 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, TB_ALIGN_32 }, 1004 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, TB_ALIGN_32 }, 1005 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, TB_ALIGN_32 }, 1006 { X86::VSUBPDYrr, X86::VSUBPDYrm, TB_ALIGN_32 }, 1007 { X86::VSUBPSYrr, X86::VSUBPSYrm, TB_ALIGN_32 }, 1008 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, TB_ALIGN_32 }, 1009 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, TB_ALIGN_32 }, 1010 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, TB_ALIGN_32 }, 1011 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, TB_ALIGN_32 }, 1012 { X86::VXORPDYrr, X86::VXORPDYrm, TB_ALIGN_32 }, 1013 { X86::VXORPSYrr, X86::VXORPSYrm, TB_ALIGN_32 }, 1014 // AVX2 foldable instructions 1015 { X86::VINSERTI128rr, X86::VINSERTI128rm, TB_ALIGN_16 }, 1016 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, TB_ALIGN_32 }, 1017 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, TB_ALIGN_32 }, 1018 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, TB_ALIGN_32 }, 1019 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, TB_ALIGN_32 }, 1020 { X86::VPADDBYrr, X86::VPADDBYrm, TB_ALIGN_32 }, 1021 { X86::VPADDDYrr, X86::VPADDDYrm, TB_ALIGN_32 }, 1022 { X86::VPADDQYrr, X86::VPADDQYrm, TB_ALIGN_32 }, 1023 { X86::VPADDSBYrr, X86::VPADDSBYrm, TB_ALIGN_32 }, 1024 { X86::VPADDSWYrr, X86::VPADDSWYrm, TB_ALIGN_32 }, 1025 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, TB_ALIGN_32 }, 1026 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, TB_ALIGN_32 }, 1027 { X86::VPADDWYrr, X86::VPADDWYrm, TB_ALIGN_32 }, 1028 { X86::VPALIGNR256rr, X86::VPALIGNR256rm, TB_ALIGN_32 }, 1029 { X86::VPANDNYrr, X86::VPANDNYrm, TB_ALIGN_32 }, 1030 { X86::VPANDYrr, X86::VPANDYrm, TB_ALIGN_32 }, 1031 { X86::VPAVGBYrr, X86::VPAVGBYrm, TB_ALIGN_32 }, 1032 { X86::VPAVGWYrr, X86::VPAVGWYrm, TB_ALIGN_32 }, 1033 { X86::VPBLENDDrri, X86::VPBLENDDrmi, TB_ALIGN_32 }, 1034 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, TB_ALIGN_32 }, 1035 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, TB_ALIGN_32 }, 1036 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, TB_ALIGN_32 }, 1037 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, TB_ALIGN_32 }, 1038 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, TB_ALIGN_32 }, 1039 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, TB_ALIGN_32 }, 1040 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, TB_ALIGN_32 }, 1041 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, TB_ALIGN_32 }, 1042 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, TB_ALIGN_32 }, 1043 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, TB_ALIGN_32 }, 1044 { X86::VPERM2I128rr, X86::VPERM2I128rm, TB_ALIGN_32 }, 1045 { X86::VPERMDYrr, X86::VPERMDYrm, TB_ALIGN_32 }, 1046 { X86::VPERMPDYri, X86::VPERMPDYmi, TB_ALIGN_32 }, 1047 { X86::VPERMPSYrr, X86::VPERMPSYrm, TB_ALIGN_32 }, 1048 { X86::VPERMQYri, X86::VPERMQYmi, TB_ALIGN_32 }, 1049 { X86::VPHADDDYrr, X86::VPHADDDYrm, TB_ALIGN_32 }, 1050 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, TB_ALIGN_32 }, 1051 { X86::VPHADDWYrr, X86::VPHADDWYrm, TB_ALIGN_32 }, 1052 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, TB_ALIGN_32 }, 1053 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, TB_ALIGN_32 }, 1054 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, TB_ALIGN_32 }, 1055 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, TB_ALIGN_32 }, 1056 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, TB_ALIGN_32 }, 1057 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, TB_ALIGN_32 }, 1058 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, TB_ALIGN_32 }, 1059 { X86::VPMINSWYrr, X86::VPMINSWYrm, TB_ALIGN_32 }, 1060 { X86::VPMINUBYrr, X86::VPMINUBYrm, TB_ALIGN_32 }, 1061 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, TB_ALIGN_32 }, 1062 { X86::VPMULDQYrr, X86::VPMULDQYrm, TB_ALIGN_32 }, 1063 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, TB_ALIGN_32 }, 1064 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, TB_ALIGN_32 }, 1065 { X86::VPMULHWYrr, X86::VPMULHWYrm, TB_ALIGN_32 }, 1066 { X86::VPMULLDYrr, X86::VPMULLDYrm, TB_ALIGN_32 }, 1067 { X86::VPMULLWYrr, X86::VPMULLWYrm, TB_ALIGN_32 }, 1068 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, TB_ALIGN_32 }, 1069 { X86::VPORYrr, X86::VPORYrm, TB_ALIGN_32 }, 1070 { X86::VPSADBWYrr, X86::VPSADBWYrm, TB_ALIGN_32 }, 1071 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, TB_ALIGN_32 }, 1072 { X86::VPSIGNBYrr, X86::VPSIGNBYrm, TB_ALIGN_32 }, 1073 { X86::VPSIGNWYrr, X86::VPSIGNWYrm, TB_ALIGN_32 }, 1074 { X86::VPSIGNDYrr, X86::VPSIGNDYrm, TB_ALIGN_32 }, 1075 { X86::VPSLLDYrr, X86::VPSLLDYrm, TB_ALIGN_16 }, 1076 { X86::VPSLLQYrr, X86::VPSLLQYrm, TB_ALIGN_16 }, 1077 { X86::VPSLLWYrr, X86::VPSLLWYrm, TB_ALIGN_16 }, 1078 { X86::VPSLLVDrr, X86::VPSLLVDrm, TB_ALIGN_16 }, 1079 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, TB_ALIGN_32 }, 1080 { X86::VPSLLVQrr, X86::VPSLLVQrm, TB_ALIGN_16 }, 1081 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, TB_ALIGN_32 }, 1082 { X86::VPSRADYrr, X86::VPSRADYrm, TB_ALIGN_16 }, 1083 { X86::VPSRAWYrr, X86::VPSRAWYrm, TB_ALIGN_16 }, 1084 { X86::VPSRAVDrr, X86::VPSRAVDrm, TB_ALIGN_16 }, 1085 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, TB_ALIGN_32 }, 1086 { X86::VPSRLDYrr, X86::VPSRLDYrm, TB_ALIGN_16 }, 1087 { X86::VPSRLQYrr, X86::VPSRLQYrm, TB_ALIGN_16 }, 1088 { X86::VPSRLWYrr, X86::VPSRLWYrm, TB_ALIGN_16 }, 1089 { X86::VPSRLVDrr, X86::VPSRLVDrm, TB_ALIGN_16 }, 1090 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, TB_ALIGN_32 }, 1091 { X86::VPSRLVQrr, X86::VPSRLVQrm, TB_ALIGN_16 }, 1092 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, TB_ALIGN_32 }, 1093 { X86::VPSUBBYrr, X86::VPSUBBYrm, TB_ALIGN_32 }, 1094 { X86::VPSUBDYrr, X86::VPSUBDYrm, TB_ALIGN_32 }, 1095 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, TB_ALIGN_32 }, 1096 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, TB_ALIGN_32 }, 1097 { X86::VPSUBWYrr, X86::VPSUBWYrm, TB_ALIGN_32 }, 1098 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, TB_ALIGN_32 }, 1099 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, TB_ALIGN_32 }, 1100 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, TB_ALIGN_16 }, 1101 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, TB_ALIGN_32 }, 1102 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, TB_ALIGN_32 }, 1103 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, TB_ALIGN_32 }, 1104 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, TB_ALIGN_32 }, 1105 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, TB_ALIGN_32 }, 1106 { X86::VPXORYrr, X86::VPXORYrm, TB_ALIGN_32 }, 1107 // FIXME: add AVX 256-bit foldable instructions 1108 }; 1109 1110 for (unsigned i = 0, e = array_lengthof(OpTbl2); i != e; ++i) { 1111 unsigned RegOp = OpTbl2[i].RegOp; 1112 unsigned MemOp = OpTbl2[i].MemOp; 1113 unsigned Flags = OpTbl2[i].Flags; 1114 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable, 1115 RegOp, MemOp, 1116 // Index 2, folded load 1117 Flags | TB_INDEX_2 | TB_FOLDED_LOAD); 1118 } 1119 1120 static const X86OpTblEntry OpTbl3[] = { 1121 // FMA foldable instructions 1122 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, 0 }, 1123 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, 0 }, 1124 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, 0 }, 1125 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, 0 }, 1126 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, 0 }, 1127 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, 0 }, 1128 { X86::VFMADDSSr132r_Int, X86::VFMADDSSr132m_Int, 0 }, 1129 { X86::VFMADDSDr132r_Int, X86::VFMADDSDr132m_Int, 0 }, 1130 1131 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_16 }, 1132 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_16 }, 1133 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_16 }, 1134 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_16 }, 1135 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_16 }, 1136 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_16 }, 1137 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_32 }, 1138 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_32 }, 1139 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_32 }, 1140 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_32 }, 1141 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_32 }, 1142 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_32 }, 1143 { X86::VFMADDPSr132r_Int, X86::VFMADDPSr132m_Int, TB_ALIGN_16 }, 1144 { X86::VFMADDPDr132r_Int, X86::VFMADDPDr132m_Int, TB_ALIGN_16 }, 1145 { X86::VFMADDPSr132rY_Int, X86::VFMADDPSr132mY_Int, TB_ALIGN_32 }, 1146 { X86::VFMADDPDr132rY_Int, X86::VFMADDPDr132mY_Int, TB_ALIGN_32 }, 1147 1148 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, 0 }, 1149 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, 0 }, 1150 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, 0 }, 1151 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, 0 }, 1152 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, 0 }, 1153 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, 0 }, 1154 { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr132m_Int, 0 }, 1155 { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr132m_Int, 0 }, 1156 1157 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_16 }, 1158 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_16 }, 1159 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_16 }, 1160 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_16 }, 1161 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_16 }, 1162 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_16 }, 1163 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_32 }, 1164 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_32 }, 1165 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_32 }, 1166 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_32 }, 1167 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_32 }, 1168 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_32 }, 1169 { X86::VFNMADDPSr132r_Int, X86::VFNMADDPSr132m_Int, TB_ALIGN_16 }, 1170 { X86::VFNMADDPDr132r_Int, X86::VFNMADDPDr132m_Int, TB_ALIGN_16 }, 1171 { X86::VFNMADDPSr132rY_Int, X86::VFNMADDPSr132mY_Int, TB_ALIGN_32 }, 1172 { X86::VFNMADDPDr132rY_Int, X86::VFNMADDPDr132mY_Int, TB_ALIGN_32 }, 1173 1174 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, 0 }, 1175 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, 0 }, 1176 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, 0 }, 1177 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, 0 }, 1178 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, 0 }, 1179 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, 0 }, 1180 { X86::VFMSUBSSr132r_Int, X86::VFMSUBSSr132m_Int, 0 }, 1181 { X86::VFMSUBSDr132r_Int, X86::VFMSUBSDr132m_Int, 0 }, 1182 1183 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_16 }, 1184 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_16 }, 1185 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_16 }, 1186 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_16 }, 1187 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_16 }, 1188 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_16 }, 1189 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_32 }, 1190 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_32 }, 1191 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_32 }, 1192 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_32 }, 1193 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_32 }, 1194 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_32 }, 1195 { X86::VFMSUBPSr132r_Int, X86::VFMSUBPSr132m_Int, TB_ALIGN_16 }, 1196 { X86::VFMSUBPDr132r_Int, X86::VFMSUBPDr132m_Int, TB_ALIGN_16 }, 1197 { X86::VFMSUBPSr132rY_Int, X86::VFMSUBPSr132mY_Int, TB_ALIGN_32 }, 1198 { X86::VFMSUBPDr132rY_Int, X86::VFMSUBPDr132mY_Int, TB_ALIGN_32 }, 1199 1200 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, 0 }, 1201 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, 0 }, 1202 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, 0 }, 1203 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, 0 }, 1204 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, 0 }, 1205 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, 0 }, 1206 { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr132m_Int, 0 }, 1207 { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr132m_Int, 0 }, 1208 1209 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_16 }, 1210 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_16 }, 1211 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_16 }, 1212 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_16 }, 1213 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_16 }, 1214 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_16 }, 1215 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_32 }, 1216 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_32 }, 1217 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_32 }, 1218 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_32 }, 1219 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_32 }, 1220 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_32 }, 1221 { X86::VFNMSUBPSr132r_Int, X86::VFNMSUBPSr132m_Int, TB_ALIGN_16 }, 1222 { X86::VFNMSUBPDr132r_Int, X86::VFNMSUBPDr132m_Int, TB_ALIGN_16 }, 1223 { X86::VFNMSUBPSr132rY_Int, X86::VFNMSUBPSr132mY_Int, TB_ALIGN_32 }, 1224 { X86::VFNMSUBPDr132rY_Int, X86::VFNMSUBPDr132mY_Int, TB_ALIGN_32 }, 1225 1226 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_16 }, 1227 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_16 }, 1228 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_16 }, 1229 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_16 }, 1230 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_16 }, 1231 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_16 }, 1232 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_32 }, 1233 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_32 }, 1234 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_32 }, 1235 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_32 }, 1236 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_32 }, 1237 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_32 }, 1238 { X86::VFMADDSUBPSr132r_Int, X86::VFMADDSUBPSr132m_Int, TB_ALIGN_16 }, 1239 { X86::VFMADDSUBPDr132r_Int, X86::VFMADDSUBPDr132m_Int, TB_ALIGN_16 }, 1240 { X86::VFMADDSUBPSr132rY_Int, X86::VFMADDSUBPSr132mY_Int, TB_ALIGN_32 }, 1241 { X86::VFMADDSUBPDr132rY_Int, X86::VFMADDSUBPDr132mY_Int, TB_ALIGN_32 }, 1242 1243 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_16 }, 1244 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_16 }, 1245 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_16 }, 1246 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_16 }, 1247 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_16 }, 1248 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_16 }, 1249 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_32 }, 1250 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_32 }, 1251 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_32 }, 1252 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_32 }, 1253 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_32 }, 1254 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_32 }, 1255 { X86::VFMSUBADDPSr132r_Int, X86::VFMSUBADDPSr132m_Int, TB_ALIGN_16 }, 1256 { X86::VFMSUBADDPDr132r_Int, X86::VFMSUBADDPDr132m_Int, TB_ALIGN_16 }, 1257 { X86::VFMSUBADDPSr132rY_Int, X86::VFMSUBADDPSr132mY_Int, TB_ALIGN_32 }, 1258 { X86::VFMSUBADDPDr132rY_Int, X86::VFMSUBADDPDr132mY_Int, TB_ALIGN_32 }, 1259 }; 1260 1261 for (unsigned i = 0, e = array_lengthof(OpTbl3); i != e; ++i) { 1262 unsigned RegOp = OpTbl3[i].RegOp; 1263 unsigned MemOp = OpTbl3[i].MemOp; 1264 unsigned Flags = OpTbl3[i].Flags; 1265 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable, 1266 RegOp, MemOp, 1267 // Index 3, folded load 1268 Flags | TB_INDEX_3 | TB_FOLDED_LOAD); 1269 } 1270 1271} 1272 1273void 1274X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable, 1275 MemOp2RegOpTableType &M2RTable, 1276 unsigned RegOp, unsigned MemOp, unsigned Flags) { 1277 if ((Flags & TB_NO_FORWARD) == 0) { 1278 assert(!R2MTable.count(RegOp) && "Duplicate entry!"); 1279 R2MTable[RegOp] = std::make_pair(MemOp, Flags); 1280 } 1281 if ((Flags & TB_NO_REVERSE) == 0) { 1282 assert(!M2RTable.count(MemOp) && 1283 "Duplicated entries in unfolding maps?"); 1284 M2RTable[MemOp] = std::make_pair(RegOp, Flags); 1285 } 1286} 1287 1288bool 1289X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI, 1290 unsigned &SrcReg, unsigned &DstReg, 1291 unsigned &SubIdx) const { 1292 switch (MI.getOpcode()) { 1293 default: break; 1294 case X86::MOVSX16rr8: 1295 case X86::MOVZX16rr8: 1296 case X86::MOVSX32rr8: 1297 case X86::MOVZX32rr8: 1298 case X86::MOVSX64rr8: 1299 case X86::MOVZX64rr8: 1300 if (!TM.getSubtarget<X86Subtarget>().is64Bit()) 1301 // It's not always legal to reference the low 8-bit of the larger 1302 // register in 32-bit mode. 1303 return false; 1304 case X86::MOVSX32rr16: 1305 case X86::MOVZX32rr16: 1306 case X86::MOVSX64rr16: 1307 case X86::MOVZX64rr16: 1308 case X86::MOVSX64rr32: 1309 case X86::MOVZX64rr32: { 1310 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg()) 1311 // Be conservative. 1312 return false; 1313 SrcReg = MI.getOperand(1).getReg(); 1314 DstReg = MI.getOperand(0).getReg(); 1315 switch (MI.getOpcode()) { 1316 default: 1317 llvm_unreachable(0); 1318 case X86::MOVSX16rr8: 1319 case X86::MOVZX16rr8: 1320 case X86::MOVSX32rr8: 1321 case X86::MOVZX32rr8: 1322 case X86::MOVSX64rr8: 1323 case X86::MOVZX64rr8: 1324 SubIdx = X86::sub_8bit; 1325 break; 1326 case X86::MOVSX32rr16: 1327 case X86::MOVZX32rr16: 1328 case X86::MOVSX64rr16: 1329 case X86::MOVZX64rr16: 1330 SubIdx = X86::sub_16bit; 1331 break; 1332 case X86::MOVSX64rr32: 1333 case X86::MOVZX64rr32: 1334 SubIdx = X86::sub_32bit; 1335 break; 1336 } 1337 return true; 1338 } 1339 } 1340 return false; 1341} 1342 1343/// isFrameOperand - Return true and the FrameIndex if the specified 1344/// operand and follow operands form a reference to the stack frame. 1345bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op, 1346 int &FrameIndex) const { 1347 if (MI->getOperand(Op).isFI() && MI->getOperand(Op+1).isImm() && 1348 MI->getOperand(Op+2).isReg() && MI->getOperand(Op+3).isImm() && 1349 MI->getOperand(Op+1).getImm() == 1 && 1350 MI->getOperand(Op+2).getReg() == 0 && 1351 MI->getOperand(Op+3).getImm() == 0) { 1352 FrameIndex = MI->getOperand(Op).getIndex(); 1353 return true; 1354 } 1355 return false; 1356} 1357 1358static bool isFrameLoadOpcode(int Opcode) { 1359 switch (Opcode) { 1360 default: 1361 return false; 1362 case X86::MOV8rm: 1363 case X86::MOV16rm: 1364 case X86::MOV32rm: 1365 case X86::MOV64rm: 1366 case X86::LD_Fp64m: 1367 case X86::MOVSSrm: 1368 case X86::MOVSDrm: 1369 case X86::MOVAPSrm: 1370 case X86::MOVAPDrm: 1371 case X86::MOVDQArm: 1372 case X86::VMOVSSrm: 1373 case X86::VMOVSDrm: 1374 case X86::VMOVAPSrm: 1375 case X86::VMOVAPDrm: 1376 case X86::VMOVDQArm: 1377 case X86::VMOVAPSYrm: 1378 case X86::VMOVAPDYrm: 1379 case X86::VMOVDQAYrm: 1380 case X86::MMX_MOVD64rm: 1381 case X86::MMX_MOVQ64rm: 1382 return true; 1383 } 1384} 1385 1386static bool isFrameStoreOpcode(int Opcode) { 1387 switch (Opcode) { 1388 default: break; 1389 case X86::MOV8mr: 1390 case X86::MOV16mr: 1391 case X86::MOV32mr: 1392 case X86::MOV64mr: 1393 case X86::ST_FpP64m: 1394 case X86::MOVSSmr: 1395 case X86::MOVSDmr: 1396 case X86::MOVAPSmr: 1397 case X86::MOVAPDmr: 1398 case X86::MOVDQAmr: 1399 case X86::VMOVSSmr: 1400 case X86::VMOVSDmr: 1401 case X86::VMOVAPSmr: 1402 case X86::VMOVAPDmr: 1403 case X86::VMOVDQAmr: 1404 case X86::VMOVAPSYmr: 1405 case X86::VMOVAPDYmr: 1406 case X86::VMOVDQAYmr: 1407 case X86::MMX_MOVD64mr: 1408 case X86::MMX_MOVQ64mr: 1409 case X86::MMX_MOVNTQmr: 1410 return true; 1411 } 1412 return false; 1413} 1414 1415unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 1416 int &FrameIndex) const { 1417 if (isFrameLoadOpcode(MI->getOpcode())) 1418 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex)) 1419 return MI->getOperand(0).getReg(); 1420 return 0; 1421} 1422 1423unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI, 1424 int &FrameIndex) const { 1425 if (isFrameLoadOpcode(MI->getOpcode())) { 1426 unsigned Reg; 1427 if ((Reg = isLoadFromStackSlot(MI, FrameIndex))) 1428 return Reg; 1429 // Check for post-frame index elimination operations 1430 const MachineMemOperand *Dummy; 1431 return hasLoadFromStackSlot(MI, Dummy, FrameIndex); 1432 } 1433 return 0; 1434} 1435 1436unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI, 1437 int &FrameIndex) const { 1438 if (isFrameStoreOpcode(MI->getOpcode())) 1439 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 && 1440 isFrameOperand(MI, 0, FrameIndex)) 1441 return MI->getOperand(X86::AddrNumOperands).getReg(); 1442 return 0; 1443} 1444 1445unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI, 1446 int &FrameIndex) const { 1447 if (isFrameStoreOpcode(MI->getOpcode())) { 1448 unsigned Reg; 1449 if ((Reg = isStoreToStackSlot(MI, FrameIndex))) 1450 return Reg; 1451 // Check for post-frame index elimination operations 1452 const MachineMemOperand *Dummy; 1453 return hasStoreToStackSlot(MI, Dummy, FrameIndex); 1454 } 1455 return 0; 1456} 1457 1458/// regIsPICBase - Return true if register is PIC base (i.e.g defined by 1459/// X86::MOVPC32r. 1460static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) { 1461 bool isPICBase = false; 1462 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg), 1463 E = MRI.def_end(); I != E; ++I) { 1464 MachineInstr *DefMI = I.getOperand().getParent(); 1465 if (DefMI->getOpcode() != X86::MOVPC32r) 1466 return false; 1467 assert(!isPICBase && "More than one PIC base?"); 1468 isPICBase = true; 1469 } 1470 return isPICBase; 1471} 1472 1473bool 1474X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI, 1475 AliasAnalysis *AA) const { 1476 switch (MI->getOpcode()) { 1477 default: break; 1478 case X86::MOV8rm: 1479 case X86::MOV16rm: 1480 case X86::MOV32rm: 1481 case X86::MOV64rm: 1482 case X86::LD_Fp64m: 1483 case X86::MOVSSrm: 1484 case X86::MOVSDrm: 1485 case X86::MOVAPSrm: 1486 case X86::MOVUPSrm: 1487 case X86::MOVAPDrm: 1488 case X86::MOVDQArm: 1489 case X86::VMOVSSrm: 1490 case X86::VMOVSDrm: 1491 case X86::VMOVAPSrm: 1492 case X86::VMOVUPSrm: 1493 case X86::VMOVAPDrm: 1494 case X86::VMOVDQArm: 1495 case X86::VMOVAPSYrm: 1496 case X86::VMOVUPSYrm: 1497 case X86::VMOVAPDYrm: 1498 case X86::VMOVDQAYrm: 1499 case X86::MMX_MOVD64rm: 1500 case X86::MMX_MOVQ64rm: 1501 case X86::FsVMOVAPSrm: 1502 case X86::FsVMOVAPDrm: 1503 case X86::FsMOVAPSrm: 1504 case X86::FsMOVAPDrm: { 1505 // Loads from constant pools are trivially rematerializable. 1506 if (MI->getOperand(1).isReg() && 1507 MI->getOperand(2).isImm() && 1508 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 && 1509 MI->isInvariantLoad(AA)) { 1510 unsigned BaseReg = MI->getOperand(1).getReg(); 1511 if (BaseReg == 0 || BaseReg == X86::RIP) 1512 return true; 1513 // Allow re-materialization of PIC load. 1514 if (!ReMatPICStubLoad && MI->getOperand(4).isGlobal()) 1515 return false; 1516 const MachineFunction &MF = *MI->getParent()->getParent(); 1517 const MachineRegisterInfo &MRI = MF.getRegInfo(); 1518 bool isPICBase = false; 1519 for (MachineRegisterInfo::def_iterator I = MRI.def_begin(BaseReg), 1520 E = MRI.def_end(); I != E; ++I) { 1521 MachineInstr *DefMI = I.getOperand().getParent(); 1522 if (DefMI->getOpcode() != X86::MOVPC32r) 1523 return false; 1524 assert(!isPICBase && "More than one PIC base?"); 1525 isPICBase = true; 1526 } 1527 return isPICBase; 1528 } 1529 return false; 1530 } 1531 1532 case X86::LEA32r: 1533 case X86::LEA64r: { 1534 if (MI->getOperand(2).isImm() && 1535 MI->getOperand(3).isReg() && MI->getOperand(3).getReg() == 0 && 1536 !MI->getOperand(4).isReg()) { 1537 // lea fi#, lea GV, etc. are all rematerializable. 1538 if (!MI->getOperand(1).isReg()) 1539 return true; 1540 unsigned BaseReg = MI->getOperand(1).getReg(); 1541 if (BaseReg == 0) 1542 return true; 1543 // Allow re-materialization of lea PICBase + x. 1544 const MachineFunction &MF = *MI->getParent()->getParent(); 1545 const MachineRegisterInfo &MRI = MF.getRegInfo(); 1546 return regIsPICBase(BaseReg, MRI); 1547 } 1548 return false; 1549 } 1550 } 1551 1552 // All other instructions marked M_REMATERIALIZABLE are always trivially 1553 // rematerializable. 1554 return true; 1555} 1556 1557/// isSafeToClobberEFLAGS - Return true if it's safe insert an instruction that 1558/// would clobber the EFLAGS condition register. Note the result may be 1559/// conservative. If it cannot definitely determine the safety after visiting 1560/// a few instructions in each direction it assumes it's not safe. 1561static bool isSafeToClobberEFLAGS(MachineBasicBlock &MBB, 1562 MachineBasicBlock::iterator I) { 1563 MachineBasicBlock::iterator E = MBB.end(); 1564 1565 // For compile time consideration, if we are not able to determine the 1566 // safety after visiting 4 instructions in each direction, we will assume 1567 // it's not safe. 1568 MachineBasicBlock::iterator Iter = I; 1569 for (unsigned i = 0; Iter != E && i < 4; ++i) { 1570 bool SeenDef = false; 1571 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 1572 MachineOperand &MO = Iter->getOperand(j); 1573 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 1574 SeenDef = true; 1575 if (!MO.isReg()) 1576 continue; 1577 if (MO.getReg() == X86::EFLAGS) { 1578 if (MO.isUse()) 1579 return false; 1580 SeenDef = true; 1581 } 1582 } 1583 1584 if (SeenDef) 1585 // This instruction defines EFLAGS, no need to look any further. 1586 return true; 1587 ++Iter; 1588 // Skip over DBG_VALUE. 1589 while (Iter != E && Iter->isDebugValue()) 1590 ++Iter; 1591 } 1592 1593 // It is safe to clobber EFLAGS at the end of a block of no successor has it 1594 // live in. 1595 if (Iter == E) { 1596 for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(), 1597 SE = MBB.succ_end(); SI != SE; ++SI) 1598 if ((*SI)->isLiveIn(X86::EFLAGS)) 1599 return false; 1600 return true; 1601 } 1602 1603 MachineBasicBlock::iterator B = MBB.begin(); 1604 Iter = I; 1605 for (unsigned i = 0; i < 4; ++i) { 1606 // If we make it to the beginning of the block, it's safe to clobber 1607 // EFLAGS iff EFLAGS is not live-in. 1608 if (Iter == B) 1609 return !MBB.isLiveIn(X86::EFLAGS); 1610 1611 --Iter; 1612 // Skip over DBG_VALUE. 1613 while (Iter != B && Iter->isDebugValue()) 1614 --Iter; 1615 1616 bool SawKill = false; 1617 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 1618 MachineOperand &MO = Iter->getOperand(j); 1619 // A register mask may clobber EFLAGS, but we should still look for a 1620 // live EFLAGS def. 1621 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 1622 SawKill = true; 1623 if (MO.isReg() && MO.getReg() == X86::EFLAGS) { 1624 if (MO.isDef()) return MO.isDead(); 1625 if (MO.isKill()) SawKill = true; 1626 } 1627 } 1628 1629 if (SawKill) 1630 // This instruction kills EFLAGS and doesn't redefine it, so 1631 // there's no need to look further. 1632 return true; 1633 } 1634 1635 // Conservative answer. 1636 return false; 1637} 1638 1639void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB, 1640 MachineBasicBlock::iterator I, 1641 unsigned DestReg, unsigned SubIdx, 1642 const MachineInstr *Orig, 1643 const TargetRegisterInfo &TRI) const { 1644 DebugLoc DL = Orig->getDebugLoc(); 1645 1646 // MOV32r0 etc. are implemented with xor which clobbers condition code. 1647 // Re-materialize them as movri instructions to avoid side effects. 1648 bool Clone = true; 1649 unsigned Opc = Orig->getOpcode(); 1650 switch (Opc) { 1651 default: break; 1652 case X86::MOV8r0: 1653 case X86::MOV16r0: 1654 case X86::MOV32r0: 1655 case X86::MOV64r0: { 1656 if (!isSafeToClobberEFLAGS(MBB, I)) { 1657 switch (Opc) { 1658 default: break; 1659 case X86::MOV8r0: Opc = X86::MOV8ri; break; 1660 case X86::MOV16r0: Opc = X86::MOV16ri; break; 1661 case X86::MOV32r0: Opc = X86::MOV32ri; break; 1662 case X86::MOV64r0: Opc = X86::MOV64ri64i32; break; 1663 } 1664 Clone = false; 1665 } 1666 break; 1667 } 1668 } 1669 1670 if (Clone) { 1671 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 1672 MBB.insert(I, MI); 1673 } else { 1674 BuildMI(MBB, I, DL, get(Opc)).addOperand(Orig->getOperand(0)).addImm(0); 1675 } 1676 1677 MachineInstr *NewMI = prior(I); 1678 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI); 1679} 1680 1681/// hasLiveCondCodeDef - True if MI has a condition code def, e.g. EFLAGS, that 1682/// is not marked dead. 1683static bool hasLiveCondCodeDef(MachineInstr *MI) { 1684 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1685 MachineOperand &MO = MI->getOperand(i); 1686 if (MO.isReg() && MO.isDef() && 1687 MO.getReg() == X86::EFLAGS && !MO.isDead()) { 1688 return true; 1689 } 1690 } 1691 return false; 1692} 1693 1694/// convertToThreeAddressWithLEA - Helper for convertToThreeAddress when 1695/// 16-bit LEA is disabled, use 32-bit LEA to form 3-address code by promoting 1696/// to a 32-bit superregister and then truncating back down to a 16-bit 1697/// subregister. 1698MachineInstr * 1699X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc, 1700 MachineFunction::iterator &MFI, 1701 MachineBasicBlock::iterator &MBBI, 1702 LiveVariables *LV) const { 1703 MachineInstr *MI = MBBI; 1704 unsigned Dest = MI->getOperand(0).getReg(); 1705 unsigned Src = MI->getOperand(1).getReg(); 1706 bool isDead = MI->getOperand(0).isDead(); 1707 bool isKill = MI->getOperand(1).isKill(); 1708 1709 unsigned Opc = TM.getSubtarget<X86Subtarget>().is64Bit() 1710 ? X86::LEA64_32r : X86::LEA32r; 1711 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo(); 1712 unsigned leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 1713 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass); 1714 1715 // Build and insert into an implicit UNDEF value. This is OK because 1716 // well be shifting and then extracting the lower 16-bits. 1717 // This has the potential to cause partial register stall. e.g. 1718 // movw (%rbp,%rcx,2), %dx 1719 // leal -65(%rdx), %esi 1720 // But testing has shown this *does* help performance in 64-bit mode (at 1721 // least on modern x86 machines). 1722 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg); 1723 MachineInstr *InsMI = 1724 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY)) 1725 .addReg(leaInReg, RegState::Define, X86::sub_16bit) 1726 .addReg(Src, getKillRegState(isKill)); 1727 1728 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(), 1729 get(Opc), leaOutReg); 1730 switch (MIOpc) { 1731 default: 1732 llvm_unreachable(0); 1733 case X86::SHL16ri: { 1734 unsigned ShAmt = MI->getOperand(2).getImm(); 1735 MIB.addReg(0).addImm(1 << ShAmt) 1736 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0); 1737 break; 1738 } 1739 case X86::INC16r: 1740 case X86::INC64_16r: 1741 addRegOffset(MIB, leaInReg, true, 1); 1742 break; 1743 case X86::DEC16r: 1744 case X86::DEC64_16r: 1745 addRegOffset(MIB, leaInReg, true, -1); 1746 break; 1747 case X86::ADD16ri: 1748 case X86::ADD16ri8: 1749 case X86::ADD16ri_DB: 1750 case X86::ADD16ri8_DB: 1751 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm()); 1752 break; 1753 case X86::ADD16rr: 1754 case X86::ADD16rr_DB: { 1755 unsigned Src2 = MI->getOperand(2).getReg(); 1756 bool isKill2 = MI->getOperand(2).isKill(); 1757 unsigned leaInReg2 = 0; 1758 MachineInstr *InsMI2 = 0; 1759 if (Src == Src2) { 1760 // ADD16rr %reg1028<kill>, %reg1028 1761 // just a single insert_subreg. 1762 addRegReg(MIB, leaInReg, true, leaInReg, false); 1763 } else { 1764 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 1765 // Build and insert into an implicit UNDEF value. This is OK because 1766 // well be shifting and then extracting the lower 16-bits. 1767 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2); 1768 InsMI2 = 1769 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY)) 1770 .addReg(leaInReg2, RegState::Define, X86::sub_16bit) 1771 .addReg(Src2, getKillRegState(isKill2)); 1772 addRegReg(MIB, leaInReg, true, leaInReg2, true); 1773 } 1774 if (LV && isKill2 && InsMI2) 1775 LV->replaceKillInstruction(Src2, MI, InsMI2); 1776 break; 1777 } 1778 } 1779 1780 MachineInstr *NewMI = MIB; 1781 MachineInstr *ExtMI = 1782 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY)) 1783 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 1784 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit); 1785 1786 if (LV) { 1787 // Update live variables 1788 LV->getVarInfo(leaInReg).Kills.push_back(NewMI); 1789 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI); 1790 if (isKill) 1791 LV->replaceKillInstruction(Src, MI, InsMI); 1792 if (isDead) 1793 LV->replaceKillInstruction(Dest, MI, ExtMI); 1794 } 1795 1796 return ExtMI; 1797} 1798 1799/// convertToThreeAddress - This method must be implemented by targets that 1800/// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target 1801/// may be able to convert a two-address instruction into a true 1802/// three-address instruction on demand. This allows the X86 target (for 1803/// example) to convert ADD and SHL instructions into LEA instructions if they 1804/// would require register copies due to two-addressness. 1805/// 1806/// This method returns a null pointer if the transformation cannot be 1807/// performed, otherwise it returns the new instruction. 1808/// 1809MachineInstr * 1810X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI, 1811 MachineBasicBlock::iterator &MBBI, 1812 LiveVariables *LV) const { 1813 MachineInstr *MI = MBBI; 1814 MachineFunction &MF = *MI->getParent()->getParent(); 1815 // All instructions input are two-addr instructions. Get the known operands. 1816 unsigned Dest = MI->getOperand(0).getReg(); 1817 unsigned Src = MI->getOperand(1).getReg(); 1818 bool isDead = MI->getOperand(0).isDead(); 1819 bool isKill = MI->getOperand(1).isKill(); 1820 1821 MachineInstr *NewMI = NULL; 1822 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When 1823 // we have better subtarget support, enable the 16-bit LEA generation here. 1824 // 16-bit LEA is also slow on Core2. 1825 bool DisableLEA16 = true; 1826 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit(); 1827 1828 unsigned MIOpc = MI->getOpcode(); 1829 switch (MIOpc) { 1830 case X86::SHUFPSrri: { 1831 assert(MI->getNumOperands() == 4 && "Unknown shufps instruction!"); 1832 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0; 1833 1834 unsigned B = MI->getOperand(1).getReg(); 1835 unsigned C = MI->getOperand(2).getReg(); 1836 if (B != C) return 0; 1837 unsigned A = MI->getOperand(0).getReg(); 1838 unsigned M = MI->getOperand(3).getImm(); 1839 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri)) 1840 .addReg(A, RegState::Define | getDeadRegState(isDead)) 1841 .addReg(B, getKillRegState(isKill)).addImm(M); 1842 break; 1843 } 1844 case X86::SHUFPDrri: { 1845 assert(MI->getNumOperands() == 4 && "Unknown shufpd instruction!"); 1846 if (!TM.getSubtarget<X86Subtarget>().hasSSE2()) return 0; 1847 1848 unsigned B = MI->getOperand(1).getReg(); 1849 unsigned C = MI->getOperand(2).getReg(); 1850 if (B != C) return 0; 1851 unsigned A = MI->getOperand(0).getReg(); 1852 unsigned M = MI->getOperand(3).getImm(); 1853 1854 // Convert to PSHUFD mask. 1855 M = ((M & 1) << 1) | ((M & 1) << 3) | ((M & 2) << 4) | ((M & 2) << 6)| 0x44; 1856 1857 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::PSHUFDri)) 1858 .addReg(A, RegState::Define | getDeadRegState(isDead)) 1859 .addReg(B, getKillRegState(isKill)).addImm(M); 1860 break; 1861 } 1862 case X86::SHL64ri: { 1863 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 1864 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses 1865 // the flags produced by a shift yet, so this is safe. 1866 unsigned ShAmt = MI->getOperand(2).getImm(); 1867 if (ShAmt == 0 || ShAmt >= 4) return 0; 1868 1869 // LEA can't handle RSP. 1870 if (TargetRegisterInfo::isVirtualRegister(Src) && 1871 !MF.getRegInfo().constrainRegClass(Src, &X86::GR64_NOSPRegClass)) 1872 return 0; 1873 1874 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r)) 1875 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 1876 .addReg(0).addImm(1 << ShAmt) 1877 .addReg(Src, getKillRegState(isKill)) 1878 .addImm(0).addReg(0); 1879 break; 1880 } 1881 case X86::SHL32ri: { 1882 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 1883 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses 1884 // the flags produced by a shift yet, so this is safe. 1885 unsigned ShAmt = MI->getOperand(2).getImm(); 1886 if (ShAmt == 0 || ShAmt >= 4) return 0; 1887 1888 // LEA can't handle ESP. 1889 if (TargetRegisterInfo::isVirtualRegister(Src) && 1890 !MF.getRegInfo().constrainRegClass(Src, &X86::GR32_NOSPRegClass)) 1891 return 0; 1892 1893 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 1894 NewMI = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 1895 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 1896 .addReg(0).addImm(1 << ShAmt) 1897 .addReg(Src, getKillRegState(isKill)).addImm(0).addReg(0); 1898 break; 1899 } 1900 case X86::SHL16ri: { 1901 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 1902 // NOTE: LEA doesn't produce flags like shift does, but LLVM never uses 1903 // the flags produced by a shift yet, so this is safe. 1904 unsigned ShAmt = MI->getOperand(2).getImm(); 1905 if (ShAmt == 0 || ShAmt >= 4) return 0; 1906 1907 if (DisableLEA16) 1908 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0; 1909 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 1910 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 1911 .addReg(0).addImm(1 << ShAmt) 1912 .addReg(Src, getKillRegState(isKill)) 1913 .addImm(0).addReg(0); 1914 break; 1915 } 1916 default: { 1917 // The following opcodes also sets the condition code register(s). Only 1918 // convert them to equivalent lea if the condition code register def's 1919 // are dead! 1920 if (hasLiveCondCodeDef(MI)) 1921 return 0; 1922 1923 switch (MIOpc) { 1924 default: return 0; 1925 case X86::INC64r: 1926 case X86::INC32r: 1927 case X86::INC64_32r: { 1928 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!"); 1929 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r 1930 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 1931 const TargetRegisterClass *RC = MIOpc == X86::INC64r ? 1932 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass : 1933 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass; 1934 1935 // LEA can't handle RSP. 1936 if (TargetRegisterInfo::isVirtualRegister(Src) && 1937 !MF.getRegInfo().constrainRegClass(Src, RC)) 1938 return 0; 1939 1940 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc)) 1941 .addReg(Dest, RegState::Define | 1942 getDeadRegState(isDead)), 1943 Src, isKill, 1); 1944 break; 1945 } 1946 case X86::INC16r: 1947 case X86::INC64_16r: 1948 if (DisableLEA16) 1949 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0; 1950 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!"); 1951 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 1952 .addReg(Dest, RegState::Define | 1953 getDeadRegState(isDead)), 1954 Src, isKill, 1); 1955 break; 1956 case X86::DEC64r: 1957 case X86::DEC32r: 1958 case X86::DEC64_32r: { 1959 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!"); 1960 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r 1961 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 1962 const TargetRegisterClass *RC = MIOpc == X86::DEC64r ? 1963 (const TargetRegisterClass*)&X86::GR64_NOSPRegClass : 1964 (const TargetRegisterClass*)&X86::GR32_NOSPRegClass; 1965 // LEA can't handle RSP. 1966 if (TargetRegisterInfo::isVirtualRegister(Src) && 1967 !MF.getRegInfo().constrainRegClass(Src, RC)) 1968 return 0; 1969 1970 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc)) 1971 .addReg(Dest, RegState::Define | 1972 getDeadRegState(isDead)), 1973 Src, isKill, -1); 1974 break; 1975 } 1976 case X86::DEC16r: 1977 case X86::DEC64_16r: 1978 if (DisableLEA16) 1979 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0; 1980 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!"); 1981 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 1982 .addReg(Dest, RegState::Define | 1983 getDeadRegState(isDead)), 1984 Src, isKill, -1); 1985 break; 1986 case X86::ADD64rr: 1987 case X86::ADD64rr_DB: 1988 case X86::ADD32rr: 1989 case X86::ADD32rr_DB: { 1990 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 1991 unsigned Opc; 1992 const TargetRegisterClass *RC; 1993 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) { 1994 Opc = X86::LEA64r; 1995 RC = &X86::GR64_NOSPRegClass; 1996 } else { 1997 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 1998 RC = &X86::GR32_NOSPRegClass; 1999 } 2000 2001 2002 unsigned Src2 = MI->getOperand(2).getReg(); 2003 bool isKill2 = MI->getOperand(2).isKill(); 2004 2005 // LEA can't handle RSP. 2006 if (TargetRegisterInfo::isVirtualRegister(Src2) && 2007 !MF.getRegInfo().constrainRegClass(Src2, RC)) 2008 return 0; 2009 2010 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2011 .addReg(Dest, RegState::Define | 2012 getDeadRegState(isDead)), 2013 Src, isKill, Src2, isKill2); 2014 if (LV && isKill2) 2015 LV->replaceKillInstruction(Src2, MI, NewMI); 2016 break; 2017 } 2018 case X86::ADD16rr: 2019 case X86::ADD16rr_DB: { 2020 if (DisableLEA16) 2021 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0; 2022 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2023 unsigned Src2 = MI->getOperand(2).getReg(); 2024 bool isKill2 = MI->getOperand(2).isKill(); 2025 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2026 .addReg(Dest, RegState::Define | 2027 getDeadRegState(isDead)), 2028 Src, isKill, Src2, isKill2); 2029 if (LV && isKill2) 2030 LV->replaceKillInstruction(Src2, MI, NewMI); 2031 break; 2032 } 2033 case X86::ADD64ri32: 2034 case X86::ADD64ri8: 2035 case X86::ADD64ri32_DB: 2036 case X86::ADD64ri8_DB: 2037 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2038 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r)) 2039 .addReg(Dest, RegState::Define | 2040 getDeadRegState(isDead)), 2041 Src, isKill, MI->getOperand(2).getImm()); 2042 break; 2043 case X86::ADD32ri: 2044 case X86::ADD32ri8: 2045 case X86::ADD32ri_DB: 2046 case X86::ADD32ri8_DB: { 2047 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2048 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 2049 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2050 .addReg(Dest, RegState::Define | 2051 getDeadRegState(isDead)), 2052 Src, isKill, MI->getOperand(2).getImm()); 2053 break; 2054 } 2055 case X86::ADD16ri: 2056 case X86::ADD16ri8: 2057 case X86::ADD16ri_DB: 2058 case X86::ADD16ri8_DB: 2059 if (DisableLEA16) 2060 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : 0; 2061 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2062 NewMI = addRegOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2063 .addReg(Dest, RegState::Define | 2064 getDeadRegState(isDead)), 2065 Src, isKill, MI->getOperand(2).getImm()); 2066 break; 2067 } 2068 } 2069 } 2070 2071 if (!NewMI) return 0; 2072 2073 if (LV) { // Update live variables 2074 if (isKill) 2075 LV->replaceKillInstruction(Src, MI, NewMI); 2076 if (isDead) 2077 LV->replaceKillInstruction(Dest, MI, NewMI); 2078 } 2079 2080 MFI->insert(MBBI, NewMI); // Insert the new inst 2081 return NewMI; 2082} 2083 2084/// commuteInstruction - We have a few instructions that must be hacked on to 2085/// commute them. 2086/// 2087MachineInstr * 2088X86InstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const { 2089 switch (MI->getOpcode()) { 2090 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I) 2091 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I) 2092 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I) 2093 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I) 2094 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I) 2095 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I) 2096 unsigned Opc; 2097 unsigned Size; 2098 switch (MI->getOpcode()) { 2099 default: llvm_unreachable("Unreachable!"); 2100 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break; 2101 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break; 2102 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break; 2103 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break; 2104 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break; 2105 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break; 2106 } 2107 unsigned Amt = MI->getOperand(3).getImm(); 2108 if (NewMI) { 2109 MachineFunction &MF = *MI->getParent()->getParent(); 2110 MI = MF.CloneMachineInstr(MI); 2111 NewMI = false; 2112 } 2113 MI->setDesc(get(Opc)); 2114 MI->getOperand(3).setImm(Size-Amt); 2115 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI); 2116 } 2117 case X86::CMOVB16rr: 2118 case X86::CMOVB32rr: 2119 case X86::CMOVB64rr: 2120 case X86::CMOVAE16rr: 2121 case X86::CMOVAE32rr: 2122 case X86::CMOVAE64rr: 2123 case X86::CMOVE16rr: 2124 case X86::CMOVE32rr: 2125 case X86::CMOVE64rr: 2126 case X86::CMOVNE16rr: 2127 case X86::CMOVNE32rr: 2128 case X86::CMOVNE64rr: 2129 case X86::CMOVBE16rr: 2130 case X86::CMOVBE32rr: 2131 case X86::CMOVBE64rr: 2132 case X86::CMOVA16rr: 2133 case X86::CMOVA32rr: 2134 case X86::CMOVA64rr: 2135 case X86::CMOVL16rr: 2136 case X86::CMOVL32rr: 2137 case X86::CMOVL64rr: 2138 case X86::CMOVGE16rr: 2139 case X86::CMOVGE32rr: 2140 case X86::CMOVGE64rr: 2141 case X86::CMOVLE16rr: 2142 case X86::CMOVLE32rr: 2143 case X86::CMOVLE64rr: 2144 case X86::CMOVG16rr: 2145 case X86::CMOVG32rr: 2146 case X86::CMOVG64rr: 2147 case X86::CMOVS16rr: 2148 case X86::CMOVS32rr: 2149 case X86::CMOVS64rr: 2150 case X86::CMOVNS16rr: 2151 case X86::CMOVNS32rr: 2152 case X86::CMOVNS64rr: 2153 case X86::CMOVP16rr: 2154 case X86::CMOVP32rr: 2155 case X86::CMOVP64rr: 2156 case X86::CMOVNP16rr: 2157 case X86::CMOVNP32rr: 2158 case X86::CMOVNP64rr: 2159 case X86::CMOVO16rr: 2160 case X86::CMOVO32rr: 2161 case X86::CMOVO64rr: 2162 case X86::CMOVNO16rr: 2163 case X86::CMOVNO32rr: 2164 case X86::CMOVNO64rr: { 2165 unsigned Opc = 0; 2166 switch (MI->getOpcode()) { 2167 default: break; 2168 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break; 2169 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break; 2170 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break; 2171 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break; 2172 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break; 2173 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break; 2174 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break; 2175 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break; 2176 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break; 2177 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break; 2178 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break; 2179 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break; 2180 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break; 2181 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break; 2182 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break; 2183 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break; 2184 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break; 2185 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break; 2186 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break; 2187 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break; 2188 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break; 2189 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break; 2190 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break; 2191 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break; 2192 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break; 2193 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break; 2194 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break; 2195 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break; 2196 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break; 2197 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break; 2198 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break; 2199 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break; 2200 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break; 2201 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break; 2202 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break; 2203 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break; 2204 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break; 2205 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break; 2206 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break; 2207 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break; 2208 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break; 2209 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break; 2210 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break; 2211 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break; 2212 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break; 2213 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break; 2214 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break; 2215 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break; 2216 } 2217 if (NewMI) { 2218 MachineFunction &MF = *MI->getParent()->getParent(); 2219 MI = MF.CloneMachineInstr(MI); 2220 NewMI = false; 2221 } 2222 MI->setDesc(get(Opc)); 2223 // Fallthrough intended. 2224 } 2225 default: 2226 return TargetInstrInfoImpl::commuteInstruction(MI, NewMI); 2227 } 2228} 2229 2230static X86::CondCode GetCondFromBranchOpc(unsigned BrOpc) { 2231 switch (BrOpc) { 2232 default: return X86::COND_INVALID; 2233 case X86::JE_4: return X86::COND_E; 2234 case X86::JNE_4: return X86::COND_NE; 2235 case X86::JL_4: return X86::COND_L; 2236 case X86::JLE_4: return X86::COND_LE; 2237 case X86::JG_4: return X86::COND_G; 2238 case X86::JGE_4: return X86::COND_GE; 2239 case X86::JB_4: return X86::COND_B; 2240 case X86::JBE_4: return X86::COND_BE; 2241 case X86::JA_4: return X86::COND_A; 2242 case X86::JAE_4: return X86::COND_AE; 2243 case X86::JS_4: return X86::COND_S; 2244 case X86::JNS_4: return X86::COND_NS; 2245 case X86::JP_4: return X86::COND_P; 2246 case X86::JNP_4: return X86::COND_NP; 2247 case X86::JO_4: return X86::COND_O; 2248 case X86::JNO_4: return X86::COND_NO; 2249 } 2250} 2251 2252unsigned X86::GetCondBranchFromCond(X86::CondCode CC) { 2253 switch (CC) { 2254 default: llvm_unreachable("Illegal condition code!"); 2255 case X86::COND_E: return X86::JE_4; 2256 case X86::COND_NE: return X86::JNE_4; 2257 case X86::COND_L: return X86::JL_4; 2258 case X86::COND_LE: return X86::JLE_4; 2259 case X86::COND_G: return X86::JG_4; 2260 case X86::COND_GE: return X86::JGE_4; 2261 case X86::COND_B: return X86::JB_4; 2262 case X86::COND_BE: return X86::JBE_4; 2263 case X86::COND_A: return X86::JA_4; 2264 case X86::COND_AE: return X86::JAE_4; 2265 case X86::COND_S: return X86::JS_4; 2266 case X86::COND_NS: return X86::JNS_4; 2267 case X86::COND_P: return X86::JP_4; 2268 case X86::COND_NP: return X86::JNP_4; 2269 case X86::COND_O: return X86::JO_4; 2270 case X86::COND_NO: return X86::JNO_4; 2271 } 2272} 2273 2274/// GetOppositeBranchCondition - Return the inverse of the specified condition, 2275/// e.g. turning COND_E to COND_NE. 2276X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) { 2277 switch (CC) { 2278 default: llvm_unreachable("Illegal condition code!"); 2279 case X86::COND_E: return X86::COND_NE; 2280 case X86::COND_NE: return X86::COND_E; 2281 case X86::COND_L: return X86::COND_GE; 2282 case X86::COND_LE: return X86::COND_G; 2283 case X86::COND_G: return X86::COND_LE; 2284 case X86::COND_GE: return X86::COND_L; 2285 case X86::COND_B: return X86::COND_AE; 2286 case X86::COND_BE: return X86::COND_A; 2287 case X86::COND_A: return X86::COND_BE; 2288 case X86::COND_AE: return X86::COND_B; 2289 case X86::COND_S: return X86::COND_NS; 2290 case X86::COND_NS: return X86::COND_S; 2291 case X86::COND_P: return X86::COND_NP; 2292 case X86::COND_NP: return X86::COND_P; 2293 case X86::COND_O: return X86::COND_NO; 2294 case X86::COND_NO: return X86::COND_O; 2295 } 2296} 2297 2298bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 2299 if (!MI->isTerminator()) return false; 2300 2301 // Conditional branch is a special case. 2302 if (MI->isBranch() && !MI->isBarrier()) 2303 return true; 2304 if (!MI->isPredicable()) 2305 return true; 2306 return !isPredicated(MI); 2307} 2308 2309bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, 2310 MachineBasicBlock *&TBB, 2311 MachineBasicBlock *&FBB, 2312 SmallVectorImpl<MachineOperand> &Cond, 2313 bool AllowModify) const { 2314 // Start from the bottom of the block and work up, examining the 2315 // terminator instructions. 2316 MachineBasicBlock::iterator I = MBB.end(); 2317 MachineBasicBlock::iterator UnCondBrIter = MBB.end(); 2318 while (I != MBB.begin()) { 2319 --I; 2320 if (I->isDebugValue()) 2321 continue; 2322 2323 // Working from the bottom, when we see a non-terminator instruction, we're 2324 // done. 2325 if (!isUnpredicatedTerminator(I)) 2326 break; 2327 2328 // A terminator that isn't a branch can't easily be handled by this 2329 // analysis. 2330 if (!I->isBranch()) 2331 return true; 2332 2333 // Handle unconditional branches. 2334 if (I->getOpcode() == X86::JMP_4) { 2335 UnCondBrIter = I; 2336 2337 if (!AllowModify) { 2338 TBB = I->getOperand(0).getMBB(); 2339 continue; 2340 } 2341 2342 // If the block has any instructions after a JMP, delete them. 2343 while (llvm::next(I) != MBB.end()) 2344 llvm::next(I)->eraseFromParent(); 2345 2346 Cond.clear(); 2347 FBB = 0; 2348 2349 // Delete the JMP if it's equivalent to a fall-through. 2350 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 2351 TBB = 0; 2352 I->eraseFromParent(); 2353 I = MBB.end(); 2354 UnCondBrIter = MBB.end(); 2355 continue; 2356 } 2357 2358 // TBB is used to indicate the unconditional destination. 2359 TBB = I->getOperand(0).getMBB(); 2360 continue; 2361 } 2362 2363 // Handle conditional branches. 2364 X86::CondCode BranchCode = GetCondFromBranchOpc(I->getOpcode()); 2365 if (BranchCode == X86::COND_INVALID) 2366 return true; // Can't handle indirect branch. 2367 2368 // Working from the bottom, handle the first conditional branch. 2369 if (Cond.empty()) { 2370 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB(); 2371 if (AllowModify && UnCondBrIter != MBB.end() && 2372 MBB.isLayoutSuccessor(TargetBB)) { 2373 // If we can modify the code and it ends in something like: 2374 // 2375 // jCC L1 2376 // jmp L2 2377 // L1: 2378 // ... 2379 // L2: 2380 // 2381 // Then we can change this to: 2382 // 2383 // jnCC L2 2384 // L1: 2385 // ... 2386 // L2: 2387 // 2388 // Which is a bit more efficient. 2389 // We conditionally jump to the fall-through block. 2390 BranchCode = GetOppositeBranchCondition(BranchCode); 2391 unsigned JNCC = GetCondBranchFromCond(BranchCode); 2392 MachineBasicBlock::iterator OldInst = I; 2393 2394 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC)) 2395 .addMBB(UnCondBrIter->getOperand(0).getMBB()); 2396 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_4)) 2397 .addMBB(TargetBB); 2398 2399 OldInst->eraseFromParent(); 2400 UnCondBrIter->eraseFromParent(); 2401 2402 // Restart the analysis. 2403 UnCondBrIter = MBB.end(); 2404 I = MBB.end(); 2405 continue; 2406 } 2407 2408 FBB = TBB; 2409 TBB = I->getOperand(0).getMBB(); 2410 Cond.push_back(MachineOperand::CreateImm(BranchCode)); 2411 continue; 2412 } 2413 2414 // Handle subsequent conditional branches. Only handle the case where all 2415 // conditional branches branch to the same destination and their condition 2416 // opcodes fit one of the special multi-branch idioms. 2417 assert(Cond.size() == 1); 2418 assert(TBB); 2419 2420 // Only handle the case where all conditional branches branch to the same 2421 // destination. 2422 if (TBB != I->getOperand(0).getMBB()) 2423 return true; 2424 2425 // If the conditions are the same, we can leave them alone. 2426 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm(); 2427 if (OldBranchCode == BranchCode) 2428 continue; 2429 2430 // If they differ, see if they fit one of the known patterns. Theoretically, 2431 // we could handle more patterns here, but we shouldn't expect to see them 2432 // if instruction selection has done a reasonable job. 2433 if ((OldBranchCode == X86::COND_NP && 2434 BranchCode == X86::COND_E) || 2435 (OldBranchCode == X86::COND_E && 2436 BranchCode == X86::COND_NP)) 2437 BranchCode = X86::COND_NP_OR_E; 2438 else if ((OldBranchCode == X86::COND_P && 2439 BranchCode == X86::COND_NE) || 2440 (OldBranchCode == X86::COND_NE && 2441 BranchCode == X86::COND_P)) 2442 BranchCode = X86::COND_NE_OR_P; 2443 else 2444 return true; 2445 2446 // Update the MachineOperand. 2447 Cond[0].setImm(BranchCode); 2448 } 2449 2450 return false; 2451} 2452 2453unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 2454 MachineBasicBlock::iterator I = MBB.end(); 2455 unsigned Count = 0; 2456 2457 while (I != MBB.begin()) { 2458 --I; 2459 if (I->isDebugValue()) 2460 continue; 2461 if (I->getOpcode() != X86::JMP_4 && 2462 GetCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID) 2463 break; 2464 // Remove the branch. 2465 I->eraseFromParent(); 2466 I = MBB.end(); 2467 ++Count; 2468 } 2469 2470 return Count; 2471} 2472 2473unsigned 2474X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 2475 MachineBasicBlock *FBB, 2476 const SmallVectorImpl<MachineOperand> &Cond, 2477 DebugLoc DL) const { 2478 // Shouldn't be a fall through. 2479 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 2480 assert((Cond.size() == 1 || Cond.size() == 0) && 2481 "X86 branch conditions have one component!"); 2482 2483 if (Cond.empty()) { 2484 // Unconditional branch? 2485 assert(!FBB && "Unconditional branch with multiple successors!"); 2486 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(TBB); 2487 return 1; 2488 } 2489 2490 // Conditional branch. 2491 unsigned Count = 0; 2492 X86::CondCode CC = (X86::CondCode)Cond[0].getImm(); 2493 switch (CC) { 2494 case X86::COND_NP_OR_E: 2495 // Synthesize NP_OR_E with two branches. 2496 BuildMI(&MBB, DL, get(X86::JNP_4)).addMBB(TBB); 2497 ++Count; 2498 BuildMI(&MBB, DL, get(X86::JE_4)).addMBB(TBB); 2499 ++Count; 2500 break; 2501 case X86::COND_NE_OR_P: 2502 // Synthesize NE_OR_P with two branches. 2503 BuildMI(&MBB, DL, get(X86::JNE_4)).addMBB(TBB); 2504 ++Count; 2505 BuildMI(&MBB, DL, get(X86::JP_4)).addMBB(TBB); 2506 ++Count; 2507 break; 2508 default: { 2509 unsigned Opc = GetCondBranchFromCond(CC); 2510 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB); 2511 ++Count; 2512 } 2513 } 2514 if (FBB) { 2515 // Two-way Conditional branch. Insert the second branch. 2516 BuildMI(&MBB, DL, get(X86::JMP_4)).addMBB(FBB); 2517 ++Count; 2518 } 2519 return Count; 2520} 2521 2522/// isHReg - Test if the given register is a physical h register. 2523static bool isHReg(unsigned Reg) { 2524 return X86::GR8_ABCD_HRegClass.contains(Reg); 2525} 2526 2527// Try and copy between VR128/VR64 and GR64 registers. 2528static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg, 2529 bool HasAVX) { 2530 // SrcReg(VR128) -> DestReg(GR64) 2531 // SrcReg(VR64) -> DestReg(GR64) 2532 // SrcReg(GR64) -> DestReg(VR128) 2533 // SrcReg(GR64) -> DestReg(VR64) 2534 2535 if (X86::GR64RegClass.contains(DestReg)) { 2536 if (X86::VR128RegClass.contains(SrcReg)) { 2537 // Copy from a VR128 register to a GR64 register. 2538 return HasAVX ? X86::VMOVPQIto64rr : X86::MOVPQIto64rr; 2539 } else if (X86::VR64RegClass.contains(SrcReg)) { 2540 // Copy from a VR64 register to a GR64 register. 2541 return X86::MOVSDto64rr; 2542 } 2543 } else if (X86::GR64RegClass.contains(SrcReg)) { 2544 // Copy from a GR64 register to a VR128 register. 2545 if (X86::VR128RegClass.contains(DestReg)) 2546 return HasAVX ? X86::VMOV64toPQIrr : X86::MOV64toPQIrr; 2547 // Copy from a GR64 register to a VR64 register. 2548 else if (X86::VR64RegClass.contains(DestReg)) 2549 return X86::MOV64toSDrr; 2550 } 2551 2552 // SrcReg(FR32) -> DestReg(GR32) 2553 // SrcReg(GR32) -> DestReg(FR32) 2554 2555 if (X86::GR32RegClass.contains(DestReg) && X86::FR32RegClass.contains(SrcReg)) 2556 // Copy from a FR32 register to a GR32 register. 2557 return HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr; 2558 2559 if (X86::FR32RegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg)) 2560 // Copy from a GR32 register to a FR32 register. 2561 return HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr; 2562 2563 return 0; 2564} 2565 2566void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB, 2567 MachineBasicBlock::iterator MI, DebugLoc DL, 2568 unsigned DestReg, unsigned SrcReg, 2569 bool KillSrc) const { 2570 // First deal with the normal symmetric copies. 2571 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX(); 2572 unsigned Opc = 0; 2573 if (X86::GR64RegClass.contains(DestReg, SrcReg)) 2574 Opc = X86::MOV64rr; 2575 else if (X86::GR32RegClass.contains(DestReg, SrcReg)) 2576 Opc = X86::MOV32rr; 2577 else if (X86::GR16RegClass.contains(DestReg, SrcReg)) 2578 Opc = X86::MOV16rr; 2579 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) { 2580 // Copying to or from a physical H register on x86-64 requires a NOREX 2581 // move. Otherwise use a normal move. 2582 if ((isHReg(DestReg) || isHReg(SrcReg)) && 2583 TM.getSubtarget<X86Subtarget>().is64Bit()) { 2584 Opc = X86::MOV8rr_NOREX; 2585 // Both operands must be encodable without an REX prefix. 2586 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) && 2587 "8-bit H register can not be copied outside GR8_NOREX"); 2588 } else 2589 Opc = X86::MOV8rr; 2590 } else if (X86::VR128RegClass.contains(DestReg, SrcReg)) 2591 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr; 2592 else if (X86::VR256RegClass.contains(DestReg, SrcReg)) 2593 Opc = X86::VMOVAPSYrr; 2594 else if (X86::VR64RegClass.contains(DestReg, SrcReg)) 2595 Opc = X86::MMX_MOVQ64rr; 2596 else 2597 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, HasAVX); 2598 2599 if (Opc) { 2600 BuildMI(MBB, MI, DL, get(Opc), DestReg) 2601 .addReg(SrcReg, getKillRegState(KillSrc)); 2602 return; 2603 } 2604 2605 // Moving EFLAGS to / from another register requires a push and a pop. 2606 if (SrcReg == X86::EFLAGS) { 2607 if (X86::GR64RegClass.contains(DestReg)) { 2608 BuildMI(MBB, MI, DL, get(X86::PUSHF64)); 2609 BuildMI(MBB, MI, DL, get(X86::POP64r), DestReg); 2610 return; 2611 } else if (X86::GR32RegClass.contains(DestReg)) { 2612 BuildMI(MBB, MI, DL, get(X86::PUSHF32)); 2613 BuildMI(MBB, MI, DL, get(X86::POP32r), DestReg); 2614 return; 2615 } 2616 } 2617 if (DestReg == X86::EFLAGS) { 2618 if (X86::GR64RegClass.contains(SrcReg)) { 2619 BuildMI(MBB, MI, DL, get(X86::PUSH64r)) 2620 .addReg(SrcReg, getKillRegState(KillSrc)); 2621 BuildMI(MBB, MI, DL, get(X86::POPF64)); 2622 return; 2623 } else if (X86::GR32RegClass.contains(SrcReg)) { 2624 BuildMI(MBB, MI, DL, get(X86::PUSH32r)) 2625 .addReg(SrcReg, getKillRegState(KillSrc)); 2626 BuildMI(MBB, MI, DL, get(X86::POPF32)); 2627 return; 2628 } 2629 } 2630 2631 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) 2632 << " to " << RI.getName(DestReg) << '\n'); 2633 llvm_unreachable("Cannot emit physreg copy instruction"); 2634} 2635 2636static unsigned getLoadStoreRegOpcode(unsigned Reg, 2637 const TargetRegisterClass *RC, 2638 bool isStackAligned, 2639 const TargetMachine &TM, 2640 bool load) { 2641 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX(); 2642 switch (RC->getSize()) { 2643 default: 2644 llvm_unreachable("Unknown spill size"); 2645 case 1: 2646 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass"); 2647 if (TM.getSubtarget<X86Subtarget>().is64Bit()) 2648 // Copying to or from a physical H register on x86-64 requires a NOREX 2649 // move. Otherwise use a normal move. 2650 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC)) 2651 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX; 2652 return load ? X86::MOV8rm : X86::MOV8mr; 2653 case 2: 2654 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass"); 2655 return load ? X86::MOV16rm : X86::MOV16mr; 2656 case 4: 2657 if (X86::GR32RegClass.hasSubClassEq(RC)) 2658 return load ? X86::MOV32rm : X86::MOV32mr; 2659 if (X86::FR32RegClass.hasSubClassEq(RC)) 2660 return load ? 2661 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) : 2662 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr); 2663 if (X86::RFP32RegClass.hasSubClassEq(RC)) 2664 return load ? X86::LD_Fp32m : X86::ST_Fp32m; 2665 llvm_unreachable("Unknown 4-byte regclass"); 2666 case 8: 2667 if (X86::GR64RegClass.hasSubClassEq(RC)) 2668 return load ? X86::MOV64rm : X86::MOV64mr; 2669 if (X86::FR64RegClass.hasSubClassEq(RC)) 2670 return load ? 2671 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) : 2672 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr); 2673 if (X86::VR64RegClass.hasSubClassEq(RC)) 2674 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr; 2675 if (X86::RFP64RegClass.hasSubClassEq(RC)) 2676 return load ? X86::LD_Fp64m : X86::ST_Fp64m; 2677 llvm_unreachable("Unknown 8-byte regclass"); 2678 case 10: 2679 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass"); 2680 return load ? X86::LD_Fp80m : X86::ST_FpP80m; 2681 case 16: { 2682 assert(X86::VR128RegClass.hasSubClassEq(RC) && "Unknown 16-byte regclass"); 2683 // If stack is realigned we can use aligned stores. 2684 if (isStackAligned) 2685 return load ? 2686 (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) : 2687 (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr); 2688 else 2689 return load ? 2690 (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) : 2691 (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr); 2692 } 2693 case 32: 2694 assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass"); 2695 // If stack is realigned we can use aligned stores. 2696 if (isStackAligned) 2697 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr; 2698 else 2699 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr; 2700 } 2701} 2702 2703static unsigned getStoreRegOpcode(unsigned SrcReg, 2704 const TargetRegisterClass *RC, 2705 bool isStackAligned, 2706 TargetMachine &TM) { 2707 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, TM, false); 2708} 2709 2710 2711static unsigned getLoadRegOpcode(unsigned DestReg, 2712 const TargetRegisterClass *RC, 2713 bool isStackAligned, 2714 const TargetMachine &TM) { 2715 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, TM, true); 2716} 2717 2718void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 2719 MachineBasicBlock::iterator MI, 2720 unsigned SrcReg, bool isKill, int FrameIdx, 2721 const TargetRegisterClass *RC, 2722 const TargetRegisterInfo *TRI) const { 2723 const MachineFunction &MF = *MBB.getParent(); 2724 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() && 2725 "Stack slot too small for store"); 2726 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 2727 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) || 2728 RI.canRealignStack(MF); 2729 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM); 2730 DebugLoc DL = MBB.findDebugLoc(MI); 2731 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx) 2732 .addReg(SrcReg, getKillRegState(isKill)); 2733} 2734 2735void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg, 2736 bool isKill, 2737 SmallVectorImpl<MachineOperand> &Addr, 2738 const TargetRegisterClass *RC, 2739 MachineInstr::mmo_iterator MMOBegin, 2740 MachineInstr::mmo_iterator MMOEnd, 2741 SmallVectorImpl<MachineInstr*> &NewMIs) const { 2742 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 2743 bool isAligned = MMOBegin != MMOEnd && 2744 (*MMOBegin)->getAlignment() >= Alignment; 2745 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, TM); 2746 DebugLoc DL; 2747 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc)); 2748 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 2749 MIB.addOperand(Addr[i]); 2750 MIB.addReg(SrcReg, getKillRegState(isKill)); 2751 (*MIB).setMemRefs(MMOBegin, MMOEnd); 2752 NewMIs.push_back(MIB); 2753} 2754 2755 2756void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 2757 MachineBasicBlock::iterator MI, 2758 unsigned DestReg, int FrameIdx, 2759 const TargetRegisterClass *RC, 2760 const TargetRegisterInfo *TRI) const { 2761 const MachineFunction &MF = *MBB.getParent(); 2762 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 2763 bool isAligned = (TM.getFrameLowering()->getStackAlignment() >= Alignment) || 2764 RI.canRealignStack(MF); 2765 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM); 2766 DebugLoc DL = MBB.findDebugLoc(MI); 2767 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx); 2768} 2769 2770void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg, 2771 SmallVectorImpl<MachineOperand> &Addr, 2772 const TargetRegisterClass *RC, 2773 MachineInstr::mmo_iterator MMOBegin, 2774 MachineInstr::mmo_iterator MMOEnd, 2775 SmallVectorImpl<MachineInstr*> &NewMIs) const { 2776 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 2777 bool isAligned = MMOBegin != MMOEnd && 2778 (*MMOBegin)->getAlignment() >= Alignment; 2779 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, TM); 2780 DebugLoc DL; 2781 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg); 2782 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 2783 MIB.addOperand(Addr[i]); 2784 (*MIB).setMemRefs(MMOBegin, MMOEnd); 2785 NewMIs.push_back(MIB); 2786} 2787 2788/// Expand2AddrUndef - Expand a single-def pseudo instruction to a two-addr 2789/// instruction with two undef reads of the register being defined. This is 2790/// used for mapping: 2791/// %xmm4 = V_SET0 2792/// to: 2793/// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef> 2794/// 2795static bool Expand2AddrUndef(MachineInstr *MI, const MCInstrDesc &Desc) { 2796 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction."); 2797 unsigned Reg = MI->getOperand(0).getReg(); 2798 MI->setDesc(Desc); 2799 2800 // MachineInstr::addOperand() will insert explicit operands before any 2801 // implicit operands. 2802 MachineInstrBuilder(MI).addReg(Reg, RegState::Undef) 2803 .addReg(Reg, RegState::Undef); 2804 // But we don't trust that. 2805 assert(MI->getOperand(1).getReg() == Reg && 2806 MI->getOperand(2).getReg() == Reg && "Misplaced operand"); 2807 return true; 2808} 2809 2810bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const { 2811 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX(); 2812 switch (MI->getOpcode()) { 2813 case X86::V_SET0: 2814 case X86::FsFLD0SS: 2815 case X86::FsFLD0SD: 2816 return Expand2AddrUndef(MI, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr)); 2817 case X86::TEST8ri_NOREX: 2818 MI->setDesc(get(X86::TEST8ri)); 2819 return true; 2820 } 2821 return false; 2822} 2823 2824MachineInstr* 2825X86InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF, 2826 int FrameIx, uint64_t Offset, 2827 const MDNode *MDPtr, 2828 DebugLoc DL) const { 2829 X86AddressMode AM; 2830 AM.BaseType = X86AddressMode::FrameIndexBase; 2831 AM.Base.FrameIndex = FrameIx; 2832 MachineInstrBuilder MIB = BuildMI(MF, DL, get(X86::DBG_VALUE)); 2833 addFullAddress(MIB, AM).addImm(Offset).addMetadata(MDPtr); 2834 return &*MIB; 2835} 2836 2837static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode, 2838 const SmallVectorImpl<MachineOperand> &MOs, 2839 MachineInstr *MI, 2840 const TargetInstrInfo &TII) { 2841 // Create the base instruction with the memory operand as the first part. 2842 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), 2843 MI->getDebugLoc(), true); 2844 MachineInstrBuilder MIB(NewMI); 2845 unsigned NumAddrOps = MOs.size(); 2846 for (unsigned i = 0; i != NumAddrOps; ++i) 2847 MIB.addOperand(MOs[i]); 2848 if (NumAddrOps < 4) // FrameIndex only 2849 addOffset(MIB, 0); 2850 2851 // Loop over the rest of the ri operands, converting them over. 2852 unsigned NumOps = MI->getDesc().getNumOperands()-2; 2853 for (unsigned i = 0; i != NumOps; ++i) { 2854 MachineOperand &MO = MI->getOperand(i+2); 2855 MIB.addOperand(MO); 2856 } 2857 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) { 2858 MachineOperand &MO = MI->getOperand(i); 2859 MIB.addOperand(MO); 2860 } 2861 return MIB; 2862} 2863 2864static MachineInstr *FuseInst(MachineFunction &MF, 2865 unsigned Opcode, unsigned OpNo, 2866 const SmallVectorImpl<MachineOperand> &MOs, 2867 MachineInstr *MI, const TargetInstrInfo &TII) { 2868 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), 2869 MI->getDebugLoc(), true); 2870 MachineInstrBuilder MIB(NewMI); 2871 2872 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 2873 MachineOperand &MO = MI->getOperand(i); 2874 if (i == OpNo) { 2875 assert(MO.isReg() && "Expected to fold into reg operand!"); 2876 unsigned NumAddrOps = MOs.size(); 2877 for (unsigned i = 0; i != NumAddrOps; ++i) 2878 MIB.addOperand(MOs[i]); 2879 if (NumAddrOps < 4) // FrameIndex only 2880 addOffset(MIB, 0); 2881 } else { 2882 MIB.addOperand(MO); 2883 } 2884 } 2885 return MIB; 2886} 2887 2888static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode, 2889 const SmallVectorImpl<MachineOperand> &MOs, 2890 MachineInstr *MI) { 2891 MachineFunction &MF = *MI->getParent()->getParent(); 2892 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), TII.get(Opcode)); 2893 2894 unsigned NumAddrOps = MOs.size(); 2895 for (unsigned i = 0; i != NumAddrOps; ++i) 2896 MIB.addOperand(MOs[i]); 2897 if (NumAddrOps < 4) // FrameIndex only 2898 addOffset(MIB, 0); 2899 return MIB.addImm(0); 2900} 2901 2902MachineInstr* 2903X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, 2904 MachineInstr *MI, unsigned i, 2905 const SmallVectorImpl<MachineOperand> &MOs, 2906 unsigned Size, unsigned Align) const { 2907 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0; 2908 bool isTwoAddrFold = false; 2909 unsigned NumOps = MI->getDesc().getNumOperands(); 2910 bool isTwoAddr = NumOps > 1 && 2911 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1; 2912 2913 // FIXME: AsmPrinter doesn't know how to handle 2914 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding. 2915 if (MI->getOpcode() == X86::ADD32ri && 2916 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS) 2917 return NULL; 2918 2919 MachineInstr *NewMI = NULL; 2920 // Folding a memory location into the two-address part of a two-address 2921 // instruction is different than folding it other places. It requires 2922 // replacing the *two* registers with the memory location. 2923 if (isTwoAddr && NumOps >= 2 && i < 2 && 2924 MI->getOperand(0).isReg() && 2925 MI->getOperand(1).isReg() && 2926 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) { 2927 OpcodeTablePtr = &RegOp2MemOpTable2Addr; 2928 isTwoAddrFold = true; 2929 } else if (i == 0) { // If operand 0 2930 if (MI->getOpcode() == X86::MOV64r0) 2931 NewMI = MakeM0Inst(*this, X86::MOV64mi32, MOs, MI); 2932 else if (MI->getOpcode() == X86::MOV32r0) 2933 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, MI); 2934 else if (MI->getOpcode() == X86::MOV16r0) 2935 NewMI = MakeM0Inst(*this, X86::MOV16mi, MOs, MI); 2936 else if (MI->getOpcode() == X86::MOV8r0) 2937 NewMI = MakeM0Inst(*this, X86::MOV8mi, MOs, MI); 2938 if (NewMI) 2939 return NewMI; 2940 2941 OpcodeTablePtr = &RegOp2MemOpTable0; 2942 } else if (i == 1) { 2943 OpcodeTablePtr = &RegOp2MemOpTable1; 2944 } else if (i == 2) { 2945 OpcodeTablePtr = &RegOp2MemOpTable2; 2946 } 2947 2948 // If table selected... 2949 if (OpcodeTablePtr) { 2950 // Find the Opcode to fuse 2951 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I = 2952 OpcodeTablePtr->find(MI->getOpcode()); 2953 if (I != OpcodeTablePtr->end()) { 2954 unsigned Opcode = I->second.first; 2955 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT; 2956 if (Align < MinAlign) 2957 return NULL; 2958 bool NarrowToMOV32rm = false; 2959 if (Size) { 2960 unsigned RCSize = getRegClass(MI->getDesc(), i, &RI, MF)->getSize(); 2961 if (Size < RCSize) { 2962 // Check if it's safe to fold the load. If the size of the object is 2963 // narrower than the load width, then it's not. 2964 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4) 2965 return NULL; 2966 // If this is a 64-bit load, but the spill slot is 32, then we can do 2967 // a 32-bit load which is implicitly zero-extended. This likely is due 2968 // to liveintervalanalysis remat'ing a load from stack slot. 2969 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg()) 2970 return NULL; 2971 Opcode = X86::MOV32rm; 2972 NarrowToMOV32rm = true; 2973 } 2974 } 2975 2976 if (isTwoAddrFold) 2977 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, MI, *this); 2978 else 2979 NewMI = FuseInst(MF, Opcode, i, MOs, MI, *this); 2980 2981 if (NarrowToMOV32rm) { 2982 // If this is the special case where we use a MOV32rm to load a 32-bit 2983 // value and zero-extend the top bits. Change the destination register 2984 // to a 32-bit one. 2985 unsigned DstReg = NewMI->getOperand(0).getReg(); 2986 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) 2987 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, 2988 X86::sub_32bit)); 2989 else 2990 NewMI->getOperand(0).setSubReg(X86::sub_32bit); 2991 } 2992 return NewMI; 2993 } 2994 } 2995 2996 // No fusion 2997 if (PrintFailedFusing && !MI->isCopy()) 2998 dbgs() << "We failed to fuse operand " << i << " in " << *MI; 2999 return NULL; 3000} 3001 3002/// hasPartialRegUpdate - Return true for all instructions that only update 3003/// the first 32 or 64-bits of the destination register and leave the rest 3004/// unmodified. This can be used to avoid folding loads if the instructions 3005/// only update part of the destination register, and the non-updated part is 3006/// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these 3007/// instructions breaks the partial register dependency and it can improve 3008/// performance. e.g.: 3009/// 3010/// movss (%rdi), %xmm0 3011/// cvtss2sd %xmm0, %xmm0 3012/// 3013/// Instead of 3014/// cvtss2sd (%rdi), %xmm0 3015/// 3016/// FIXME: This should be turned into a TSFlags. 3017/// 3018static bool hasPartialRegUpdate(unsigned Opcode) { 3019 switch (Opcode) { 3020 case X86::CVTSI2SSrr: 3021 case X86::CVTSI2SS64rr: 3022 case X86::CVTSI2SDrr: 3023 case X86::CVTSI2SD64rr: 3024 case X86::CVTSD2SSrr: 3025 case X86::Int_CVTSD2SSrr: 3026 case X86::CVTSS2SDrr: 3027 case X86::Int_CVTSS2SDrr: 3028 case X86::RCPSSr: 3029 case X86::RCPSSr_Int: 3030 case X86::ROUNDSDr: 3031 case X86::ROUNDSDr_Int: 3032 case X86::ROUNDSSr: 3033 case X86::ROUNDSSr_Int: 3034 case X86::RSQRTSSr: 3035 case X86::RSQRTSSr_Int: 3036 case X86::SQRTSSr: 3037 case X86::SQRTSSr_Int: 3038 // AVX encoded versions 3039 case X86::VCVTSD2SSrr: 3040 case X86::Int_VCVTSD2SSrr: 3041 case X86::VCVTSS2SDrr: 3042 case X86::Int_VCVTSS2SDrr: 3043 case X86::VRCPSSr: 3044 case X86::VROUNDSDr: 3045 case X86::VROUNDSDr_Int: 3046 case X86::VROUNDSSr: 3047 case X86::VROUNDSSr_Int: 3048 case X86::VRSQRTSSr: 3049 case X86::VSQRTSSr: 3050 return true; 3051 } 3052 3053 return false; 3054} 3055 3056/// getPartialRegUpdateClearance - Inform the ExeDepsFix pass how many idle 3057/// instructions we would like before a partial register update. 3058unsigned X86InstrInfo:: 3059getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum, 3060 const TargetRegisterInfo *TRI) const { 3061 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode())) 3062 return 0; 3063 3064 // If MI is marked as reading Reg, the partial register update is wanted. 3065 const MachineOperand &MO = MI->getOperand(0); 3066 unsigned Reg = MO.getReg(); 3067 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 3068 if (MO.readsReg() || MI->readsVirtualRegister(Reg)) 3069 return 0; 3070 } else { 3071 if (MI->readsRegister(Reg, TRI)) 3072 return 0; 3073 } 3074 3075 // If any of the preceding 16 instructions are reading Reg, insert a 3076 // dependency breaking instruction. The magic number is based on a few 3077 // Nehalem experiments. 3078 return 16; 3079} 3080 3081void X86InstrInfo:: 3082breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum, 3083 const TargetRegisterInfo *TRI) const { 3084 unsigned Reg = MI->getOperand(OpNum).getReg(); 3085 if (X86::VR128RegClass.contains(Reg)) { 3086 // These instructions are all floating point domain, so xorps is the best 3087 // choice. 3088 bool HasAVX = TM.getSubtarget<X86Subtarget>().hasAVX(); 3089 unsigned Opc = HasAVX ? X86::VXORPSrr : X86::XORPSrr; 3090 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg) 3091 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 3092 } else if (X86::VR256RegClass.contains(Reg)) { 3093 // Use vxorps to clear the full ymm register. 3094 // It wants to read and write the xmm sub-register. 3095 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm); 3096 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg) 3097 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef) 3098 .addReg(Reg, RegState::ImplicitDefine); 3099 } else 3100 return; 3101 MI->addRegisterKilled(Reg, TRI, true); 3102} 3103 3104MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, 3105 MachineInstr *MI, 3106 const SmallVectorImpl<unsigned> &Ops, 3107 int FrameIndex) const { 3108 // Check switch flag 3109 if (NoFusing) return NULL; 3110 3111 // Unless optimizing for size, don't fold to avoid partial 3112 // register update stalls 3113 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 3114 hasPartialRegUpdate(MI->getOpcode())) 3115 return 0; 3116 3117 const MachineFrameInfo *MFI = MF.getFrameInfo(); 3118 unsigned Size = MFI->getObjectSize(FrameIndex); 3119 unsigned Alignment = MFI->getObjectAlignment(FrameIndex); 3120 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 3121 unsigned NewOpc = 0; 3122 unsigned RCSize = 0; 3123 switch (MI->getOpcode()) { 3124 default: return NULL; 3125 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break; 3126 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break; 3127 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break; 3128 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break; 3129 } 3130 // Check if it's safe to fold the load. If the size of the object is 3131 // narrower than the load width, then it's not. 3132 if (Size < RCSize) 3133 return NULL; 3134 // Change to CMPXXri r, 0 first. 3135 MI->setDesc(get(NewOpc)); 3136 MI->getOperand(1).ChangeToImmediate(0); 3137 } else if (Ops.size() != 1) 3138 return NULL; 3139 3140 SmallVector<MachineOperand,4> MOs; 3141 MOs.push_back(MachineOperand::CreateFI(FrameIndex)); 3142 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, Size, Alignment); 3143} 3144 3145MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF, 3146 MachineInstr *MI, 3147 const SmallVectorImpl<unsigned> &Ops, 3148 MachineInstr *LoadMI) const { 3149 // Check switch flag 3150 if (NoFusing) return NULL; 3151 3152 // Unless optimizing for size, don't fold to avoid partial 3153 // register update stalls 3154 if (!MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 3155 hasPartialRegUpdate(MI->getOpcode())) 3156 return 0; 3157 3158 // Determine the alignment of the load. 3159 unsigned Alignment = 0; 3160 if (LoadMI->hasOneMemOperand()) 3161 Alignment = (*LoadMI->memoperands_begin())->getAlignment(); 3162 else 3163 switch (LoadMI->getOpcode()) { 3164 case X86::AVX_SET0PSY: 3165 case X86::AVX_SET0PDY: 3166 case X86::AVX2_SETALLONES: 3167 case X86::AVX2_SET0: 3168 Alignment = 32; 3169 break; 3170 case X86::V_SET0: 3171 case X86::V_SETALLONES: 3172 case X86::AVX_SETALLONES: 3173 Alignment = 16; 3174 break; 3175 case X86::FsFLD0SD: 3176 Alignment = 8; 3177 break; 3178 case X86::FsFLD0SS: 3179 Alignment = 4; 3180 break; 3181 default: 3182 return 0; 3183 } 3184 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 3185 unsigned NewOpc = 0; 3186 switch (MI->getOpcode()) { 3187 default: return NULL; 3188 case X86::TEST8rr: NewOpc = X86::CMP8ri; break; 3189 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break; 3190 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break; 3191 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break; 3192 } 3193 // Change to CMPXXri r, 0 first. 3194 MI->setDesc(get(NewOpc)); 3195 MI->getOperand(1).ChangeToImmediate(0); 3196 } else if (Ops.size() != 1) 3197 return NULL; 3198 3199 // Make sure the subregisters match. 3200 // Otherwise we risk changing the size of the load. 3201 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg()) 3202 return NULL; 3203 3204 SmallVector<MachineOperand,X86::AddrNumOperands> MOs; 3205 switch (LoadMI->getOpcode()) { 3206 case X86::V_SET0: 3207 case X86::V_SETALLONES: 3208 case X86::AVX_SET0PSY: 3209 case X86::AVX_SET0PDY: 3210 case X86::AVX_SETALLONES: 3211 case X86::AVX2_SETALLONES: 3212 case X86::AVX2_SET0: 3213 case X86::FsFLD0SD: 3214 case X86::FsFLD0SS: { 3215 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure. 3216 // Create a constant-pool entry and operands to load from it. 3217 3218 // Medium and large mode can't fold loads this way. 3219 if (TM.getCodeModel() != CodeModel::Small && 3220 TM.getCodeModel() != CodeModel::Kernel) 3221 return NULL; 3222 3223 // x86-32 PIC requires a PIC base register for constant pools. 3224 unsigned PICBase = 0; 3225 if (TM.getRelocationModel() == Reloc::PIC_) { 3226 if (TM.getSubtarget<X86Subtarget>().is64Bit()) 3227 PICBase = X86::RIP; 3228 else 3229 // FIXME: PICBase = getGlobalBaseReg(&MF); 3230 // This doesn't work for several reasons. 3231 // 1. GlobalBaseReg may have been spilled. 3232 // 2. It may not be live at MI. 3233 return NULL; 3234 } 3235 3236 // Create a constant-pool entry. 3237 MachineConstantPool &MCP = *MF.getConstantPool(); 3238 Type *Ty; 3239 unsigned Opc = LoadMI->getOpcode(); 3240 if (Opc == X86::FsFLD0SS) 3241 Ty = Type::getFloatTy(MF.getFunction()->getContext()); 3242 else if (Opc == X86::FsFLD0SD) 3243 Ty = Type::getDoubleTy(MF.getFunction()->getContext()); 3244 else if (Opc == X86::AVX_SET0PSY || Opc == X86::AVX_SET0PDY) 3245 Ty = VectorType::get(Type::getFloatTy(MF.getFunction()->getContext()), 8); 3246 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX2_SET0) 3247 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8); 3248 else 3249 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4); 3250 3251 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX_SETALLONES || 3252 Opc == X86::AVX2_SETALLONES); 3253 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) : 3254 Constant::getNullValue(Ty); 3255 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment); 3256 3257 // Create operands to load from the constant pool entry. 3258 MOs.push_back(MachineOperand::CreateReg(PICBase, false)); 3259 MOs.push_back(MachineOperand::CreateImm(1)); 3260 MOs.push_back(MachineOperand::CreateReg(0, false)); 3261 MOs.push_back(MachineOperand::CreateCPI(CPI, 0)); 3262 MOs.push_back(MachineOperand::CreateReg(0, false)); 3263 break; 3264 } 3265 default: { 3266 // Folding a normal load. Just copy the load's address operands. 3267 unsigned NumOps = LoadMI->getDesc().getNumOperands(); 3268 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i) 3269 MOs.push_back(LoadMI->getOperand(i)); 3270 break; 3271 } 3272 } 3273 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, 0, Alignment); 3274} 3275 3276 3277bool X86InstrInfo::canFoldMemoryOperand(const MachineInstr *MI, 3278 const SmallVectorImpl<unsigned> &Ops) const { 3279 // Check switch flag 3280 if (NoFusing) return 0; 3281 3282 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 3283 switch (MI->getOpcode()) { 3284 default: return false; 3285 case X86::TEST8rr: 3286 case X86::TEST16rr: 3287 case X86::TEST32rr: 3288 case X86::TEST64rr: 3289 return true; 3290 case X86::ADD32ri: 3291 // FIXME: AsmPrinter doesn't know how to handle 3292 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding. 3293 if (MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS) 3294 return false; 3295 break; 3296 } 3297 } 3298 3299 if (Ops.size() != 1) 3300 return false; 3301 3302 unsigned OpNum = Ops[0]; 3303 unsigned Opc = MI->getOpcode(); 3304 unsigned NumOps = MI->getDesc().getNumOperands(); 3305 bool isTwoAddr = NumOps > 1 && 3306 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1; 3307 3308 // Folding a memory location into the two-address part of a two-address 3309 // instruction is different than folding it other places. It requires 3310 // replacing the *two* registers with the memory location. 3311 const DenseMap<unsigned, std::pair<unsigned,unsigned> > *OpcodeTablePtr = 0; 3312 if (isTwoAddr && NumOps >= 2 && OpNum < 2) { 3313 OpcodeTablePtr = &RegOp2MemOpTable2Addr; 3314 } else if (OpNum == 0) { // If operand 0 3315 switch (Opc) { 3316 case X86::MOV8r0: 3317 case X86::MOV16r0: 3318 case X86::MOV32r0: 3319 case X86::MOV64r0: return true; 3320 default: break; 3321 } 3322 OpcodeTablePtr = &RegOp2MemOpTable0; 3323 } else if (OpNum == 1) { 3324 OpcodeTablePtr = &RegOp2MemOpTable1; 3325 } else if (OpNum == 2) { 3326 OpcodeTablePtr = &RegOp2MemOpTable2; 3327 } 3328 3329 if (OpcodeTablePtr && OpcodeTablePtr->count(Opc)) 3330 return true; 3331 return TargetInstrInfoImpl::canFoldMemoryOperand(MI, Ops); 3332} 3333 3334bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI, 3335 unsigned Reg, bool UnfoldLoad, bool UnfoldStore, 3336 SmallVectorImpl<MachineInstr*> &NewMIs) const { 3337 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I = 3338 MemOp2RegOpTable.find(MI->getOpcode()); 3339 if (I == MemOp2RegOpTable.end()) 3340 return false; 3341 unsigned Opc = I->second.first; 3342 unsigned Index = I->second.second & TB_INDEX_MASK; 3343 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 3344 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 3345 if (UnfoldLoad && !FoldedLoad) 3346 return false; 3347 UnfoldLoad &= FoldedLoad; 3348 if (UnfoldStore && !FoldedStore) 3349 return false; 3350 UnfoldStore &= FoldedStore; 3351 3352 const MCInstrDesc &MCID = get(Opc); 3353 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 3354 if (!MI->hasOneMemOperand() && 3355 RC == &X86::VR128RegClass && 3356 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast()) 3357 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will 3358 // conservatively assume the address is unaligned. That's bad for 3359 // performance. 3360 return false; 3361 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps; 3362 SmallVector<MachineOperand,2> BeforeOps; 3363 SmallVector<MachineOperand,2> AfterOps; 3364 SmallVector<MachineOperand,4> ImpOps; 3365 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 3366 MachineOperand &Op = MI->getOperand(i); 3367 if (i >= Index && i < Index + X86::AddrNumOperands) 3368 AddrOps.push_back(Op); 3369 else if (Op.isReg() && Op.isImplicit()) 3370 ImpOps.push_back(Op); 3371 else if (i < Index) 3372 BeforeOps.push_back(Op); 3373 else if (i > Index) 3374 AfterOps.push_back(Op); 3375 } 3376 3377 // Emit the load instruction. 3378 if (UnfoldLoad) { 3379 std::pair<MachineInstr::mmo_iterator, 3380 MachineInstr::mmo_iterator> MMOs = 3381 MF.extractLoadMemRefs(MI->memoperands_begin(), 3382 MI->memoperands_end()); 3383 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs); 3384 if (UnfoldStore) { 3385 // Address operands cannot be marked isKill. 3386 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) { 3387 MachineOperand &MO = NewMIs[0]->getOperand(i); 3388 if (MO.isReg()) 3389 MO.setIsKill(false); 3390 } 3391 } 3392 } 3393 3394 // Emit the data processing instruction. 3395 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true); 3396 MachineInstrBuilder MIB(DataMI); 3397 3398 if (FoldedStore) 3399 MIB.addReg(Reg, RegState::Define); 3400 for (unsigned i = 0, e = BeforeOps.size(); i != e; ++i) 3401 MIB.addOperand(BeforeOps[i]); 3402 if (FoldedLoad) 3403 MIB.addReg(Reg); 3404 for (unsigned i = 0, e = AfterOps.size(); i != e; ++i) 3405 MIB.addOperand(AfterOps[i]); 3406 for (unsigned i = 0, e = ImpOps.size(); i != e; ++i) { 3407 MachineOperand &MO = ImpOps[i]; 3408 MIB.addReg(MO.getReg(), 3409 getDefRegState(MO.isDef()) | 3410 RegState::Implicit | 3411 getKillRegState(MO.isKill()) | 3412 getDeadRegState(MO.isDead()) | 3413 getUndefRegState(MO.isUndef())); 3414 } 3415 // Change CMP32ri r, 0 back to TEST32rr r, r, etc. 3416 unsigned NewOpc = 0; 3417 switch (DataMI->getOpcode()) { 3418 default: break; 3419 case X86::CMP64ri32: 3420 case X86::CMP64ri8: 3421 case X86::CMP32ri: 3422 case X86::CMP32ri8: 3423 case X86::CMP16ri: 3424 case X86::CMP16ri8: 3425 case X86::CMP8ri: { 3426 MachineOperand &MO0 = DataMI->getOperand(0); 3427 MachineOperand &MO1 = DataMI->getOperand(1); 3428 if (MO1.getImm() == 0) { 3429 switch (DataMI->getOpcode()) { 3430 default: break; 3431 case X86::CMP64ri8: 3432 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break; 3433 case X86::CMP32ri8: 3434 case X86::CMP32ri: NewOpc = X86::TEST32rr; break; 3435 case X86::CMP16ri8: 3436 case X86::CMP16ri: NewOpc = X86::TEST16rr; break; 3437 case X86::CMP8ri: NewOpc = X86::TEST8rr; break; 3438 } 3439 DataMI->setDesc(get(NewOpc)); 3440 MO1.ChangeToRegister(MO0.getReg(), false); 3441 } 3442 } 3443 } 3444 NewMIs.push_back(DataMI); 3445 3446 // Emit the store instruction. 3447 if (UnfoldStore) { 3448 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF); 3449 std::pair<MachineInstr::mmo_iterator, 3450 MachineInstr::mmo_iterator> MMOs = 3451 MF.extractStoreMemRefs(MI->memoperands_begin(), 3452 MI->memoperands_end()); 3453 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs); 3454 } 3455 3456 return true; 3457} 3458 3459bool 3460X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, 3461 SmallVectorImpl<SDNode*> &NewNodes) const { 3462 if (!N->isMachineOpcode()) 3463 return false; 3464 3465 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I = 3466 MemOp2RegOpTable.find(N->getMachineOpcode()); 3467 if (I == MemOp2RegOpTable.end()) 3468 return false; 3469 unsigned Opc = I->second.first; 3470 unsigned Index = I->second.second & TB_INDEX_MASK; 3471 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 3472 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 3473 const MCInstrDesc &MCID = get(Opc); 3474 MachineFunction &MF = DAG.getMachineFunction(); 3475 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 3476 unsigned NumDefs = MCID.NumDefs; 3477 std::vector<SDValue> AddrOps; 3478 std::vector<SDValue> BeforeOps; 3479 std::vector<SDValue> AfterOps; 3480 DebugLoc dl = N->getDebugLoc(); 3481 unsigned NumOps = N->getNumOperands(); 3482 for (unsigned i = 0; i != NumOps-1; ++i) { 3483 SDValue Op = N->getOperand(i); 3484 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands) 3485 AddrOps.push_back(Op); 3486 else if (i < Index-NumDefs) 3487 BeforeOps.push_back(Op); 3488 else if (i > Index-NumDefs) 3489 AfterOps.push_back(Op); 3490 } 3491 SDValue Chain = N->getOperand(NumOps-1); 3492 AddrOps.push_back(Chain); 3493 3494 // Emit the load instruction. 3495 SDNode *Load = 0; 3496 if (FoldedLoad) { 3497 EVT VT = *RC->vt_begin(); 3498 std::pair<MachineInstr::mmo_iterator, 3499 MachineInstr::mmo_iterator> MMOs = 3500 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(), 3501 cast<MachineSDNode>(N)->memoperands_end()); 3502 if (!(*MMOs.first) && 3503 RC == &X86::VR128RegClass && 3504 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast()) 3505 // Do not introduce a slow unaligned load. 3506 return false; 3507 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 3508 bool isAligned = (*MMOs.first) && 3509 (*MMOs.first)->getAlignment() >= Alignment; 3510 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, TM), dl, 3511 VT, MVT::Other, &AddrOps[0], AddrOps.size()); 3512 NewNodes.push_back(Load); 3513 3514 // Preserve memory reference information. 3515 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second); 3516 } 3517 3518 // Emit the data processing instruction. 3519 std::vector<EVT> VTs; 3520 const TargetRegisterClass *DstRC = 0; 3521 if (MCID.getNumDefs() > 0) { 3522 DstRC = getRegClass(MCID, 0, &RI, MF); 3523 VTs.push_back(*DstRC->vt_begin()); 3524 } 3525 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) { 3526 EVT VT = N->getValueType(i); 3527 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs()) 3528 VTs.push_back(VT); 3529 } 3530 if (Load) 3531 BeforeOps.push_back(SDValue(Load, 0)); 3532 std::copy(AfterOps.begin(), AfterOps.end(), std::back_inserter(BeforeOps)); 3533 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, &BeforeOps[0], 3534 BeforeOps.size()); 3535 NewNodes.push_back(NewNode); 3536 3537 // Emit the store instruction. 3538 if (FoldedStore) { 3539 AddrOps.pop_back(); 3540 AddrOps.push_back(SDValue(NewNode, 0)); 3541 AddrOps.push_back(Chain); 3542 std::pair<MachineInstr::mmo_iterator, 3543 MachineInstr::mmo_iterator> MMOs = 3544 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(), 3545 cast<MachineSDNode>(N)->memoperands_end()); 3546 if (!(*MMOs.first) && 3547 RC == &X86::VR128RegClass && 3548 !TM.getSubtarget<X86Subtarget>().isUnalignedMemAccessFast()) 3549 // Do not introduce a slow unaligned store. 3550 return false; 3551 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 3552 bool isAligned = (*MMOs.first) && 3553 (*MMOs.first)->getAlignment() >= Alignment; 3554 SDNode *Store = DAG.getMachineNode(getStoreRegOpcode(0, DstRC, 3555 isAligned, TM), 3556 dl, MVT::Other, 3557 &AddrOps[0], AddrOps.size()); 3558 NewNodes.push_back(Store); 3559 3560 // Preserve memory reference information. 3561 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second); 3562 } 3563 3564 return true; 3565} 3566 3567unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc, 3568 bool UnfoldLoad, bool UnfoldStore, 3569 unsigned *LoadRegIndex) const { 3570 DenseMap<unsigned, std::pair<unsigned,unsigned> >::const_iterator I = 3571 MemOp2RegOpTable.find(Opc); 3572 if (I == MemOp2RegOpTable.end()) 3573 return 0; 3574 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 3575 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 3576 if (UnfoldLoad && !FoldedLoad) 3577 return 0; 3578 if (UnfoldStore && !FoldedStore) 3579 return 0; 3580 if (LoadRegIndex) 3581 *LoadRegIndex = I->second.second & TB_INDEX_MASK; 3582 return I->second.first; 3583} 3584 3585bool 3586X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 3587 int64_t &Offset1, int64_t &Offset2) const { 3588 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode()) 3589 return false; 3590 unsigned Opc1 = Load1->getMachineOpcode(); 3591 unsigned Opc2 = Load2->getMachineOpcode(); 3592 switch (Opc1) { 3593 default: return false; 3594 case X86::MOV8rm: 3595 case X86::MOV16rm: 3596 case X86::MOV32rm: 3597 case X86::MOV64rm: 3598 case X86::LD_Fp32m: 3599 case X86::LD_Fp64m: 3600 case X86::LD_Fp80m: 3601 case X86::MOVSSrm: 3602 case X86::MOVSDrm: 3603 case X86::MMX_MOVD64rm: 3604 case X86::MMX_MOVQ64rm: 3605 case X86::FsMOVAPSrm: 3606 case X86::FsMOVAPDrm: 3607 case X86::MOVAPSrm: 3608 case X86::MOVUPSrm: 3609 case X86::MOVAPDrm: 3610 case X86::MOVDQArm: 3611 case X86::MOVDQUrm: 3612 // AVX load instructions 3613 case X86::VMOVSSrm: 3614 case X86::VMOVSDrm: 3615 case X86::FsVMOVAPSrm: 3616 case X86::FsVMOVAPDrm: 3617 case X86::VMOVAPSrm: 3618 case X86::VMOVUPSrm: 3619 case X86::VMOVAPDrm: 3620 case X86::VMOVDQArm: 3621 case X86::VMOVDQUrm: 3622 case X86::VMOVAPSYrm: 3623 case X86::VMOVUPSYrm: 3624 case X86::VMOVAPDYrm: 3625 case X86::VMOVDQAYrm: 3626 case X86::VMOVDQUYrm: 3627 break; 3628 } 3629 switch (Opc2) { 3630 default: return false; 3631 case X86::MOV8rm: 3632 case X86::MOV16rm: 3633 case X86::MOV32rm: 3634 case X86::MOV64rm: 3635 case X86::LD_Fp32m: 3636 case X86::LD_Fp64m: 3637 case X86::LD_Fp80m: 3638 case X86::MOVSSrm: 3639 case X86::MOVSDrm: 3640 case X86::MMX_MOVD64rm: 3641 case X86::MMX_MOVQ64rm: 3642 case X86::FsMOVAPSrm: 3643 case X86::FsMOVAPDrm: 3644 case X86::MOVAPSrm: 3645 case X86::MOVUPSrm: 3646 case X86::MOVAPDrm: 3647 case X86::MOVDQArm: 3648 case X86::MOVDQUrm: 3649 // AVX load instructions 3650 case X86::VMOVSSrm: 3651 case X86::VMOVSDrm: 3652 case X86::FsVMOVAPSrm: 3653 case X86::FsVMOVAPDrm: 3654 case X86::VMOVAPSrm: 3655 case X86::VMOVUPSrm: 3656 case X86::VMOVAPDrm: 3657 case X86::VMOVDQArm: 3658 case X86::VMOVDQUrm: 3659 case X86::VMOVAPSYrm: 3660 case X86::VMOVUPSYrm: 3661 case X86::VMOVAPDYrm: 3662 case X86::VMOVDQAYrm: 3663 case X86::VMOVDQUYrm: 3664 break; 3665 } 3666 3667 // Check if chain operands and base addresses match. 3668 if (Load1->getOperand(0) != Load2->getOperand(0) || 3669 Load1->getOperand(5) != Load2->getOperand(5)) 3670 return false; 3671 // Segment operands should match as well. 3672 if (Load1->getOperand(4) != Load2->getOperand(4)) 3673 return false; 3674 // Scale should be 1, Index should be Reg0. 3675 if (Load1->getOperand(1) == Load2->getOperand(1) && 3676 Load1->getOperand(2) == Load2->getOperand(2)) { 3677 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1) 3678 return false; 3679 3680 // Now let's examine the displacements. 3681 if (isa<ConstantSDNode>(Load1->getOperand(3)) && 3682 isa<ConstantSDNode>(Load2->getOperand(3))) { 3683 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue(); 3684 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue(); 3685 return true; 3686 } 3687 } 3688 return false; 3689} 3690 3691bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 3692 int64_t Offset1, int64_t Offset2, 3693 unsigned NumLoads) const { 3694 assert(Offset2 > Offset1); 3695 if ((Offset2 - Offset1) / 8 > 64) 3696 return false; 3697 3698 unsigned Opc1 = Load1->getMachineOpcode(); 3699 unsigned Opc2 = Load2->getMachineOpcode(); 3700 if (Opc1 != Opc2) 3701 return false; // FIXME: overly conservative? 3702 3703 switch (Opc1) { 3704 default: break; 3705 case X86::LD_Fp32m: 3706 case X86::LD_Fp64m: 3707 case X86::LD_Fp80m: 3708 case X86::MMX_MOVD64rm: 3709 case X86::MMX_MOVQ64rm: 3710 return false; 3711 } 3712 3713 EVT VT = Load1->getValueType(0); 3714 switch (VT.getSimpleVT().SimpleTy) { 3715 default: 3716 // XMM registers. In 64-bit mode we can be a bit more aggressive since we 3717 // have 16 of them to play with. 3718 if (TM.getSubtargetImpl()->is64Bit()) { 3719 if (NumLoads >= 3) 3720 return false; 3721 } else if (NumLoads) { 3722 return false; 3723 } 3724 break; 3725 case MVT::i8: 3726 case MVT::i16: 3727 case MVT::i32: 3728 case MVT::i64: 3729 case MVT::f32: 3730 case MVT::f64: 3731 if (NumLoads) 3732 return false; 3733 break; 3734 } 3735 3736 return true; 3737} 3738 3739 3740bool X86InstrInfo:: 3741ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 3742 assert(Cond.size() == 1 && "Invalid X86 branch condition!"); 3743 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm()); 3744 if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E) 3745 return true; 3746 Cond[0].setImm(GetOppositeBranchCondition(CC)); 3747 return false; 3748} 3749 3750bool X86InstrInfo:: 3751isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { 3752 // FIXME: Return false for x87 stack register classes for now. We can't 3753 // allow any loads of these registers before FpGet_ST0_80. 3754 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass || 3755 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass); 3756} 3757 3758/// getGlobalBaseReg - Return a virtual register initialized with the 3759/// the global base register value. Output instructions required to 3760/// initialize the register in the function entry block, if necessary. 3761/// 3762/// TODO: Eliminate this and move the code to X86MachineFunctionInfo. 3763/// 3764unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const { 3765 assert(!TM.getSubtarget<X86Subtarget>().is64Bit() && 3766 "X86-64 PIC uses RIP relative addressing"); 3767 3768 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); 3769 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 3770 if (GlobalBaseReg != 0) 3771 return GlobalBaseReg; 3772 3773 // Create the register. The code to initialize it is inserted 3774 // later, by the CGBR pass (below). 3775 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 3776 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 3777 X86FI->setGlobalBaseReg(GlobalBaseReg); 3778 return GlobalBaseReg; 3779} 3780 3781// These are the replaceable SSE instructions. Some of these have Int variants 3782// that we don't include here. We don't want to replace instructions selected 3783// by intrinsics. 3784static const uint16_t ReplaceableInstrs[][3] = { 3785 //PackedSingle PackedDouble PackedInt 3786 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr }, 3787 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm }, 3788 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr }, 3789 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr }, 3790 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm }, 3791 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr }, 3792 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm }, 3793 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr }, 3794 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm }, 3795 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr }, 3796 { X86::ORPSrm, X86::ORPDrm, X86::PORrm }, 3797 { X86::ORPSrr, X86::ORPDrr, X86::PORrr }, 3798 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm }, 3799 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr }, 3800 // AVX 128-bit support 3801 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr }, 3802 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm }, 3803 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr }, 3804 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr }, 3805 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm }, 3806 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr }, 3807 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm }, 3808 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr }, 3809 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm }, 3810 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr }, 3811 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm }, 3812 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr }, 3813 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm }, 3814 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr }, 3815 // AVX 256-bit support 3816 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr }, 3817 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm }, 3818 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr }, 3819 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr }, 3820 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm }, 3821 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr } 3822}; 3823 3824static const uint16_t ReplaceableInstrsAVX2[][3] = { 3825 //PackedSingle PackedDouble PackedInt 3826 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm }, 3827 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr }, 3828 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm }, 3829 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr }, 3830 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm }, 3831 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr }, 3832 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm }, 3833 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr }, 3834 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr }, 3835 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr }, 3836 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm }, 3837 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr }, 3838 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm }, 3839 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr } 3840}; 3841 3842// FIXME: Some shuffle and unpack instructions have equivalents in different 3843// domains, but they require a bit more work than just switching opcodes. 3844 3845static const uint16_t *lookup(unsigned opcode, unsigned domain) { 3846 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrs); i != e; ++i) 3847 if (ReplaceableInstrs[i][domain-1] == opcode) 3848 return ReplaceableInstrs[i]; 3849 return 0; 3850} 3851 3852static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) { 3853 for (unsigned i = 0, e = array_lengthof(ReplaceableInstrsAVX2); i != e; ++i) 3854 if (ReplaceableInstrsAVX2[i][domain-1] == opcode) 3855 return ReplaceableInstrsAVX2[i]; 3856 return 0; 3857} 3858 3859std::pair<uint16_t, uint16_t> 3860X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const { 3861 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 3862 bool hasAVX2 = TM.getSubtarget<X86Subtarget>().hasAVX2(); 3863 uint16_t validDomains = 0; 3864 if (domain && lookup(MI->getOpcode(), domain)) 3865 validDomains = 0xe; 3866 else if (domain && lookupAVX2(MI->getOpcode(), domain)) 3867 validDomains = hasAVX2 ? 0xe : 0x6; 3868 return std::make_pair(domain, validDomains); 3869} 3870 3871void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const { 3872 assert(Domain>0 && Domain<4 && "Invalid execution domain"); 3873 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 3874 assert(dom && "Not an SSE instruction"); 3875 const uint16_t *table = lookup(MI->getOpcode(), dom); 3876 if (!table) { // try the other table 3877 assert((TM.getSubtarget<X86Subtarget>().hasAVX2() || Domain < 3) && 3878 "256-bit vector operations only available in AVX2"); 3879 table = lookupAVX2(MI->getOpcode(), dom); 3880 } 3881 assert(table && "Cannot change domain"); 3882 MI->setDesc(get(table[Domain-1])); 3883} 3884 3885/// getNoopForMachoTarget - Return the noop instruction to use for a noop. 3886void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { 3887 NopInst.setOpcode(X86::NOOP); 3888} 3889 3890bool X86InstrInfo::isHighLatencyDef(int opc) const { 3891 switch (opc) { 3892 default: return false; 3893 case X86::DIVSDrm: 3894 case X86::DIVSDrm_Int: 3895 case X86::DIVSDrr: 3896 case X86::DIVSDrr_Int: 3897 case X86::DIVSSrm: 3898 case X86::DIVSSrm_Int: 3899 case X86::DIVSSrr: 3900 case X86::DIVSSrr_Int: 3901 case X86::SQRTPDm: 3902 case X86::SQRTPDm_Int: 3903 case X86::SQRTPDr: 3904 case X86::SQRTPDr_Int: 3905 case X86::SQRTPSm: 3906 case X86::SQRTPSm_Int: 3907 case X86::SQRTPSr: 3908 case X86::SQRTPSr_Int: 3909 case X86::SQRTSDm: 3910 case X86::SQRTSDm_Int: 3911 case X86::SQRTSDr: 3912 case X86::SQRTSDr_Int: 3913 case X86::SQRTSSm: 3914 case X86::SQRTSSm_Int: 3915 case X86::SQRTSSr: 3916 case X86::SQRTSSr_Int: 3917 // AVX instructions with high latency 3918 case X86::VDIVSDrm: 3919 case X86::VDIVSDrm_Int: 3920 case X86::VDIVSDrr: 3921 case X86::VDIVSDrr_Int: 3922 case X86::VDIVSSrm: 3923 case X86::VDIVSSrm_Int: 3924 case X86::VDIVSSrr: 3925 case X86::VDIVSSrr_Int: 3926 case X86::VSQRTPDm: 3927 case X86::VSQRTPDm_Int: 3928 case X86::VSQRTPDr: 3929 case X86::VSQRTPDr_Int: 3930 case X86::VSQRTPSm: 3931 case X86::VSQRTPSm_Int: 3932 case X86::VSQRTPSr: 3933 case X86::VSQRTPSr_Int: 3934 case X86::VSQRTSDm: 3935 case X86::VSQRTSDm_Int: 3936 case X86::VSQRTSDr: 3937 case X86::VSQRTSSm: 3938 case X86::VSQRTSSm_Int: 3939 case X86::VSQRTSSr: 3940 return true; 3941 } 3942} 3943 3944bool X86InstrInfo:: 3945hasHighOperandLatency(const InstrItineraryData *ItinData, 3946 const MachineRegisterInfo *MRI, 3947 const MachineInstr *DefMI, unsigned DefIdx, 3948 const MachineInstr *UseMI, unsigned UseIdx) const { 3949 return isHighLatencyDef(DefMI->getOpcode()); 3950} 3951 3952namespace { 3953 /// CGBR - Create Global Base Reg pass. This initializes the PIC 3954 /// global base register for x86-32. 3955 struct CGBR : public MachineFunctionPass { 3956 static char ID; 3957 CGBR() : MachineFunctionPass(ID) {} 3958 3959 virtual bool runOnMachineFunction(MachineFunction &MF) { 3960 const X86TargetMachine *TM = 3961 static_cast<const X86TargetMachine *>(&MF.getTarget()); 3962 3963 assert(!TM->getSubtarget<X86Subtarget>().is64Bit() && 3964 "X86-64 PIC uses RIP relative addressing"); 3965 3966 // Only emit a global base reg in PIC mode. 3967 if (TM->getRelocationModel() != Reloc::PIC_) 3968 return false; 3969 3970 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 3971 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 3972 3973 // If we didn't need a GlobalBaseReg, don't insert code. 3974 if (GlobalBaseReg == 0) 3975 return false; 3976 3977 // Insert the set of GlobalBaseReg into the first MBB of the function 3978 MachineBasicBlock &FirstMBB = MF.front(); 3979 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 3980 DebugLoc DL = FirstMBB.findDebugLoc(MBBI); 3981 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 3982 const X86InstrInfo *TII = TM->getInstrInfo(); 3983 3984 unsigned PC; 3985 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) 3986 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass); 3987 else 3988 PC = GlobalBaseReg; 3989 3990 // Operand of MovePCtoStack is completely ignored by asm printer. It's 3991 // only used in JIT code emission as displacement to pc. 3992 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0); 3993 3994 // If we're using vanilla 'GOT' PIC style, we should use relative addressing 3995 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external. 3996 if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) { 3997 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register 3998 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg) 3999 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_", 4000 X86II::MO_GOT_ABSOLUTE_ADDRESS); 4001 } 4002 4003 return true; 4004 } 4005 4006 virtual const char *getPassName() const { 4007 return "X86 PIC Global Base Reg Initialization"; 4008 } 4009 4010 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 4011 AU.setPreservesCFG(); 4012 MachineFunctionPass::getAnalysisUsage(AU); 4013 } 4014 }; 4015} 4016 4017char CGBR::ID = 0; 4018FunctionPass* 4019llvm::createGlobalBaseRegPass() { return new CGBR(); } 4020 4021namespace { 4022 struct LDTLSCleanup : public MachineFunctionPass { 4023 static char ID; 4024 LDTLSCleanup() : MachineFunctionPass(ID) {} 4025 4026 virtual bool runOnMachineFunction(MachineFunction &MF) { 4027 X86MachineFunctionInfo* MFI = MF.getInfo<X86MachineFunctionInfo>(); 4028 if (MFI->getNumLocalDynamicTLSAccesses() < 2) { 4029 // No point folding accesses if there isn't at least two. 4030 return false; 4031 } 4032 4033 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>(); 4034 return VisitNode(DT->getRootNode(), 0); 4035 } 4036 4037 // Visit the dominator subtree rooted at Node in pre-order. 4038 // If TLSBaseAddrReg is non-null, then use that to replace any 4039 // TLS_base_addr instructions. Otherwise, create the register 4040 // when the first such instruction is seen, and then use it 4041 // as we encounter more instructions. 4042 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) { 4043 MachineBasicBlock *BB = Node->getBlock(); 4044 bool Changed = false; 4045 4046 // Traverse the current block. 4047 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; 4048 ++I) { 4049 switch (I->getOpcode()) { 4050 case X86::TLS_base_addr32: 4051 case X86::TLS_base_addr64: 4052 if (TLSBaseAddrReg) 4053 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg); 4054 else 4055 I = SetRegister(I, &TLSBaseAddrReg); 4056 Changed = true; 4057 break; 4058 default: 4059 break; 4060 } 4061 } 4062 4063 // Visit the children of this block in the dominator tree. 4064 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end(); 4065 I != E; ++I) { 4066 Changed |= VisitNode(*I, TLSBaseAddrReg); 4067 } 4068 4069 return Changed; 4070 } 4071 4072 // Replace the TLS_base_addr instruction I with a copy from 4073 // TLSBaseAddrReg, returning the new instruction. 4074 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I, 4075 unsigned TLSBaseAddrReg) { 4076 MachineFunction *MF = I->getParent()->getParent(); 4077 const X86TargetMachine *TM = 4078 static_cast<const X86TargetMachine *>(&MF->getTarget()); 4079 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit(); 4080 const X86InstrInfo *TII = TM->getInstrInfo(); 4081 4082 // Insert a Copy from TLSBaseAddrReg to RAX/EAX. 4083 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(), 4084 TII->get(TargetOpcode::COPY), 4085 is64Bit ? X86::RAX : X86::EAX) 4086 .addReg(TLSBaseAddrReg); 4087 4088 // Erase the TLS_base_addr instruction. 4089 I->eraseFromParent(); 4090 4091 return Copy; 4092 } 4093 4094 // Create a virtal register in *TLSBaseAddrReg, and populate it by 4095 // inserting a copy instruction after I. Returns the new instruction. 4096 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) { 4097 MachineFunction *MF = I->getParent()->getParent(); 4098 const X86TargetMachine *TM = 4099 static_cast<const X86TargetMachine *>(&MF->getTarget()); 4100 const bool is64Bit = TM->getSubtarget<X86Subtarget>().is64Bit(); 4101 const X86InstrInfo *TII = TM->getInstrInfo(); 4102 4103 // Create a virtual register for the TLS base address. 4104 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 4105 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit 4106 ? &X86::GR64RegClass 4107 : &X86::GR32RegClass); 4108 4109 // Insert a copy from RAX/EAX to TLSBaseAddrReg. 4110 MachineInstr *Next = I->getNextNode(); 4111 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(), 4112 TII->get(TargetOpcode::COPY), 4113 *TLSBaseAddrReg) 4114 .addReg(is64Bit ? X86::RAX : X86::EAX); 4115 4116 return Copy; 4117 } 4118 4119 virtual const char *getPassName() const { 4120 return "Local Dynamic TLS Access Clean-up"; 4121 } 4122 4123 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 4124 AU.setPreservesCFG(); 4125 AU.addRequired<MachineDominatorTree>(); 4126 MachineFunctionPass::getAnalysisUsage(AU); 4127 } 4128 }; 4129} 4130 4131char LDTLSCleanup::ID = 0; 4132FunctionPass* 4133llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); } 4134