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/ADT/STLExtras.h" 21 #include "llvm/CodeGen/LivePhysRegs.h" 22 #include "llvm/CodeGen/LiveVariables.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/MachineModuleInfo.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 #include "llvm/CodeGen/StackMaps.h" 30 #include "llvm/IR/DerivedTypes.h" 31 #include "llvm/IR/Function.h" 32 #include "llvm/IR/LLVMContext.h" 33 #include "llvm/MC/MCAsmInfo.h" 34 #include "llvm/MC/MCExpr.h" 35 #include "llvm/MC/MCInst.h" 36 #include "llvm/Support/CommandLine.h" 37 #include "llvm/Support/Debug.h" 38 #include "llvm/Support/ErrorHandling.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include "llvm/Target/TargetOptions.h" 41 42 using namespace llvm; 43 44 #define DEBUG_TYPE "x86-instr-info" 45 46 #define GET_INSTRINFO_CTOR_DTOR 47 #include "X86GenInstrInfo.inc" 48 49 static cl::opt<bool> 50 NoFusing("disable-spill-fusing", 51 cl::desc("Disable fusing of spill code into instructions")); 52 static cl::opt<bool> 53 PrintFailedFusing("print-failed-fuse-candidates", 54 cl::desc("Print instructions that the allocator wants to" 55 " fuse, but the X86 backend currently can't"), 56 cl::Hidden); 57 static cl::opt<bool> 58 ReMatPICStubLoad("remat-pic-stub-load", 59 cl::desc("Re-materialize load from stub in PIC mode"), 60 cl::init(false), cl::Hidden); 61 static cl::opt<unsigned> 62 PartialRegUpdateClearance("partial-reg-update-clearance", 63 cl::desc("Clearance between two register writes " 64 "for inserting XOR to avoid partial " 65 "register update"), 66 cl::init(64), cl::Hidden); 67 static cl::opt<unsigned> 68 UndefRegClearance("undef-reg-clearance", 69 cl::desc("How many idle instructions we would like before " 70 "certain undef register reads"), 71 cl::init(64), cl::Hidden); 72 73 enum { 74 // Select which memory operand is being unfolded. 75 // (stored in bits 0 - 3) 76 TB_INDEX_0 = 0, 77 TB_INDEX_1 = 1, 78 TB_INDEX_2 = 2, 79 TB_INDEX_3 = 3, 80 TB_INDEX_4 = 4, 81 TB_INDEX_MASK = 0xf, 82 83 // Do not insert the reverse map (MemOp -> RegOp) into the table. 84 // This may be needed because there is a many -> one mapping. 85 TB_NO_REVERSE = 1 << 4, 86 87 // Do not insert the forward map (RegOp -> MemOp) into the table. 88 // This is needed for Native Client, which prohibits branch 89 // instructions from using a memory operand. 90 TB_NO_FORWARD = 1 << 5, 91 92 TB_FOLDED_LOAD = 1 << 6, 93 TB_FOLDED_STORE = 1 << 7, 94 95 // Minimum alignment required for load/store. 96 // Used for RegOp->MemOp conversion. 97 // (stored in bits 8 - 15) 98 TB_ALIGN_SHIFT = 8, 99 TB_ALIGN_NONE = 0 << TB_ALIGN_SHIFT, 100 TB_ALIGN_16 = 16 << TB_ALIGN_SHIFT, 101 TB_ALIGN_32 = 32 << TB_ALIGN_SHIFT, 102 TB_ALIGN_64 = 64 << TB_ALIGN_SHIFT, 103 TB_ALIGN_MASK = 0xff << TB_ALIGN_SHIFT 104 }; 105 106 struct X86MemoryFoldTableEntry { 107 uint16_t RegOp; 108 uint16_t MemOp; 109 uint16_t Flags; 110 }; 111 112 // Pin the vtable to this file. 113 void X86InstrInfo::anchor() {} 114 115 X86InstrInfo::X86InstrInfo(X86Subtarget &STI) 116 : X86GenInstrInfo((STI.isTarget64BitLP64() ? X86::ADJCALLSTACKDOWN64 117 : X86::ADJCALLSTACKDOWN32), 118 (STI.isTarget64BitLP64() ? X86::ADJCALLSTACKUP64 119 : X86::ADJCALLSTACKUP32), 120 X86::CATCHRET), 121 Subtarget(STI), RI(STI.getTargetTriple()) { 122 123 static const X86MemoryFoldTableEntry MemoryFoldTable2Addr[] = { 124 { X86::ADC32ri, X86::ADC32mi, 0 }, 125 { X86::ADC32ri8, X86::ADC32mi8, 0 }, 126 { X86::ADC32rr, X86::ADC32mr, 0 }, 127 { X86::ADC64ri32, X86::ADC64mi32, 0 }, 128 { X86::ADC64ri8, X86::ADC64mi8, 0 }, 129 { X86::ADC64rr, X86::ADC64mr, 0 }, 130 { X86::ADD16ri, X86::ADD16mi, 0 }, 131 { X86::ADD16ri8, X86::ADD16mi8, 0 }, 132 { X86::ADD16ri_DB, X86::ADD16mi, TB_NO_REVERSE }, 133 { X86::ADD16ri8_DB, X86::ADD16mi8, TB_NO_REVERSE }, 134 { X86::ADD16rr, X86::ADD16mr, 0 }, 135 { X86::ADD16rr_DB, X86::ADD16mr, TB_NO_REVERSE }, 136 { X86::ADD32ri, X86::ADD32mi, 0 }, 137 { X86::ADD32ri8, X86::ADD32mi8, 0 }, 138 { X86::ADD32ri_DB, X86::ADD32mi, TB_NO_REVERSE }, 139 { X86::ADD32ri8_DB, X86::ADD32mi8, TB_NO_REVERSE }, 140 { X86::ADD32rr, X86::ADD32mr, 0 }, 141 { X86::ADD32rr_DB, X86::ADD32mr, TB_NO_REVERSE }, 142 { X86::ADD64ri32, X86::ADD64mi32, 0 }, 143 { X86::ADD64ri8, X86::ADD64mi8, 0 }, 144 { X86::ADD64ri32_DB,X86::ADD64mi32, TB_NO_REVERSE }, 145 { X86::ADD64ri8_DB, X86::ADD64mi8, TB_NO_REVERSE }, 146 { X86::ADD64rr, X86::ADD64mr, 0 }, 147 { X86::ADD64rr_DB, X86::ADD64mr, TB_NO_REVERSE }, 148 { X86::ADD8ri, X86::ADD8mi, 0 }, 149 { X86::ADD8rr, X86::ADD8mr, 0 }, 150 { X86::AND16ri, X86::AND16mi, 0 }, 151 { X86::AND16ri8, X86::AND16mi8, 0 }, 152 { X86::AND16rr, X86::AND16mr, 0 }, 153 { X86::AND32ri, X86::AND32mi, 0 }, 154 { X86::AND32ri8, X86::AND32mi8, 0 }, 155 { X86::AND32rr, X86::AND32mr, 0 }, 156 { X86::AND64ri32, X86::AND64mi32, 0 }, 157 { X86::AND64ri8, X86::AND64mi8, 0 }, 158 { X86::AND64rr, X86::AND64mr, 0 }, 159 { X86::AND8ri, X86::AND8mi, 0 }, 160 { X86::AND8rr, X86::AND8mr, 0 }, 161 { X86::DEC16r, X86::DEC16m, 0 }, 162 { X86::DEC32r, X86::DEC32m, 0 }, 163 { X86::DEC64r, X86::DEC64m, 0 }, 164 { X86::DEC8r, X86::DEC8m, 0 }, 165 { X86::INC16r, X86::INC16m, 0 }, 166 { X86::INC32r, X86::INC32m, 0 }, 167 { X86::INC64r, X86::INC64m, 0 }, 168 { X86::INC8r, X86::INC8m, 0 }, 169 { X86::NEG16r, X86::NEG16m, 0 }, 170 { X86::NEG32r, X86::NEG32m, 0 }, 171 { X86::NEG64r, X86::NEG64m, 0 }, 172 { X86::NEG8r, X86::NEG8m, 0 }, 173 { X86::NOT16r, X86::NOT16m, 0 }, 174 { X86::NOT32r, X86::NOT32m, 0 }, 175 { X86::NOT64r, X86::NOT64m, 0 }, 176 { X86::NOT8r, X86::NOT8m, 0 }, 177 { X86::OR16ri, X86::OR16mi, 0 }, 178 { X86::OR16ri8, X86::OR16mi8, 0 }, 179 { X86::OR16rr, X86::OR16mr, 0 }, 180 { X86::OR32ri, X86::OR32mi, 0 }, 181 { X86::OR32ri8, X86::OR32mi8, 0 }, 182 { X86::OR32rr, X86::OR32mr, 0 }, 183 { X86::OR64ri32, X86::OR64mi32, 0 }, 184 { X86::OR64ri8, X86::OR64mi8, 0 }, 185 { X86::OR64rr, X86::OR64mr, 0 }, 186 { X86::OR8ri, X86::OR8mi, 0 }, 187 { X86::OR8rr, X86::OR8mr, 0 }, 188 { X86::ROL16r1, X86::ROL16m1, 0 }, 189 { X86::ROL16rCL, X86::ROL16mCL, 0 }, 190 { X86::ROL16ri, X86::ROL16mi, 0 }, 191 { X86::ROL32r1, X86::ROL32m1, 0 }, 192 { X86::ROL32rCL, X86::ROL32mCL, 0 }, 193 { X86::ROL32ri, X86::ROL32mi, 0 }, 194 { X86::ROL64r1, X86::ROL64m1, 0 }, 195 { X86::ROL64rCL, X86::ROL64mCL, 0 }, 196 { X86::ROL64ri, X86::ROL64mi, 0 }, 197 { X86::ROL8r1, X86::ROL8m1, 0 }, 198 { X86::ROL8rCL, X86::ROL8mCL, 0 }, 199 { X86::ROL8ri, X86::ROL8mi, 0 }, 200 { X86::ROR16r1, X86::ROR16m1, 0 }, 201 { X86::ROR16rCL, X86::ROR16mCL, 0 }, 202 { X86::ROR16ri, X86::ROR16mi, 0 }, 203 { X86::ROR32r1, X86::ROR32m1, 0 }, 204 { X86::ROR32rCL, X86::ROR32mCL, 0 }, 205 { X86::ROR32ri, X86::ROR32mi, 0 }, 206 { X86::ROR64r1, X86::ROR64m1, 0 }, 207 { X86::ROR64rCL, X86::ROR64mCL, 0 }, 208 { X86::ROR64ri, X86::ROR64mi, 0 }, 209 { X86::ROR8r1, X86::ROR8m1, 0 }, 210 { X86::ROR8rCL, X86::ROR8mCL, 0 }, 211 { X86::ROR8ri, X86::ROR8mi, 0 }, 212 { X86::SAR16r1, X86::SAR16m1, 0 }, 213 { X86::SAR16rCL, X86::SAR16mCL, 0 }, 214 { X86::SAR16ri, X86::SAR16mi, 0 }, 215 { X86::SAR32r1, X86::SAR32m1, 0 }, 216 { X86::SAR32rCL, X86::SAR32mCL, 0 }, 217 { X86::SAR32ri, X86::SAR32mi, 0 }, 218 { X86::SAR64r1, X86::SAR64m1, 0 }, 219 { X86::SAR64rCL, X86::SAR64mCL, 0 }, 220 { X86::SAR64ri, X86::SAR64mi, 0 }, 221 { X86::SAR8r1, X86::SAR8m1, 0 }, 222 { X86::SAR8rCL, X86::SAR8mCL, 0 }, 223 { X86::SAR8ri, X86::SAR8mi, 0 }, 224 { X86::SBB32ri, X86::SBB32mi, 0 }, 225 { X86::SBB32ri8, X86::SBB32mi8, 0 }, 226 { X86::SBB32rr, X86::SBB32mr, 0 }, 227 { X86::SBB64ri32, X86::SBB64mi32, 0 }, 228 { X86::SBB64ri8, X86::SBB64mi8, 0 }, 229 { X86::SBB64rr, X86::SBB64mr, 0 }, 230 { X86::SHL16rCL, X86::SHL16mCL, 0 }, 231 { X86::SHL16ri, X86::SHL16mi, 0 }, 232 { X86::SHL32rCL, X86::SHL32mCL, 0 }, 233 { X86::SHL32ri, X86::SHL32mi, 0 }, 234 { X86::SHL64rCL, X86::SHL64mCL, 0 }, 235 { X86::SHL64ri, X86::SHL64mi, 0 }, 236 { X86::SHL8rCL, X86::SHL8mCL, 0 }, 237 { X86::SHL8ri, X86::SHL8mi, 0 }, 238 { X86::SHLD16rrCL, X86::SHLD16mrCL, 0 }, 239 { X86::SHLD16rri8, X86::SHLD16mri8, 0 }, 240 { X86::SHLD32rrCL, X86::SHLD32mrCL, 0 }, 241 { X86::SHLD32rri8, X86::SHLD32mri8, 0 }, 242 { X86::SHLD64rrCL, X86::SHLD64mrCL, 0 }, 243 { X86::SHLD64rri8, X86::SHLD64mri8, 0 }, 244 { X86::SHR16r1, X86::SHR16m1, 0 }, 245 { X86::SHR16rCL, X86::SHR16mCL, 0 }, 246 { X86::SHR16ri, X86::SHR16mi, 0 }, 247 { X86::SHR32r1, X86::SHR32m1, 0 }, 248 { X86::SHR32rCL, X86::SHR32mCL, 0 }, 249 { X86::SHR32ri, X86::SHR32mi, 0 }, 250 { X86::SHR64r1, X86::SHR64m1, 0 }, 251 { X86::SHR64rCL, X86::SHR64mCL, 0 }, 252 { X86::SHR64ri, X86::SHR64mi, 0 }, 253 { X86::SHR8r1, X86::SHR8m1, 0 }, 254 { X86::SHR8rCL, X86::SHR8mCL, 0 }, 255 { X86::SHR8ri, X86::SHR8mi, 0 }, 256 { X86::SHRD16rrCL, X86::SHRD16mrCL, 0 }, 257 { X86::SHRD16rri8, X86::SHRD16mri8, 0 }, 258 { X86::SHRD32rrCL, X86::SHRD32mrCL, 0 }, 259 { X86::SHRD32rri8, X86::SHRD32mri8, 0 }, 260 { X86::SHRD64rrCL, X86::SHRD64mrCL, 0 }, 261 { X86::SHRD64rri8, X86::SHRD64mri8, 0 }, 262 { X86::SUB16ri, X86::SUB16mi, 0 }, 263 { X86::SUB16ri8, X86::SUB16mi8, 0 }, 264 { X86::SUB16rr, X86::SUB16mr, 0 }, 265 { X86::SUB32ri, X86::SUB32mi, 0 }, 266 { X86::SUB32ri8, X86::SUB32mi8, 0 }, 267 { X86::SUB32rr, X86::SUB32mr, 0 }, 268 { X86::SUB64ri32, X86::SUB64mi32, 0 }, 269 { X86::SUB64ri8, X86::SUB64mi8, 0 }, 270 { X86::SUB64rr, X86::SUB64mr, 0 }, 271 { X86::SUB8ri, X86::SUB8mi, 0 }, 272 { X86::SUB8rr, X86::SUB8mr, 0 }, 273 { X86::XOR16ri, X86::XOR16mi, 0 }, 274 { X86::XOR16ri8, X86::XOR16mi8, 0 }, 275 { X86::XOR16rr, X86::XOR16mr, 0 }, 276 { X86::XOR32ri, X86::XOR32mi, 0 }, 277 { X86::XOR32ri8, X86::XOR32mi8, 0 }, 278 { X86::XOR32rr, X86::XOR32mr, 0 }, 279 { X86::XOR64ri32, X86::XOR64mi32, 0 }, 280 { X86::XOR64ri8, X86::XOR64mi8, 0 }, 281 { X86::XOR64rr, X86::XOR64mr, 0 }, 282 { X86::XOR8ri, X86::XOR8mi, 0 }, 283 { X86::XOR8rr, X86::XOR8mr, 0 } 284 }; 285 286 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2Addr) { 287 AddTableEntry(RegOp2MemOpTable2Addr, MemOp2RegOpTable, 288 Entry.RegOp, Entry.MemOp, 289 // Index 0, folded load and store, no alignment requirement. 290 Entry.Flags | TB_INDEX_0 | TB_FOLDED_LOAD | TB_FOLDED_STORE); 291 } 292 293 static const X86MemoryFoldTableEntry MemoryFoldTable0[] = { 294 { X86::BT16ri8, X86::BT16mi8, TB_FOLDED_LOAD }, 295 { X86::BT32ri8, X86::BT32mi8, TB_FOLDED_LOAD }, 296 { X86::BT64ri8, X86::BT64mi8, TB_FOLDED_LOAD }, 297 { X86::CALL32r, X86::CALL32m, TB_FOLDED_LOAD }, 298 { X86::CALL64r, X86::CALL64m, TB_FOLDED_LOAD }, 299 { X86::CMP16ri, X86::CMP16mi, TB_FOLDED_LOAD }, 300 { X86::CMP16ri8, X86::CMP16mi8, TB_FOLDED_LOAD }, 301 { X86::CMP16rr, X86::CMP16mr, TB_FOLDED_LOAD }, 302 { X86::CMP32ri, X86::CMP32mi, TB_FOLDED_LOAD }, 303 { X86::CMP32ri8, X86::CMP32mi8, TB_FOLDED_LOAD }, 304 { X86::CMP32rr, X86::CMP32mr, TB_FOLDED_LOAD }, 305 { X86::CMP64ri32, X86::CMP64mi32, TB_FOLDED_LOAD }, 306 { X86::CMP64ri8, X86::CMP64mi8, TB_FOLDED_LOAD }, 307 { X86::CMP64rr, X86::CMP64mr, TB_FOLDED_LOAD }, 308 { X86::CMP8ri, X86::CMP8mi, TB_FOLDED_LOAD }, 309 { X86::CMP8rr, X86::CMP8mr, TB_FOLDED_LOAD }, 310 { X86::DIV16r, X86::DIV16m, TB_FOLDED_LOAD }, 311 { X86::DIV32r, X86::DIV32m, TB_FOLDED_LOAD }, 312 { X86::DIV64r, X86::DIV64m, TB_FOLDED_LOAD }, 313 { X86::DIV8r, X86::DIV8m, TB_FOLDED_LOAD }, 314 { X86::EXTRACTPSrr, X86::EXTRACTPSmr, TB_FOLDED_STORE }, 315 { X86::IDIV16r, X86::IDIV16m, TB_FOLDED_LOAD }, 316 { X86::IDIV32r, X86::IDIV32m, TB_FOLDED_LOAD }, 317 { X86::IDIV64r, X86::IDIV64m, TB_FOLDED_LOAD }, 318 { X86::IDIV8r, X86::IDIV8m, TB_FOLDED_LOAD }, 319 { X86::IMUL16r, X86::IMUL16m, TB_FOLDED_LOAD }, 320 { X86::IMUL32r, X86::IMUL32m, TB_FOLDED_LOAD }, 321 { X86::IMUL64r, X86::IMUL64m, TB_FOLDED_LOAD }, 322 { X86::IMUL8r, X86::IMUL8m, TB_FOLDED_LOAD }, 323 { X86::JMP32r, X86::JMP32m, TB_FOLDED_LOAD }, 324 { X86::JMP64r, X86::JMP64m, TB_FOLDED_LOAD }, 325 { X86::MOV16ri, X86::MOV16mi, TB_FOLDED_STORE }, 326 { X86::MOV16rr, X86::MOV16mr, TB_FOLDED_STORE }, 327 { X86::MOV32ri, X86::MOV32mi, TB_FOLDED_STORE }, 328 { X86::MOV32rr, X86::MOV32mr, TB_FOLDED_STORE }, 329 { X86::MOV64ri32, X86::MOV64mi32, TB_FOLDED_STORE }, 330 { X86::MOV64rr, X86::MOV64mr, TB_FOLDED_STORE }, 331 { X86::MOV8ri, X86::MOV8mi, TB_FOLDED_STORE }, 332 { X86::MOV8rr, X86::MOV8mr, TB_FOLDED_STORE }, 333 { X86::MOV8rr_NOREX, X86::MOV8mr_NOREX, TB_FOLDED_STORE }, 334 { X86::MOVAPDrr, X86::MOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 335 { X86::MOVAPSrr, X86::MOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 336 { X86::MOVDQArr, X86::MOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 337 { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, TB_FOLDED_STORE }, 338 { X86::MOVPQIto64rr,X86::MOVPQI2QImr, TB_FOLDED_STORE }, 339 { X86::MOVSDto64rr, X86::MOVSDto64mr, TB_FOLDED_STORE }, 340 { X86::MOVSS2DIrr, X86::MOVSS2DImr, TB_FOLDED_STORE }, 341 { X86::MOVUPDrr, X86::MOVUPDmr, TB_FOLDED_STORE }, 342 { X86::MOVUPSrr, X86::MOVUPSmr, TB_FOLDED_STORE }, 343 { X86::MUL16r, X86::MUL16m, TB_FOLDED_LOAD }, 344 { X86::MUL32r, X86::MUL32m, TB_FOLDED_LOAD }, 345 { X86::MUL64r, X86::MUL64m, TB_FOLDED_LOAD }, 346 { X86::MUL8r, X86::MUL8m, TB_FOLDED_LOAD }, 347 { X86::PEXTRDrr, X86::PEXTRDmr, TB_FOLDED_STORE }, 348 { X86::PEXTRQrr, X86::PEXTRQmr, TB_FOLDED_STORE }, 349 { X86::PUSH16r, X86::PUSH16rmm, TB_FOLDED_LOAD }, 350 { X86::PUSH32r, X86::PUSH32rmm, TB_FOLDED_LOAD }, 351 { X86::PUSH64r, X86::PUSH64rmm, TB_FOLDED_LOAD }, 352 { X86::SETAEr, X86::SETAEm, TB_FOLDED_STORE }, 353 { X86::SETAr, X86::SETAm, TB_FOLDED_STORE }, 354 { X86::SETBEr, X86::SETBEm, TB_FOLDED_STORE }, 355 { X86::SETBr, X86::SETBm, TB_FOLDED_STORE }, 356 { X86::SETEr, X86::SETEm, TB_FOLDED_STORE }, 357 { X86::SETGEr, X86::SETGEm, TB_FOLDED_STORE }, 358 { X86::SETGr, X86::SETGm, TB_FOLDED_STORE }, 359 { X86::SETLEr, X86::SETLEm, TB_FOLDED_STORE }, 360 { X86::SETLr, X86::SETLm, TB_FOLDED_STORE }, 361 { X86::SETNEr, X86::SETNEm, TB_FOLDED_STORE }, 362 { X86::SETNOr, X86::SETNOm, TB_FOLDED_STORE }, 363 { X86::SETNPr, X86::SETNPm, TB_FOLDED_STORE }, 364 { X86::SETNSr, X86::SETNSm, TB_FOLDED_STORE }, 365 { X86::SETOr, X86::SETOm, TB_FOLDED_STORE }, 366 { X86::SETPr, X86::SETPm, TB_FOLDED_STORE }, 367 { X86::SETSr, X86::SETSm, TB_FOLDED_STORE }, 368 { X86::TAILJMPr, X86::TAILJMPm, TB_FOLDED_LOAD }, 369 { X86::TAILJMPr64, X86::TAILJMPm64, TB_FOLDED_LOAD }, 370 { X86::TAILJMPr64_REX, X86::TAILJMPm64_REX, TB_FOLDED_LOAD }, 371 { X86::TEST16ri, X86::TEST16mi, TB_FOLDED_LOAD }, 372 { X86::TEST32ri, X86::TEST32mi, TB_FOLDED_LOAD }, 373 { X86::TEST64ri32, X86::TEST64mi32, TB_FOLDED_LOAD }, 374 { X86::TEST8ri, X86::TEST8mi, TB_FOLDED_LOAD }, 375 376 // AVX 128-bit versions of foldable instructions 377 { X86::VEXTRACTPSrr,X86::VEXTRACTPSmr, TB_FOLDED_STORE }, 378 { X86::VEXTRACTF128rr, X86::VEXTRACTF128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 379 { X86::VMOVAPDrr, X86::VMOVAPDmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 380 { X86::VMOVAPSrr, X86::VMOVAPSmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 381 { X86::VMOVDQArr, X86::VMOVDQAmr, TB_FOLDED_STORE | TB_ALIGN_16 }, 382 { X86::VMOVPDI2DIrr,X86::VMOVPDI2DImr, TB_FOLDED_STORE }, 383 { X86::VMOVPQIto64rr, X86::VMOVPQI2QImr,TB_FOLDED_STORE }, 384 { X86::VMOVSDto64rr,X86::VMOVSDto64mr, TB_FOLDED_STORE }, 385 { X86::VMOVSS2DIrr, X86::VMOVSS2DImr, TB_FOLDED_STORE }, 386 { X86::VMOVUPDrr, X86::VMOVUPDmr, TB_FOLDED_STORE }, 387 { X86::VMOVUPSrr, X86::VMOVUPSmr, TB_FOLDED_STORE }, 388 { X86::VPEXTRDrr, X86::VPEXTRDmr, TB_FOLDED_STORE }, 389 { X86::VPEXTRQrr, X86::VPEXTRQmr, TB_FOLDED_STORE }, 390 391 // AVX 256-bit foldable instructions 392 { X86::VEXTRACTI128rr, X86::VEXTRACTI128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 393 { X86::VMOVAPDYrr, X86::VMOVAPDYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 394 { X86::VMOVAPSYrr, X86::VMOVAPSYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 395 { X86::VMOVDQAYrr, X86::VMOVDQAYmr, TB_FOLDED_STORE | TB_ALIGN_32 }, 396 { X86::VMOVUPDYrr, X86::VMOVUPDYmr, TB_FOLDED_STORE }, 397 { X86::VMOVUPSYrr, X86::VMOVUPSYmr, TB_FOLDED_STORE }, 398 399 // AVX-512 foldable instructions 400 { X86::VMOVPDI2DIZrr, X86::VMOVPDI2DIZmr, TB_FOLDED_STORE }, 401 { X86::VMOVAPDZrr, X86::VMOVAPDZmr, TB_FOLDED_STORE | TB_ALIGN_64 }, 402 { X86::VMOVAPSZrr, X86::VMOVAPSZmr, TB_FOLDED_STORE | TB_ALIGN_64 }, 403 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zmr, TB_FOLDED_STORE | TB_ALIGN_64 }, 404 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zmr, TB_FOLDED_STORE | TB_ALIGN_64 }, 405 { X86::VMOVUPDZrr, X86::VMOVUPDZmr, TB_FOLDED_STORE }, 406 { X86::VMOVUPSZrr, X86::VMOVUPSZmr, TB_FOLDED_STORE }, 407 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zmr, TB_FOLDED_STORE }, 408 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zmr, TB_FOLDED_STORE }, 409 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zmr, TB_FOLDED_STORE }, 410 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zmr, TB_FOLDED_STORE }, 411 412 // AVX-512 foldable instructions (256-bit versions) 413 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 }, 414 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256mr, TB_FOLDED_STORE | TB_ALIGN_32 }, 415 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 }, 416 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256mr, TB_FOLDED_STORE | TB_ALIGN_32 }, 417 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256mr, TB_FOLDED_STORE }, 418 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256mr, TB_FOLDED_STORE }, 419 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256mr, TB_FOLDED_STORE }, 420 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256mr, TB_FOLDED_STORE }, 421 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256mr, TB_FOLDED_STORE }, 422 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256mr, TB_FOLDED_STORE }, 423 424 // AVX-512 foldable instructions (128-bit versions) 425 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 426 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 427 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 428 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128mr, TB_FOLDED_STORE | TB_ALIGN_16 }, 429 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128mr, TB_FOLDED_STORE }, 430 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128mr, TB_FOLDED_STORE }, 431 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128mr, TB_FOLDED_STORE }, 432 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128mr, TB_FOLDED_STORE }, 433 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128mr, TB_FOLDED_STORE }, 434 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128mr, TB_FOLDED_STORE }, 435 436 // F16C foldable instructions 437 { X86::VCVTPS2PHrr, X86::VCVTPS2PHmr, TB_FOLDED_STORE }, 438 { X86::VCVTPS2PHYrr, X86::VCVTPS2PHYmr, TB_FOLDED_STORE } 439 }; 440 441 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable0) { 442 AddTableEntry(RegOp2MemOpTable0, MemOp2RegOpTable, 443 Entry.RegOp, Entry.MemOp, TB_INDEX_0 | Entry.Flags); 444 } 445 446 static const X86MemoryFoldTableEntry MemoryFoldTable1[] = { 447 { X86::BSF16rr, X86::BSF16rm, 0 }, 448 { X86::BSF32rr, X86::BSF32rm, 0 }, 449 { X86::BSF64rr, X86::BSF64rm, 0 }, 450 { X86::BSR16rr, X86::BSR16rm, 0 }, 451 { X86::BSR32rr, X86::BSR32rm, 0 }, 452 { X86::BSR64rr, X86::BSR64rm, 0 }, 453 { X86::CMP16rr, X86::CMP16rm, 0 }, 454 { X86::CMP32rr, X86::CMP32rm, 0 }, 455 { X86::CMP64rr, X86::CMP64rm, 0 }, 456 { X86::CMP8rr, X86::CMP8rm, 0 }, 457 { X86::CVTSD2SSrr, X86::CVTSD2SSrm, 0 }, 458 { X86::CVTSI2SD64rr, X86::CVTSI2SD64rm, 0 }, 459 { X86::CVTSI2SDrr, X86::CVTSI2SDrm, 0 }, 460 { X86::CVTSI2SS64rr, X86::CVTSI2SS64rm, 0 }, 461 { X86::CVTSI2SSrr, X86::CVTSI2SSrm, 0 }, 462 { X86::CVTSS2SDrr, X86::CVTSS2SDrm, 0 }, 463 { X86::CVTTSD2SI64rr, X86::CVTTSD2SI64rm, 0 }, 464 { X86::CVTTSD2SIrr, X86::CVTTSD2SIrm, 0 }, 465 { X86::CVTTSS2SI64rr, X86::CVTTSS2SI64rm, 0 }, 466 { X86::CVTTSS2SIrr, X86::CVTTSS2SIrm, 0 }, 467 { X86::IMUL16rri, X86::IMUL16rmi, 0 }, 468 { X86::IMUL16rri8, X86::IMUL16rmi8, 0 }, 469 { X86::IMUL32rri, X86::IMUL32rmi, 0 }, 470 { X86::IMUL32rri8, X86::IMUL32rmi8, 0 }, 471 { X86::IMUL64rri32, X86::IMUL64rmi32, 0 }, 472 { X86::IMUL64rri8, X86::IMUL64rmi8, 0 }, 473 { X86::Int_COMISDrr, X86::Int_COMISDrm, 0 }, 474 { X86::Int_COMISSrr, X86::Int_COMISSrm, 0 }, 475 { X86::CVTSD2SI64rr, X86::CVTSD2SI64rm, 0 }, 476 { X86::CVTSD2SIrr, X86::CVTSD2SIrm, 0 }, 477 { X86::CVTSS2SI64rr, X86::CVTSS2SI64rm, 0 }, 478 { X86::CVTSS2SIrr, X86::CVTSS2SIrm, 0 }, 479 { X86::CVTDQ2PDrr, X86::CVTDQ2PDrm, TB_ALIGN_16 }, 480 { X86::CVTDQ2PSrr, X86::CVTDQ2PSrm, TB_ALIGN_16 }, 481 { X86::CVTPD2DQrr, X86::CVTPD2DQrm, TB_ALIGN_16 }, 482 { X86::CVTPD2PSrr, X86::CVTPD2PSrm, TB_ALIGN_16 }, 483 { X86::CVTPS2DQrr, X86::CVTPS2DQrm, TB_ALIGN_16 }, 484 { X86::CVTPS2PDrr, X86::CVTPS2PDrm, TB_ALIGN_16 }, 485 { X86::CVTTPD2DQrr, X86::CVTTPD2DQrm, TB_ALIGN_16 }, 486 { X86::CVTTPS2DQrr, X86::CVTTPS2DQrm, TB_ALIGN_16 }, 487 { X86::Int_CVTTSD2SI64rr,X86::Int_CVTTSD2SI64rm, 0 }, 488 { X86::Int_CVTTSD2SIrr, X86::Int_CVTTSD2SIrm, 0 }, 489 { X86::Int_CVTTSS2SI64rr,X86::Int_CVTTSS2SI64rm, 0 }, 490 { X86::Int_CVTTSS2SIrr, X86::Int_CVTTSS2SIrm, 0 }, 491 { X86::Int_UCOMISDrr, X86::Int_UCOMISDrm, 0 }, 492 { X86::Int_UCOMISSrr, X86::Int_UCOMISSrm, 0 }, 493 { X86::MOV16rr, X86::MOV16rm, 0 }, 494 { X86::MOV32rr, X86::MOV32rm, 0 }, 495 { X86::MOV64rr, X86::MOV64rm, 0 }, 496 { X86::MOV64toPQIrr, X86::MOVQI2PQIrm, 0 }, 497 { X86::MOV64toSDrr, X86::MOV64toSDrm, 0 }, 498 { X86::MOV8rr, X86::MOV8rm, 0 }, 499 { X86::MOVAPDrr, X86::MOVAPDrm, TB_ALIGN_16 }, 500 { X86::MOVAPSrr, X86::MOVAPSrm, TB_ALIGN_16 }, 501 { X86::MOVDDUPrr, X86::MOVDDUPrm, 0 }, 502 { X86::MOVDI2PDIrr, X86::MOVDI2PDIrm, 0 }, 503 { X86::MOVDI2SSrr, X86::MOVDI2SSrm, 0 }, 504 { X86::MOVDQArr, X86::MOVDQArm, TB_ALIGN_16 }, 505 { X86::MOVSHDUPrr, X86::MOVSHDUPrm, TB_ALIGN_16 }, 506 { X86::MOVSLDUPrr, X86::MOVSLDUPrm, TB_ALIGN_16 }, 507 { X86::MOVSX16rr8, X86::MOVSX16rm8, 0 }, 508 { X86::MOVSX32rr16, X86::MOVSX32rm16, 0 }, 509 { X86::MOVSX32rr8, X86::MOVSX32rm8, 0 }, 510 { X86::MOVSX64rr16, X86::MOVSX64rm16, 0 }, 511 { X86::MOVSX64rr32, X86::MOVSX64rm32, 0 }, 512 { X86::MOVSX64rr8, X86::MOVSX64rm8, 0 }, 513 { X86::MOVUPDrr, X86::MOVUPDrm, TB_ALIGN_16 }, 514 { X86::MOVUPSrr, X86::MOVUPSrm, 0 }, 515 { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, TB_ALIGN_16 }, 516 { X86::MOVZX16rr8, X86::MOVZX16rm8, 0 }, 517 { X86::MOVZX32rr16, X86::MOVZX32rm16, 0 }, 518 { X86::MOVZX32_NOREXrr8, X86::MOVZX32_NOREXrm8, 0 }, 519 { X86::MOVZX32rr8, X86::MOVZX32rm8, 0 }, 520 { X86::PABSBrr128, X86::PABSBrm128, TB_ALIGN_16 }, 521 { X86::PABSDrr128, X86::PABSDrm128, TB_ALIGN_16 }, 522 { X86::PABSWrr128, X86::PABSWrm128, TB_ALIGN_16 }, 523 { X86::PCMPESTRIrr, X86::PCMPESTRIrm, TB_ALIGN_16 }, 524 { X86::PCMPESTRM128rr, X86::PCMPESTRM128rm, TB_ALIGN_16 }, 525 { X86::PCMPISTRIrr, X86::PCMPISTRIrm, TB_ALIGN_16 }, 526 { X86::PCMPISTRM128rr, X86::PCMPISTRM128rm, TB_ALIGN_16 }, 527 { X86::PHMINPOSUWrr128, X86::PHMINPOSUWrm128, TB_ALIGN_16 }, 528 { X86::PMOVSXBDrr, X86::PMOVSXBDrm, TB_ALIGN_16 }, 529 { X86::PMOVSXBQrr, X86::PMOVSXBQrm, TB_ALIGN_16 }, 530 { X86::PMOVSXBWrr, X86::PMOVSXBWrm, TB_ALIGN_16 }, 531 { X86::PMOVSXDQrr, X86::PMOVSXDQrm, TB_ALIGN_16 }, 532 { X86::PMOVSXWDrr, X86::PMOVSXWDrm, TB_ALIGN_16 }, 533 { X86::PMOVSXWQrr, X86::PMOVSXWQrm, TB_ALIGN_16 }, 534 { X86::PMOVZXBDrr, X86::PMOVZXBDrm, TB_ALIGN_16 }, 535 { X86::PMOVZXBQrr, X86::PMOVZXBQrm, TB_ALIGN_16 }, 536 { X86::PMOVZXBWrr, X86::PMOVZXBWrm, TB_ALIGN_16 }, 537 { X86::PMOVZXDQrr, X86::PMOVZXDQrm, TB_ALIGN_16 }, 538 { X86::PMOVZXWDrr, X86::PMOVZXWDrm, TB_ALIGN_16 }, 539 { X86::PMOVZXWQrr, X86::PMOVZXWQrm, TB_ALIGN_16 }, 540 { X86::PSHUFDri, X86::PSHUFDmi, TB_ALIGN_16 }, 541 { X86::PSHUFHWri, X86::PSHUFHWmi, TB_ALIGN_16 }, 542 { X86::PSHUFLWri, X86::PSHUFLWmi, TB_ALIGN_16 }, 543 { X86::PTESTrr, X86::PTESTrm, TB_ALIGN_16 }, 544 { X86::RCPPSr, X86::RCPPSm, TB_ALIGN_16 }, 545 { X86::RCPSSr, X86::RCPSSm, 0 }, 546 { X86::RCPSSr_Int, X86::RCPSSm_Int, 0 }, 547 { X86::ROUNDPDr, X86::ROUNDPDm, TB_ALIGN_16 }, 548 { X86::ROUNDPSr, X86::ROUNDPSm, TB_ALIGN_16 }, 549 { X86::RSQRTPSr, X86::RSQRTPSm, TB_ALIGN_16 }, 550 { X86::RSQRTSSr, X86::RSQRTSSm, 0 }, 551 { X86::RSQRTSSr_Int, X86::RSQRTSSm_Int, 0 }, 552 { X86::SQRTPDr, X86::SQRTPDm, TB_ALIGN_16 }, 553 { X86::SQRTPSr, X86::SQRTPSm, TB_ALIGN_16 }, 554 { X86::SQRTSDr, X86::SQRTSDm, 0 }, 555 { X86::SQRTSDr_Int, X86::SQRTSDm_Int, 0 }, 556 { X86::SQRTSSr, X86::SQRTSSm, 0 }, 557 { X86::SQRTSSr_Int, X86::SQRTSSm_Int, 0 }, 558 { X86::TEST16rr, X86::TEST16rm, 0 }, 559 { X86::TEST32rr, X86::TEST32rm, 0 }, 560 { X86::TEST64rr, X86::TEST64rm, 0 }, 561 { X86::TEST8rr, X86::TEST8rm, 0 }, 562 // FIXME: TEST*rr EAX,EAX ---> CMP [mem], 0 563 { X86::UCOMISDrr, X86::UCOMISDrm, 0 }, 564 { X86::UCOMISSrr, X86::UCOMISSrm, 0 }, 565 566 // MMX version of foldable instructions 567 { X86::MMX_CVTPD2PIirr, X86::MMX_CVTPD2PIirm, 0 }, 568 { X86::MMX_CVTPI2PDirr, X86::MMX_CVTPI2PDirm, 0 }, 569 { X86::MMX_CVTPS2PIirr, X86::MMX_CVTPS2PIirm, 0 }, 570 { X86::MMX_CVTTPD2PIirr, X86::MMX_CVTTPD2PIirm, 0 }, 571 { X86::MMX_CVTTPS2PIirr, X86::MMX_CVTTPS2PIirm, 0 }, 572 { X86::MMX_MOVD64to64rr, X86::MMX_MOVQ64rm, 0 }, 573 { X86::MMX_PABSBrr64, X86::MMX_PABSBrm64, 0 }, 574 { X86::MMX_PABSDrr64, X86::MMX_PABSDrm64, 0 }, 575 { X86::MMX_PABSWrr64, X86::MMX_PABSWrm64, 0 }, 576 { X86::MMX_PSHUFWri, X86::MMX_PSHUFWmi, 0 }, 577 578 // 3DNow! version of foldable instructions 579 { X86::PF2IDrr, X86::PF2IDrm, 0 }, 580 { X86::PF2IWrr, X86::PF2IWrm, 0 }, 581 { X86::PFRCPrr, X86::PFRCPrm, 0 }, 582 { X86::PFRSQRTrr, X86::PFRSQRTrm, 0 }, 583 { X86::PI2FDrr, X86::PI2FDrm, 0 }, 584 { X86::PI2FWrr, X86::PI2FWrm, 0 }, 585 { X86::PSWAPDrr, X86::PSWAPDrm, 0 }, 586 587 // AVX 128-bit versions of foldable instructions 588 { X86::Int_VCOMISDrr, X86::Int_VCOMISDrm, 0 }, 589 { X86::Int_VCOMISSrr, X86::Int_VCOMISSrm, 0 }, 590 { X86::Int_VUCOMISDrr, X86::Int_VUCOMISDrm, 0 }, 591 { X86::Int_VUCOMISSrr, X86::Int_VUCOMISSrm, 0 }, 592 { X86::VCVTTSD2SI64rr, X86::VCVTTSD2SI64rm, 0 }, 593 { X86::Int_VCVTTSD2SI64rr,X86::Int_VCVTTSD2SI64rm,0 }, 594 { X86::VCVTTSD2SIrr, X86::VCVTTSD2SIrm, 0 }, 595 { X86::Int_VCVTTSD2SIrr,X86::Int_VCVTTSD2SIrm, 0 }, 596 { X86::VCVTTSS2SI64rr, X86::VCVTTSS2SI64rm, 0 }, 597 { X86::Int_VCVTTSS2SI64rr,X86::Int_VCVTTSS2SI64rm,0 }, 598 { X86::VCVTTSS2SIrr, X86::VCVTTSS2SIrm, 0 }, 599 { X86::Int_VCVTTSS2SIrr,X86::Int_VCVTTSS2SIrm, 0 }, 600 { X86::VCVTSD2SI64rr, X86::VCVTSD2SI64rm, 0 }, 601 { X86::VCVTSD2SIrr, X86::VCVTSD2SIrm, 0 }, 602 { X86::VCVTSS2SI64rr, X86::VCVTSS2SI64rm, 0 }, 603 { X86::VCVTSS2SIrr, X86::VCVTSS2SIrm, 0 }, 604 { X86::VCVTDQ2PDrr, X86::VCVTDQ2PDrm, 0 }, 605 { X86::VCVTDQ2PSrr, X86::VCVTDQ2PSrm, 0 }, 606 { X86::VCVTPD2DQrr, X86::VCVTPD2DQXrm, 0 }, 607 { X86::VCVTPD2PSrr, X86::VCVTPD2PSXrm, 0 }, 608 { X86::VCVTPS2DQrr, X86::VCVTPS2DQrm, 0 }, 609 { X86::VCVTPS2PDrr, X86::VCVTPS2PDrm, 0 }, 610 { X86::VCVTTPD2DQrr, X86::VCVTTPD2DQXrm, 0 }, 611 { X86::VCVTTPS2DQrr, X86::VCVTTPS2DQrm, 0 }, 612 { X86::VMOV64toPQIrr, X86::VMOVQI2PQIrm, 0 }, 613 { X86::VMOV64toSDrr, X86::VMOV64toSDrm, 0 }, 614 { X86::VMOVAPDrr, X86::VMOVAPDrm, TB_ALIGN_16 }, 615 { X86::VMOVAPSrr, X86::VMOVAPSrm, TB_ALIGN_16 }, 616 { X86::VMOVDDUPrr, X86::VMOVDDUPrm, 0 }, 617 { X86::VMOVDI2PDIrr, X86::VMOVDI2PDIrm, 0 }, 618 { X86::VMOVDI2SSrr, X86::VMOVDI2SSrm, 0 }, 619 { X86::VMOVDQArr, X86::VMOVDQArm, TB_ALIGN_16 }, 620 { X86::VMOVSLDUPrr, X86::VMOVSLDUPrm, 0 }, 621 { X86::VMOVSHDUPrr, X86::VMOVSHDUPrm, 0 }, 622 { X86::VMOVUPDrr, X86::VMOVUPDrm, 0 }, 623 { X86::VMOVUPSrr, X86::VMOVUPSrm, 0 }, 624 { X86::VMOVZPQILo2PQIrr,X86::VMOVZPQILo2PQIrm, TB_ALIGN_16 }, 625 { X86::VPABSBrr128, X86::VPABSBrm128, 0 }, 626 { X86::VPABSDrr128, X86::VPABSDrm128, 0 }, 627 { X86::VPABSWrr128, X86::VPABSWrm128, 0 }, 628 { X86::VPCMPESTRIrr, X86::VPCMPESTRIrm, 0 }, 629 { X86::VPCMPESTRM128rr, X86::VPCMPESTRM128rm, 0 }, 630 { X86::VPCMPISTRIrr, X86::VPCMPISTRIrm, 0 }, 631 { X86::VPCMPISTRM128rr, X86::VPCMPISTRM128rm, 0 }, 632 { X86::VPHMINPOSUWrr128, X86::VPHMINPOSUWrm128, 0 }, 633 { X86::VPERMILPDri, X86::VPERMILPDmi, 0 }, 634 { X86::VPERMILPSri, X86::VPERMILPSmi, 0 }, 635 { X86::VPMOVSXBDrr, X86::VPMOVSXBDrm, 0 }, 636 { X86::VPMOVSXBQrr, X86::VPMOVSXBQrm, 0 }, 637 { X86::VPMOVSXBWrr, X86::VPMOVSXBWrm, 0 }, 638 { X86::VPMOVSXDQrr, X86::VPMOVSXDQrm, 0 }, 639 { X86::VPMOVSXWDrr, X86::VPMOVSXWDrm, 0 }, 640 { X86::VPMOVSXWQrr, X86::VPMOVSXWQrm, 0 }, 641 { X86::VPMOVZXBDrr, X86::VPMOVZXBDrm, 0 }, 642 { X86::VPMOVZXBQrr, X86::VPMOVZXBQrm, 0 }, 643 { X86::VPMOVZXBWrr, X86::VPMOVZXBWrm, 0 }, 644 { X86::VPMOVZXDQrr, X86::VPMOVZXDQrm, 0 }, 645 { X86::VPMOVZXWDrr, X86::VPMOVZXWDrm, 0 }, 646 { X86::VPMOVZXWQrr, X86::VPMOVZXWQrm, 0 }, 647 { X86::VPSHUFDri, X86::VPSHUFDmi, 0 }, 648 { X86::VPSHUFHWri, X86::VPSHUFHWmi, 0 }, 649 { X86::VPSHUFLWri, X86::VPSHUFLWmi, 0 }, 650 { X86::VPTESTrr, X86::VPTESTrm, 0 }, 651 { X86::VRCPPSr, X86::VRCPPSm, 0 }, 652 { X86::VROUNDPDr, X86::VROUNDPDm, 0 }, 653 { X86::VROUNDPSr, X86::VROUNDPSm, 0 }, 654 { X86::VRSQRTPSr, X86::VRSQRTPSm, 0 }, 655 { X86::VSQRTPDr, X86::VSQRTPDm, 0 }, 656 { X86::VSQRTPSr, X86::VSQRTPSm, 0 }, 657 { X86::VTESTPDrr, X86::VTESTPDrm, 0 }, 658 { X86::VTESTPSrr, X86::VTESTPSrm, 0 }, 659 { X86::VUCOMISDrr, X86::VUCOMISDrm, 0 }, 660 { X86::VUCOMISSrr, X86::VUCOMISSrm, 0 }, 661 662 // AVX 256-bit foldable instructions 663 { X86::VCVTDQ2PDYrr, X86::VCVTDQ2PDYrm, 0 }, 664 { X86::VCVTDQ2PSYrr, X86::VCVTDQ2PSYrm, 0 }, 665 { X86::VCVTPD2DQYrr, X86::VCVTPD2DQYrm, 0 }, 666 { X86::VCVTPD2PSYrr, X86::VCVTPD2PSYrm, 0 }, 667 { X86::VCVTPS2DQYrr, X86::VCVTPS2DQYrm, 0 }, 668 { X86::VCVTPS2PDYrr, X86::VCVTPS2PDYrm, 0 }, 669 { X86::VCVTTPD2DQYrr, X86::VCVTTPD2DQYrm, 0 }, 670 { X86::VCVTTPS2DQYrr, X86::VCVTTPS2DQYrm, 0 }, 671 { X86::VMOVAPDYrr, X86::VMOVAPDYrm, TB_ALIGN_32 }, 672 { X86::VMOVAPSYrr, X86::VMOVAPSYrm, TB_ALIGN_32 }, 673 { X86::VMOVDDUPYrr, X86::VMOVDDUPYrm, 0 }, 674 { X86::VMOVDQAYrr, X86::VMOVDQAYrm, TB_ALIGN_32 }, 675 { X86::VMOVSLDUPYrr, X86::VMOVSLDUPYrm, 0 }, 676 { X86::VMOVSHDUPYrr, X86::VMOVSHDUPYrm, 0 }, 677 { X86::VMOVUPDYrr, X86::VMOVUPDYrm, 0 }, 678 { X86::VMOVUPSYrr, X86::VMOVUPSYrm, 0 }, 679 { X86::VPERMILPDYri, X86::VPERMILPDYmi, 0 }, 680 { X86::VPERMILPSYri, X86::VPERMILPSYmi, 0 }, 681 { X86::VPTESTYrr, X86::VPTESTYrm, 0 }, 682 { X86::VRCPPSYr, X86::VRCPPSYm, 0 }, 683 { X86::VROUNDYPDr, X86::VROUNDYPDm, 0 }, 684 { X86::VROUNDYPSr, X86::VROUNDYPSm, 0 }, 685 { X86::VRSQRTPSYr, X86::VRSQRTPSYm, 0 }, 686 { X86::VSQRTPDYr, X86::VSQRTPDYm, 0 }, 687 { X86::VSQRTPSYr, X86::VSQRTPSYm, 0 }, 688 { X86::VTESTPDYrr, X86::VTESTPDYrm, 0 }, 689 { X86::VTESTPSYrr, X86::VTESTPSYrm, 0 }, 690 691 // AVX2 foldable instructions 692 693 // VBROADCASTS{SD}rr register instructions were an AVX2 addition while the 694 // VBROADCASTS{SD}rm memory instructions were available from AVX1. 695 // TB_NO_REVERSE prevents unfolding from introducing an illegal instruction 696 // on AVX1 targets. The VPBROADCAST instructions are all AVX2 instructions 697 // so they don't need an equivalent limitation. 698 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrm, TB_NO_REVERSE }, 699 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrm, TB_NO_REVERSE }, 700 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrm, TB_NO_REVERSE }, 701 { X86::VPABSBrr256, X86::VPABSBrm256, 0 }, 702 { X86::VPABSDrr256, X86::VPABSDrm256, 0 }, 703 { X86::VPABSWrr256, X86::VPABSWrm256, 0 }, 704 { X86::VPBROADCASTBrr, X86::VPBROADCASTBrm, 0 }, 705 { X86::VPBROADCASTBYrr, X86::VPBROADCASTBYrm, 0 }, 706 { X86::VPBROADCASTDrr, X86::VPBROADCASTDrm, 0 }, 707 { X86::VPBROADCASTDYrr, X86::VPBROADCASTDYrm, 0 }, 708 { X86::VPBROADCASTQrr, X86::VPBROADCASTQrm, 0 }, 709 { X86::VPBROADCASTQYrr, X86::VPBROADCASTQYrm, 0 }, 710 { X86::VPBROADCASTWrr, X86::VPBROADCASTWrm, 0 }, 711 { X86::VPBROADCASTWYrr, X86::VPBROADCASTWYrm, 0 }, 712 { X86::VPERMPDYri, X86::VPERMPDYmi, 0 }, 713 { X86::VPERMQYri, X86::VPERMQYmi, 0 }, 714 { X86::VPMOVSXBDYrr, X86::VPMOVSXBDYrm, 0 }, 715 { X86::VPMOVSXBQYrr, X86::VPMOVSXBQYrm, 0 }, 716 { X86::VPMOVSXBWYrr, X86::VPMOVSXBWYrm, 0 }, 717 { X86::VPMOVSXDQYrr, X86::VPMOVSXDQYrm, 0 }, 718 { X86::VPMOVSXWDYrr, X86::VPMOVSXWDYrm, 0 }, 719 { X86::VPMOVSXWQYrr, X86::VPMOVSXWQYrm, 0 }, 720 { X86::VPMOVZXBDYrr, X86::VPMOVZXBDYrm, 0 }, 721 { X86::VPMOVZXBQYrr, X86::VPMOVZXBQYrm, 0 }, 722 { X86::VPMOVZXBWYrr, X86::VPMOVZXBWYrm, 0 }, 723 { X86::VPMOVZXDQYrr, X86::VPMOVZXDQYrm, 0 }, 724 { X86::VPMOVZXWDYrr, X86::VPMOVZXWDYrm, 0 }, 725 { X86::VPMOVZXWQYrr, X86::VPMOVZXWQYrm, 0 }, 726 { X86::VPSHUFDYri, X86::VPSHUFDYmi, 0 }, 727 { X86::VPSHUFHWYri, X86::VPSHUFHWYmi, 0 }, 728 { X86::VPSHUFLWYri, X86::VPSHUFLWYmi, 0 }, 729 730 // XOP foldable instructions 731 { X86::VFRCZPDrr, X86::VFRCZPDrm, 0 }, 732 { X86::VFRCZPDrrY, X86::VFRCZPDrmY, 0 }, 733 { X86::VFRCZPSrr, X86::VFRCZPSrm, 0 }, 734 { X86::VFRCZPSrrY, X86::VFRCZPSrmY, 0 }, 735 { X86::VFRCZSDrr, X86::VFRCZSDrm, 0 }, 736 { X86::VFRCZSSrr, X86::VFRCZSSrm, 0 }, 737 { X86::VPHADDBDrr, X86::VPHADDBDrm, 0 }, 738 { X86::VPHADDBQrr, X86::VPHADDBQrm, 0 }, 739 { X86::VPHADDBWrr, X86::VPHADDBWrm, 0 }, 740 { X86::VPHADDDQrr, X86::VPHADDDQrm, 0 }, 741 { X86::VPHADDWDrr, X86::VPHADDWDrm, 0 }, 742 { X86::VPHADDWQrr, X86::VPHADDWQrm, 0 }, 743 { X86::VPHADDUBDrr, X86::VPHADDUBDrm, 0 }, 744 { X86::VPHADDUBQrr, X86::VPHADDUBQrm, 0 }, 745 { X86::VPHADDUBWrr, X86::VPHADDUBWrm, 0 }, 746 { X86::VPHADDUDQrr, X86::VPHADDUDQrm, 0 }, 747 { X86::VPHADDUWDrr, X86::VPHADDUWDrm, 0 }, 748 { X86::VPHADDUWQrr, X86::VPHADDUWQrm, 0 }, 749 { X86::VPHSUBBWrr, X86::VPHSUBBWrm, 0 }, 750 { X86::VPHSUBDQrr, X86::VPHSUBDQrm, 0 }, 751 { X86::VPHSUBWDrr, X86::VPHSUBWDrm, 0 }, 752 { X86::VPROTBri, X86::VPROTBmi, 0 }, 753 { X86::VPROTBrr, X86::VPROTBmr, 0 }, 754 { X86::VPROTDri, X86::VPROTDmi, 0 }, 755 { X86::VPROTDrr, X86::VPROTDmr, 0 }, 756 { X86::VPROTQri, X86::VPROTQmi, 0 }, 757 { X86::VPROTQrr, X86::VPROTQmr, 0 }, 758 { X86::VPROTWri, X86::VPROTWmi, 0 }, 759 { X86::VPROTWrr, X86::VPROTWmr, 0 }, 760 { X86::VPSHABrr, X86::VPSHABmr, 0 }, 761 { X86::VPSHADrr, X86::VPSHADmr, 0 }, 762 { X86::VPSHAQrr, X86::VPSHAQmr, 0 }, 763 { X86::VPSHAWrr, X86::VPSHAWmr, 0 }, 764 { X86::VPSHLBrr, X86::VPSHLBmr, 0 }, 765 { X86::VPSHLDrr, X86::VPSHLDmr, 0 }, 766 { X86::VPSHLQrr, X86::VPSHLQmr, 0 }, 767 { X86::VPSHLWrr, X86::VPSHLWmr, 0 }, 768 769 // BMI/BMI2/LZCNT/POPCNT/TBM foldable instructions 770 { X86::BEXTR32rr, X86::BEXTR32rm, 0 }, 771 { X86::BEXTR64rr, X86::BEXTR64rm, 0 }, 772 { X86::BEXTRI32ri, X86::BEXTRI32mi, 0 }, 773 { X86::BEXTRI64ri, X86::BEXTRI64mi, 0 }, 774 { X86::BLCFILL32rr, X86::BLCFILL32rm, 0 }, 775 { X86::BLCFILL64rr, X86::BLCFILL64rm, 0 }, 776 { X86::BLCI32rr, X86::BLCI32rm, 0 }, 777 { X86::BLCI64rr, X86::BLCI64rm, 0 }, 778 { X86::BLCIC32rr, X86::BLCIC32rm, 0 }, 779 { X86::BLCIC64rr, X86::BLCIC64rm, 0 }, 780 { X86::BLCMSK32rr, X86::BLCMSK32rm, 0 }, 781 { X86::BLCMSK64rr, X86::BLCMSK64rm, 0 }, 782 { X86::BLCS32rr, X86::BLCS32rm, 0 }, 783 { X86::BLCS64rr, X86::BLCS64rm, 0 }, 784 { X86::BLSFILL32rr, X86::BLSFILL32rm, 0 }, 785 { X86::BLSFILL64rr, X86::BLSFILL64rm, 0 }, 786 { X86::BLSI32rr, X86::BLSI32rm, 0 }, 787 { X86::BLSI64rr, X86::BLSI64rm, 0 }, 788 { X86::BLSIC32rr, X86::BLSIC32rm, 0 }, 789 { X86::BLSIC64rr, X86::BLSIC64rm, 0 }, 790 { X86::BLSMSK32rr, X86::BLSMSK32rm, 0 }, 791 { X86::BLSMSK64rr, X86::BLSMSK64rm, 0 }, 792 { X86::BLSR32rr, X86::BLSR32rm, 0 }, 793 { X86::BLSR64rr, X86::BLSR64rm, 0 }, 794 { X86::BZHI32rr, X86::BZHI32rm, 0 }, 795 { X86::BZHI64rr, X86::BZHI64rm, 0 }, 796 { X86::LZCNT16rr, X86::LZCNT16rm, 0 }, 797 { X86::LZCNT32rr, X86::LZCNT32rm, 0 }, 798 { X86::LZCNT64rr, X86::LZCNT64rm, 0 }, 799 { X86::POPCNT16rr, X86::POPCNT16rm, 0 }, 800 { X86::POPCNT32rr, X86::POPCNT32rm, 0 }, 801 { X86::POPCNT64rr, X86::POPCNT64rm, 0 }, 802 { X86::RORX32ri, X86::RORX32mi, 0 }, 803 { X86::RORX64ri, X86::RORX64mi, 0 }, 804 { X86::SARX32rr, X86::SARX32rm, 0 }, 805 { X86::SARX64rr, X86::SARX64rm, 0 }, 806 { X86::SHRX32rr, X86::SHRX32rm, 0 }, 807 { X86::SHRX64rr, X86::SHRX64rm, 0 }, 808 { X86::SHLX32rr, X86::SHLX32rm, 0 }, 809 { X86::SHLX64rr, X86::SHLX64rm, 0 }, 810 { X86::T1MSKC32rr, X86::T1MSKC32rm, 0 }, 811 { X86::T1MSKC64rr, X86::T1MSKC64rm, 0 }, 812 { X86::TZCNT16rr, X86::TZCNT16rm, 0 }, 813 { X86::TZCNT32rr, X86::TZCNT32rm, 0 }, 814 { X86::TZCNT64rr, X86::TZCNT64rm, 0 }, 815 { X86::TZMSK32rr, X86::TZMSK32rm, 0 }, 816 { X86::TZMSK64rr, X86::TZMSK64rm, 0 }, 817 818 // AVX-512 foldable instructions 819 { X86::VMOV64toPQIZrr, X86::VMOVQI2PQIZrm, 0 }, 820 { X86::VMOVDI2SSZrr, X86::VMOVDI2SSZrm, 0 }, 821 { X86::VMOVAPDZrr, X86::VMOVAPDZrm, TB_ALIGN_64 }, 822 { X86::VMOVAPSZrr, X86::VMOVAPSZrm, TB_ALIGN_64 }, 823 { X86::VMOVDQA32Zrr, X86::VMOVDQA32Zrm, TB_ALIGN_64 }, 824 { X86::VMOVDQA64Zrr, X86::VMOVDQA64Zrm, TB_ALIGN_64 }, 825 { X86::VMOVDQU8Zrr, X86::VMOVDQU8Zrm, 0 }, 826 { X86::VMOVDQU16Zrr, X86::VMOVDQU16Zrm, 0 }, 827 { X86::VMOVDQU32Zrr, X86::VMOVDQU32Zrm, 0 }, 828 { X86::VMOVDQU64Zrr, X86::VMOVDQU64Zrm, 0 }, 829 { X86::VMOVUPDZrr, X86::VMOVUPDZrm, 0 }, 830 { X86::VMOVUPSZrr, X86::VMOVUPSZrm, 0 }, 831 { X86::VPABSDZrr, X86::VPABSDZrm, 0 }, 832 { X86::VPABSQZrr, X86::VPABSQZrm, 0 }, 833 { X86::VBROADCASTSSZr, X86::VBROADCASTSSZm, TB_NO_REVERSE }, 834 { X86::VBROADCASTSSZr_s, X86::VBROADCASTSSZm, TB_NO_REVERSE }, 835 { X86::VBROADCASTSDZr, X86::VBROADCASTSDZm, TB_NO_REVERSE }, 836 { X86::VBROADCASTSDZr_s, X86::VBROADCASTSDZm, TB_NO_REVERSE }, 837 838 // AVX-512 foldable instructions (256-bit versions) 839 { X86::VMOVAPDZ256rr, X86::VMOVAPDZ256rm, TB_ALIGN_32 }, 840 { X86::VMOVAPSZ256rr, X86::VMOVAPSZ256rm, TB_ALIGN_32 }, 841 { X86::VMOVDQA32Z256rr, X86::VMOVDQA32Z256rm, TB_ALIGN_32 }, 842 { X86::VMOVDQA64Z256rr, X86::VMOVDQA64Z256rm, TB_ALIGN_32 }, 843 { X86::VMOVDQU8Z256rr, X86::VMOVDQU8Z256rm, 0 }, 844 { X86::VMOVDQU16Z256rr, X86::VMOVDQU16Z256rm, 0 }, 845 { X86::VMOVDQU32Z256rr, X86::VMOVDQU32Z256rm, 0 }, 846 { X86::VMOVDQU64Z256rr, X86::VMOVDQU64Z256rm, 0 }, 847 { X86::VMOVUPDZ256rr, X86::VMOVUPDZ256rm, 0 }, 848 { X86::VMOVUPSZ256rr, X86::VMOVUPSZ256rm, 0 }, 849 { X86::VBROADCASTSSZ256r, X86::VBROADCASTSSZ256m, TB_NO_REVERSE }, 850 { X86::VBROADCASTSSZ256r_s, X86::VBROADCASTSSZ256m, TB_NO_REVERSE }, 851 { X86::VBROADCASTSDZ256r, X86::VBROADCASTSDZ256m, TB_NO_REVERSE }, 852 { X86::VBROADCASTSDZ256r_s, X86::VBROADCASTSDZ256m, TB_NO_REVERSE }, 853 854 // AVX-512 foldable instructions (128-bit versions) 855 { X86::VMOVAPDZ128rr, X86::VMOVAPDZ128rm, TB_ALIGN_16 }, 856 { X86::VMOVAPSZ128rr, X86::VMOVAPSZ128rm, TB_ALIGN_16 }, 857 { X86::VMOVDQA32Z128rr, X86::VMOVDQA32Z128rm, TB_ALIGN_16 }, 858 { X86::VMOVDQA64Z128rr, X86::VMOVDQA64Z128rm, TB_ALIGN_16 }, 859 { X86::VMOVDQU8Z128rr, X86::VMOVDQU8Z128rm, 0 }, 860 { X86::VMOVDQU16Z128rr, X86::VMOVDQU16Z128rm, 0 }, 861 { X86::VMOVDQU32Z128rr, X86::VMOVDQU32Z128rm, 0 }, 862 { X86::VMOVDQU64Z128rr, X86::VMOVDQU64Z128rm, 0 }, 863 { X86::VMOVUPDZ128rr, X86::VMOVUPDZ128rm, 0 }, 864 { X86::VMOVUPSZ128rr, X86::VMOVUPSZ128rm, 0 }, 865 { X86::VBROADCASTSSZ128r, X86::VBROADCASTSSZ128m, TB_NO_REVERSE }, 866 { X86::VBROADCASTSSZ128r_s, X86::VBROADCASTSSZ128m, TB_NO_REVERSE }, 867 // F16C foldable instructions 868 { X86::VCVTPH2PSrr, X86::VCVTPH2PSrm, 0 }, 869 { X86::VCVTPH2PSYrr, X86::VCVTPH2PSYrm, 0 }, 870 871 // AES foldable instructions 872 { X86::AESIMCrr, X86::AESIMCrm, TB_ALIGN_16 }, 873 { X86::AESKEYGENASSIST128rr, X86::AESKEYGENASSIST128rm, TB_ALIGN_16 }, 874 { X86::VAESIMCrr, X86::VAESIMCrm, 0 }, 875 { X86::VAESKEYGENASSIST128rr, X86::VAESKEYGENASSIST128rm, 0 } 876 }; 877 878 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable1) { 879 AddTableEntry(RegOp2MemOpTable1, MemOp2RegOpTable, 880 Entry.RegOp, Entry.MemOp, 881 // Index 1, folded load 882 Entry.Flags | TB_INDEX_1 | TB_FOLDED_LOAD); 883 } 884 885 static const X86MemoryFoldTableEntry MemoryFoldTable2[] = { 886 { X86::ADC32rr, X86::ADC32rm, 0 }, 887 { X86::ADC64rr, X86::ADC64rm, 0 }, 888 { X86::ADD16rr, X86::ADD16rm, 0 }, 889 { X86::ADD16rr_DB, X86::ADD16rm, TB_NO_REVERSE }, 890 { X86::ADD32rr, X86::ADD32rm, 0 }, 891 { X86::ADD32rr_DB, X86::ADD32rm, TB_NO_REVERSE }, 892 { X86::ADD64rr, X86::ADD64rm, 0 }, 893 { X86::ADD64rr_DB, X86::ADD64rm, TB_NO_REVERSE }, 894 { X86::ADD8rr, X86::ADD8rm, 0 }, 895 { X86::ADDPDrr, X86::ADDPDrm, TB_ALIGN_16 }, 896 { X86::ADDPSrr, X86::ADDPSrm, TB_ALIGN_16 }, 897 { X86::ADDSDrr, X86::ADDSDrm, 0 }, 898 { X86::ADDSDrr_Int, X86::ADDSDrm_Int, 0 }, 899 { X86::ADDSSrr, X86::ADDSSrm, 0 }, 900 { X86::ADDSSrr_Int, X86::ADDSSrm_Int, 0 }, 901 { X86::ADDSUBPDrr, X86::ADDSUBPDrm, TB_ALIGN_16 }, 902 { X86::ADDSUBPSrr, X86::ADDSUBPSrm, TB_ALIGN_16 }, 903 { X86::AND16rr, X86::AND16rm, 0 }, 904 { X86::AND32rr, X86::AND32rm, 0 }, 905 { X86::AND64rr, X86::AND64rm, 0 }, 906 { X86::AND8rr, X86::AND8rm, 0 }, 907 { X86::ANDNPDrr, X86::ANDNPDrm, TB_ALIGN_16 }, 908 { X86::ANDNPSrr, X86::ANDNPSrm, TB_ALIGN_16 }, 909 { X86::ANDPDrr, X86::ANDPDrm, TB_ALIGN_16 }, 910 { X86::ANDPSrr, X86::ANDPSrm, TB_ALIGN_16 }, 911 { X86::BLENDPDrri, X86::BLENDPDrmi, TB_ALIGN_16 }, 912 { X86::BLENDPSrri, X86::BLENDPSrmi, TB_ALIGN_16 }, 913 { X86::BLENDVPDrr0, X86::BLENDVPDrm0, TB_ALIGN_16 }, 914 { X86::BLENDVPSrr0, X86::BLENDVPSrm0, TB_ALIGN_16 }, 915 { X86::CMOVA16rr, X86::CMOVA16rm, 0 }, 916 { X86::CMOVA32rr, X86::CMOVA32rm, 0 }, 917 { X86::CMOVA64rr, X86::CMOVA64rm, 0 }, 918 { X86::CMOVAE16rr, X86::CMOVAE16rm, 0 }, 919 { X86::CMOVAE32rr, X86::CMOVAE32rm, 0 }, 920 { X86::CMOVAE64rr, X86::CMOVAE64rm, 0 }, 921 { X86::CMOVB16rr, X86::CMOVB16rm, 0 }, 922 { X86::CMOVB32rr, X86::CMOVB32rm, 0 }, 923 { X86::CMOVB64rr, X86::CMOVB64rm, 0 }, 924 { X86::CMOVBE16rr, X86::CMOVBE16rm, 0 }, 925 { X86::CMOVBE32rr, X86::CMOVBE32rm, 0 }, 926 { X86::CMOVBE64rr, X86::CMOVBE64rm, 0 }, 927 { X86::CMOVE16rr, X86::CMOVE16rm, 0 }, 928 { X86::CMOVE32rr, X86::CMOVE32rm, 0 }, 929 { X86::CMOVE64rr, X86::CMOVE64rm, 0 }, 930 { X86::CMOVG16rr, X86::CMOVG16rm, 0 }, 931 { X86::CMOVG32rr, X86::CMOVG32rm, 0 }, 932 { X86::CMOVG64rr, X86::CMOVG64rm, 0 }, 933 { X86::CMOVGE16rr, X86::CMOVGE16rm, 0 }, 934 { X86::CMOVGE32rr, X86::CMOVGE32rm, 0 }, 935 { X86::CMOVGE64rr, X86::CMOVGE64rm, 0 }, 936 { X86::CMOVL16rr, X86::CMOVL16rm, 0 }, 937 { X86::CMOVL32rr, X86::CMOVL32rm, 0 }, 938 { X86::CMOVL64rr, X86::CMOVL64rm, 0 }, 939 { X86::CMOVLE16rr, X86::CMOVLE16rm, 0 }, 940 { X86::CMOVLE32rr, X86::CMOVLE32rm, 0 }, 941 { X86::CMOVLE64rr, X86::CMOVLE64rm, 0 }, 942 { X86::CMOVNE16rr, X86::CMOVNE16rm, 0 }, 943 { X86::CMOVNE32rr, X86::CMOVNE32rm, 0 }, 944 { X86::CMOVNE64rr, X86::CMOVNE64rm, 0 }, 945 { X86::CMOVNO16rr, X86::CMOVNO16rm, 0 }, 946 { X86::CMOVNO32rr, X86::CMOVNO32rm, 0 }, 947 { X86::CMOVNO64rr, X86::CMOVNO64rm, 0 }, 948 { X86::CMOVNP16rr, X86::CMOVNP16rm, 0 }, 949 { X86::CMOVNP32rr, X86::CMOVNP32rm, 0 }, 950 { X86::CMOVNP64rr, X86::CMOVNP64rm, 0 }, 951 { X86::CMOVNS16rr, X86::CMOVNS16rm, 0 }, 952 { X86::CMOVNS32rr, X86::CMOVNS32rm, 0 }, 953 { X86::CMOVNS64rr, X86::CMOVNS64rm, 0 }, 954 { X86::CMOVO16rr, X86::CMOVO16rm, 0 }, 955 { X86::CMOVO32rr, X86::CMOVO32rm, 0 }, 956 { X86::CMOVO64rr, X86::CMOVO64rm, 0 }, 957 { X86::CMOVP16rr, X86::CMOVP16rm, 0 }, 958 { X86::CMOVP32rr, X86::CMOVP32rm, 0 }, 959 { X86::CMOVP64rr, X86::CMOVP64rm, 0 }, 960 { X86::CMOVS16rr, X86::CMOVS16rm, 0 }, 961 { X86::CMOVS32rr, X86::CMOVS32rm, 0 }, 962 { X86::CMOVS64rr, X86::CMOVS64rm, 0 }, 963 { X86::CMPPDrri, X86::CMPPDrmi, TB_ALIGN_16 }, 964 { X86::CMPPSrri, X86::CMPPSrmi, TB_ALIGN_16 }, 965 { X86::CMPSDrr, X86::CMPSDrm, 0 }, 966 { X86::CMPSSrr, X86::CMPSSrm, 0 }, 967 { X86::CRC32r32r32, X86::CRC32r32m32, 0 }, 968 { X86::CRC32r64r64, X86::CRC32r64m64, 0 }, 969 { X86::DIVPDrr, X86::DIVPDrm, TB_ALIGN_16 }, 970 { X86::DIVPSrr, X86::DIVPSrm, TB_ALIGN_16 }, 971 { X86::DIVSDrr, X86::DIVSDrm, 0 }, 972 { X86::DIVSDrr_Int, X86::DIVSDrm_Int, 0 }, 973 { X86::DIVSSrr, X86::DIVSSrm, 0 }, 974 { X86::DIVSSrr_Int, X86::DIVSSrm_Int, 0 }, 975 { X86::DPPDrri, X86::DPPDrmi, TB_ALIGN_16 }, 976 { X86::DPPSrri, X86::DPPSrmi, TB_ALIGN_16 }, 977 978 // Do not fold Fs* scalar logical op loads because there are no scalar 979 // load variants for these instructions. When folded, the load is required 980 // to be 128-bits, so the load size would not match. 981 982 { X86::FvANDNPDrr, X86::FvANDNPDrm, TB_ALIGN_16 }, 983 { X86::FvANDNPSrr, X86::FvANDNPSrm, TB_ALIGN_16 }, 984 { X86::FvANDPDrr, X86::FvANDPDrm, TB_ALIGN_16 }, 985 { X86::FvANDPSrr, X86::FvANDPSrm, TB_ALIGN_16 }, 986 { X86::FvORPDrr, X86::FvORPDrm, TB_ALIGN_16 }, 987 { X86::FvORPSrr, X86::FvORPSrm, TB_ALIGN_16 }, 988 { X86::FvXORPDrr, X86::FvXORPDrm, TB_ALIGN_16 }, 989 { X86::FvXORPSrr, X86::FvXORPSrm, TB_ALIGN_16 }, 990 { X86::HADDPDrr, X86::HADDPDrm, TB_ALIGN_16 }, 991 { X86::HADDPSrr, X86::HADDPSrm, TB_ALIGN_16 }, 992 { X86::HSUBPDrr, X86::HSUBPDrm, TB_ALIGN_16 }, 993 { X86::HSUBPSrr, X86::HSUBPSrm, TB_ALIGN_16 }, 994 { X86::IMUL16rr, X86::IMUL16rm, 0 }, 995 { X86::IMUL32rr, X86::IMUL32rm, 0 }, 996 { X86::IMUL64rr, X86::IMUL64rm, 0 }, 997 { X86::Int_CMPSDrr, X86::Int_CMPSDrm, 0 }, 998 { X86::Int_CMPSSrr, X86::Int_CMPSSrm, 0 }, 999 { X86::Int_CVTSD2SSrr, X86::Int_CVTSD2SSrm, 0 }, 1000 { X86::Int_CVTSI2SD64rr,X86::Int_CVTSI2SD64rm, 0 }, 1001 { X86::Int_CVTSI2SDrr, X86::Int_CVTSI2SDrm, 0 }, 1002 { X86::Int_CVTSI2SS64rr,X86::Int_CVTSI2SS64rm, 0 }, 1003 { X86::Int_CVTSI2SSrr, X86::Int_CVTSI2SSrm, 0 }, 1004 { X86::Int_CVTSS2SDrr, X86::Int_CVTSS2SDrm, 0 }, 1005 { X86::MAXPDrr, X86::MAXPDrm, TB_ALIGN_16 }, 1006 { X86::MAXPSrr, X86::MAXPSrm, TB_ALIGN_16 }, 1007 { X86::MAXSDrr, X86::MAXSDrm, 0 }, 1008 { X86::MAXSDrr_Int, X86::MAXSDrm_Int, 0 }, 1009 { X86::MAXSSrr, X86::MAXSSrm, 0 }, 1010 { X86::MAXSSrr_Int, X86::MAXSSrm_Int, 0 }, 1011 { X86::MINPDrr, X86::MINPDrm, TB_ALIGN_16 }, 1012 { X86::MINPSrr, X86::MINPSrm, TB_ALIGN_16 }, 1013 { X86::MINSDrr, X86::MINSDrm, 0 }, 1014 { X86::MINSDrr_Int, X86::MINSDrm_Int, 0 }, 1015 { X86::MINSSrr, X86::MINSSrm, 0 }, 1016 { X86::MINSSrr_Int, X86::MINSSrm_Int, 0 }, 1017 { X86::MOVLHPSrr, X86::MOVHPSrm, TB_NO_REVERSE }, 1018 { X86::MPSADBWrri, X86::MPSADBWrmi, TB_ALIGN_16 }, 1019 { X86::MULPDrr, X86::MULPDrm, TB_ALIGN_16 }, 1020 { X86::MULPSrr, X86::MULPSrm, TB_ALIGN_16 }, 1021 { X86::MULSDrr, X86::MULSDrm, 0 }, 1022 { X86::MULSDrr_Int, X86::MULSDrm_Int, 0 }, 1023 { X86::MULSSrr, X86::MULSSrm, 0 }, 1024 { X86::MULSSrr_Int, X86::MULSSrm_Int, 0 }, 1025 { X86::OR16rr, X86::OR16rm, 0 }, 1026 { X86::OR32rr, X86::OR32rm, 0 }, 1027 { X86::OR64rr, X86::OR64rm, 0 }, 1028 { X86::OR8rr, X86::OR8rm, 0 }, 1029 { X86::ORPDrr, X86::ORPDrm, TB_ALIGN_16 }, 1030 { X86::ORPSrr, X86::ORPSrm, TB_ALIGN_16 }, 1031 { X86::PACKSSDWrr, X86::PACKSSDWrm, TB_ALIGN_16 }, 1032 { X86::PACKSSWBrr, X86::PACKSSWBrm, TB_ALIGN_16 }, 1033 { X86::PACKUSDWrr, X86::PACKUSDWrm, TB_ALIGN_16 }, 1034 { X86::PACKUSWBrr, X86::PACKUSWBrm, TB_ALIGN_16 }, 1035 { X86::PADDBrr, X86::PADDBrm, TB_ALIGN_16 }, 1036 { X86::PADDDrr, X86::PADDDrm, TB_ALIGN_16 }, 1037 { X86::PADDQrr, X86::PADDQrm, TB_ALIGN_16 }, 1038 { X86::PADDSBrr, X86::PADDSBrm, TB_ALIGN_16 }, 1039 { X86::PADDSWrr, X86::PADDSWrm, TB_ALIGN_16 }, 1040 { X86::PADDUSBrr, X86::PADDUSBrm, TB_ALIGN_16 }, 1041 { X86::PADDUSWrr, X86::PADDUSWrm, TB_ALIGN_16 }, 1042 { X86::PADDWrr, X86::PADDWrm, TB_ALIGN_16 }, 1043 { X86::PALIGNRrri, X86::PALIGNRrmi, TB_ALIGN_16 }, 1044 { X86::PANDNrr, X86::PANDNrm, TB_ALIGN_16 }, 1045 { X86::PANDrr, X86::PANDrm, TB_ALIGN_16 }, 1046 { X86::PAVGBrr, X86::PAVGBrm, TB_ALIGN_16 }, 1047 { X86::PAVGWrr, X86::PAVGWrm, TB_ALIGN_16 }, 1048 { X86::PBLENDVBrr0, X86::PBLENDVBrm0, TB_ALIGN_16 }, 1049 { X86::PBLENDWrri, X86::PBLENDWrmi, TB_ALIGN_16 }, 1050 { X86::PCLMULQDQrr, X86::PCLMULQDQrm, TB_ALIGN_16 }, 1051 { X86::PCMPEQBrr, X86::PCMPEQBrm, TB_ALIGN_16 }, 1052 { X86::PCMPEQDrr, X86::PCMPEQDrm, TB_ALIGN_16 }, 1053 { X86::PCMPEQQrr, X86::PCMPEQQrm, TB_ALIGN_16 }, 1054 { X86::PCMPEQWrr, X86::PCMPEQWrm, TB_ALIGN_16 }, 1055 { X86::PCMPGTBrr, X86::PCMPGTBrm, TB_ALIGN_16 }, 1056 { X86::PCMPGTDrr, X86::PCMPGTDrm, TB_ALIGN_16 }, 1057 { X86::PCMPGTQrr, X86::PCMPGTQrm, TB_ALIGN_16 }, 1058 { X86::PCMPGTWrr, X86::PCMPGTWrm, TB_ALIGN_16 }, 1059 { X86::PHADDDrr, X86::PHADDDrm, TB_ALIGN_16 }, 1060 { X86::PHADDWrr, X86::PHADDWrm, TB_ALIGN_16 }, 1061 { X86::PHADDSWrr128, X86::PHADDSWrm128, TB_ALIGN_16 }, 1062 { X86::PHSUBDrr, X86::PHSUBDrm, TB_ALIGN_16 }, 1063 { X86::PHSUBSWrr128, X86::PHSUBSWrm128, TB_ALIGN_16 }, 1064 { X86::PHSUBWrr, X86::PHSUBWrm, TB_ALIGN_16 }, 1065 { X86::PINSRBrr, X86::PINSRBrm, 0 }, 1066 { X86::PINSRDrr, X86::PINSRDrm, 0 }, 1067 { X86::PINSRQrr, X86::PINSRQrm, 0 }, 1068 { X86::PINSRWrri, X86::PINSRWrmi, 0 }, 1069 { X86::PMADDUBSWrr128, X86::PMADDUBSWrm128, TB_ALIGN_16 }, 1070 { X86::PMADDWDrr, X86::PMADDWDrm, TB_ALIGN_16 }, 1071 { X86::PMAXSWrr, X86::PMAXSWrm, TB_ALIGN_16 }, 1072 { X86::PMAXUBrr, X86::PMAXUBrm, TB_ALIGN_16 }, 1073 { X86::PMINSWrr, X86::PMINSWrm, TB_ALIGN_16 }, 1074 { X86::PMINUBrr, X86::PMINUBrm, TB_ALIGN_16 }, 1075 { X86::PMINSBrr, X86::PMINSBrm, TB_ALIGN_16 }, 1076 { X86::PMINSDrr, X86::PMINSDrm, TB_ALIGN_16 }, 1077 { X86::PMINUDrr, X86::PMINUDrm, TB_ALIGN_16 }, 1078 { X86::PMINUWrr, X86::PMINUWrm, TB_ALIGN_16 }, 1079 { X86::PMAXSBrr, X86::PMAXSBrm, TB_ALIGN_16 }, 1080 { X86::PMAXSDrr, X86::PMAXSDrm, TB_ALIGN_16 }, 1081 { X86::PMAXUDrr, X86::PMAXUDrm, TB_ALIGN_16 }, 1082 { X86::PMAXUWrr, X86::PMAXUWrm, TB_ALIGN_16 }, 1083 { X86::PMULDQrr, X86::PMULDQrm, TB_ALIGN_16 }, 1084 { X86::PMULHRSWrr128, X86::PMULHRSWrm128, TB_ALIGN_16 }, 1085 { X86::PMULHUWrr, X86::PMULHUWrm, TB_ALIGN_16 }, 1086 { X86::PMULHWrr, X86::PMULHWrm, TB_ALIGN_16 }, 1087 { X86::PMULLDrr, X86::PMULLDrm, TB_ALIGN_16 }, 1088 { X86::PMULLWrr, X86::PMULLWrm, TB_ALIGN_16 }, 1089 { X86::PMULUDQrr, X86::PMULUDQrm, TB_ALIGN_16 }, 1090 { X86::PORrr, X86::PORrm, TB_ALIGN_16 }, 1091 { X86::PSADBWrr, X86::PSADBWrm, TB_ALIGN_16 }, 1092 { X86::PSHUFBrr, X86::PSHUFBrm, TB_ALIGN_16 }, 1093 { X86::PSIGNBrr128, X86::PSIGNBrm128, TB_ALIGN_16 }, 1094 { X86::PSIGNWrr128, X86::PSIGNWrm128, TB_ALIGN_16 }, 1095 { X86::PSIGNDrr128, X86::PSIGNDrm128, TB_ALIGN_16 }, 1096 { X86::PSLLDrr, X86::PSLLDrm, TB_ALIGN_16 }, 1097 { X86::PSLLQrr, X86::PSLLQrm, TB_ALIGN_16 }, 1098 { X86::PSLLWrr, X86::PSLLWrm, TB_ALIGN_16 }, 1099 { X86::PSRADrr, X86::PSRADrm, TB_ALIGN_16 }, 1100 { X86::PSRAWrr, X86::PSRAWrm, TB_ALIGN_16 }, 1101 { X86::PSRLDrr, X86::PSRLDrm, TB_ALIGN_16 }, 1102 { X86::PSRLQrr, X86::PSRLQrm, TB_ALIGN_16 }, 1103 { X86::PSRLWrr, X86::PSRLWrm, TB_ALIGN_16 }, 1104 { X86::PSUBBrr, X86::PSUBBrm, TB_ALIGN_16 }, 1105 { X86::PSUBDrr, X86::PSUBDrm, TB_ALIGN_16 }, 1106 { X86::PSUBQrr, X86::PSUBQrm, TB_ALIGN_16 }, 1107 { X86::PSUBSBrr, X86::PSUBSBrm, TB_ALIGN_16 }, 1108 { X86::PSUBSWrr, X86::PSUBSWrm, TB_ALIGN_16 }, 1109 { X86::PSUBUSBrr, X86::PSUBUSBrm, TB_ALIGN_16 }, 1110 { X86::PSUBUSWrr, X86::PSUBUSWrm, TB_ALIGN_16 }, 1111 { X86::PSUBWrr, X86::PSUBWrm, TB_ALIGN_16 }, 1112 { X86::PUNPCKHBWrr, X86::PUNPCKHBWrm, TB_ALIGN_16 }, 1113 { X86::PUNPCKHDQrr, X86::PUNPCKHDQrm, TB_ALIGN_16 }, 1114 { X86::PUNPCKHQDQrr, X86::PUNPCKHQDQrm, TB_ALIGN_16 }, 1115 { X86::PUNPCKHWDrr, X86::PUNPCKHWDrm, TB_ALIGN_16 }, 1116 { X86::PUNPCKLBWrr, X86::PUNPCKLBWrm, TB_ALIGN_16 }, 1117 { X86::PUNPCKLDQrr, X86::PUNPCKLDQrm, TB_ALIGN_16 }, 1118 { X86::PUNPCKLQDQrr, X86::PUNPCKLQDQrm, TB_ALIGN_16 }, 1119 { X86::PUNPCKLWDrr, X86::PUNPCKLWDrm, TB_ALIGN_16 }, 1120 { X86::PXORrr, X86::PXORrm, TB_ALIGN_16 }, 1121 { X86::ROUNDSDr, X86::ROUNDSDm, 0 }, 1122 { X86::ROUNDSSr, X86::ROUNDSSm, 0 }, 1123 { X86::SBB32rr, X86::SBB32rm, 0 }, 1124 { X86::SBB64rr, X86::SBB64rm, 0 }, 1125 { X86::SHUFPDrri, X86::SHUFPDrmi, TB_ALIGN_16 }, 1126 { X86::SHUFPSrri, X86::SHUFPSrmi, TB_ALIGN_16 }, 1127 { X86::SUB16rr, X86::SUB16rm, 0 }, 1128 { X86::SUB32rr, X86::SUB32rm, 0 }, 1129 { X86::SUB64rr, X86::SUB64rm, 0 }, 1130 { X86::SUB8rr, X86::SUB8rm, 0 }, 1131 { X86::SUBPDrr, X86::SUBPDrm, TB_ALIGN_16 }, 1132 { X86::SUBPSrr, X86::SUBPSrm, TB_ALIGN_16 }, 1133 { X86::SUBSDrr, X86::SUBSDrm, 0 }, 1134 { X86::SUBSDrr_Int, X86::SUBSDrm_Int, 0 }, 1135 { X86::SUBSSrr, X86::SUBSSrm, 0 }, 1136 { X86::SUBSSrr_Int, X86::SUBSSrm_Int, 0 }, 1137 // FIXME: TEST*rr -> swapped operand of TEST*mr. 1138 { X86::UNPCKHPDrr, X86::UNPCKHPDrm, TB_ALIGN_16 }, 1139 { X86::UNPCKHPSrr, X86::UNPCKHPSrm, TB_ALIGN_16 }, 1140 { X86::UNPCKLPDrr, X86::UNPCKLPDrm, TB_ALIGN_16 }, 1141 { X86::UNPCKLPSrr, X86::UNPCKLPSrm, TB_ALIGN_16 }, 1142 { X86::XOR16rr, X86::XOR16rm, 0 }, 1143 { X86::XOR32rr, X86::XOR32rm, 0 }, 1144 { X86::XOR64rr, X86::XOR64rm, 0 }, 1145 { X86::XOR8rr, X86::XOR8rm, 0 }, 1146 { X86::XORPDrr, X86::XORPDrm, TB_ALIGN_16 }, 1147 { X86::XORPSrr, X86::XORPSrm, TB_ALIGN_16 }, 1148 1149 // MMX version of foldable instructions 1150 { X86::MMX_CVTPI2PSirr, X86::MMX_CVTPI2PSirm, 0 }, 1151 { X86::MMX_PACKSSDWirr, X86::MMX_PACKSSDWirm, 0 }, 1152 { X86::MMX_PACKSSWBirr, X86::MMX_PACKSSWBirm, 0 }, 1153 { X86::MMX_PACKUSWBirr, X86::MMX_PACKUSWBirm, 0 }, 1154 { X86::MMX_PADDBirr, X86::MMX_PADDBirm, 0 }, 1155 { X86::MMX_PADDDirr, X86::MMX_PADDDirm, 0 }, 1156 { X86::MMX_PADDQirr, X86::MMX_PADDQirm, 0 }, 1157 { X86::MMX_PADDSBirr, X86::MMX_PADDSBirm, 0 }, 1158 { X86::MMX_PADDSWirr, X86::MMX_PADDSWirm, 0 }, 1159 { X86::MMX_PADDUSBirr, X86::MMX_PADDUSBirm, 0 }, 1160 { X86::MMX_PADDUSWirr, X86::MMX_PADDUSWirm, 0 }, 1161 { X86::MMX_PADDWirr, X86::MMX_PADDWirm, 0 }, 1162 { X86::MMX_PALIGNR64irr, X86::MMX_PALIGNR64irm, 0 }, 1163 { X86::MMX_PANDNirr, X86::MMX_PANDNirm, 0 }, 1164 { X86::MMX_PANDirr, X86::MMX_PANDirm, 0 }, 1165 { X86::MMX_PAVGBirr, X86::MMX_PAVGBirm, 0 }, 1166 { X86::MMX_PAVGWirr, X86::MMX_PAVGWirm, 0 }, 1167 { X86::MMX_PCMPEQBirr, X86::MMX_PCMPEQBirm, 0 }, 1168 { X86::MMX_PCMPEQDirr, X86::MMX_PCMPEQDirm, 0 }, 1169 { X86::MMX_PCMPEQWirr, X86::MMX_PCMPEQWirm, 0 }, 1170 { X86::MMX_PCMPGTBirr, X86::MMX_PCMPGTBirm, 0 }, 1171 { X86::MMX_PCMPGTDirr, X86::MMX_PCMPGTDirm, 0 }, 1172 { X86::MMX_PCMPGTWirr, X86::MMX_PCMPGTWirm, 0 }, 1173 { X86::MMX_PHADDSWrr64, X86::MMX_PHADDSWrm64, 0 }, 1174 { X86::MMX_PHADDWrr64, X86::MMX_PHADDWrm64, 0 }, 1175 { X86::MMX_PHADDrr64, X86::MMX_PHADDrm64, 0 }, 1176 { X86::MMX_PHSUBDrr64, X86::MMX_PHSUBDrm64, 0 }, 1177 { X86::MMX_PHSUBSWrr64, X86::MMX_PHSUBSWrm64, 0 }, 1178 { X86::MMX_PHSUBWrr64, X86::MMX_PHSUBWrm64, 0 }, 1179 { X86::MMX_PINSRWirri, X86::MMX_PINSRWirmi, 0 }, 1180 { X86::MMX_PMADDUBSWrr64, X86::MMX_PMADDUBSWrm64, 0 }, 1181 { X86::MMX_PMADDWDirr, X86::MMX_PMADDWDirm, 0 }, 1182 { X86::MMX_PMAXSWirr, X86::MMX_PMAXSWirm, 0 }, 1183 { X86::MMX_PMAXUBirr, X86::MMX_PMAXUBirm, 0 }, 1184 { X86::MMX_PMINSWirr, X86::MMX_PMINSWirm, 0 }, 1185 { X86::MMX_PMINUBirr, X86::MMX_PMINUBirm, 0 }, 1186 { X86::MMX_PMULHRSWrr64, X86::MMX_PMULHRSWrm64, 0 }, 1187 { X86::MMX_PMULHUWirr, X86::MMX_PMULHUWirm, 0 }, 1188 { X86::MMX_PMULHWirr, X86::MMX_PMULHWirm, 0 }, 1189 { X86::MMX_PMULLWirr, X86::MMX_PMULLWirm, 0 }, 1190 { X86::MMX_PMULUDQirr, X86::MMX_PMULUDQirm, 0 }, 1191 { X86::MMX_PORirr, X86::MMX_PORirm, 0 }, 1192 { X86::MMX_PSADBWirr, X86::MMX_PSADBWirm, 0 }, 1193 { X86::MMX_PSHUFBrr64, X86::MMX_PSHUFBrm64, 0 }, 1194 { X86::MMX_PSIGNBrr64, X86::MMX_PSIGNBrm64, 0 }, 1195 { X86::MMX_PSIGNDrr64, X86::MMX_PSIGNDrm64, 0 }, 1196 { X86::MMX_PSIGNWrr64, X86::MMX_PSIGNWrm64, 0 }, 1197 { X86::MMX_PSLLDrr, X86::MMX_PSLLDrm, 0 }, 1198 { X86::MMX_PSLLQrr, X86::MMX_PSLLQrm, 0 }, 1199 { X86::MMX_PSLLWrr, X86::MMX_PSLLWrm, 0 }, 1200 { X86::MMX_PSRADrr, X86::MMX_PSRADrm, 0 }, 1201 { X86::MMX_PSRAWrr, X86::MMX_PSRAWrm, 0 }, 1202 { X86::MMX_PSRLDrr, X86::MMX_PSRLDrm, 0 }, 1203 { X86::MMX_PSRLQrr, X86::MMX_PSRLQrm, 0 }, 1204 { X86::MMX_PSRLWrr, X86::MMX_PSRLWrm, 0 }, 1205 { X86::MMX_PSUBBirr, X86::MMX_PSUBBirm, 0 }, 1206 { X86::MMX_PSUBDirr, X86::MMX_PSUBDirm, 0 }, 1207 { X86::MMX_PSUBQirr, X86::MMX_PSUBQirm, 0 }, 1208 { X86::MMX_PSUBSBirr, X86::MMX_PSUBSBirm, 0 }, 1209 { X86::MMX_PSUBSWirr, X86::MMX_PSUBSWirm, 0 }, 1210 { X86::MMX_PSUBUSBirr, X86::MMX_PSUBUSBirm, 0 }, 1211 { X86::MMX_PSUBUSWirr, X86::MMX_PSUBUSWirm, 0 }, 1212 { X86::MMX_PSUBWirr, X86::MMX_PSUBWirm, 0 }, 1213 { X86::MMX_PUNPCKHBWirr, X86::MMX_PUNPCKHBWirm, 0 }, 1214 { X86::MMX_PUNPCKHDQirr, X86::MMX_PUNPCKHDQirm, 0 }, 1215 { X86::MMX_PUNPCKHWDirr, X86::MMX_PUNPCKHWDirm, 0 }, 1216 { X86::MMX_PUNPCKLBWirr, X86::MMX_PUNPCKLBWirm, 0 }, 1217 { X86::MMX_PUNPCKLDQirr, X86::MMX_PUNPCKLDQirm, 0 }, 1218 { X86::MMX_PUNPCKLWDirr, X86::MMX_PUNPCKLWDirm, 0 }, 1219 { X86::MMX_PXORirr, X86::MMX_PXORirm, 0 }, 1220 1221 // 3DNow! version of foldable instructions 1222 { X86::PAVGUSBrr, X86::PAVGUSBrm, 0 }, 1223 { X86::PFACCrr, X86::PFACCrm, 0 }, 1224 { X86::PFADDrr, X86::PFADDrm, 0 }, 1225 { X86::PFCMPEQrr, X86::PFCMPEQrm, 0 }, 1226 { X86::PFCMPGErr, X86::PFCMPGErm, 0 }, 1227 { X86::PFCMPGTrr, X86::PFCMPGTrm, 0 }, 1228 { X86::PFMAXrr, X86::PFMAXrm, 0 }, 1229 { X86::PFMINrr, X86::PFMINrm, 0 }, 1230 { X86::PFMULrr, X86::PFMULrm, 0 }, 1231 { X86::PFNACCrr, X86::PFNACCrm, 0 }, 1232 { X86::PFPNACCrr, X86::PFPNACCrm, 0 }, 1233 { X86::PFRCPIT1rr, X86::PFRCPIT1rm, 0 }, 1234 { X86::PFRCPIT2rr, X86::PFRCPIT2rm, 0 }, 1235 { X86::PFRSQIT1rr, X86::PFRSQIT1rm, 0 }, 1236 { X86::PFSUBrr, X86::PFSUBrm, 0 }, 1237 { X86::PFSUBRrr, X86::PFSUBRrm, 0 }, 1238 { X86::PMULHRWrr, X86::PMULHRWrm, 0 }, 1239 1240 // AVX 128-bit versions of foldable instructions 1241 { X86::VCVTSD2SSrr, X86::VCVTSD2SSrm, 0 }, 1242 { X86::Int_VCVTSD2SSrr, X86::Int_VCVTSD2SSrm, 0 }, 1243 { X86::VCVTSI2SD64rr, X86::VCVTSI2SD64rm, 0 }, 1244 { X86::Int_VCVTSI2SD64rr, X86::Int_VCVTSI2SD64rm, 0 }, 1245 { X86::VCVTSI2SDrr, X86::VCVTSI2SDrm, 0 }, 1246 { X86::Int_VCVTSI2SDrr, X86::Int_VCVTSI2SDrm, 0 }, 1247 { X86::VCVTSI2SS64rr, X86::VCVTSI2SS64rm, 0 }, 1248 { X86::Int_VCVTSI2SS64rr, X86::Int_VCVTSI2SS64rm, 0 }, 1249 { X86::VCVTSI2SSrr, X86::VCVTSI2SSrm, 0 }, 1250 { X86::Int_VCVTSI2SSrr, X86::Int_VCVTSI2SSrm, 0 }, 1251 { X86::VCVTSS2SDrr, X86::VCVTSS2SDrm, 0 }, 1252 { X86::Int_VCVTSS2SDrr, X86::Int_VCVTSS2SDrm, 0 }, 1253 { X86::VRCPSSr, X86::VRCPSSm, 0 }, 1254 { X86::VRCPSSr_Int, X86::VRCPSSm_Int, 0 }, 1255 { X86::VRSQRTSSr, X86::VRSQRTSSm, 0 }, 1256 { X86::VRSQRTSSr_Int, X86::VRSQRTSSm_Int, 0 }, 1257 { X86::VSQRTSDr, X86::VSQRTSDm, 0 }, 1258 { X86::VSQRTSDr_Int, X86::VSQRTSDm_Int, 0 }, 1259 { X86::VSQRTSSr, X86::VSQRTSSm, 0 }, 1260 { X86::VSQRTSSr_Int, X86::VSQRTSSm_Int, 0 }, 1261 { X86::VADDPDrr, X86::VADDPDrm, 0 }, 1262 { X86::VADDPSrr, X86::VADDPSrm, 0 }, 1263 { X86::VADDSDrr, X86::VADDSDrm, 0 }, 1264 { X86::VADDSDrr_Int, X86::VADDSDrm_Int, 0 }, 1265 { X86::VADDSSrr, X86::VADDSSrm, 0 }, 1266 { X86::VADDSSrr_Int, X86::VADDSSrm_Int, 0 }, 1267 { X86::VADDSUBPDrr, X86::VADDSUBPDrm, 0 }, 1268 { X86::VADDSUBPSrr, X86::VADDSUBPSrm, 0 }, 1269 { X86::VANDNPDrr, X86::VANDNPDrm, 0 }, 1270 { X86::VANDNPSrr, X86::VANDNPSrm, 0 }, 1271 { X86::VANDPDrr, X86::VANDPDrm, 0 }, 1272 { X86::VANDPSrr, X86::VANDPSrm, 0 }, 1273 { X86::VBLENDPDrri, X86::VBLENDPDrmi, 0 }, 1274 { X86::VBLENDPSrri, X86::VBLENDPSrmi, 0 }, 1275 { X86::VBLENDVPDrr, X86::VBLENDVPDrm, 0 }, 1276 { X86::VBLENDVPSrr, X86::VBLENDVPSrm, 0 }, 1277 { X86::VCMPPDrri, X86::VCMPPDrmi, 0 }, 1278 { X86::VCMPPSrri, X86::VCMPPSrmi, 0 }, 1279 { X86::VCMPSDrr, X86::VCMPSDrm, 0 }, 1280 { X86::VCMPSSrr, X86::VCMPSSrm, 0 }, 1281 { X86::VDIVPDrr, X86::VDIVPDrm, 0 }, 1282 { X86::VDIVPSrr, X86::VDIVPSrm, 0 }, 1283 { X86::VDIVSDrr, X86::VDIVSDrm, 0 }, 1284 { X86::VDIVSDrr_Int, X86::VDIVSDrm_Int, 0 }, 1285 { X86::VDIVSSrr, X86::VDIVSSrm, 0 }, 1286 { X86::VDIVSSrr_Int, X86::VDIVSSrm_Int, 0 }, 1287 { X86::VDPPDrri, X86::VDPPDrmi, 0 }, 1288 { X86::VDPPSrri, X86::VDPPSrmi, 0 }, 1289 // Do not fold VFs* loads because there are no scalar load variants for 1290 // these instructions. When folded, the load is required to be 128-bits, so 1291 // the load size would not match. 1292 { X86::VFvANDNPDrr, X86::VFvANDNPDrm, 0 }, 1293 { X86::VFvANDNPSrr, X86::VFvANDNPSrm, 0 }, 1294 { X86::VFvANDPDrr, X86::VFvANDPDrm, 0 }, 1295 { X86::VFvANDPSrr, X86::VFvANDPSrm, 0 }, 1296 { X86::VFvORPDrr, X86::VFvORPDrm, 0 }, 1297 { X86::VFvORPSrr, X86::VFvORPSrm, 0 }, 1298 { X86::VFvXORPDrr, X86::VFvXORPDrm, 0 }, 1299 { X86::VFvXORPSrr, X86::VFvXORPSrm, 0 }, 1300 { X86::VHADDPDrr, X86::VHADDPDrm, 0 }, 1301 { X86::VHADDPSrr, X86::VHADDPSrm, 0 }, 1302 { X86::VHSUBPDrr, X86::VHSUBPDrm, 0 }, 1303 { X86::VHSUBPSrr, X86::VHSUBPSrm, 0 }, 1304 { X86::Int_VCMPSDrr, X86::Int_VCMPSDrm, 0 }, 1305 { X86::Int_VCMPSSrr, X86::Int_VCMPSSrm, 0 }, 1306 { X86::VMAXPDrr, X86::VMAXPDrm, 0 }, 1307 { X86::VMAXPSrr, X86::VMAXPSrm, 0 }, 1308 { X86::VMAXSDrr, X86::VMAXSDrm, 0 }, 1309 { X86::VMAXSDrr_Int, X86::VMAXSDrm_Int, 0 }, 1310 { X86::VMAXSSrr, X86::VMAXSSrm, 0 }, 1311 { X86::VMAXSSrr_Int, X86::VMAXSSrm_Int, 0 }, 1312 { X86::VMINPDrr, X86::VMINPDrm, 0 }, 1313 { X86::VMINPSrr, X86::VMINPSrm, 0 }, 1314 { X86::VMINSDrr, X86::VMINSDrm, 0 }, 1315 { X86::VMINSDrr_Int, X86::VMINSDrm_Int, 0 }, 1316 { X86::VMINSSrr, X86::VMINSSrm, 0 }, 1317 { X86::VMINSSrr_Int, X86::VMINSSrm_Int, 0 }, 1318 { X86::VMOVLHPSrr, X86::VMOVHPSrm, TB_NO_REVERSE }, 1319 { X86::VMPSADBWrri, X86::VMPSADBWrmi, 0 }, 1320 { X86::VMULPDrr, X86::VMULPDrm, 0 }, 1321 { X86::VMULPSrr, X86::VMULPSrm, 0 }, 1322 { X86::VMULSDrr, X86::VMULSDrm, 0 }, 1323 { X86::VMULSDrr_Int, X86::VMULSDrm_Int, 0 }, 1324 { X86::VMULSSrr, X86::VMULSSrm, 0 }, 1325 { X86::VMULSSrr_Int, X86::VMULSSrm_Int, 0 }, 1326 { X86::VORPDrr, X86::VORPDrm, 0 }, 1327 { X86::VORPSrr, X86::VORPSrm, 0 }, 1328 { X86::VPACKSSDWrr, X86::VPACKSSDWrm, 0 }, 1329 { X86::VPACKSSWBrr, X86::VPACKSSWBrm, 0 }, 1330 { X86::VPACKUSDWrr, X86::VPACKUSDWrm, 0 }, 1331 { X86::VPACKUSWBrr, X86::VPACKUSWBrm, 0 }, 1332 { X86::VPADDBrr, X86::VPADDBrm, 0 }, 1333 { X86::VPADDDrr, X86::VPADDDrm, 0 }, 1334 { X86::VPADDQrr, X86::VPADDQrm, 0 }, 1335 { X86::VPADDSBrr, X86::VPADDSBrm, 0 }, 1336 { X86::VPADDSWrr, X86::VPADDSWrm, 0 }, 1337 { X86::VPADDUSBrr, X86::VPADDUSBrm, 0 }, 1338 { X86::VPADDUSWrr, X86::VPADDUSWrm, 0 }, 1339 { X86::VPADDWrr, X86::VPADDWrm, 0 }, 1340 { X86::VPALIGNRrri, X86::VPALIGNRrmi, 0 }, 1341 { X86::VPANDNrr, X86::VPANDNrm, 0 }, 1342 { X86::VPANDrr, X86::VPANDrm, 0 }, 1343 { X86::VPAVGBrr, X86::VPAVGBrm, 0 }, 1344 { X86::VPAVGWrr, X86::VPAVGWrm, 0 }, 1345 { X86::VPBLENDVBrr, X86::VPBLENDVBrm, 0 }, 1346 { X86::VPBLENDWrri, X86::VPBLENDWrmi, 0 }, 1347 { X86::VPCLMULQDQrr, X86::VPCLMULQDQrm, 0 }, 1348 { X86::VPCMPEQBrr, X86::VPCMPEQBrm, 0 }, 1349 { X86::VPCMPEQDrr, X86::VPCMPEQDrm, 0 }, 1350 { X86::VPCMPEQQrr, X86::VPCMPEQQrm, 0 }, 1351 { X86::VPCMPEQWrr, X86::VPCMPEQWrm, 0 }, 1352 { X86::VPCMPGTBrr, X86::VPCMPGTBrm, 0 }, 1353 { X86::VPCMPGTDrr, X86::VPCMPGTDrm, 0 }, 1354 { X86::VPCMPGTQrr, X86::VPCMPGTQrm, 0 }, 1355 { X86::VPCMPGTWrr, X86::VPCMPGTWrm, 0 }, 1356 { X86::VPHADDDrr, X86::VPHADDDrm, 0 }, 1357 { X86::VPHADDSWrr128, X86::VPHADDSWrm128, 0 }, 1358 { X86::VPHADDWrr, X86::VPHADDWrm, 0 }, 1359 { X86::VPHSUBDrr, X86::VPHSUBDrm, 0 }, 1360 { X86::VPHSUBSWrr128, X86::VPHSUBSWrm128, 0 }, 1361 { X86::VPHSUBWrr, X86::VPHSUBWrm, 0 }, 1362 { X86::VPERMILPDrr, X86::VPERMILPDrm, 0 }, 1363 { X86::VPERMILPSrr, X86::VPERMILPSrm, 0 }, 1364 { X86::VPINSRBrr, X86::VPINSRBrm, 0 }, 1365 { X86::VPINSRDrr, X86::VPINSRDrm, 0 }, 1366 { X86::VPINSRQrr, X86::VPINSRQrm, 0 }, 1367 { X86::VPINSRWrri, X86::VPINSRWrmi, 0 }, 1368 { X86::VPMADDUBSWrr128, X86::VPMADDUBSWrm128, 0 }, 1369 { X86::VPMADDWDrr, X86::VPMADDWDrm, 0 }, 1370 { X86::VPMAXSWrr, X86::VPMAXSWrm, 0 }, 1371 { X86::VPMAXUBrr, X86::VPMAXUBrm, 0 }, 1372 { X86::VPMINSWrr, X86::VPMINSWrm, 0 }, 1373 { X86::VPMINUBrr, X86::VPMINUBrm, 0 }, 1374 { X86::VPMINSBrr, X86::VPMINSBrm, 0 }, 1375 { X86::VPMINSDrr, X86::VPMINSDrm, 0 }, 1376 { X86::VPMINUDrr, X86::VPMINUDrm, 0 }, 1377 { X86::VPMINUWrr, X86::VPMINUWrm, 0 }, 1378 { X86::VPMAXSBrr, X86::VPMAXSBrm, 0 }, 1379 { X86::VPMAXSDrr, X86::VPMAXSDrm, 0 }, 1380 { X86::VPMAXUDrr, X86::VPMAXUDrm, 0 }, 1381 { X86::VPMAXUWrr, X86::VPMAXUWrm, 0 }, 1382 { X86::VPMULDQrr, X86::VPMULDQrm, 0 }, 1383 { X86::VPMULHRSWrr128, X86::VPMULHRSWrm128, 0 }, 1384 { X86::VPMULHUWrr, X86::VPMULHUWrm, 0 }, 1385 { X86::VPMULHWrr, X86::VPMULHWrm, 0 }, 1386 { X86::VPMULLDrr, X86::VPMULLDrm, 0 }, 1387 { X86::VPMULLWrr, X86::VPMULLWrm, 0 }, 1388 { X86::VPMULUDQrr, X86::VPMULUDQrm, 0 }, 1389 { X86::VPORrr, X86::VPORrm, 0 }, 1390 { X86::VPSADBWrr, X86::VPSADBWrm, 0 }, 1391 { X86::VPSHUFBrr, X86::VPSHUFBrm, 0 }, 1392 { X86::VPSIGNBrr128, X86::VPSIGNBrm128, 0 }, 1393 { X86::VPSIGNWrr128, X86::VPSIGNWrm128, 0 }, 1394 { X86::VPSIGNDrr128, X86::VPSIGNDrm128, 0 }, 1395 { X86::VPSLLDrr, X86::VPSLLDrm, 0 }, 1396 { X86::VPSLLQrr, X86::VPSLLQrm, 0 }, 1397 { X86::VPSLLWrr, X86::VPSLLWrm, 0 }, 1398 { X86::VPSRADrr, X86::VPSRADrm, 0 }, 1399 { X86::VPSRAWrr, X86::VPSRAWrm, 0 }, 1400 { X86::VPSRLDrr, X86::VPSRLDrm, 0 }, 1401 { X86::VPSRLQrr, X86::VPSRLQrm, 0 }, 1402 { X86::VPSRLWrr, X86::VPSRLWrm, 0 }, 1403 { X86::VPSUBBrr, X86::VPSUBBrm, 0 }, 1404 { X86::VPSUBDrr, X86::VPSUBDrm, 0 }, 1405 { X86::VPSUBQrr, X86::VPSUBQrm, 0 }, 1406 { X86::VPSUBSBrr, X86::VPSUBSBrm, 0 }, 1407 { X86::VPSUBSWrr, X86::VPSUBSWrm, 0 }, 1408 { X86::VPSUBUSBrr, X86::VPSUBUSBrm, 0 }, 1409 { X86::VPSUBUSWrr, X86::VPSUBUSWrm, 0 }, 1410 { X86::VPSUBWrr, X86::VPSUBWrm, 0 }, 1411 { X86::VPUNPCKHBWrr, X86::VPUNPCKHBWrm, 0 }, 1412 { X86::VPUNPCKHDQrr, X86::VPUNPCKHDQrm, 0 }, 1413 { X86::VPUNPCKHQDQrr, X86::VPUNPCKHQDQrm, 0 }, 1414 { X86::VPUNPCKHWDrr, X86::VPUNPCKHWDrm, 0 }, 1415 { X86::VPUNPCKLBWrr, X86::VPUNPCKLBWrm, 0 }, 1416 { X86::VPUNPCKLDQrr, X86::VPUNPCKLDQrm, 0 }, 1417 { X86::VPUNPCKLQDQrr, X86::VPUNPCKLQDQrm, 0 }, 1418 { X86::VPUNPCKLWDrr, X86::VPUNPCKLWDrm, 0 }, 1419 { X86::VPXORrr, X86::VPXORrm, 0 }, 1420 { X86::VROUNDSDr, X86::VROUNDSDm, 0 }, 1421 { X86::VROUNDSSr, X86::VROUNDSSm, 0 }, 1422 { X86::VSHUFPDrri, X86::VSHUFPDrmi, 0 }, 1423 { X86::VSHUFPSrri, X86::VSHUFPSrmi, 0 }, 1424 { X86::VSUBPDrr, X86::VSUBPDrm, 0 }, 1425 { X86::VSUBPSrr, X86::VSUBPSrm, 0 }, 1426 { X86::VSUBSDrr, X86::VSUBSDrm, 0 }, 1427 { X86::VSUBSDrr_Int, X86::VSUBSDrm_Int, 0 }, 1428 { X86::VSUBSSrr, X86::VSUBSSrm, 0 }, 1429 { X86::VSUBSSrr_Int, X86::VSUBSSrm_Int, 0 }, 1430 { X86::VUNPCKHPDrr, X86::VUNPCKHPDrm, 0 }, 1431 { X86::VUNPCKHPSrr, X86::VUNPCKHPSrm, 0 }, 1432 { X86::VUNPCKLPDrr, X86::VUNPCKLPDrm, 0 }, 1433 { X86::VUNPCKLPSrr, X86::VUNPCKLPSrm, 0 }, 1434 { X86::VXORPDrr, X86::VXORPDrm, 0 }, 1435 { X86::VXORPSrr, X86::VXORPSrm, 0 }, 1436 1437 // AVX 256-bit foldable instructions 1438 { X86::VADDPDYrr, X86::VADDPDYrm, 0 }, 1439 { X86::VADDPSYrr, X86::VADDPSYrm, 0 }, 1440 { X86::VADDSUBPDYrr, X86::VADDSUBPDYrm, 0 }, 1441 { X86::VADDSUBPSYrr, X86::VADDSUBPSYrm, 0 }, 1442 { X86::VANDNPDYrr, X86::VANDNPDYrm, 0 }, 1443 { X86::VANDNPSYrr, X86::VANDNPSYrm, 0 }, 1444 { X86::VANDPDYrr, X86::VANDPDYrm, 0 }, 1445 { X86::VANDPSYrr, X86::VANDPSYrm, 0 }, 1446 { X86::VBLENDPDYrri, X86::VBLENDPDYrmi, 0 }, 1447 { X86::VBLENDPSYrri, X86::VBLENDPSYrmi, 0 }, 1448 { X86::VBLENDVPDYrr, X86::VBLENDVPDYrm, 0 }, 1449 { X86::VBLENDVPSYrr, X86::VBLENDVPSYrm, 0 }, 1450 { X86::VCMPPDYrri, X86::VCMPPDYrmi, 0 }, 1451 { X86::VCMPPSYrri, X86::VCMPPSYrmi, 0 }, 1452 { X86::VDIVPDYrr, X86::VDIVPDYrm, 0 }, 1453 { X86::VDIVPSYrr, X86::VDIVPSYrm, 0 }, 1454 { X86::VDPPSYrri, X86::VDPPSYrmi, 0 }, 1455 { X86::VHADDPDYrr, X86::VHADDPDYrm, 0 }, 1456 { X86::VHADDPSYrr, X86::VHADDPSYrm, 0 }, 1457 { X86::VHSUBPDYrr, X86::VHSUBPDYrm, 0 }, 1458 { X86::VHSUBPSYrr, X86::VHSUBPSYrm, 0 }, 1459 { X86::VINSERTF128rr, X86::VINSERTF128rm, 0 }, 1460 { X86::VMAXPDYrr, X86::VMAXPDYrm, 0 }, 1461 { X86::VMAXPSYrr, X86::VMAXPSYrm, 0 }, 1462 { X86::VMINPDYrr, X86::VMINPDYrm, 0 }, 1463 { X86::VMINPSYrr, X86::VMINPSYrm, 0 }, 1464 { X86::VMULPDYrr, X86::VMULPDYrm, 0 }, 1465 { X86::VMULPSYrr, X86::VMULPSYrm, 0 }, 1466 { X86::VORPDYrr, X86::VORPDYrm, 0 }, 1467 { X86::VORPSYrr, X86::VORPSYrm, 0 }, 1468 { X86::VPERM2F128rr, X86::VPERM2F128rm, 0 }, 1469 { X86::VPERMILPDYrr, X86::VPERMILPDYrm, 0 }, 1470 { X86::VPERMILPSYrr, X86::VPERMILPSYrm, 0 }, 1471 { X86::VSHUFPDYrri, X86::VSHUFPDYrmi, 0 }, 1472 { X86::VSHUFPSYrri, X86::VSHUFPSYrmi, 0 }, 1473 { X86::VSUBPDYrr, X86::VSUBPDYrm, 0 }, 1474 { X86::VSUBPSYrr, X86::VSUBPSYrm, 0 }, 1475 { X86::VUNPCKHPDYrr, X86::VUNPCKHPDYrm, 0 }, 1476 { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrm, 0 }, 1477 { X86::VUNPCKLPDYrr, X86::VUNPCKLPDYrm, 0 }, 1478 { X86::VUNPCKLPSYrr, X86::VUNPCKLPSYrm, 0 }, 1479 { X86::VXORPDYrr, X86::VXORPDYrm, 0 }, 1480 { X86::VXORPSYrr, X86::VXORPSYrm, 0 }, 1481 1482 // AVX2 foldable instructions 1483 { X86::VINSERTI128rr, X86::VINSERTI128rm, 0 }, 1484 { X86::VPACKSSDWYrr, X86::VPACKSSDWYrm, 0 }, 1485 { X86::VPACKSSWBYrr, X86::VPACKSSWBYrm, 0 }, 1486 { X86::VPACKUSDWYrr, X86::VPACKUSDWYrm, 0 }, 1487 { X86::VPACKUSWBYrr, X86::VPACKUSWBYrm, 0 }, 1488 { X86::VPADDBYrr, X86::VPADDBYrm, 0 }, 1489 { X86::VPADDDYrr, X86::VPADDDYrm, 0 }, 1490 { X86::VPADDQYrr, X86::VPADDQYrm, 0 }, 1491 { X86::VPADDSBYrr, X86::VPADDSBYrm, 0 }, 1492 { X86::VPADDSWYrr, X86::VPADDSWYrm, 0 }, 1493 { X86::VPADDUSBYrr, X86::VPADDUSBYrm, 0 }, 1494 { X86::VPADDUSWYrr, X86::VPADDUSWYrm, 0 }, 1495 { X86::VPADDWYrr, X86::VPADDWYrm, 0 }, 1496 { X86::VPALIGNRYrri, X86::VPALIGNRYrmi, 0 }, 1497 { X86::VPANDNYrr, X86::VPANDNYrm, 0 }, 1498 { X86::VPANDYrr, X86::VPANDYrm, 0 }, 1499 { X86::VPAVGBYrr, X86::VPAVGBYrm, 0 }, 1500 { X86::VPAVGWYrr, X86::VPAVGWYrm, 0 }, 1501 { X86::VPBLENDDrri, X86::VPBLENDDrmi, 0 }, 1502 { X86::VPBLENDDYrri, X86::VPBLENDDYrmi, 0 }, 1503 { X86::VPBLENDVBYrr, X86::VPBLENDVBYrm, 0 }, 1504 { X86::VPBLENDWYrri, X86::VPBLENDWYrmi, 0 }, 1505 { X86::VPCMPEQBYrr, X86::VPCMPEQBYrm, 0 }, 1506 { X86::VPCMPEQDYrr, X86::VPCMPEQDYrm, 0 }, 1507 { X86::VPCMPEQQYrr, X86::VPCMPEQQYrm, 0 }, 1508 { X86::VPCMPEQWYrr, X86::VPCMPEQWYrm, 0 }, 1509 { X86::VPCMPGTBYrr, X86::VPCMPGTBYrm, 0 }, 1510 { X86::VPCMPGTDYrr, X86::VPCMPGTDYrm, 0 }, 1511 { X86::VPCMPGTQYrr, X86::VPCMPGTQYrm, 0 }, 1512 { X86::VPCMPGTWYrr, X86::VPCMPGTWYrm, 0 }, 1513 { X86::VPERM2I128rr, X86::VPERM2I128rm, 0 }, 1514 { X86::VPERMDYrr, X86::VPERMDYrm, 0 }, 1515 { X86::VPERMPSYrr, X86::VPERMPSYrm, 0 }, 1516 { X86::VPHADDDYrr, X86::VPHADDDYrm, 0 }, 1517 { X86::VPHADDSWrr256, X86::VPHADDSWrm256, 0 }, 1518 { X86::VPHADDWYrr, X86::VPHADDWYrm, 0 }, 1519 { X86::VPHSUBDYrr, X86::VPHSUBDYrm, 0 }, 1520 { X86::VPHSUBSWrr256, X86::VPHSUBSWrm256, 0 }, 1521 { X86::VPHSUBWYrr, X86::VPHSUBWYrm, 0 }, 1522 { X86::VPMADDUBSWrr256, X86::VPMADDUBSWrm256, 0 }, 1523 { X86::VPMADDWDYrr, X86::VPMADDWDYrm, 0 }, 1524 { X86::VPMAXSWYrr, X86::VPMAXSWYrm, 0 }, 1525 { X86::VPMAXUBYrr, X86::VPMAXUBYrm, 0 }, 1526 { X86::VPMINSWYrr, X86::VPMINSWYrm, 0 }, 1527 { X86::VPMINUBYrr, X86::VPMINUBYrm, 0 }, 1528 { X86::VPMINSBYrr, X86::VPMINSBYrm, 0 }, 1529 { X86::VPMINSDYrr, X86::VPMINSDYrm, 0 }, 1530 { X86::VPMINUDYrr, X86::VPMINUDYrm, 0 }, 1531 { X86::VPMINUWYrr, X86::VPMINUWYrm, 0 }, 1532 { X86::VPMAXSBYrr, X86::VPMAXSBYrm, 0 }, 1533 { X86::VPMAXSDYrr, X86::VPMAXSDYrm, 0 }, 1534 { X86::VPMAXUDYrr, X86::VPMAXUDYrm, 0 }, 1535 { X86::VPMAXUWYrr, X86::VPMAXUWYrm, 0 }, 1536 { X86::VMPSADBWYrri, X86::VMPSADBWYrmi, 0 }, 1537 { X86::VPMULDQYrr, X86::VPMULDQYrm, 0 }, 1538 { X86::VPMULHRSWrr256, X86::VPMULHRSWrm256, 0 }, 1539 { X86::VPMULHUWYrr, X86::VPMULHUWYrm, 0 }, 1540 { X86::VPMULHWYrr, X86::VPMULHWYrm, 0 }, 1541 { X86::VPMULLDYrr, X86::VPMULLDYrm, 0 }, 1542 { X86::VPMULLWYrr, X86::VPMULLWYrm, 0 }, 1543 { X86::VPMULUDQYrr, X86::VPMULUDQYrm, 0 }, 1544 { X86::VPORYrr, X86::VPORYrm, 0 }, 1545 { X86::VPSADBWYrr, X86::VPSADBWYrm, 0 }, 1546 { X86::VPSHUFBYrr, X86::VPSHUFBYrm, 0 }, 1547 { X86::VPSIGNBYrr256, X86::VPSIGNBYrm256, 0 }, 1548 { X86::VPSIGNWYrr256, X86::VPSIGNWYrm256, 0 }, 1549 { X86::VPSIGNDYrr256, X86::VPSIGNDYrm256, 0 }, 1550 { X86::VPSLLDYrr, X86::VPSLLDYrm, 0 }, 1551 { X86::VPSLLQYrr, X86::VPSLLQYrm, 0 }, 1552 { X86::VPSLLWYrr, X86::VPSLLWYrm, 0 }, 1553 { X86::VPSLLVDrr, X86::VPSLLVDrm, 0 }, 1554 { X86::VPSLLVDYrr, X86::VPSLLVDYrm, 0 }, 1555 { X86::VPSLLVQrr, X86::VPSLLVQrm, 0 }, 1556 { X86::VPSLLVQYrr, X86::VPSLLVQYrm, 0 }, 1557 { X86::VPSRADYrr, X86::VPSRADYrm, 0 }, 1558 { X86::VPSRAWYrr, X86::VPSRAWYrm, 0 }, 1559 { X86::VPSRAVDrr, X86::VPSRAVDrm, 0 }, 1560 { X86::VPSRAVDYrr, X86::VPSRAVDYrm, 0 }, 1561 { X86::VPSRAVD_Intrr, X86::VPSRAVD_Intrm, 0 }, 1562 { X86::VPSRAVD_IntYrr, X86::VPSRAVD_IntYrm, 0 }, 1563 { X86::VPSRLDYrr, X86::VPSRLDYrm, 0 }, 1564 { X86::VPSRLQYrr, X86::VPSRLQYrm, 0 }, 1565 { X86::VPSRLWYrr, X86::VPSRLWYrm, 0 }, 1566 { X86::VPSRLVDrr, X86::VPSRLVDrm, 0 }, 1567 { X86::VPSRLVDYrr, X86::VPSRLVDYrm, 0 }, 1568 { X86::VPSRLVQrr, X86::VPSRLVQrm, 0 }, 1569 { X86::VPSRLVQYrr, X86::VPSRLVQYrm, 0 }, 1570 { X86::VPSUBBYrr, X86::VPSUBBYrm, 0 }, 1571 { X86::VPSUBDYrr, X86::VPSUBDYrm, 0 }, 1572 { X86::VPSUBQYrr, X86::VPSUBQYrm, 0 }, 1573 { X86::VPSUBSBYrr, X86::VPSUBSBYrm, 0 }, 1574 { X86::VPSUBSWYrr, X86::VPSUBSWYrm, 0 }, 1575 { X86::VPSUBUSBYrr, X86::VPSUBUSBYrm, 0 }, 1576 { X86::VPSUBUSWYrr, X86::VPSUBUSWYrm, 0 }, 1577 { X86::VPSUBWYrr, X86::VPSUBWYrm, 0 }, 1578 { X86::VPUNPCKHBWYrr, X86::VPUNPCKHBWYrm, 0 }, 1579 { X86::VPUNPCKHDQYrr, X86::VPUNPCKHDQYrm, 0 }, 1580 { X86::VPUNPCKHQDQYrr, X86::VPUNPCKHQDQYrm, 0 }, 1581 { X86::VPUNPCKHWDYrr, X86::VPUNPCKHWDYrm, 0 }, 1582 { X86::VPUNPCKLBWYrr, X86::VPUNPCKLBWYrm, 0 }, 1583 { X86::VPUNPCKLDQYrr, X86::VPUNPCKLDQYrm, 0 }, 1584 { X86::VPUNPCKLQDQYrr, X86::VPUNPCKLQDQYrm, 0 }, 1585 { X86::VPUNPCKLWDYrr, X86::VPUNPCKLWDYrm, 0 }, 1586 { X86::VPXORYrr, X86::VPXORYrm, 0 }, 1587 1588 // FMA4 foldable patterns 1589 { X86::VFMADDSS4rr, X86::VFMADDSS4mr, TB_ALIGN_NONE }, 1590 { X86::VFMADDSD4rr, X86::VFMADDSD4mr, TB_ALIGN_NONE }, 1591 { X86::VFMADDPS4rr, X86::VFMADDPS4mr, TB_ALIGN_NONE }, 1592 { X86::VFMADDPD4rr, X86::VFMADDPD4mr, TB_ALIGN_NONE }, 1593 { X86::VFMADDPS4rrY, X86::VFMADDPS4mrY, TB_ALIGN_NONE }, 1594 { X86::VFMADDPD4rrY, X86::VFMADDPD4mrY, TB_ALIGN_NONE }, 1595 { X86::VFNMADDSS4rr, X86::VFNMADDSS4mr, TB_ALIGN_NONE }, 1596 { X86::VFNMADDSD4rr, X86::VFNMADDSD4mr, TB_ALIGN_NONE }, 1597 { X86::VFNMADDPS4rr, X86::VFNMADDPS4mr, TB_ALIGN_NONE }, 1598 { X86::VFNMADDPD4rr, X86::VFNMADDPD4mr, TB_ALIGN_NONE }, 1599 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4mrY, TB_ALIGN_NONE }, 1600 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4mrY, TB_ALIGN_NONE }, 1601 { X86::VFMSUBSS4rr, X86::VFMSUBSS4mr, TB_ALIGN_NONE }, 1602 { X86::VFMSUBSD4rr, X86::VFMSUBSD4mr, TB_ALIGN_NONE }, 1603 { X86::VFMSUBPS4rr, X86::VFMSUBPS4mr, TB_ALIGN_NONE }, 1604 { X86::VFMSUBPD4rr, X86::VFMSUBPD4mr, TB_ALIGN_NONE }, 1605 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4mrY, TB_ALIGN_NONE }, 1606 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4mrY, TB_ALIGN_NONE }, 1607 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4mr, TB_ALIGN_NONE }, 1608 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4mr, TB_ALIGN_NONE }, 1609 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4mr, TB_ALIGN_NONE }, 1610 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4mr, TB_ALIGN_NONE }, 1611 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4mrY, TB_ALIGN_NONE }, 1612 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4mrY, TB_ALIGN_NONE }, 1613 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4mr, TB_ALIGN_NONE }, 1614 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4mr, TB_ALIGN_NONE }, 1615 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4mrY, TB_ALIGN_NONE }, 1616 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4mrY, TB_ALIGN_NONE }, 1617 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4mr, TB_ALIGN_NONE }, 1618 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4mr, TB_ALIGN_NONE }, 1619 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4mrY, TB_ALIGN_NONE }, 1620 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4mrY, TB_ALIGN_NONE }, 1621 1622 // XOP foldable instructions 1623 { X86::VPCMOVrrr, X86::VPCMOVrmr, 0 }, 1624 { X86::VPCMOVrrrY, X86::VPCMOVrmrY, 0 }, 1625 { X86::VPCOMBri, X86::VPCOMBmi, 0 }, 1626 { X86::VPCOMDri, X86::VPCOMDmi, 0 }, 1627 { X86::VPCOMQri, X86::VPCOMQmi, 0 }, 1628 { X86::VPCOMWri, X86::VPCOMWmi, 0 }, 1629 { X86::VPCOMUBri, X86::VPCOMUBmi, 0 }, 1630 { X86::VPCOMUDri, X86::VPCOMUDmi, 0 }, 1631 { X86::VPCOMUQri, X86::VPCOMUQmi, 0 }, 1632 { X86::VPCOMUWri, X86::VPCOMUWmi, 0 }, 1633 { X86::VPERMIL2PDrr, X86::VPERMIL2PDmr, 0 }, 1634 { X86::VPERMIL2PDrrY, X86::VPERMIL2PDmrY, 0 }, 1635 { X86::VPERMIL2PSrr, X86::VPERMIL2PSmr, 0 }, 1636 { X86::VPERMIL2PSrrY, X86::VPERMIL2PSmrY, 0 }, 1637 { X86::VPMACSDDrr, X86::VPMACSDDrm, 0 }, 1638 { X86::VPMACSDQHrr, X86::VPMACSDQHrm, 0 }, 1639 { X86::VPMACSDQLrr, X86::VPMACSDQLrm, 0 }, 1640 { X86::VPMACSSDDrr, X86::VPMACSSDDrm, 0 }, 1641 { X86::VPMACSSDQHrr, X86::VPMACSSDQHrm, 0 }, 1642 { X86::VPMACSSDQLrr, X86::VPMACSSDQLrm, 0 }, 1643 { X86::VPMACSSWDrr, X86::VPMACSSWDrm, 0 }, 1644 { X86::VPMACSSWWrr, X86::VPMACSSWWrm, 0 }, 1645 { X86::VPMACSWDrr, X86::VPMACSWDrm, 0 }, 1646 { X86::VPMACSWWrr, X86::VPMACSWWrm, 0 }, 1647 { X86::VPMADCSSWDrr, X86::VPMADCSSWDrm, 0 }, 1648 { X86::VPMADCSWDrr, X86::VPMADCSWDrm, 0 }, 1649 { X86::VPPERMrrr, X86::VPPERMrmr, 0 }, 1650 { X86::VPROTBrr, X86::VPROTBrm, 0 }, 1651 { X86::VPROTDrr, X86::VPROTDrm, 0 }, 1652 { X86::VPROTQrr, X86::VPROTQrm, 0 }, 1653 { X86::VPROTWrr, X86::VPROTWrm, 0 }, 1654 { X86::VPSHABrr, X86::VPSHABrm, 0 }, 1655 { X86::VPSHADrr, X86::VPSHADrm, 0 }, 1656 { X86::VPSHAQrr, X86::VPSHAQrm, 0 }, 1657 { X86::VPSHAWrr, X86::VPSHAWrm, 0 }, 1658 { X86::VPSHLBrr, X86::VPSHLBrm, 0 }, 1659 { X86::VPSHLDrr, X86::VPSHLDrm, 0 }, 1660 { X86::VPSHLQrr, X86::VPSHLQrm, 0 }, 1661 { X86::VPSHLWrr, X86::VPSHLWrm, 0 }, 1662 1663 // BMI/BMI2 foldable instructions 1664 { X86::ANDN32rr, X86::ANDN32rm, 0 }, 1665 { X86::ANDN64rr, X86::ANDN64rm, 0 }, 1666 { X86::MULX32rr, X86::MULX32rm, 0 }, 1667 { X86::MULX64rr, X86::MULX64rm, 0 }, 1668 { X86::PDEP32rr, X86::PDEP32rm, 0 }, 1669 { X86::PDEP64rr, X86::PDEP64rm, 0 }, 1670 { X86::PEXT32rr, X86::PEXT32rm, 0 }, 1671 { X86::PEXT64rr, X86::PEXT64rm, 0 }, 1672 1673 // ADX foldable instructions 1674 { X86::ADCX32rr, X86::ADCX32rm, 0 }, 1675 { X86::ADCX64rr, X86::ADCX64rm, 0 }, 1676 { X86::ADOX32rr, X86::ADOX32rm, 0 }, 1677 { X86::ADOX64rr, X86::ADOX64rm, 0 }, 1678 1679 // AVX-512 foldable instructions 1680 { X86::VADDPSZrr, X86::VADDPSZrm, 0 }, 1681 { X86::VADDPDZrr, X86::VADDPDZrm, 0 }, 1682 { X86::VSUBPSZrr, X86::VSUBPSZrm, 0 }, 1683 { X86::VSUBPDZrr, X86::VSUBPDZrm, 0 }, 1684 { X86::VMULPSZrr, X86::VMULPSZrm, 0 }, 1685 { X86::VMULPDZrr, X86::VMULPDZrm, 0 }, 1686 { X86::VDIVPSZrr, X86::VDIVPSZrm, 0 }, 1687 { X86::VDIVPDZrr, X86::VDIVPDZrm, 0 }, 1688 { X86::VMINPSZrr, X86::VMINPSZrm, 0 }, 1689 { X86::VMINPDZrr, X86::VMINPDZrm, 0 }, 1690 { X86::VMAXPSZrr, X86::VMAXPSZrm, 0 }, 1691 { X86::VMAXPDZrr, X86::VMAXPDZrm, 0 }, 1692 { X86::VPADDDZrr, X86::VPADDDZrm, 0 }, 1693 { X86::VPADDQZrr, X86::VPADDQZrm, 0 }, 1694 { X86::VPERMPDZri, X86::VPERMPDZmi, 0 }, 1695 { X86::VPERMPSZrr, X86::VPERMPSZrm, 0 }, 1696 { X86::VPMAXSDZrr, X86::VPMAXSDZrm, 0 }, 1697 { X86::VPMAXSQZrr, X86::VPMAXSQZrm, 0 }, 1698 { X86::VPMAXUDZrr, X86::VPMAXUDZrm, 0 }, 1699 { X86::VPMAXUQZrr, X86::VPMAXUQZrm, 0 }, 1700 { X86::VPMINSDZrr, X86::VPMINSDZrm, 0 }, 1701 { X86::VPMINSQZrr, X86::VPMINSQZrm, 0 }, 1702 { X86::VPMINUDZrr, X86::VPMINUDZrm, 0 }, 1703 { X86::VPMINUQZrr, X86::VPMINUQZrm, 0 }, 1704 { X86::VPMULDQZrr, X86::VPMULDQZrm, 0 }, 1705 { X86::VPSLLVDZrr, X86::VPSLLVDZrm, 0 }, 1706 { X86::VPSLLVQZrr, X86::VPSLLVQZrm, 0 }, 1707 { X86::VPSRAVDZrr, X86::VPSRAVDZrm, 0 }, 1708 { X86::VPSRLVDZrr, X86::VPSRLVDZrm, 0 }, 1709 { X86::VPSRLVQZrr, X86::VPSRLVQZrm, 0 }, 1710 { X86::VPSUBDZrr, X86::VPSUBDZrm, 0 }, 1711 { X86::VPSUBQZrr, X86::VPSUBQZrm, 0 }, 1712 { X86::VSHUFPDZrri, X86::VSHUFPDZrmi, 0 }, 1713 { X86::VSHUFPSZrri, X86::VSHUFPSZrmi, 0 }, 1714 { X86::VALIGNQZrri, X86::VALIGNQZrmi, 0 }, 1715 { X86::VALIGNDZrri, X86::VALIGNDZrmi, 0 }, 1716 { X86::VPMULUDQZrr, X86::VPMULUDQZrm, 0 }, 1717 { X86::VBROADCASTSSZrkz, X86::VBROADCASTSSZmkz, TB_NO_REVERSE }, 1718 { X86::VBROADCASTSDZrkz, X86::VBROADCASTSDZmkz, TB_NO_REVERSE }, 1719 1720 // AVX-512{F,VL} foldable instructions 1721 { X86::VBROADCASTSSZ256rkz, X86::VBROADCASTSSZ256mkz, TB_NO_REVERSE }, 1722 { X86::VBROADCASTSDZ256rkz, X86::VBROADCASTSDZ256mkz, TB_NO_REVERSE }, 1723 { X86::VBROADCASTSSZ128rkz, X86::VBROADCASTSSZ128mkz, TB_NO_REVERSE }, 1724 1725 // AVX-512{F,VL} foldable instructions 1726 { X86::VADDPDZ128rr, X86::VADDPDZ128rm, 0 }, 1727 { X86::VADDPDZ256rr, X86::VADDPDZ256rm, 0 }, 1728 { X86::VADDPSZ128rr, X86::VADDPSZ128rm, 0 }, 1729 { X86::VADDPSZ256rr, X86::VADDPSZ256rm, 0 }, 1730 1731 // AES foldable instructions 1732 { X86::AESDECLASTrr, X86::AESDECLASTrm, TB_ALIGN_16 }, 1733 { X86::AESDECrr, X86::AESDECrm, TB_ALIGN_16 }, 1734 { X86::AESENCLASTrr, X86::AESENCLASTrm, TB_ALIGN_16 }, 1735 { X86::AESENCrr, X86::AESENCrm, TB_ALIGN_16 }, 1736 { X86::VAESDECLASTrr, X86::VAESDECLASTrm, 0 }, 1737 { X86::VAESDECrr, X86::VAESDECrm, 0 }, 1738 { X86::VAESENCLASTrr, X86::VAESENCLASTrm, 0 }, 1739 { X86::VAESENCrr, X86::VAESENCrm, 0 }, 1740 1741 // SHA foldable instructions 1742 { X86::SHA1MSG1rr, X86::SHA1MSG1rm, TB_ALIGN_16 }, 1743 { X86::SHA1MSG2rr, X86::SHA1MSG2rm, TB_ALIGN_16 }, 1744 { X86::SHA1NEXTErr, X86::SHA1NEXTErm, TB_ALIGN_16 }, 1745 { X86::SHA1RNDS4rri, X86::SHA1RNDS4rmi, TB_ALIGN_16 }, 1746 { X86::SHA256MSG1rr, X86::SHA256MSG1rm, TB_ALIGN_16 }, 1747 { X86::SHA256MSG2rr, X86::SHA256MSG2rm, TB_ALIGN_16 }, 1748 { X86::SHA256RNDS2rr, X86::SHA256RNDS2rm, TB_ALIGN_16 } 1749 }; 1750 1751 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable2) { 1752 AddTableEntry(RegOp2MemOpTable2, MemOp2RegOpTable, 1753 Entry.RegOp, Entry.MemOp, 1754 // Index 2, folded load 1755 Entry.Flags | TB_INDEX_2 | TB_FOLDED_LOAD); 1756 } 1757 1758 static const X86MemoryFoldTableEntry MemoryFoldTable3[] = { 1759 // FMA foldable instructions 1760 { X86::VFMADDSSr231r, X86::VFMADDSSr231m, TB_ALIGN_NONE }, 1761 { X86::VFMADDSSr231r_Int, X86::VFMADDSSr231m_Int, TB_ALIGN_NONE }, 1762 { X86::VFMADDSDr231r, X86::VFMADDSDr231m, TB_ALIGN_NONE }, 1763 { X86::VFMADDSDr231r_Int, X86::VFMADDSDr231m_Int, TB_ALIGN_NONE }, 1764 { X86::VFMADDSSr132r, X86::VFMADDSSr132m, TB_ALIGN_NONE }, 1765 { X86::VFMADDSSr132r_Int, X86::VFMADDSSr132m_Int, TB_ALIGN_NONE }, 1766 { X86::VFMADDSDr132r, X86::VFMADDSDr132m, TB_ALIGN_NONE }, 1767 { X86::VFMADDSDr132r_Int, X86::VFMADDSDr132m_Int, TB_ALIGN_NONE }, 1768 { X86::VFMADDSSr213r, X86::VFMADDSSr213m, TB_ALIGN_NONE }, 1769 { X86::VFMADDSSr213r_Int, X86::VFMADDSSr213m_Int, TB_ALIGN_NONE }, 1770 { X86::VFMADDSDr213r, X86::VFMADDSDr213m, TB_ALIGN_NONE }, 1771 { X86::VFMADDSDr213r_Int, X86::VFMADDSDr213m_Int, TB_ALIGN_NONE }, 1772 1773 { X86::VFMADDPSr231r, X86::VFMADDPSr231m, TB_ALIGN_NONE }, 1774 { X86::VFMADDPDr231r, X86::VFMADDPDr231m, TB_ALIGN_NONE }, 1775 { X86::VFMADDPSr132r, X86::VFMADDPSr132m, TB_ALIGN_NONE }, 1776 { X86::VFMADDPDr132r, X86::VFMADDPDr132m, TB_ALIGN_NONE }, 1777 { X86::VFMADDPSr213r, X86::VFMADDPSr213m, TB_ALIGN_NONE }, 1778 { X86::VFMADDPDr213r, X86::VFMADDPDr213m, TB_ALIGN_NONE }, 1779 { X86::VFMADDPSr231rY, X86::VFMADDPSr231mY, TB_ALIGN_NONE }, 1780 { X86::VFMADDPDr231rY, X86::VFMADDPDr231mY, TB_ALIGN_NONE }, 1781 { X86::VFMADDPSr132rY, X86::VFMADDPSr132mY, TB_ALIGN_NONE }, 1782 { X86::VFMADDPDr132rY, X86::VFMADDPDr132mY, TB_ALIGN_NONE }, 1783 { X86::VFMADDPSr213rY, X86::VFMADDPSr213mY, TB_ALIGN_NONE }, 1784 { X86::VFMADDPDr213rY, X86::VFMADDPDr213mY, TB_ALIGN_NONE }, 1785 1786 { X86::VFNMADDSSr231r, X86::VFNMADDSSr231m, TB_ALIGN_NONE }, 1787 { X86::VFNMADDSSr231r_Int, X86::VFNMADDSSr231m_Int, TB_ALIGN_NONE }, 1788 { X86::VFNMADDSDr231r, X86::VFNMADDSDr231m, TB_ALIGN_NONE }, 1789 { X86::VFNMADDSDr231r_Int, X86::VFNMADDSDr231m_Int, TB_ALIGN_NONE }, 1790 { X86::VFNMADDSSr132r, X86::VFNMADDSSr132m, TB_ALIGN_NONE }, 1791 { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr132m_Int, TB_ALIGN_NONE }, 1792 { X86::VFNMADDSDr132r, X86::VFNMADDSDr132m, TB_ALIGN_NONE }, 1793 { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr132m_Int, TB_ALIGN_NONE }, 1794 { X86::VFNMADDSSr213r, X86::VFNMADDSSr213m, TB_ALIGN_NONE }, 1795 { X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr213m_Int, TB_ALIGN_NONE }, 1796 { X86::VFNMADDSDr213r, X86::VFNMADDSDr213m, TB_ALIGN_NONE }, 1797 { X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr213m_Int, TB_ALIGN_NONE }, 1798 1799 { X86::VFNMADDPSr231r, X86::VFNMADDPSr231m, TB_ALIGN_NONE }, 1800 { X86::VFNMADDPDr231r, X86::VFNMADDPDr231m, TB_ALIGN_NONE }, 1801 { X86::VFNMADDPSr132r, X86::VFNMADDPSr132m, TB_ALIGN_NONE }, 1802 { X86::VFNMADDPDr132r, X86::VFNMADDPDr132m, TB_ALIGN_NONE }, 1803 { X86::VFNMADDPSr213r, X86::VFNMADDPSr213m, TB_ALIGN_NONE }, 1804 { X86::VFNMADDPDr213r, X86::VFNMADDPDr213m, TB_ALIGN_NONE }, 1805 { X86::VFNMADDPSr231rY, X86::VFNMADDPSr231mY, TB_ALIGN_NONE }, 1806 { X86::VFNMADDPDr231rY, X86::VFNMADDPDr231mY, TB_ALIGN_NONE }, 1807 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr132mY, TB_ALIGN_NONE }, 1808 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr132mY, TB_ALIGN_NONE }, 1809 { X86::VFNMADDPSr213rY, X86::VFNMADDPSr213mY, TB_ALIGN_NONE }, 1810 { X86::VFNMADDPDr213rY, X86::VFNMADDPDr213mY, TB_ALIGN_NONE }, 1811 1812 { X86::VFMSUBSSr231r, X86::VFMSUBSSr231m, TB_ALIGN_NONE }, 1813 { X86::VFMSUBSSr231r_Int, X86::VFMSUBSSr231m_Int, TB_ALIGN_NONE }, 1814 { X86::VFMSUBSDr231r, X86::VFMSUBSDr231m, TB_ALIGN_NONE }, 1815 { X86::VFMSUBSDr231r_Int, X86::VFMSUBSDr231m_Int, TB_ALIGN_NONE }, 1816 { X86::VFMSUBSSr132r, X86::VFMSUBSSr132m, TB_ALIGN_NONE }, 1817 { X86::VFMSUBSSr132r_Int, X86::VFMSUBSSr132m_Int, TB_ALIGN_NONE }, 1818 { X86::VFMSUBSDr132r, X86::VFMSUBSDr132m, TB_ALIGN_NONE }, 1819 { X86::VFMSUBSDr132r_Int, X86::VFMSUBSDr132m_Int, TB_ALIGN_NONE }, 1820 { X86::VFMSUBSSr213r, X86::VFMSUBSSr213m, TB_ALIGN_NONE }, 1821 { X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr213m_Int, TB_ALIGN_NONE }, 1822 { X86::VFMSUBSDr213r, X86::VFMSUBSDr213m, TB_ALIGN_NONE }, 1823 { X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr213m_Int, TB_ALIGN_NONE }, 1824 1825 { X86::VFMSUBPSr231r, X86::VFMSUBPSr231m, TB_ALIGN_NONE }, 1826 { X86::VFMSUBPDr231r, X86::VFMSUBPDr231m, TB_ALIGN_NONE }, 1827 { X86::VFMSUBPSr132r, X86::VFMSUBPSr132m, TB_ALIGN_NONE }, 1828 { X86::VFMSUBPDr132r, X86::VFMSUBPDr132m, TB_ALIGN_NONE }, 1829 { X86::VFMSUBPSr213r, X86::VFMSUBPSr213m, TB_ALIGN_NONE }, 1830 { X86::VFMSUBPDr213r, X86::VFMSUBPDr213m, TB_ALIGN_NONE }, 1831 { X86::VFMSUBPSr231rY, X86::VFMSUBPSr231mY, TB_ALIGN_NONE }, 1832 { X86::VFMSUBPDr231rY, X86::VFMSUBPDr231mY, TB_ALIGN_NONE }, 1833 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr132mY, TB_ALIGN_NONE }, 1834 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr132mY, TB_ALIGN_NONE }, 1835 { X86::VFMSUBPSr213rY, X86::VFMSUBPSr213mY, TB_ALIGN_NONE }, 1836 { X86::VFMSUBPDr213rY, X86::VFMSUBPDr213mY, TB_ALIGN_NONE }, 1837 1838 { X86::VFNMSUBSSr231r, X86::VFNMSUBSSr231m, TB_ALIGN_NONE }, 1839 { X86::VFNMSUBSSr231r_Int, X86::VFNMSUBSSr231m_Int, TB_ALIGN_NONE }, 1840 { X86::VFNMSUBSDr231r, X86::VFNMSUBSDr231m, TB_ALIGN_NONE }, 1841 { X86::VFNMSUBSDr231r_Int, X86::VFNMSUBSDr231m_Int, TB_ALIGN_NONE }, 1842 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr132m, TB_ALIGN_NONE }, 1843 { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr132m_Int, TB_ALIGN_NONE }, 1844 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr132m, TB_ALIGN_NONE }, 1845 { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr132m_Int, TB_ALIGN_NONE }, 1846 { X86::VFNMSUBSSr213r, X86::VFNMSUBSSr213m, TB_ALIGN_NONE }, 1847 { X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr213m_Int, TB_ALIGN_NONE }, 1848 { X86::VFNMSUBSDr213r, X86::VFNMSUBSDr213m, TB_ALIGN_NONE }, 1849 { X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr213m_Int, TB_ALIGN_NONE }, 1850 1851 { X86::VFNMSUBPSr231r, X86::VFNMSUBPSr231m, TB_ALIGN_NONE }, 1852 { X86::VFNMSUBPDr231r, X86::VFNMSUBPDr231m, TB_ALIGN_NONE }, 1853 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr132m, TB_ALIGN_NONE }, 1854 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr132m, TB_ALIGN_NONE }, 1855 { X86::VFNMSUBPSr213r, X86::VFNMSUBPSr213m, TB_ALIGN_NONE }, 1856 { X86::VFNMSUBPDr213r, X86::VFNMSUBPDr213m, TB_ALIGN_NONE }, 1857 { X86::VFNMSUBPSr231rY, X86::VFNMSUBPSr231mY, TB_ALIGN_NONE }, 1858 { X86::VFNMSUBPDr231rY, X86::VFNMSUBPDr231mY, TB_ALIGN_NONE }, 1859 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr132mY, TB_ALIGN_NONE }, 1860 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr132mY, TB_ALIGN_NONE }, 1861 { X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr213mY, TB_ALIGN_NONE }, 1862 { X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr213mY, TB_ALIGN_NONE }, 1863 1864 { X86::VFMADDSUBPSr231r, X86::VFMADDSUBPSr231m, TB_ALIGN_NONE }, 1865 { X86::VFMADDSUBPDr231r, X86::VFMADDSUBPDr231m, TB_ALIGN_NONE }, 1866 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr132m, TB_ALIGN_NONE }, 1867 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr132m, TB_ALIGN_NONE }, 1868 { X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr213m, TB_ALIGN_NONE }, 1869 { X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr213m, TB_ALIGN_NONE }, 1870 { X86::VFMADDSUBPSr231rY, X86::VFMADDSUBPSr231mY, TB_ALIGN_NONE }, 1871 { X86::VFMADDSUBPDr231rY, X86::VFMADDSUBPDr231mY, TB_ALIGN_NONE }, 1872 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr132mY, TB_ALIGN_NONE }, 1873 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr132mY, TB_ALIGN_NONE }, 1874 { X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr213mY, TB_ALIGN_NONE }, 1875 { X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr213mY, TB_ALIGN_NONE }, 1876 1877 { X86::VFMSUBADDPSr231r, X86::VFMSUBADDPSr231m, TB_ALIGN_NONE }, 1878 { X86::VFMSUBADDPDr231r, X86::VFMSUBADDPDr231m, TB_ALIGN_NONE }, 1879 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr132m, TB_ALIGN_NONE }, 1880 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr132m, TB_ALIGN_NONE }, 1881 { X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr213m, TB_ALIGN_NONE }, 1882 { X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr213m, TB_ALIGN_NONE }, 1883 { X86::VFMSUBADDPSr231rY, X86::VFMSUBADDPSr231mY, TB_ALIGN_NONE }, 1884 { X86::VFMSUBADDPDr231rY, X86::VFMSUBADDPDr231mY, TB_ALIGN_NONE }, 1885 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr132mY, TB_ALIGN_NONE }, 1886 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr132mY, TB_ALIGN_NONE }, 1887 { X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr213mY, TB_ALIGN_NONE }, 1888 { X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr213mY, TB_ALIGN_NONE }, 1889 1890 // FMA4 foldable patterns 1891 { X86::VFMADDSS4rr, X86::VFMADDSS4rm, TB_ALIGN_NONE }, 1892 { X86::VFMADDSD4rr, X86::VFMADDSD4rm, TB_ALIGN_NONE }, 1893 { X86::VFMADDPS4rr, X86::VFMADDPS4rm, TB_ALIGN_NONE }, 1894 { X86::VFMADDPD4rr, X86::VFMADDPD4rm, TB_ALIGN_NONE }, 1895 { X86::VFMADDPS4rrY, X86::VFMADDPS4rmY, TB_ALIGN_NONE }, 1896 { X86::VFMADDPD4rrY, X86::VFMADDPD4rmY, TB_ALIGN_NONE }, 1897 { X86::VFNMADDSS4rr, X86::VFNMADDSS4rm, TB_ALIGN_NONE }, 1898 { X86::VFNMADDSD4rr, X86::VFNMADDSD4rm, TB_ALIGN_NONE }, 1899 { X86::VFNMADDPS4rr, X86::VFNMADDPS4rm, TB_ALIGN_NONE }, 1900 { X86::VFNMADDPD4rr, X86::VFNMADDPD4rm, TB_ALIGN_NONE }, 1901 { X86::VFNMADDPS4rrY, X86::VFNMADDPS4rmY, TB_ALIGN_NONE }, 1902 { X86::VFNMADDPD4rrY, X86::VFNMADDPD4rmY, TB_ALIGN_NONE }, 1903 { X86::VFMSUBSS4rr, X86::VFMSUBSS4rm, TB_ALIGN_NONE }, 1904 { X86::VFMSUBSD4rr, X86::VFMSUBSD4rm, TB_ALIGN_NONE }, 1905 { X86::VFMSUBPS4rr, X86::VFMSUBPS4rm, TB_ALIGN_NONE }, 1906 { X86::VFMSUBPD4rr, X86::VFMSUBPD4rm, TB_ALIGN_NONE }, 1907 { X86::VFMSUBPS4rrY, X86::VFMSUBPS4rmY, TB_ALIGN_NONE }, 1908 { X86::VFMSUBPD4rrY, X86::VFMSUBPD4rmY, TB_ALIGN_NONE }, 1909 { X86::VFNMSUBSS4rr, X86::VFNMSUBSS4rm, TB_ALIGN_NONE }, 1910 { X86::VFNMSUBSD4rr, X86::VFNMSUBSD4rm, TB_ALIGN_NONE }, 1911 { X86::VFNMSUBPS4rr, X86::VFNMSUBPS4rm, TB_ALIGN_NONE }, 1912 { X86::VFNMSUBPD4rr, X86::VFNMSUBPD4rm, TB_ALIGN_NONE }, 1913 { X86::VFNMSUBPS4rrY, X86::VFNMSUBPS4rmY, TB_ALIGN_NONE }, 1914 { X86::VFNMSUBPD4rrY, X86::VFNMSUBPD4rmY, TB_ALIGN_NONE }, 1915 { X86::VFMADDSUBPS4rr, X86::VFMADDSUBPS4rm, TB_ALIGN_NONE }, 1916 { X86::VFMADDSUBPD4rr, X86::VFMADDSUBPD4rm, TB_ALIGN_NONE }, 1917 { X86::VFMADDSUBPS4rrY, X86::VFMADDSUBPS4rmY, TB_ALIGN_NONE }, 1918 { X86::VFMADDSUBPD4rrY, X86::VFMADDSUBPD4rmY, TB_ALIGN_NONE }, 1919 { X86::VFMSUBADDPS4rr, X86::VFMSUBADDPS4rm, TB_ALIGN_NONE }, 1920 { X86::VFMSUBADDPD4rr, X86::VFMSUBADDPD4rm, TB_ALIGN_NONE }, 1921 { X86::VFMSUBADDPS4rrY, X86::VFMSUBADDPS4rmY, TB_ALIGN_NONE }, 1922 { X86::VFMSUBADDPD4rrY, X86::VFMSUBADDPD4rmY, TB_ALIGN_NONE }, 1923 1924 // XOP foldable instructions 1925 { X86::VPCMOVrrr, X86::VPCMOVrrm, 0 }, 1926 { X86::VPCMOVrrrY, X86::VPCMOVrrmY, 0 }, 1927 { X86::VPERMIL2PDrr, X86::VPERMIL2PDrm, 0 }, 1928 { X86::VPERMIL2PDrrY, X86::VPERMIL2PDrmY, 0 }, 1929 { X86::VPERMIL2PSrr, X86::VPERMIL2PSrm, 0 }, 1930 { X86::VPERMIL2PSrrY, X86::VPERMIL2PSrmY, 0 }, 1931 { X86::VPPERMrrr, X86::VPPERMrrm, 0 }, 1932 1933 // AVX-512 VPERMI instructions with 3 source operands. 1934 { X86::VPERMI2Drr, X86::VPERMI2Drm, 0 }, 1935 { X86::VPERMI2Qrr, X86::VPERMI2Qrm, 0 }, 1936 { X86::VPERMI2PSrr, X86::VPERMI2PSrm, 0 }, 1937 { X86::VPERMI2PDrr, X86::VPERMI2PDrm, 0 }, 1938 { X86::VBLENDMPDZrr, X86::VBLENDMPDZrm, 0 }, 1939 { X86::VBLENDMPSZrr, X86::VBLENDMPSZrm, 0 }, 1940 { X86::VPBLENDMDZrr, X86::VPBLENDMDZrm, 0 }, 1941 { X86::VPBLENDMQZrr, X86::VPBLENDMQZrm, 0 }, 1942 { X86::VBROADCASTSSZrk, X86::VBROADCASTSSZmk, TB_NO_REVERSE }, 1943 { X86::VBROADCASTSDZrk, X86::VBROADCASTSDZmk, TB_NO_REVERSE }, 1944 { X86::VBROADCASTSSZ256rk, X86::VBROADCASTSSZ256mk, TB_NO_REVERSE }, 1945 { X86::VBROADCASTSDZ256rk, X86::VBROADCASTSDZ256mk, TB_NO_REVERSE }, 1946 { X86::VBROADCASTSSZ128rk, X86::VBROADCASTSSZ128mk, TB_NO_REVERSE }, 1947 // AVX-512 arithmetic instructions 1948 { X86::VADDPSZrrkz, X86::VADDPSZrmkz, 0 }, 1949 { X86::VADDPDZrrkz, X86::VADDPDZrmkz, 0 }, 1950 { X86::VSUBPSZrrkz, X86::VSUBPSZrmkz, 0 }, 1951 { X86::VSUBPDZrrkz, X86::VSUBPDZrmkz, 0 }, 1952 { X86::VMULPSZrrkz, X86::VMULPSZrmkz, 0 }, 1953 { X86::VMULPDZrrkz, X86::VMULPDZrmkz, 0 }, 1954 { X86::VDIVPSZrrkz, X86::VDIVPSZrmkz, 0 }, 1955 { X86::VDIVPDZrrkz, X86::VDIVPDZrmkz, 0 }, 1956 { X86::VMINPSZrrkz, X86::VMINPSZrmkz, 0 }, 1957 { X86::VMINPDZrrkz, X86::VMINPDZrmkz, 0 }, 1958 { X86::VMAXPSZrrkz, X86::VMAXPSZrmkz, 0 }, 1959 { X86::VMAXPDZrrkz, X86::VMAXPDZrmkz, 0 }, 1960 // AVX-512{F,VL} arithmetic instructions 256-bit 1961 { X86::VADDPSZ256rrkz, X86::VADDPSZ256rmkz, 0 }, 1962 { X86::VADDPDZ256rrkz, X86::VADDPDZ256rmkz, 0 }, 1963 { X86::VSUBPSZ256rrkz, X86::VSUBPSZ256rmkz, 0 }, 1964 { X86::VSUBPDZ256rrkz, X86::VSUBPDZ256rmkz, 0 }, 1965 { X86::VMULPSZ256rrkz, X86::VMULPSZ256rmkz, 0 }, 1966 { X86::VMULPDZ256rrkz, X86::VMULPDZ256rmkz, 0 }, 1967 { X86::VDIVPSZ256rrkz, X86::VDIVPSZ256rmkz, 0 }, 1968 { X86::VDIVPDZ256rrkz, X86::VDIVPDZ256rmkz, 0 }, 1969 { X86::VMINPSZ256rrkz, X86::VMINPSZ256rmkz, 0 }, 1970 { X86::VMINPDZ256rrkz, X86::VMINPDZ256rmkz, 0 }, 1971 { X86::VMAXPSZ256rrkz, X86::VMAXPSZ256rmkz, 0 }, 1972 { X86::VMAXPDZ256rrkz, X86::VMAXPDZ256rmkz, 0 }, 1973 // AVX-512{F,VL} arithmetic instructions 128-bit 1974 { X86::VADDPSZ128rrkz, X86::VADDPSZ128rmkz, 0 }, 1975 { X86::VADDPDZ128rrkz, X86::VADDPDZ128rmkz, 0 }, 1976 { X86::VSUBPSZ128rrkz, X86::VSUBPSZ128rmkz, 0 }, 1977 { X86::VSUBPDZ128rrkz, X86::VSUBPDZ128rmkz, 0 }, 1978 { X86::VMULPSZ128rrkz, X86::VMULPSZ128rmkz, 0 }, 1979 { X86::VMULPDZ128rrkz, X86::VMULPDZ128rmkz, 0 }, 1980 { X86::VDIVPSZ128rrkz, X86::VDIVPSZ128rmkz, 0 }, 1981 { X86::VDIVPDZ128rrkz, X86::VDIVPDZ128rmkz, 0 }, 1982 { X86::VMINPSZ128rrkz, X86::VMINPSZ128rmkz, 0 }, 1983 { X86::VMINPDZ128rrkz, X86::VMINPDZ128rmkz, 0 }, 1984 { X86::VMAXPSZ128rrkz, X86::VMAXPSZ128rmkz, 0 }, 1985 { X86::VMAXPDZ128rrkz, X86::VMAXPDZ128rmkz, 0 } 1986 }; 1987 1988 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable3) { 1989 AddTableEntry(RegOp2MemOpTable3, MemOp2RegOpTable, 1990 Entry.RegOp, Entry.MemOp, 1991 // Index 3, folded load 1992 Entry.Flags | TB_INDEX_3 | TB_FOLDED_LOAD); 1993 } 1994 1995 static const X86MemoryFoldTableEntry MemoryFoldTable4[] = { 1996 // AVX-512 foldable instructions 1997 { X86::VADDPSZrrk, X86::VADDPSZrmk, 0 }, 1998 { X86::VADDPDZrrk, X86::VADDPDZrmk, 0 }, 1999 { X86::VSUBPSZrrk, X86::VSUBPSZrmk, 0 }, 2000 { X86::VSUBPDZrrk, X86::VSUBPDZrmk, 0 }, 2001 { X86::VMULPSZrrk, X86::VMULPSZrmk, 0 }, 2002 { X86::VMULPDZrrk, X86::VMULPDZrmk, 0 }, 2003 { X86::VDIVPSZrrk, X86::VDIVPSZrmk, 0 }, 2004 { X86::VDIVPDZrrk, X86::VDIVPDZrmk, 0 }, 2005 { X86::VMINPSZrrk, X86::VMINPSZrmk, 0 }, 2006 { X86::VMINPDZrrk, X86::VMINPDZrmk, 0 }, 2007 { X86::VMAXPSZrrk, X86::VMAXPSZrmk, 0 }, 2008 { X86::VMAXPDZrrk, X86::VMAXPDZrmk, 0 }, 2009 // AVX-512{F,VL} foldable instructions 256-bit 2010 { X86::VADDPSZ256rrk, X86::VADDPSZ256rmk, 0 }, 2011 { X86::VADDPDZ256rrk, X86::VADDPDZ256rmk, 0 }, 2012 { X86::VSUBPSZ256rrk, X86::VSUBPSZ256rmk, 0 }, 2013 { X86::VSUBPDZ256rrk, X86::VSUBPDZ256rmk, 0 }, 2014 { X86::VMULPSZ256rrk, X86::VMULPSZ256rmk, 0 }, 2015 { X86::VMULPDZ256rrk, X86::VMULPDZ256rmk, 0 }, 2016 { X86::VDIVPSZ256rrk, X86::VDIVPSZ256rmk, 0 }, 2017 { X86::VDIVPDZ256rrk, X86::VDIVPDZ256rmk, 0 }, 2018 { X86::VMINPSZ256rrk, X86::VMINPSZ256rmk, 0 }, 2019 { X86::VMINPDZ256rrk, X86::VMINPDZ256rmk, 0 }, 2020 { X86::VMAXPSZ256rrk, X86::VMAXPSZ256rmk, 0 }, 2021 { X86::VMAXPDZ256rrk, X86::VMAXPDZ256rmk, 0 }, 2022 // AVX-512{F,VL} foldable instructions 128-bit 2023 { X86::VADDPSZ128rrk, X86::VADDPSZ128rmk, 0 }, 2024 { X86::VADDPDZ128rrk, X86::VADDPDZ128rmk, 0 }, 2025 { X86::VSUBPSZ128rrk, X86::VSUBPSZ128rmk, 0 }, 2026 { X86::VSUBPDZ128rrk, X86::VSUBPDZ128rmk, 0 }, 2027 { X86::VMULPSZ128rrk, X86::VMULPSZ128rmk, 0 }, 2028 { X86::VMULPDZ128rrk, X86::VMULPDZ128rmk, 0 }, 2029 { X86::VDIVPSZ128rrk, X86::VDIVPSZ128rmk, 0 }, 2030 { X86::VDIVPDZ128rrk, X86::VDIVPDZ128rmk, 0 }, 2031 { X86::VMINPSZ128rrk, X86::VMINPSZ128rmk, 0 }, 2032 { X86::VMINPDZ128rrk, X86::VMINPDZ128rmk, 0 }, 2033 { X86::VMAXPSZ128rrk, X86::VMAXPSZ128rmk, 0 }, 2034 { X86::VMAXPDZ128rrk, X86::VMAXPDZ128rmk, 0 } 2035 }; 2036 2037 for (X86MemoryFoldTableEntry Entry : MemoryFoldTable4) { 2038 AddTableEntry(RegOp2MemOpTable4, MemOp2RegOpTable, 2039 Entry.RegOp, Entry.MemOp, 2040 // Index 4, folded load 2041 Entry.Flags | TB_INDEX_4 | TB_FOLDED_LOAD); 2042 } 2043 } 2044 2045 void 2046 X86InstrInfo::AddTableEntry(RegOp2MemOpTableType &R2MTable, 2047 MemOp2RegOpTableType &M2RTable, 2048 uint16_t RegOp, uint16_t MemOp, uint16_t Flags) { 2049 if ((Flags & TB_NO_FORWARD) == 0) { 2050 assert(!R2MTable.count(RegOp) && "Duplicate entry!"); 2051 R2MTable[RegOp] = std::make_pair(MemOp, Flags); 2052 } 2053 if ((Flags & TB_NO_REVERSE) == 0) { 2054 assert(!M2RTable.count(MemOp) && 2055 "Duplicated entries in unfolding maps?"); 2056 M2RTable[MemOp] = std::make_pair(RegOp, Flags); 2057 } 2058 } 2059 2060 bool 2061 X86InstrInfo::isCoalescableExtInstr(const MachineInstr &MI, 2062 unsigned &SrcReg, unsigned &DstReg, 2063 unsigned &SubIdx) const { 2064 switch (MI.getOpcode()) { 2065 default: break; 2066 case X86::MOVSX16rr8: 2067 case X86::MOVZX16rr8: 2068 case X86::MOVSX32rr8: 2069 case X86::MOVZX32rr8: 2070 case X86::MOVSX64rr8: 2071 if (!Subtarget.is64Bit()) 2072 // It's not always legal to reference the low 8-bit of the larger 2073 // register in 32-bit mode. 2074 return false; 2075 case X86::MOVSX32rr16: 2076 case X86::MOVZX32rr16: 2077 case X86::MOVSX64rr16: 2078 case X86::MOVSX64rr32: { 2079 if (MI.getOperand(0).getSubReg() || MI.getOperand(1).getSubReg()) 2080 // Be conservative. 2081 return false; 2082 SrcReg = MI.getOperand(1).getReg(); 2083 DstReg = MI.getOperand(0).getReg(); 2084 switch (MI.getOpcode()) { 2085 default: llvm_unreachable("Unreachable!"); 2086 case X86::MOVSX16rr8: 2087 case X86::MOVZX16rr8: 2088 case X86::MOVSX32rr8: 2089 case X86::MOVZX32rr8: 2090 case X86::MOVSX64rr8: 2091 SubIdx = X86::sub_8bit; 2092 break; 2093 case X86::MOVSX32rr16: 2094 case X86::MOVZX32rr16: 2095 case X86::MOVSX64rr16: 2096 SubIdx = X86::sub_16bit; 2097 break; 2098 case X86::MOVSX64rr32: 2099 SubIdx = X86::sub_32bit; 2100 break; 2101 } 2102 return true; 2103 } 2104 } 2105 return false; 2106 } 2107 2108 int X86InstrInfo::getSPAdjust(const MachineInstr *MI) const { 2109 const MachineFunction *MF = MI->getParent()->getParent(); 2110 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 2111 2112 if (MI->getOpcode() == getCallFrameSetupOpcode() || 2113 MI->getOpcode() == getCallFrameDestroyOpcode()) { 2114 unsigned StackAlign = TFI->getStackAlignment(); 2115 int SPAdj = (MI->getOperand(0).getImm() + StackAlign - 1) / StackAlign * 2116 StackAlign; 2117 2118 SPAdj -= MI->getOperand(1).getImm(); 2119 2120 if (MI->getOpcode() == getCallFrameSetupOpcode()) 2121 return SPAdj; 2122 else 2123 return -SPAdj; 2124 } 2125 2126 // To know whether a call adjusts the stack, we need information 2127 // that is bound to the following ADJCALLSTACKUP pseudo. 2128 // Look for the next ADJCALLSTACKUP that follows the call. 2129 if (MI->isCall()) { 2130 const MachineBasicBlock* MBB = MI->getParent(); 2131 auto I = ++MachineBasicBlock::const_iterator(MI); 2132 for (auto E = MBB->end(); I != E; ++I) { 2133 if (I->getOpcode() == getCallFrameDestroyOpcode() || 2134 I->isCall()) 2135 break; 2136 } 2137 2138 // If we could not find a frame destroy opcode, then it has already 2139 // been simplified, so we don't care. 2140 if (I->getOpcode() != getCallFrameDestroyOpcode()) 2141 return 0; 2142 2143 return -(I->getOperand(1).getImm()); 2144 } 2145 2146 // Currently handle only PUSHes we can reasonably expect to see 2147 // in call sequences 2148 switch (MI->getOpcode()) { 2149 default: 2150 return 0; 2151 case X86::PUSH32i8: 2152 case X86::PUSH32r: 2153 case X86::PUSH32rmm: 2154 case X86::PUSH32rmr: 2155 case X86::PUSHi32: 2156 return 4; 2157 case X86::PUSH64i8: 2158 case X86::PUSH64r: 2159 case X86::PUSH64rmm: 2160 case X86::PUSH64rmr: 2161 case X86::PUSH64i32: 2162 return 8; 2163 } 2164 } 2165 2166 /// Return true and the FrameIndex if the specified 2167 /// operand and follow operands form a reference to the stack frame. 2168 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op, 2169 int &FrameIndex) const { 2170 if (MI->getOperand(Op+X86::AddrBaseReg).isFI() && 2171 MI->getOperand(Op+X86::AddrScaleAmt).isImm() && 2172 MI->getOperand(Op+X86::AddrIndexReg).isReg() && 2173 MI->getOperand(Op+X86::AddrDisp).isImm() && 2174 MI->getOperand(Op+X86::AddrScaleAmt).getImm() == 1 && 2175 MI->getOperand(Op+X86::AddrIndexReg).getReg() == 0 && 2176 MI->getOperand(Op+X86::AddrDisp).getImm() == 0) { 2177 FrameIndex = MI->getOperand(Op+X86::AddrBaseReg).getIndex(); 2178 return true; 2179 } 2180 return false; 2181 } 2182 2183 static bool isFrameLoadOpcode(int Opcode) { 2184 switch (Opcode) { 2185 default: 2186 return false; 2187 case X86::MOV8rm: 2188 case X86::MOV16rm: 2189 case X86::MOV32rm: 2190 case X86::MOV64rm: 2191 case X86::LD_Fp64m: 2192 case X86::MOVSSrm: 2193 case X86::MOVSDrm: 2194 case X86::MOVAPSrm: 2195 case X86::MOVAPDrm: 2196 case X86::MOVDQArm: 2197 case X86::VMOVSSrm: 2198 case X86::VMOVSDrm: 2199 case X86::VMOVAPSrm: 2200 case X86::VMOVAPDrm: 2201 case X86::VMOVDQArm: 2202 case X86::VMOVUPSYrm: 2203 case X86::VMOVAPSYrm: 2204 case X86::VMOVUPDYrm: 2205 case X86::VMOVAPDYrm: 2206 case X86::VMOVDQUYrm: 2207 case X86::VMOVDQAYrm: 2208 case X86::MMX_MOVD64rm: 2209 case X86::MMX_MOVQ64rm: 2210 case X86::VMOVAPSZrm: 2211 case X86::VMOVUPSZrm: 2212 return true; 2213 } 2214 } 2215 2216 static bool isFrameStoreOpcode(int Opcode) { 2217 switch (Opcode) { 2218 default: break; 2219 case X86::MOV8mr: 2220 case X86::MOV16mr: 2221 case X86::MOV32mr: 2222 case X86::MOV64mr: 2223 case X86::ST_FpP64m: 2224 case X86::MOVSSmr: 2225 case X86::MOVSDmr: 2226 case X86::MOVAPSmr: 2227 case X86::MOVAPDmr: 2228 case X86::MOVDQAmr: 2229 case X86::VMOVSSmr: 2230 case X86::VMOVSDmr: 2231 case X86::VMOVAPSmr: 2232 case X86::VMOVAPDmr: 2233 case X86::VMOVDQAmr: 2234 case X86::VMOVUPSYmr: 2235 case X86::VMOVAPSYmr: 2236 case X86::VMOVUPDYmr: 2237 case X86::VMOVAPDYmr: 2238 case X86::VMOVDQUYmr: 2239 case X86::VMOVDQAYmr: 2240 case X86::VMOVUPSZmr: 2241 case X86::VMOVAPSZmr: 2242 case X86::MMX_MOVD64mr: 2243 case X86::MMX_MOVQ64mr: 2244 case X86::MMX_MOVNTQmr: 2245 return true; 2246 } 2247 return false; 2248 } 2249 2250 unsigned X86InstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 2251 int &FrameIndex) const { 2252 if (isFrameLoadOpcode(MI->getOpcode())) 2253 if (MI->getOperand(0).getSubReg() == 0 && isFrameOperand(MI, 1, FrameIndex)) 2254 return MI->getOperand(0).getReg(); 2255 return 0; 2256 } 2257 2258 unsigned X86InstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI, 2259 int &FrameIndex) const { 2260 if (isFrameLoadOpcode(MI->getOpcode())) { 2261 unsigned Reg; 2262 if ((Reg = isLoadFromStackSlot(MI, FrameIndex))) 2263 return Reg; 2264 // Check for post-frame index elimination operations 2265 const MachineMemOperand *Dummy; 2266 return hasLoadFromStackSlot(MI, Dummy, FrameIndex); 2267 } 2268 return 0; 2269 } 2270 2271 unsigned X86InstrInfo::isStoreToStackSlot(const MachineInstr *MI, 2272 int &FrameIndex) const { 2273 if (isFrameStoreOpcode(MI->getOpcode())) 2274 if (MI->getOperand(X86::AddrNumOperands).getSubReg() == 0 && 2275 isFrameOperand(MI, 0, FrameIndex)) 2276 return MI->getOperand(X86::AddrNumOperands).getReg(); 2277 return 0; 2278 } 2279 2280 unsigned X86InstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI, 2281 int &FrameIndex) const { 2282 if (isFrameStoreOpcode(MI->getOpcode())) { 2283 unsigned Reg; 2284 if ((Reg = isStoreToStackSlot(MI, FrameIndex))) 2285 return Reg; 2286 // Check for post-frame index elimination operations 2287 const MachineMemOperand *Dummy; 2288 return hasStoreToStackSlot(MI, Dummy, FrameIndex); 2289 } 2290 return 0; 2291 } 2292 2293 /// Return true if register is PIC base; i.e.g defined by X86::MOVPC32r. 2294 static bool regIsPICBase(unsigned BaseReg, const MachineRegisterInfo &MRI) { 2295 // Don't waste compile time scanning use-def chains of physregs. 2296 if (!TargetRegisterInfo::isVirtualRegister(BaseReg)) 2297 return false; 2298 bool isPICBase = false; 2299 for (MachineRegisterInfo::def_instr_iterator I = MRI.def_instr_begin(BaseReg), 2300 E = MRI.def_instr_end(); I != E; ++I) { 2301 MachineInstr *DefMI = &*I; 2302 if (DefMI->getOpcode() != X86::MOVPC32r) 2303 return false; 2304 assert(!isPICBase && "More than one PIC base?"); 2305 isPICBase = true; 2306 } 2307 return isPICBase; 2308 } 2309 2310 bool 2311 X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI, 2312 AliasAnalysis *AA) const { 2313 switch (MI->getOpcode()) { 2314 default: break; 2315 case X86::MOV8rm: 2316 case X86::MOV16rm: 2317 case X86::MOV32rm: 2318 case X86::MOV64rm: 2319 case X86::LD_Fp64m: 2320 case X86::MOVSSrm: 2321 case X86::MOVSDrm: 2322 case X86::MOVAPSrm: 2323 case X86::MOVUPSrm: 2324 case X86::MOVAPDrm: 2325 case X86::MOVDQArm: 2326 case X86::MOVDQUrm: 2327 case X86::VMOVSSrm: 2328 case X86::VMOVSDrm: 2329 case X86::VMOVAPSrm: 2330 case X86::VMOVUPSrm: 2331 case X86::VMOVAPDrm: 2332 case X86::VMOVDQArm: 2333 case X86::VMOVDQUrm: 2334 case X86::VMOVAPSYrm: 2335 case X86::VMOVUPSYrm: 2336 case X86::VMOVAPDYrm: 2337 case X86::VMOVDQAYrm: 2338 case X86::VMOVDQUYrm: 2339 case X86::MMX_MOVD64rm: 2340 case X86::MMX_MOVQ64rm: 2341 case X86::FsVMOVAPSrm: 2342 case X86::FsVMOVAPDrm: 2343 case X86::FsMOVAPSrm: 2344 case X86::FsMOVAPDrm: 2345 // AVX-512 2346 case X86::VMOVAPDZ128rm: 2347 case X86::VMOVAPDZ256rm: 2348 case X86::VMOVAPDZrm: 2349 case X86::VMOVAPSZ128rm: 2350 case X86::VMOVAPSZ256rm: 2351 case X86::VMOVAPSZrm: 2352 case X86::VMOVDQA32Z128rm: 2353 case X86::VMOVDQA32Z256rm: 2354 case X86::VMOVDQA32Zrm: 2355 case X86::VMOVDQA64Z128rm: 2356 case X86::VMOVDQA64Z256rm: 2357 case X86::VMOVDQA64Zrm: 2358 case X86::VMOVDQU16Z128rm: 2359 case X86::VMOVDQU16Z256rm: 2360 case X86::VMOVDQU16Zrm: 2361 case X86::VMOVDQU32Z128rm: 2362 case X86::VMOVDQU32Z256rm: 2363 case X86::VMOVDQU32Zrm: 2364 case X86::VMOVDQU64Z128rm: 2365 case X86::VMOVDQU64Z256rm: 2366 case X86::VMOVDQU64Zrm: 2367 case X86::VMOVDQU8Z128rm: 2368 case X86::VMOVDQU8Z256rm: 2369 case X86::VMOVDQU8Zrm: 2370 case X86::VMOVUPSZ128rm: 2371 case X86::VMOVUPSZ256rm: 2372 case X86::VMOVUPSZrm: { 2373 // Loads from constant pools are trivially rematerializable. 2374 if (MI->getOperand(1+X86::AddrBaseReg).isReg() && 2375 MI->getOperand(1+X86::AddrScaleAmt).isImm() && 2376 MI->getOperand(1+X86::AddrIndexReg).isReg() && 2377 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 && 2378 MI->isInvariantLoad(AA)) { 2379 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg(); 2380 if (BaseReg == 0 || BaseReg == X86::RIP) 2381 return true; 2382 // Allow re-materialization of PIC load. 2383 if (!ReMatPICStubLoad && MI->getOperand(1+X86::AddrDisp).isGlobal()) 2384 return false; 2385 const MachineFunction &MF = *MI->getParent()->getParent(); 2386 const MachineRegisterInfo &MRI = MF.getRegInfo(); 2387 return regIsPICBase(BaseReg, MRI); 2388 } 2389 return false; 2390 } 2391 2392 case X86::LEA32r: 2393 case X86::LEA64r: { 2394 if (MI->getOperand(1+X86::AddrScaleAmt).isImm() && 2395 MI->getOperand(1+X86::AddrIndexReg).isReg() && 2396 MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 && 2397 !MI->getOperand(1+X86::AddrDisp).isReg()) { 2398 // lea fi#, lea GV, etc. are all rematerializable. 2399 if (!MI->getOperand(1+X86::AddrBaseReg).isReg()) 2400 return true; 2401 unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg(); 2402 if (BaseReg == 0) 2403 return true; 2404 // Allow re-materialization of lea PICBase + x. 2405 const MachineFunction &MF = *MI->getParent()->getParent(); 2406 const MachineRegisterInfo &MRI = MF.getRegInfo(); 2407 return regIsPICBase(BaseReg, MRI); 2408 } 2409 return false; 2410 } 2411 } 2412 2413 // All other instructions marked M_REMATERIALIZABLE are always trivially 2414 // rematerializable. 2415 return true; 2416 } 2417 2418 bool X86InstrInfo::isSafeToClobberEFLAGS(MachineBasicBlock &MBB, 2419 MachineBasicBlock::iterator I) const { 2420 MachineBasicBlock::iterator E = MBB.end(); 2421 2422 // For compile time consideration, if we are not able to determine the 2423 // safety after visiting 4 instructions in each direction, we will assume 2424 // it's not safe. 2425 MachineBasicBlock::iterator Iter = I; 2426 for (unsigned i = 0; Iter != E && i < 4; ++i) { 2427 bool SeenDef = false; 2428 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 2429 MachineOperand &MO = Iter->getOperand(j); 2430 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 2431 SeenDef = true; 2432 if (!MO.isReg()) 2433 continue; 2434 if (MO.getReg() == X86::EFLAGS) { 2435 if (MO.isUse()) 2436 return false; 2437 SeenDef = true; 2438 } 2439 } 2440 2441 if (SeenDef) 2442 // This instruction defines EFLAGS, no need to look any further. 2443 return true; 2444 ++Iter; 2445 // Skip over DBG_VALUE. 2446 while (Iter != E && Iter->isDebugValue()) 2447 ++Iter; 2448 } 2449 2450 // It is safe to clobber EFLAGS at the end of a block of no successor has it 2451 // live in. 2452 if (Iter == E) { 2453 for (MachineBasicBlock *S : MBB.successors()) 2454 if (S->isLiveIn(X86::EFLAGS)) 2455 return false; 2456 return true; 2457 } 2458 2459 MachineBasicBlock::iterator B = MBB.begin(); 2460 Iter = I; 2461 for (unsigned i = 0; i < 4; ++i) { 2462 // If we make it to the beginning of the block, it's safe to clobber 2463 // EFLAGS iff EFLAGS is not live-in. 2464 if (Iter == B) 2465 return !MBB.isLiveIn(X86::EFLAGS); 2466 2467 --Iter; 2468 // Skip over DBG_VALUE. 2469 while (Iter != B && Iter->isDebugValue()) 2470 --Iter; 2471 2472 bool SawKill = false; 2473 for (unsigned j = 0, e = Iter->getNumOperands(); j != e; ++j) { 2474 MachineOperand &MO = Iter->getOperand(j); 2475 // A register mask may clobber EFLAGS, but we should still look for a 2476 // live EFLAGS def. 2477 if (MO.isRegMask() && MO.clobbersPhysReg(X86::EFLAGS)) 2478 SawKill = true; 2479 if (MO.isReg() && MO.getReg() == X86::EFLAGS) { 2480 if (MO.isDef()) return MO.isDead(); 2481 if (MO.isKill()) SawKill = true; 2482 } 2483 } 2484 2485 if (SawKill) 2486 // This instruction kills EFLAGS and doesn't redefine it, so 2487 // there's no need to look further. 2488 return true; 2489 } 2490 2491 // Conservative answer. 2492 return false; 2493 } 2494 2495 void X86InstrInfo::reMaterialize(MachineBasicBlock &MBB, 2496 MachineBasicBlock::iterator I, 2497 unsigned DestReg, unsigned SubIdx, 2498 const MachineInstr *Orig, 2499 const TargetRegisterInfo &TRI) const { 2500 bool ClobbersEFLAGS = false; 2501 for (const MachineOperand &MO : Orig->operands()) { 2502 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) { 2503 ClobbersEFLAGS = true; 2504 break; 2505 } 2506 } 2507 2508 if (ClobbersEFLAGS && !isSafeToClobberEFLAGS(MBB, I)) { 2509 // The instruction clobbers EFLAGS. Re-materialize as MOV32ri to avoid side 2510 // effects. 2511 int Value; 2512 switch (Orig->getOpcode()) { 2513 case X86::MOV32r0: Value = 0; break; 2514 case X86::MOV32r1: Value = 1; break; 2515 case X86::MOV32r_1: Value = -1; break; 2516 default: 2517 llvm_unreachable("Unexpected instruction!"); 2518 } 2519 2520 const DebugLoc &DL = Orig->getDebugLoc(); 2521 BuildMI(MBB, I, DL, get(X86::MOV32ri)).addOperand(Orig->getOperand(0)) 2522 .addImm(Value); 2523 } else { 2524 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 2525 MBB.insert(I, MI); 2526 } 2527 2528 MachineInstr *NewMI = std::prev(I); 2529 NewMI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI); 2530 } 2531 2532 /// True if MI has a condition code def, e.g. EFLAGS, that is not marked dead. 2533 bool X86InstrInfo::hasLiveCondCodeDef(MachineInstr *MI) const { 2534 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 2535 MachineOperand &MO = MI->getOperand(i); 2536 if (MO.isReg() && MO.isDef() && 2537 MO.getReg() == X86::EFLAGS && !MO.isDead()) { 2538 return true; 2539 } 2540 } 2541 return false; 2542 } 2543 2544 /// Check whether the shift count for a machine operand is non-zero. 2545 inline static unsigned getTruncatedShiftCount(MachineInstr *MI, 2546 unsigned ShiftAmtOperandIdx) { 2547 // The shift count is six bits with the REX.W prefix and five bits without. 2548 unsigned ShiftCountMask = (MI->getDesc().TSFlags & X86II::REX_W) ? 63 : 31; 2549 unsigned Imm = MI->getOperand(ShiftAmtOperandIdx).getImm(); 2550 return Imm & ShiftCountMask; 2551 } 2552 2553 /// Check whether the given shift count is appropriate 2554 /// can be represented by a LEA instruction. 2555 inline static bool isTruncatedShiftCountForLEA(unsigned ShAmt) { 2556 // Left shift instructions can be transformed into load-effective-address 2557 // instructions if we can encode them appropriately. 2558 // A LEA instruction utilizes a SIB byte to encode its scale factor. 2559 // The SIB.scale field is two bits wide which means that we can encode any 2560 // shift amount less than 4. 2561 return ShAmt < 4 && ShAmt > 0; 2562 } 2563 2564 bool X86InstrInfo::classifyLEAReg(MachineInstr *MI, const MachineOperand &Src, 2565 unsigned Opc, bool AllowSP, 2566 unsigned &NewSrc, bool &isKill, bool &isUndef, 2567 MachineOperand &ImplicitOp) const { 2568 MachineFunction &MF = *MI->getParent()->getParent(); 2569 const TargetRegisterClass *RC; 2570 if (AllowSP) { 2571 RC = Opc != X86::LEA32r ? &X86::GR64RegClass : &X86::GR32RegClass; 2572 } else { 2573 RC = Opc != X86::LEA32r ? 2574 &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass; 2575 } 2576 unsigned SrcReg = Src.getReg(); 2577 2578 // For both LEA64 and LEA32 the register already has essentially the right 2579 // type (32-bit or 64-bit) we may just need to forbid SP. 2580 if (Opc != X86::LEA64_32r) { 2581 NewSrc = SrcReg; 2582 isKill = Src.isKill(); 2583 isUndef = Src.isUndef(); 2584 2585 if (TargetRegisterInfo::isVirtualRegister(NewSrc) && 2586 !MF.getRegInfo().constrainRegClass(NewSrc, RC)) 2587 return false; 2588 2589 return true; 2590 } 2591 2592 // This is for an LEA64_32r and incoming registers are 32-bit. One way or 2593 // another we need to add 64-bit registers to the final MI. 2594 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)) { 2595 ImplicitOp = Src; 2596 ImplicitOp.setImplicit(); 2597 2598 NewSrc = getX86SubSuperRegister(Src.getReg(), 64); 2599 MachineBasicBlock::LivenessQueryResult LQR = 2600 MI->getParent()->computeRegisterLiveness(&getRegisterInfo(), NewSrc, MI); 2601 2602 switch (LQR) { 2603 case MachineBasicBlock::LQR_Unknown: 2604 // We can't give sane liveness flags to the instruction, abandon LEA 2605 // formation. 2606 return false; 2607 case MachineBasicBlock::LQR_Live: 2608 isKill = MI->killsRegister(SrcReg); 2609 isUndef = false; 2610 break; 2611 default: 2612 // The physreg itself is dead, so we have to use it as an <undef>. 2613 isKill = false; 2614 isUndef = true; 2615 break; 2616 } 2617 } else { 2618 // Virtual register of the wrong class, we have to create a temporary 64-bit 2619 // vreg to feed into the LEA. 2620 NewSrc = MF.getRegInfo().createVirtualRegister(RC); 2621 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 2622 get(TargetOpcode::COPY)) 2623 .addReg(NewSrc, RegState::Define | RegState::Undef, X86::sub_32bit) 2624 .addOperand(Src); 2625 2626 // Which is obviously going to be dead after we're done with it. 2627 isKill = true; 2628 isUndef = false; 2629 } 2630 2631 // We've set all the parameters without issue. 2632 return true; 2633 } 2634 2635 /// Helper for convertToThreeAddress when 16-bit LEA is disabled, use 32-bit 2636 /// LEA to form 3-address code by promoting to a 32-bit superregister and then 2637 /// truncating back down to a 16-bit subregister. 2638 MachineInstr * 2639 X86InstrInfo::convertToThreeAddressWithLEA(unsigned MIOpc, 2640 MachineFunction::iterator &MFI, 2641 MachineBasicBlock::iterator &MBBI, 2642 LiveVariables *LV) const { 2643 MachineInstr *MI = MBBI; 2644 unsigned Dest = MI->getOperand(0).getReg(); 2645 unsigned Src = MI->getOperand(1).getReg(); 2646 bool isDead = MI->getOperand(0).isDead(); 2647 bool isKill = MI->getOperand(1).isKill(); 2648 2649 MachineRegisterInfo &RegInfo = MFI->getParent()->getRegInfo(); 2650 unsigned leaOutReg = RegInfo.createVirtualRegister(&X86::GR32RegClass); 2651 unsigned Opc, leaInReg; 2652 if (Subtarget.is64Bit()) { 2653 Opc = X86::LEA64_32r; 2654 leaInReg = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass); 2655 } else { 2656 Opc = X86::LEA32r; 2657 leaInReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 2658 } 2659 2660 // Build and insert into an implicit UNDEF value. This is OK because 2661 // well be shifting and then extracting the lower 16-bits. 2662 // This has the potential to cause partial register stall. e.g. 2663 // movw (%rbp,%rcx,2), %dx 2664 // leal -65(%rdx), %esi 2665 // But testing has shown this *does* help performance in 64-bit mode (at 2666 // least on modern x86 machines). 2667 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(X86::IMPLICIT_DEF), leaInReg); 2668 MachineInstr *InsMI = 2669 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY)) 2670 .addReg(leaInReg, RegState::Define, X86::sub_16bit) 2671 .addReg(Src, getKillRegState(isKill)); 2672 2673 MachineInstrBuilder MIB = BuildMI(*MFI, MBBI, MI->getDebugLoc(), 2674 get(Opc), leaOutReg); 2675 switch (MIOpc) { 2676 default: llvm_unreachable("Unreachable!"); 2677 case X86::SHL16ri: { 2678 unsigned ShAmt = MI->getOperand(2).getImm(); 2679 MIB.addReg(0).addImm(1ULL << ShAmt) 2680 .addReg(leaInReg, RegState::Kill).addImm(0).addReg(0); 2681 break; 2682 } 2683 case X86::INC16r: 2684 addRegOffset(MIB, leaInReg, true, 1); 2685 break; 2686 case X86::DEC16r: 2687 addRegOffset(MIB, leaInReg, true, -1); 2688 break; 2689 case X86::ADD16ri: 2690 case X86::ADD16ri8: 2691 case X86::ADD16ri_DB: 2692 case X86::ADD16ri8_DB: 2693 addRegOffset(MIB, leaInReg, true, MI->getOperand(2).getImm()); 2694 break; 2695 case X86::ADD16rr: 2696 case X86::ADD16rr_DB: { 2697 unsigned Src2 = MI->getOperand(2).getReg(); 2698 bool isKill2 = MI->getOperand(2).isKill(); 2699 unsigned leaInReg2 = 0; 2700 MachineInstr *InsMI2 = nullptr; 2701 if (Src == Src2) { 2702 // ADD16rr %reg1028<kill>, %reg1028 2703 // just a single insert_subreg. 2704 addRegReg(MIB, leaInReg, true, leaInReg, false); 2705 } else { 2706 if (Subtarget.is64Bit()) 2707 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR64_NOSPRegClass); 2708 else 2709 leaInReg2 = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 2710 // Build and insert into an implicit UNDEF value. This is OK because 2711 // well be shifting and then extracting the lower 16-bits. 2712 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(X86::IMPLICIT_DEF),leaInReg2); 2713 InsMI2 = 2714 BuildMI(*MFI, &*MIB, MI->getDebugLoc(), get(TargetOpcode::COPY)) 2715 .addReg(leaInReg2, RegState::Define, X86::sub_16bit) 2716 .addReg(Src2, getKillRegState(isKill2)); 2717 addRegReg(MIB, leaInReg, true, leaInReg2, true); 2718 } 2719 if (LV && isKill2 && InsMI2) 2720 LV->replaceKillInstruction(Src2, MI, InsMI2); 2721 break; 2722 } 2723 } 2724 2725 MachineInstr *NewMI = MIB; 2726 MachineInstr *ExtMI = 2727 BuildMI(*MFI, MBBI, MI->getDebugLoc(), get(TargetOpcode::COPY)) 2728 .addReg(Dest, RegState::Define | getDeadRegState(isDead)) 2729 .addReg(leaOutReg, RegState::Kill, X86::sub_16bit); 2730 2731 if (LV) { 2732 // Update live variables 2733 LV->getVarInfo(leaInReg).Kills.push_back(NewMI); 2734 LV->getVarInfo(leaOutReg).Kills.push_back(ExtMI); 2735 if (isKill) 2736 LV->replaceKillInstruction(Src, MI, InsMI); 2737 if (isDead) 2738 LV->replaceKillInstruction(Dest, MI, ExtMI); 2739 } 2740 2741 return ExtMI; 2742 } 2743 2744 /// This method must be implemented by targets that 2745 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target 2746 /// may be able to convert a two-address instruction into a true 2747 /// three-address instruction on demand. This allows the X86 target (for 2748 /// example) to convert ADD and SHL instructions into LEA instructions if they 2749 /// would require register copies due to two-addressness. 2750 /// 2751 /// This method returns a null pointer if the transformation cannot be 2752 /// performed, otherwise it returns the new instruction. 2753 /// 2754 MachineInstr * 2755 X86InstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI, 2756 MachineBasicBlock::iterator &MBBI, 2757 LiveVariables *LV) const { 2758 MachineInstr *MI = MBBI; 2759 2760 // The following opcodes also sets the condition code register(s). Only 2761 // convert them to equivalent lea if the condition code register def's 2762 // are dead! 2763 if (hasLiveCondCodeDef(MI)) 2764 return nullptr; 2765 2766 MachineFunction &MF = *MI->getParent()->getParent(); 2767 // All instructions input are two-addr instructions. Get the known operands. 2768 const MachineOperand &Dest = MI->getOperand(0); 2769 const MachineOperand &Src = MI->getOperand(1); 2770 2771 MachineInstr *NewMI = nullptr; 2772 // FIXME: 16-bit LEA's are really slow on Athlons, but not bad on P4's. When 2773 // we have better subtarget support, enable the 16-bit LEA generation here. 2774 // 16-bit LEA is also slow on Core2. 2775 bool DisableLEA16 = true; 2776 bool is64Bit = Subtarget.is64Bit(); 2777 2778 unsigned MIOpc = MI->getOpcode(); 2779 switch (MIOpc) { 2780 default: return nullptr; 2781 case X86::SHL64ri: { 2782 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 2783 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 2784 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 2785 2786 // LEA can't handle RSP. 2787 if (TargetRegisterInfo::isVirtualRegister(Src.getReg()) && 2788 !MF.getRegInfo().constrainRegClass(Src.getReg(), 2789 &X86::GR64_NOSPRegClass)) 2790 return nullptr; 2791 2792 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r)) 2793 .addOperand(Dest) 2794 .addReg(0).addImm(1ULL << ShAmt).addOperand(Src).addImm(0).addReg(0); 2795 break; 2796 } 2797 case X86::SHL32ri: { 2798 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 2799 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 2800 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 2801 2802 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 2803 2804 // LEA can't handle ESP. 2805 bool isKill, isUndef; 2806 unsigned SrcReg; 2807 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 2808 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 2809 SrcReg, isKill, isUndef, ImplicitOp)) 2810 return nullptr; 2811 2812 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2813 .addOperand(Dest) 2814 .addReg(0).addImm(1ULL << ShAmt) 2815 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef)) 2816 .addImm(0).addReg(0); 2817 if (ImplicitOp.getReg() != 0) 2818 MIB.addOperand(ImplicitOp); 2819 NewMI = MIB; 2820 2821 break; 2822 } 2823 case X86::SHL16ri: { 2824 assert(MI->getNumOperands() >= 3 && "Unknown shift instruction!"); 2825 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 2826 if (!isTruncatedShiftCountForLEA(ShAmt)) return nullptr; 2827 2828 if (DisableLEA16) 2829 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) : nullptr; 2830 NewMI = BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2831 .addOperand(Dest) 2832 .addReg(0).addImm(1ULL << ShAmt).addOperand(Src).addImm(0).addReg(0); 2833 break; 2834 } 2835 case X86::INC64r: 2836 case X86::INC32r: { 2837 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!"); 2838 unsigned Opc = MIOpc == X86::INC64r ? X86::LEA64r 2839 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 2840 bool isKill, isUndef; 2841 unsigned SrcReg; 2842 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 2843 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 2844 SrcReg, isKill, isUndef, ImplicitOp)) 2845 return nullptr; 2846 2847 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2848 .addOperand(Dest) 2849 .addReg(SrcReg, getKillRegState(isKill) | getUndefRegState(isUndef)); 2850 if (ImplicitOp.getReg() != 0) 2851 MIB.addOperand(ImplicitOp); 2852 2853 NewMI = addOffset(MIB, 1); 2854 break; 2855 } 2856 case X86::INC16r: 2857 if (DisableLEA16) 2858 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) 2859 : nullptr; 2860 assert(MI->getNumOperands() >= 2 && "Unknown inc instruction!"); 2861 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2862 .addOperand(Dest).addOperand(Src), 1); 2863 break; 2864 case X86::DEC64r: 2865 case X86::DEC32r: { 2866 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!"); 2867 unsigned Opc = MIOpc == X86::DEC64r ? X86::LEA64r 2868 : (is64Bit ? X86::LEA64_32r : X86::LEA32r); 2869 2870 bool isKill, isUndef; 2871 unsigned SrcReg; 2872 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 2873 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ false, 2874 SrcReg, isKill, isUndef, ImplicitOp)) 2875 return nullptr; 2876 2877 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2878 .addOperand(Dest) 2879 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill)); 2880 if (ImplicitOp.getReg() != 0) 2881 MIB.addOperand(ImplicitOp); 2882 2883 NewMI = addOffset(MIB, -1); 2884 2885 break; 2886 } 2887 case X86::DEC16r: 2888 if (DisableLEA16) 2889 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) 2890 : nullptr; 2891 assert(MI->getNumOperands() >= 2 && "Unknown dec instruction!"); 2892 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2893 .addOperand(Dest).addOperand(Src), -1); 2894 break; 2895 case X86::ADD64rr: 2896 case X86::ADD64rr_DB: 2897 case X86::ADD32rr: 2898 case X86::ADD32rr_DB: { 2899 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2900 unsigned Opc; 2901 if (MIOpc == X86::ADD64rr || MIOpc == X86::ADD64rr_DB) 2902 Opc = X86::LEA64r; 2903 else 2904 Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 2905 2906 bool isKill, isUndef; 2907 unsigned SrcReg; 2908 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 2909 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true, 2910 SrcReg, isKill, isUndef, ImplicitOp)) 2911 return nullptr; 2912 2913 const MachineOperand &Src2 = MI->getOperand(2); 2914 bool isKill2, isUndef2; 2915 unsigned SrcReg2; 2916 MachineOperand ImplicitOp2 = MachineOperand::CreateReg(0, false); 2917 if (!classifyLEAReg(MI, Src2, Opc, /*AllowSP=*/ false, 2918 SrcReg2, isKill2, isUndef2, ImplicitOp2)) 2919 return nullptr; 2920 2921 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2922 .addOperand(Dest); 2923 if (ImplicitOp.getReg() != 0) 2924 MIB.addOperand(ImplicitOp); 2925 if (ImplicitOp2.getReg() != 0) 2926 MIB.addOperand(ImplicitOp2); 2927 2928 NewMI = addRegReg(MIB, SrcReg, isKill, SrcReg2, isKill2); 2929 2930 // Preserve undefness of the operands. 2931 NewMI->getOperand(1).setIsUndef(isUndef); 2932 NewMI->getOperand(3).setIsUndef(isUndef2); 2933 2934 if (LV && Src2.isKill()) 2935 LV->replaceKillInstruction(SrcReg2, MI, NewMI); 2936 break; 2937 } 2938 case X86::ADD16rr: 2939 case X86::ADD16rr_DB: { 2940 if (DisableLEA16) 2941 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) 2942 : nullptr; 2943 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2944 unsigned Src2 = MI->getOperand(2).getReg(); 2945 bool isKill2 = MI->getOperand(2).isKill(); 2946 NewMI = addRegReg(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 2947 .addOperand(Dest), 2948 Src.getReg(), Src.isKill(), Src2, isKill2); 2949 2950 // Preserve undefness of the operands. 2951 bool isUndef = MI->getOperand(1).isUndef(); 2952 bool isUndef2 = MI->getOperand(2).isUndef(); 2953 NewMI->getOperand(1).setIsUndef(isUndef); 2954 NewMI->getOperand(3).setIsUndef(isUndef2); 2955 2956 if (LV && isKill2) 2957 LV->replaceKillInstruction(Src2, MI, NewMI); 2958 break; 2959 } 2960 case X86::ADD64ri32: 2961 case X86::ADD64ri8: 2962 case X86::ADD64ri32_DB: 2963 case X86::ADD64ri8_DB: 2964 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2965 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA64r)) 2966 .addOperand(Dest).addOperand(Src), 2967 MI->getOperand(2).getImm()); 2968 break; 2969 case X86::ADD32ri: 2970 case X86::ADD32ri8: 2971 case X86::ADD32ri_DB: 2972 case X86::ADD32ri8_DB: { 2973 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 2974 unsigned Opc = is64Bit ? X86::LEA64_32r : X86::LEA32r; 2975 2976 bool isKill, isUndef; 2977 unsigned SrcReg; 2978 MachineOperand ImplicitOp = MachineOperand::CreateReg(0, false); 2979 if (!classifyLEAReg(MI, Src, Opc, /*AllowSP=*/ true, 2980 SrcReg, isKill, isUndef, ImplicitOp)) 2981 return nullptr; 2982 2983 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), get(Opc)) 2984 .addOperand(Dest) 2985 .addReg(SrcReg, getUndefRegState(isUndef) | getKillRegState(isKill)); 2986 if (ImplicitOp.getReg() != 0) 2987 MIB.addOperand(ImplicitOp); 2988 2989 NewMI = addOffset(MIB, MI->getOperand(2).getImm()); 2990 break; 2991 } 2992 case X86::ADD16ri: 2993 case X86::ADD16ri8: 2994 case X86::ADD16ri_DB: 2995 case X86::ADD16ri8_DB: 2996 if (DisableLEA16) 2997 return is64Bit ? convertToThreeAddressWithLEA(MIOpc, MFI, MBBI, LV) 2998 : nullptr; 2999 assert(MI->getNumOperands() >= 3 && "Unknown add instruction!"); 3000 NewMI = addOffset(BuildMI(MF, MI->getDebugLoc(), get(X86::LEA16r)) 3001 .addOperand(Dest).addOperand(Src), 3002 MI->getOperand(2).getImm()); 3003 break; 3004 } 3005 3006 if (!NewMI) return nullptr; 3007 3008 if (LV) { // Update live variables 3009 if (Src.isKill()) 3010 LV->replaceKillInstruction(Src.getReg(), MI, NewMI); 3011 if (Dest.isDead()) 3012 LV->replaceKillInstruction(Dest.getReg(), MI, NewMI); 3013 } 3014 3015 MFI->insert(MBBI, NewMI); // Insert the new inst 3016 return NewMI; 3017 } 3018 3019 /// Returns true if the given instruction opcode is FMA3. 3020 /// Otherwise, returns false. 3021 /// The second parameter is optional and is used as the second return from 3022 /// the function. It is set to true if the given instruction has FMA3 opcode 3023 /// that is used for lowering of scalar FMA intrinsics, and it is set to false 3024 /// otherwise. 3025 static bool isFMA3(unsigned Opcode, bool *IsIntrinsic = nullptr) { 3026 if (IsIntrinsic) 3027 *IsIntrinsic = false; 3028 3029 switch (Opcode) { 3030 case X86::VFMADDSDr132r: case X86::VFMADDSDr132m: 3031 case X86::VFMADDSSr132r: case X86::VFMADDSSr132m: 3032 case X86::VFMSUBSDr132r: case X86::VFMSUBSDr132m: 3033 case X86::VFMSUBSSr132r: case X86::VFMSUBSSr132m: 3034 case X86::VFNMADDSDr132r: case X86::VFNMADDSDr132m: 3035 case X86::VFNMADDSSr132r: case X86::VFNMADDSSr132m: 3036 case X86::VFNMSUBSDr132r: case X86::VFNMSUBSDr132m: 3037 case X86::VFNMSUBSSr132r: case X86::VFNMSUBSSr132m: 3038 3039 case X86::VFMADDSDr213r: case X86::VFMADDSDr213m: 3040 case X86::VFMADDSSr213r: case X86::VFMADDSSr213m: 3041 case X86::VFMSUBSDr213r: case X86::VFMSUBSDr213m: 3042 case X86::VFMSUBSSr213r: case X86::VFMSUBSSr213m: 3043 case X86::VFNMADDSDr213r: case X86::VFNMADDSDr213m: 3044 case X86::VFNMADDSSr213r: case X86::VFNMADDSSr213m: 3045 case X86::VFNMSUBSDr213r: case X86::VFNMSUBSDr213m: 3046 case X86::VFNMSUBSSr213r: case X86::VFNMSUBSSr213m: 3047 3048 case X86::VFMADDSDr231r: case X86::VFMADDSDr231m: 3049 case X86::VFMADDSSr231r: case X86::VFMADDSSr231m: 3050 case X86::VFMSUBSDr231r: case X86::VFMSUBSDr231m: 3051 case X86::VFMSUBSSr231r: case X86::VFMSUBSSr231m: 3052 case X86::VFNMADDSDr231r: case X86::VFNMADDSDr231m: 3053 case X86::VFNMADDSSr231r: case X86::VFNMADDSSr231m: 3054 case X86::VFNMSUBSDr231r: case X86::VFNMSUBSDr231m: 3055 case X86::VFNMSUBSSr231r: case X86::VFNMSUBSSr231m: 3056 3057 case X86::VFMADDSUBPDr132r: case X86::VFMADDSUBPDr132m: 3058 case X86::VFMADDSUBPSr132r: case X86::VFMADDSUBPSr132m: 3059 case X86::VFMSUBADDPDr132r: case X86::VFMSUBADDPDr132m: 3060 case X86::VFMSUBADDPSr132r: case X86::VFMSUBADDPSr132m: 3061 case X86::VFMADDSUBPDr132rY: case X86::VFMADDSUBPDr132mY: 3062 case X86::VFMADDSUBPSr132rY: case X86::VFMADDSUBPSr132mY: 3063 case X86::VFMSUBADDPDr132rY: case X86::VFMSUBADDPDr132mY: 3064 case X86::VFMSUBADDPSr132rY: case X86::VFMSUBADDPSr132mY: 3065 3066 case X86::VFMADDPDr132r: case X86::VFMADDPDr132m: 3067 case X86::VFMADDPSr132r: case X86::VFMADDPSr132m: 3068 case X86::VFMSUBPDr132r: case X86::VFMSUBPDr132m: 3069 case X86::VFMSUBPSr132r: case X86::VFMSUBPSr132m: 3070 case X86::VFNMADDPDr132r: case X86::VFNMADDPDr132m: 3071 case X86::VFNMADDPSr132r: case X86::VFNMADDPSr132m: 3072 case X86::VFNMSUBPDr132r: case X86::VFNMSUBPDr132m: 3073 case X86::VFNMSUBPSr132r: case X86::VFNMSUBPSr132m: 3074 case X86::VFMADDPDr132rY: case X86::VFMADDPDr132mY: 3075 case X86::VFMADDPSr132rY: case X86::VFMADDPSr132mY: 3076 case X86::VFMSUBPDr132rY: case X86::VFMSUBPDr132mY: 3077 case X86::VFMSUBPSr132rY: case X86::VFMSUBPSr132mY: 3078 case X86::VFNMADDPDr132rY: case X86::VFNMADDPDr132mY: 3079 case X86::VFNMADDPSr132rY: case X86::VFNMADDPSr132mY: 3080 case X86::VFNMSUBPDr132rY: case X86::VFNMSUBPDr132mY: 3081 case X86::VFNMSUBPSr132rY: case X86::VFNMSUBPSr132mY: 3082 3083 case X86::VFMADDSUBPDr213r: case X86::VFMADDSUBPDr213m: 3084 case X86::VFMADDSUBPSr213r: case X86::VFMADDSUBPSr213m: 3085 case X86::VFMSUBADDPDr213r: case X86::VFMSUBADDPDr213m: 3086 case X86::VFMSUBADDPSr213r: case X86::VFMSUBADDPSr213m: 3087 case X86::VFMADDSUBPDr213rY: case X86::VFMADDSUBPDr213mY: 3088 case X86::VFMADDSUBPSr213rY: case X86::VFMADDSUBPSr213mY: 3089 case X86::VFMSUBADDPDr213rY: case X86::VFMSUBADDPDr213mY: 3090 case X86::VFMSUBADDPSr213rY: case X86::VFMSUBADDPSr213mY: 3091 3092 case X86::VFMADDPDr213r: case X86::VFMADDPDr213m: 3093 case X86::VFMADDPSr213r: case X86::VFMADDPSr213m: 3094 case X86::VFMSUBPDr213r: case X86::VFMSUBPDr213m: 3095 case X86::VFMSUBPSr213r: case X86::VFMSUBPSr213m: 3096 case X86::VFNMADDPDr213r: case X86::VFNMADDPDr213m: 3097 case X86::VFNMADDPSr213r: case X86::VFNMADDPSr213m: 3098 case X86::VFNMSUBPDr213r: case X86::VFNMSUBPDr213m: 3099 case X86::VFNMSUBPSr213r: case X86::VFNMSUBPSr213m: 3100 case X86::VFMADDPDr213rY: case X86::VFMADDPDr213mY: 3101 case X86::VFMADDPSr213rY: case X86::VFMADDPSr213mY: 3102 case X86::VFMSUBPDr213rY: case X86::VFMSUBPDr213mY: 3103 case X86::VFMSUBPSr213rY: case X86::VFMSUBPSr213mY: 3104 case X86::VFNMADDPDr213rY: case X86::VFNMADDPDr213mY: 3105 case X86::VFNMADDPSr213rY: case X86::VFNMADDPSr213mY: 3106 case X86::VFNMSUBPDr213rY: case X86::VFNMSUBPDr213mY: 3107 case X86::VFNMSUBPSr213rY: case X86::VFNMSUBPSr213mY: 3108 3109 case X86::VFMADDSUBPDr231r: case X86::VFMADDSUBPDr231m: 3110 case X86::VFMADDSUBPSr231r: case X86::VFMADDSUBPSr231m: 3111 case X86::VFMSUBADDPDr231r: case X86::VFMSUBADDPDr231m: 3112 case X86::VFMSUBADDPSr231r: case X86::VFMSUBADDPSr231m: 3113 case X86::VFMADDSUBPDr231rY: case X86::VFMADDSUBPDr231mY: 3114 case X86::VFMADDSUBPSr231rY: case X86::VFMADDSUBPSr231mY: 3115 case X86::VFMSUBADDPDr231rY: case X86::VFMSUBADDPDr231mY: 3116 case X86::VFMSUBADDPSr231rY: case X86::VFMSUBADDPSr231mY: 3117 3118 case X86::VFMADDPDr231r: case X86::VFMADDPDr231m: 3119 case X86::VFMADDPSr231r: case X86::VFMADDPSr231m: 3120 case X86::VFMSUBPDr231r: case X86::VFMSUBPDr231m: 3121 case X86::VFMSUBPSr231r: case X86::VFMSUBPSr231m: 3122 case X86::VFNMADDPDr231r: case X86::VFNMADDPDr231m: 3123 case X86::VFNMADDPSr231r: case X86::VFNMADDPSr231m: 3124 case X86::VFNMSUBPDr231r: case X86::VFNMSUBPDr231m: 3125 case X86::VFNMSUBPSr231r: case X86::VFNMSUBPSr231m: 3126 case X86::VFMADDPDr231rY: case X86::VFMADDPDr231mY: 3127 case X86::VFMADDPSr231rY: case X86::VFMADDPSr231mY: 3128 case X86::VFMSUBPDr231rY: case X86::VFMSUBPDr231mY: 3129 case X86::VFMSUBPSr231rY: case X86::VFMSUBPSr231mY: 3130 case X86::VFNMADDPDr231rY: case X86::VFNMADDPDr231mY: 3131 case X86::VFNMADDPSr231rY: case X86::VFNMADDPSr231mY: 3132 case X86::VFNMSUBPDr231rY: case X86::VFNMSUBPDr231mY: 3133 case X86::VFNMSUBPSr231rY: case X86::VFNMSUBPSr231mY: 3134 return true; 3135 3136 case X86::VFMADDSDr132r_Int: case X86::VFMADDSDr132m_Int: 3137 case X86::VFMADDSSr132r_Int: case X86::VFMADDSSr132m_Int: 3138 case X86::VFMSUBSDr132r_Int: case X86::VFMSUBSDr132m_Int: 3139 case X86::VFMSUBSSr132r_Int: case X86::VFMSUBSSr132m_Int: 3140 case X86::VFNMADDSDr132r_Int: case X86::VFNMADDSDr132m_Int: 3141 case X86::VFNMADDSSr132r_Int: case X86::VFNMADDSSr132m_Int: 3142 case X86::VFNMSUBSDr132r_Int: case X86::VFNMSUBSDr132m_Int: 3143 case X86::VFNMSUBSSr132r_Int: case X86::VFNMSUBSSr132m_Int: 3144 3145 case X86::VFMADDSDr213r_Int: case X86::VFMADDSDr213m_Int: 3146 case X86::VFMADDSSr213r_Int: case X86::VFMADDSSr213m_Int: 3147 case X86::VFMSUBSDr213r_Int: case X86::VFMSUBSDr213m_Int: 3148 case X86::VFMSUBSSr213r_Int: case X86::VFMSUBSSr213m_Int: 3149 case X86::VFNMADDSDr213r_Int: case X86::VFNMADDSDr213m_Int: 3150 case X86::VFNMADDSSr213r_Int: case X86::VFNMADDSSr213m_Int: 3151 case X86::VFNMSUBSDr213r_Int: case X86::VFNMSUBSDr213m_Int: 3152 case X86::VFNMSUBSSr213r_Int: case X86::VFNMSUBSSr213m_Int: 3153 3154 case X86::VFMADDSDr231r_Int: case X86::VFMADDSDr231m_Int: 3155 case X86::VFMADDSSr231r_Int: case X86::VFMADDSSr231m_Int: 3156 case X86::VFMSUBSDr231r_Int: case X86::VFMSUBSDr231m_Int: 3157 case X86::VFMSUBSSr231r_Int: case X86::VFMSUBSSr231m_Int: 3158 case X86::VFNMADDSDr231r_Int: case X86::VFNMADDSDr231m_Int: 3159 case X86::VFNMADDSSr231r_Int: case X86::VFNMADDSSr231m_Int: 3160 case X86::VFNMSUBSDr231r_Int: case X86::VFNMSUBSDr231m_Int: 3161 case X86::VFNMSUBSSr231r_Int: case X86::VFNMSUBSSr231m_Int: 3162 if (IsIntrinsic) 3163 *IsIntrinsic = true; 3164 return true; 3165 default: 3166 return false; 3167 } 3168 llvm_unreachable("Opcode not handled by the switch"); 3169 } 3170 3171 MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr *MI, 3172 bool NewMI, 3173 unsigned OpIdx1, 3174 unsigned OpIdx2) const { 3175 switch (MI->getOpcode()) { 3176 case X86::SHRD16rri8: // A = SHRD16rri8 B, C, I -> A = SHLD16rri8 C, B, (16-I) 3177 case X86::SHLD16rri8: // A = SHLD16rri8 B, C, I -> A = SHRD16rri8 C, B, (16-I) 3178 case X86::SHRD32rri8: // A = SHRD32rri8 B, C, I -> A = SHLD32rri8 C, B, (32-I) 3179 case X86::SHLD32rri8: // A = SHLD32rri8 B, C, I -> A = SHRD32rri8 C, B, (32-I) 3180 case X86::SHRD64rri8: // A = SHRD64rri8 B, C, I -> A = SHLD64rri8 C, B, (64-I) 3181 case X86::SHLD64rri8:{// A = SHLD64rri8 B, C, I -> A = SHRD64rri8 C, B, (64-I) 3182 unsigned Opc; 3183 unsigned Size; 3184 switch (MI->getOpcode()) { 3185 default: llvm_unreachable("Unreachable!"); 3186 case X86::SHRD16rri8: Size = 16; Opc = X86::SHLD16rri8; break; 3187 case X86::SHLD16rri8: Size = 16; Opc = X86::SHRD16rri8; break; 3188 case X86::SHRD32rri8: Size = 32; Opc = X86::SHLD32rri8; break; 3189 case X86::SHLD32rri8: Size = 32; Opc = X86::SHRD32rri8; break; 3190 case X86::SHRD64rri8: Size = 64; Opc = X86::SHLD64rri8; break; 3191 case X86::SHLD64rri8: Size = 64; Opc = X86::SHRD64rri8; break; 3192 } 3193 unsigned Amt = MI->getOperand(3).getImm(); 3194 if (NewMI) { 3195 MachineFunction &MF = *MI->getParent()->getParent(); 3196 MI = MF.CloneMachineInstr(MI); 3197 NewMI = false; 3198 } 3199 MI->setDesc(get(Opc)); 3200 MI->getOperand(3).setImm(Size-Amt); 3201 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3202 } 3203 case X86::BLENDPDrri: 3204 case X86::BLENDPSrri: 3205 case X86::PBLENDWrri: 3206 case X86::VBLENDPDrri: 3207 case X86::VBLENDPSrri: 3208 case X86::VBLENDPDYrri: 3209 case X86::VBLENDPSYrri: 3210 case X86::VPBLENDDrri: 3211 case X86::VPBLENDWrri: 3212 case X86::VPBLENDDYrri: 3213 case X86::VPBLENDWYrri:{ 3214 unsigned Mask; 3215 switch (MI->getOpcode()) { 3216 default: llvm_unreachable("Unreachable!"); 3217 case X86::BLENDPDrri: Mask = 0x03; break; 3218 case X86::BLENDPSrri: Mask = 0x0F; break; 3219 case X86::PBLENDWrri: Mask = 0xFF; break; 3220 case X86::VBLENDPDrri: Mask = 0x03; break; 3221 case X86::VBLENDPSrri: Mask = 0x0F; break; 3222 case X86::VBLENDPDYrri: Mask = 0x0F; break; 3223 case X86::VBLENDPSYrri: Mask = 0xFF; break; 3224 case X86::VPBLENDDrri: Mask = 0x0F; break; 3225 case X86::VPBLENDWrri: Mask = 0xFF; break; 3226 case X86::VPBLENDDYrri: Mask = 0xFF; break; 3227 case X86::VPBLENDWYrri: Mask = 0xFF; break; 3228 } 3229 // Only the least significant bits of Imm are used. 3230 unsigned Imm = MI->getOperand(3).getImm() & Mask; 3231 if (NewMI) { 3232 MachineFunction &MF = *MI->getParent()->getParent(); 3233 MI = MF.CloneMachineInstr(MI); 3234 NewMI = false; 3235 } 3236 MI->getOperand(3).setImm(Mask ^ Imm); 3237 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3238 } 3239 case X86::PCLMULQDQrr: 3240 case X86::VPCLMULQDQrr:{ 3241 // SRC1 64bits = Imm[0] ? SRC1[127:64] : SRC1[63:0] 3242 // SRC2 64bits = Imm[4] ? SRC2[127:64] : SRC2[63:0] 3243 unsigned Imm = MI->getOperand(3).getImm(); 3244 unsigned Src1Hi = Imm & 0x01; 3245 unsigned Src2Hi = Imm & 0x10; 3246 if (NewMI) { 3247 MachineFunction &MF = *MI->getParent()->getParent(); 3248 MI = MF.CloneMachineInstr(MI); 3249 NewMI = false; 3250 } 3251 MI->getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4)); 3252 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3253 } 3254 case X86::CMPPDrri: 3255 case X86::CMPPSrri: 3256 case X86::VCMPPDrri: 3257 case X86::VCMPPSrri: 3258 case X86::VCMPPDYrri: 3259 case X86::VCMPPSYrri: { 3260 // Float comparison can be safely commuted for 3261 // Ordered/Unordered/Equal/NotEqual tests 3262 unsigned Imm = MI->getOperand(3).getImm() & 0x7; 3263 switch (Imm) { 3264 case 0x00: // EQUAL 3265 case 0x03: // UNORDERED 3266 case 0x04: // NOT EQUAL 3267 case 0x07: // ORDERED 3268 if (NewMI) { 3269 MachineFunction &MF = *MI->getParent()->getParent(); 3270 MI = MF.CloneMachineInstr(MI); 3271 NewMI = false; 3272 } 3273 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3274 default: 3275 return nullptr; 3276 } 3277 } 3278 case X86::VPCOMBri: case X86::VPCOMUBri: 3279 case X86::VPCOMDri: case X86::VPCOMUDri: 3280 case X86::VPCOMQri: case X86::VPCOMUQri: 3281 case X86::VPCOMWri: case X86::VPCOMUWri: { 3282 // Flip comparison mode immediate (if necessary). 3283 unsigned Imm = MI->getOperand(3).getImm() & 0x7; 3284 switch (Imm) { 3285 case 0x00: Imm = 0x02; break; // LT -> GT 3286 case 0x01: Imm = 0x03; break; // LE -> GE 3287 case 0x02: Imm = 0x00; break; // GT -> LT 3288 case 0x03: Imm = 0x01; break; // GE -> LE 3289 case 0x04: // EQ 3290 case 0x05: // NE 3291 case 0x06: // FALSE 3292 case 0x07: // TRUE 3293 default: 3294 break; 3295 } 3296 if (NewMI) { 3297 MachineFunction &MF = *MI->getParent()->getParent(); 3298 MI = MF.CloneMachineInstr(MI); 3299 NewMI = false; 3300 } 3301 MI->getOperand(3).setImm(Imm); 3302 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3303 } 3304 case X86::VPERM2F128rr: 3305 case X86::VPERM2I128rr: { 3306 // Flip permute source immediate. 3307 // Imm & 0x02: lo = if set, select Op1.lo/hi else Op0.lo/hi. 3308 // Imm & 0x20: hi = if set, select Op1.lo/hi else Op0.lo/hi. 3309 unsigned Imm = MI->getOperand(3).getImm() & 0xFF; 3310 if (NewMI) { 3311 MachineFunction &MF = *MI->getParent()->getParent(); 3312 MI = MF.CloneMachineInstr(MI); 3313 NewMI = false; 3314 } 3315 MI->getOperand(3).setImm(Imm ^ 0x22); 3316 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3317 } 3318 case X86::CMOVB16rr: case X86::CMOVB32rr: case X86::CMOVB64rr: 3319 case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr: 3320 case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr: 3321 case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr: 3322 case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr: 3323 case X86::CMOVA16rr: case X86::CMOVA32rr: case X86::CMOVA64rr: 3324 case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr: 3325 case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr: 3326 case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr: 3327 case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr: 3328 case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr: 3329 case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr: 3330 case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr: 3331 case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr: 3332 case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr: 3333 case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr: { 3334 unsigned Opc; 3335 switch (MI->getOpcode()) { 3336 default: llvm_unreachable("Unreachable!"); 3337 case X86::CMOVB16rr: Opc = X86::CMOVAE16rr; break; 3338 case X86::CMOVB32rr: Opc = X86::CMOVAE32rr; break; 3339 case X86::CMOVB64rr: Opc = X86::CMOVAE64rr; break; 3340 case X86::CMOVAE16rr: Opc = X86::CMOVB16rr; break; 3341 case X86::CMOVAE32rr: Opc = X86::CMOVB32rr; break; 3342 case X86::CMOVAE64rr: Opc = X86::CMOVB64rr; break; 3343 case X86::CMOVE16rr: Opc = X86::CMOVNE16rr; break; 3344 case X86::CMOVE32rr: Opc = X86::CMOVNE32rr; break; 3345 case X86::CMOVE64rr: Opc = X86::CMOVNE64rr; break; 3346 case X86::CMOVNE16rr: Opc = X86::CMOVE16rr; break; 3347 case X86::CMOVNE32rr: Opc = X86::CMOVE32rr; break; 3348 case X86::CMOVNE64rr: Opc = X86::CMOVE64rr; break; 3349 case X86::CMOVBE16rr: Opc = X86::CMOVA16rr; break; 3350 case X86::CMOVBE32rr: Opc = X86::CMOVA32rr; break; 3351 case X86::CMOVBE64rr: Opc = X86::CMOVA64rr; break; 3352 case X86::CMOVA16rr: Opc = X86::CMOVBE16rr; break; 3353 case X86::CMOVA32rr: Opc = X86::CMOVBE32rr; break; 3354 case X86::CMOVA64rr: Opc = X86::CMOVBE64rr; break; 3355 case X86::CMOVL16rr: Opc = X86::CMOVGE16rr; break; 3356 case X86::CMOVL32rr: Opc = X86::CMOVGE32rr; break; 3357 case X86::CMOVL64rr: Opc = X86::CMOVGE64rr; break; 3358 case X86::CMOVGE16rr: Opc = X86::CMOVL16rr; break; 3359 case X86::CMOVGE32rr: Opc = X86::CMOVL32rr; break; 3360 case X86::CMOVGE64rr: Opc = X86::CMOVL64rr; break; 3361 case X86::CMOVLE16rr: Opc = X86::CMOVG16rr; break; 3362 case X86::CMOVLE32rr: Opc = X86::CMOVG32rr; break; 3363 case X86::CMOVLE64rr: Opc = X86::CMOVG64rr; break; 3364 case X86::CMOVG16rr: Opc = X86::CMOVLE16rr; break; 3365 case X86::CMOVG32rr: Opc = X86::CMOVLE32rr; break; 3366 case X86::CMOVG64rr: Opc = X86::CMOVLE64rr; break; 3367 case X86::CMOVS16rr: Opc = X86::CMOVNS16rr; break; 3368 case X86::CMOVS32rr: Opc = X86::CMOVNS32rr; break; 3369 case X86::CMOVS64rr: Opc = X86::CMOVNS64rr; break; 3370 case X86::CMOVNS16rr: Opc = X86::CMOVS16rr; break; 3371 case X86::CMOVNS32rr: Opc = X86::CMOVS32rr; break; 3372 case X86::CMOVNS64rr: Opc = X86::CMOVS64rr; break; 3373 case X86::CMOVP16rr: Opc = X86::CMOVNP16rr; break; 3374 case X86::CMOVP32rr: Opc = X86::CMOVNP32rr; break; 3375 case X86::CMOVP64rr: Opc = X86::CMOVNP64rr; break; 3376 case X86::CMOVNP16rr: Opc = X86::CMOVP16rr; break; 3377 case X86::CMOVNP32rr: Opc = X86::CMOVP32rr; break; 3378 case X86::CMOVNP64rr: Opc = X86::CMOVP64rr; break; 3379 case X86::CMOVO16rr: Opc = X86::CMOVNO16rr; break; 3380 case X86::CMOVO32rr: Opc = X86::CMOVNO32rr; break; 3381 case X86::CMOVO64rr: Opc = X86::CMOVNO64rr; break; 3382 case X86::CMOVNO16rr: Opc = X86::CMOVO16rr; break; 3383 case X86::CMOVNO32rr: Opc = X86::CMOVO32rr; break; 3384 case X86::CMOVNO64rr: Opc = X86::CMOVO64rr; break; 3385 } 3386 if (NewMI) { 3387 MachineFunction &MF = *MI->getParent()->getParent(); 3388 MI = MF.CloneMachineInstr(MI); 3389 NewMI = false; 3390 } 3391 MI->setDesc(get(Opc)); 3392 // Fallthrough intended. 3393 } 3394 default: 3395 if (isFMA3(MI->getOpcode())) { 3396 unsigned Opc = getFMA3OpcodeToCommuteOperands(MI, OpIdx1, OpIdx2); 3397 if (Opc == 0) 3398 return nullptr; 3399 if (NewMI) { 3400 MachineFunction &MF = *MI->getParent()->getParent(); 3401 MI = MF.CloneMachineInstr(MI); 3402 NewMI = false; 3403 } 3404 MI->setDesc(get(Opc)); 3405 } 3406 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 3407 } 3408 } 3409 3410 bool X86InstrInfo::findFMA3CommutedOpIndices(MachineInstr *MI, 3411 unsigned &SrcOpIdx1, 3412 unsigned &SrcOpIdx2) const { 3413 3414 unsigned RegOpsNum = isMem(MI, 3) ? 2 : 3; 3415 3416 // Only the first RegOpsNum operands are commutable. 3417 // Also, the value 'CommuteAnyOperandIndex' is valid here as it means 3418 // that the operand is not specified/fixed. 3419 if (SrcOpIdx1 != CommuteAnyOperandIndex && 3420 (SrcOpIdx1 < 1 || SrcOpIdx1 > RegOpsNum)) 3421 return false; 3422 if (SrcOpIdx2 != CommuteAnyOperandIndex && 3423 (SrcOpIdx2 < 1 || SrcOpIdx2 > RegOpsNum)) 3424 return false; 3425 3426 // Look for two different register operands assumed to be commutable 3427 // regardless of the FMA opcode. The FMA opcode is adjusted later. 3428 if (SrcOpIdx1 == CommuteAnyOperandIndex || 3429 SrcOpIdx2 == CommuteAnyOperandIndex) { 3430 unsigned CommutableOpIdx1 = SrcOpIdx1; 3431 unsigned CommutableOpIdx2 = SrcOpIdx2; 3432 3433 // At least one of operands to be commuted is not specified and 3434 // this method is free to choose appropriate commutable operands. 3435 if (SrcOpIdx1 == SrcOpIdx2) 3436 // Both of operands are not fixed. By default set one of commutable 3437 // operands to the last register operand of the instruction. 3438 CommutableOpIdx2 = RegOpsNum; 3439 else if (SrcOpIdx2 == CommuteAnyOperandIndex) 3440 // Only one of operands is not fixed. 3441 CommutableOpIdx2 = SrcOpIdx1; 3442 3443 // CommutableOpIdx2 is well defined now. Let's choose another commutable 3444 // operand and assign its index to CommutableOpIdx1. 3445 unsigned Op2Reg = MI->getOperand(CommutableOpIdx2).getReg(); 3446 for (CommutableOpIdx1 = RegOpsNum; CommutableOpIdx1 > 0; CommutableOpIdx1--) { 3447 // The commuted operands must have different registers. 3448 // Otherwise, the commute transformation does not change anything and 3449 // is useless then. 3450 if (Op2Reg != MI->getOperand(CommutableOpIdx1).getReg()) 3451 break; 3452 } 3453 3454 // No appropriate commutable operands were found. 3455 if (CommutableOpIdx1 == 0) 3456 return false; 3457 3458 // Assign the found pair of commutable indices to SrcOpIdx1 and SrcOpidx2 3459 // to return those values. 3460 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 3461 CommutableOpIdx1, CommutableOpIdx2)) 3462 return false; 3463 } 3464 3465 // Check if we can adjust the opcode to preserve the semantics when 3466 // commute the register operands. 3467 return getFMA3OpcodeToCommuteOperands(MI, SrcOpIdx1, SrcOpIdx2) != 0; 3468 } 3469 3470 unsigned X86InstrInfo::getFMA3OpcodeToCommuteOperands(MachineInstr *MI, 3471 unsigned SrcOpIdx1, 3472 unsigned SrcOpIdx2) const { 3473 unsigned Opc = MI->getOpcode(); 3474 3475 // Define the array that holds FMA opcodes in groups 3476 // of 3 opcodes(132, 213, 231) in each group. 3477 static const uint16_t RegularOpcodeGroups[][3] = { 3478 { X86::VFMADDSSr132r, X86::VFMADDSSr213r, X86::VFMADDSSr231r }, 3479 { X86::VFMADDSDr132r, X86::VFMADDSDr213r, X86::VFMADDSDr231r }, 3480 { X86::VFMADDPSr132r, X86::VFMADDPSr213r, X86::VFMADDPSr231r }, 3481 { X86::VFMADDPDr132r, X86::VFMADDPDr213r, X86::VFMADDPDr231r }, 3482 { X86::VFMADDPSr132rY, X86::VFMADDPSr213rY, X86::VFMADDPSr231rY }, 3483 { X86::VFMADDPDr132rY, X86::VFMADDPDr213rY, X86::VFMADDPDr231rY }, 3484 { X86::VFMADDSSr132m, X86::VFMADDSSr213m, X86::VFMADDSSr231m }, 3485 { X86::VFMADDSDr132m, X86::VFMADDSDr213m, X86::VFMADDSDr231m }, 3486 { X86::VFMADDPSr132m, X86::VFMADDPSr213m, X86::VFMADDPSr231m }, 3487 { X86::VFMADDPDr132m, X86::VFMADDPDr213m, X86::VFMADDPDr231m }, 3488 { X86::VFMADDPSr132mY, X86::VFMADDPSr213mY, X86::VFMADDPSr231mY }, 3489 { X86::VFMADDPDr132mY, X86::VFMADDPDr213mY, X86::VFMADDPDr231mY }, 3490 3491 { X86::VFMSUBSSr132r, X86::VFMSUBSSr213r, X86::VFMSUBSSr231r }, 3492 { X86::VFMSUBSDr132r, X86::VFMSUBSDr213r, X86::VFMSUBSDr231r }, 3493 { X86::VFMSUBPSr132r, X86::VFMSUBPSr213r, X86::VFMSUBPSr231r }, 3494 { X86::VFMSUBPDr132r, X86::VFMSUBPDr213r, X86::VFMSUBPDr231r }, 3495 { X86::VFMSUBPSr132rY, X86::VFMSUBPSr213rY, X86::VFMSUBPSr231rY }, 3496 { X86::VFMSUBPDr132rY, X86::VFMSUBPDr213rY, X86::VFMSUBPDr231rY }, 3497 { X86::VFMSUBSSr132m, X86::VFMSUBSSr213m, X86::VFMSUBSSr231m }, 3498 { X86::VFMSUBSDr132m, X86::VFMSUBSDr213m, X86::VFMSUBSDr231m }, 3499 { X86::VFMSUBPSr132m, X86::VFMSUBPSr213m, X86::VFMSUBPSr231m }, 3500 { X86::VFMSUBPDr132m, X86::VFMSUBPDr213m, X86::VFMSUBPDr231m }, 3501 { X86::VFMSUBPSr132mY, X86::VFMSUBPSr213mY, X86::VFMSUBPSr231mY }, 3502 { X86::VFMSUBPDr132mY, X86::VFMSUBPDr213mY, X86::VFMSUBPDr231mY }, 3503 3504 { X86::VFNMADDSSr132r, X86::VFNMADDSSr213r, X86::VFNMADDSSr231r }, 3505 { X86::VFNMADDSDr132r, X86::VFNMADDSDr213r, X86::VFNMADDSDr231r }, 3506 { X86::VFNMADDPSr132r, X86::VFNMADDPSr213r, X86::VFNMADDPSr231r }, 3507 { X86::VFNMADDPDr132r, X86::VFNMADDPDr213r, X86::VFNMADDPDr231r }, 3508 { X86::VFNMADDPSr132rY, X86::VFNMADDPSr213rY, X86::VFNMADDPSr231rY }, 3509 { X86::VFNMADDPDr132rY, X86::VFNMADDPDr213rY, X86::VFNMADDPDr231rY }, 3510 { X86::VFNMADDSSr132m, X86::VFNMADDSSr213m, X86::VFNMADDSSr231m }, 3511 { X86::VFNMADDSDr132m, X86::VFNMADDSDr213m, X86::VFNMADDSDr231m }, 3512 { X86::VFNMADDPSr132m, X86::VFNMADDPSr213m, X86::VFNMADDPSr231m }, 3513 { X86::VFNMADDPDr132m, X86::VFNMADDPDr213m, X86::VFNMADDPDr231m }, 3514 { X86::VFNMADDPSr132mY, X86::VFNMADDPSr213mY, X86::VFNMADDPSr231mY }, 3515 { X86::VFNMADDPDr132mY, X86::VFNMADDPDr213mY, X86::VFNMADDPDr231mY }, 3516 3517 { X86::VFNMSUBSSr132r, X86::VFNMSUBSSr213r, X86::VFNMSUBSSr231r }, 3518 { X86::VFNMSUBSDr132r, X86::VFNMSUBSDr213r, X86::VFNMSUBSDr231r }, 3519 { X86::VFNMSUBPSr132r, X86::VFNMSUBPSr213r, X86::VFNMSUBPSr231r }, 3520 { X86::VFNMSUBPDr132r, X86::VFNMSUBPDr213r, X86::VFNMSUBPDr231r }, 3521 { X86::VFNMSUBPSr132rY, X86::VFNMSUBPSr213rY, X86::VFNMSUBPSr231rY }, 3522 { X86::VFNMSUBPDr132rY, X86::VFNMSUBPDr213rY, X86::VFNMSUBPDr231rY }, 3523 { X86::VFNMSUBSSr132m, X86::VFNMSUBSSr213m, X86::VFNMSUBSSr231m }, 3524 { X86::VFNMSUBSDr132m, X86::VFNMSUBSDr213m, X86::VFNMSUBSDr231m }, 3525 { X86::VFNMSUBPSr132m, X86::VFNMSUBPSr213m, X86::VFNMSUBPSr231m }, 3526 { X86::VFNMSUBPDr132m, X86::VFNMSUBPDr213m, X86::VFNMSUBPDr231m }, 3527 { X86::VFNMSUBPSr132mY, X86::VFNMSUBPSr213mY, X86::VFNMSUBPSr231mY }, 3528 { X86::VFNMSUBPDr132mY, X86::VFNMSUBPDr213mY, X86::VFNMSUBPDr231mY }, 3529 3530 { X86::VFMADDSUBPSr132r, X86::VFMADDSUBPSr213r, X86::VFMADDSUBPSr231r }, 3531 { X86::VFMADDSUBPDr132r, X86::VFMADDSUBPDr213r, X86::VFMADDSUBPDr231r }, 3532 { X86::VFMADDSUBPSr132rY, X86::VFMADDSUBPSr213rY, X86::VFMADDSUBPSr231rY }, 3533 { X86::VFMADDSUBPDr132rY, X86::VFMADDSUBPDr213rY, X86::VFMADDSUBPDr231rY }, 3534 { X86::VFMADDSUBPSr132m, X86::VFMADDSUBPSr213m, X86::VFMADDSUBPSr231m }, 3535 { X86::VFMADDSUBPDr132m, X86::VFMADDSUBPDr213m, X86::VFMADDSUBPDr231m }, 3536 { X86::VFMADDSUBPSr132mY, X86::VFMADDSUBPSr213mY, X86::VFMADDSUBPSr231mY }, 3537 { X86::VFMADDSUBPDr132mY, X86::VFMADDSUBPDr213mY, X86::VFMADDSUBPDr231mY }, 3538 3539 { X86::VFMSUBADDPSr132r, X86::VFMSUBADDPSr213r, X86::VFMSUBADDPSr231r }, 3540 { X86::VFMSUBADDPDr132r, X86::VFMSUBADDPDr213r, X86::VFMSUBADDPDr231r }, 3541 { X86::VFMSUBADDPSr132rY, X86::VFMSUBADDPSr213rY, X86::VFMSUBADDPSr231rY }, 3542 { X86::VFMSUBADDPDr132rY, X86::VFMSUBADDPDr213rY, X86::VFMSUBADDPDr231rY }, 3543 { X86::VFMSUBADDPSr132m, X86::VFMSUBADDPSr213m, X86::VFMSUBADDPSr231m }, 3544 { X86::VFMSUBADDPDr132m, X86::VFMSUBADDPDr213m, X86::VFMSUBADDPDr231m }, 3545 { X86::VFMSUBADDPSr132mY, X86::VFMSUBADDPSr213mY, X86::VFMSUBADDPSr231mY }, 3546 { X86::VFMSUBADDPDr132mY, X86::VFMSUBADDPDr213mY, X86::VFMSUBADDPDr231mY } 3547 }; 3548 3549 // Define the array that holds FMA*_Int opcodes in groups 3550 // of 3 opcodes(132, 213, 231) in each group. 3551 static const uint16_t IntrinOpcodeGroups[][3] = { 3552 { X86::VFMADDSSr132r_Int, X86::VFMADDSSr213r_Int, X86::VFMADDSSr231r_Int }, 3553 { X86::VFMADDSDr132r_Int, X86::VFMADDSDr213r_Int, X86::VFMADDSDr231r_Int }, 3554 { X86::VFMADDSSr132m_Int, X86::VFMADDSSr213m_Int, X86::VFMADDSSr231m_Int }, 3555 { X86::VFMADDSDr132m_Int, X86::VFMADDSDr213m_Int, X86::VFMADDSDr231m_Int }, 3556 3557 { X86::VFMSUBSSr132r_Int, X86::VFMSUBSSr213r_Int, X86::VFMSUBSSr231r_Int }, 3558 { X86::VFMSUBSDr132r_Int, X86::VFMSUBSDr213r_Int, X86::VFMSUBSDr231r_Int }, 3559 { X86::VFMSUBSSr132m_Int, X86::VFMSUBSSr213m_Int, X86::VFMSUBSSr231m_Int }, 3560 { X86::VFMSUBSDr132m_Int, X86::VFMSUBSDr213m_Int, X86::VFMSUBSDr231m_Int }, 3561 3562 { X86::VFNMADDSSr132r_Int, X86::VFNMADDSSr213r_Int, X86::VFNMADDSSr231r_Int }, 3563 { X86::VFNMADDSDr132r_Int, X86::VFNMADDSDr213r_Int, X86::VFNMADDSDr231r_Int }, 3564 { X86::VFNMADDSSr132m_Int, X86::VFNMADDSSr213m_Int, X86::VFNMADDSSr231m_Int }, 3565 { X86::VFNMADDSDr132m_Int, X86::VFNMADDSDr213m_Int, X86::VFNMADDSDr231m_Int }, 3566 3567 { X86::VFNMSUBSSr132r_Int, X86::VFNMSUBSSr213r_Int, X86::VFNMSUBSSr231r_Int }, 3568 { X86::VFNMSUBSDr132r_Int, X86::VFNMSUBSDr213r_Int, X86::VFNMSUBSDr231r_Int }, 3569 { X86::VFNMSUBSSr132m_Int, X86::VFNMSUBSSr213m_Int, X86::VFNMSUBSSr231m_Int }, 3570 { X86::VFNMSUBSDr132m_Int, X86::VFNMSUBSDr213m_Int, X86::VFNMSUBSDr231m_Int }, 3571 }; 3572 3573 const unsigned Form132Index = 0; 3574 const unsigned Form213Index = 1; 3575 const unsigned Form231Index = 2; 3576 const unsigned FormsNum = 3; 3577 3578 bool IsIntrinOpcode; 3579 isFMA3(Opc, &IsIntrinOpcode); 3580 3581 size_t GroupsNum; 3582 const uint16_t (*OpcodeGroups)[3]; 3583 if (IsIntrinOpcode) { 3584 GroupsNum = array_lengthof(IntrinOpcodeGroups); 3585 OpcodeGroups = IntrinOpcodeGroups; 3586 } else { 3587 GroupsNum = array_lengthof(RegularOpcodeGroups); 3588 OpcodeGroups = RegularOpcodeGroups; 3589 } 3590 3591 const uint16_t *FoundOpcodesGroup = nullptr; 3592 size_t FormIndex; 3593 3594 // Look for the input opcode in the corresponding opcodes table. 3595 for (size_t GroupIndex = 0; GroupIndex < GroupsNum && !FoundOpcodesGroup; 3596 ++GroupIndex) { 3597 for (FormIndex = 0; FormIndex < FormsNum; ++FormIndex) { 3598 if (OpcodeGroups[GroupIndex][FormIndex] == Opc) { 3599 FoundOpcodesGroup = OpcodeGroups[GroupIndex]; 3600 break; 3601 } 3602 } 3603 } 3604 3605 // The input opcode does not match with any of the opcodes from the tables. 3606 // The unsupported FMA opcode must be added to one of the two opcode groups 3607 // defined above. 3608 assert(FoundOpcodesGroup != nullptr && "Unexpected FMA3 opcode"); 3609 3610 // Put the lowest index to SrcOpIdx1 to simplify the checks below. 3611 if (SrcOpIdx1 > SrcOpIdx2) 3612 std::swap(SrcOpIdx1, SrcOpIdx2); 3613 3614 // TODO: Commuting the 1st operand of FMA*_Int requires some additional 3615 // analysis. The commute optimization is legal only if all users of FMA*_Int 3616 // use only the lowest element of the FMA*_Int instruction. Such analysis are 3617 // not implemented yet. So, just return 0 in that case. 3618 // When such analysis are available this place will be the right place for 3619 // calling it. 3620 if (IsIntrinOpcode && SrcOpIdx1 == 1) 3621 return 0; 3622 3623 unsigned Case; 3624 if (SrcOpIdx1 == 1 && SrcOpIdx2 == 2) 3625 Case = 0; 3626 else if (SrcOpIdx1 == 1 && SrcOpIdx2 == 3) 3627 Case = 1; 3628 else if (SrcOpIdx1 == 2 && SrcOpIdx2 == 3) 3629 Case = 2; 3630 else 3631 return 0; 3632 3633 // Define the FMA forms mapping array that helps to map input FMA form 3634 // to output FMA form to preserve the operation semantics after 3635 // commuting the operands. 3636 static const unsigned FormMapping[][3] = { 3637 // 0: SrcOpIdx1 == 1 && SrcOpIdx2 == 2; 3638 // FMA132 A, C, b; ==> FMA231 C, A, b; 3639 // FMA213 B, A, c; ==> FMA213 A, B, c; 3640 // FMA231 C, A, b; ==> FMA132 A, C, b; 3641 { Form231Index, Form213Index, Form132Index }, 3642 // 1: SrcOpIdx1 == 1 && SrcOpIdx2 == 3; 3643 // FMA132 A, c, B; ==> FMA132 B, c, A; 3644 // FMA213 B, a, C; ==> FMA231 C, a, B; 3645 // FMA231 C, a, B; ==> FMA213 B, a, C; 3646 { Form132Index, Form231Index, Form213Index }, 3647 // 2: SrcOpIdx1 == 2 && SrcOpIdx2 == 3; 3648 // FMA132 a, C, B; ==> FMA213 a, B, C; 3649 // FMA213 b, A, C; ==> FMA132 b, C, A; 3650 // FMA231 c, A, B; ==> FMA231 c, B, A; 3651 { Form213Index, Form132Index, Form231Index } 3652 }; 3653 3654 // Everything is ready, just adjust the FMA opcode and return it. 3655 FormIndex = FormMapping[Case][FormIndex]; 3656 return FoundOpcodesGroup[FormIndex]; 3657 } 3658 3659 bool X86InstrInfo::findCommutedOpIndices(MachineInstr *MI, 3660 unsigned &SrcOpIdx1, 3661 unsigned &SrcOpIdx2) const { 3662 switch (MI->getOpcode()) { 3663 case X86::CMPPDrri: 3664 case X86::CMPPSrri: 3665 case X86::VCMPPDrri: 3666 case X86::VCMPPSrri: 3667 case X86::VCMPPDYrri: 3668 case X86::VCMPPSYrri: { 3669 // Float comparison can be safely commuted for 3670 // Ordered/Unordered/Equal/NotEqual tests 3671 unsigned Imm = MI->getOperand(3).getImm() & 0x7; 3672 switch (Imm) { 3673 case 0x00: // EQUAL 3674 case 0x03: // UNORDERED 3675 case 0x04: // NOT EQUAL 3676 case 0x07: // ORDERED 3677 // The indices of the commutable operands are 1 and 2. 3678 // Assign them to the returned operand indices here. 3679 return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 1, 2); 3680 } 3681 return false; 3682 } 3683 default: 3684 if (isFMA3(MI->getOpcode())) 3685 return findFMA3CommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 3686 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 3687 } 3688 return false; 3689 } 3690 3691 static X86::CondCode getCondFromBranchOpc(unsigned BrOpc) { 3692 switch (BrOpc) { 3693 default: return X86::COND_INVALID; 3694 case X86::JE_1: return X86::COND_E; 3695 case X86::JNE_1: return X86::COND_NE; 3696 case X86::JL_1: return X86::COND_L; 3697 case X86::JLE_1: return X86::COND_LE; 3698 case X86::JG_1: return X86::COND_G; 3699 case X86::JGE_1: return X86::COND_GE; 3700 case X86::JB_1: return X86::COND_B; 3701 case X86::JBE_1: return X86::COND_BE; 3702 case X86::JA_1: return X86::COND_A; 3703 case X86::JAE_1: return X86::COND_AE; 3704 case X86::JS_1: return X86::COND_S; 3705 case X86::JNS_1: return X86::COND_NS; 3706 case X86::JP_1: return X86::COND_P; 3707 case X86::JNP_1: return X86::COND_NP; 3708 case X86::JO_1: return X86::COND_O; 3709 case X86::JNO_1: return X86::COND_NO; 3710 } 3711 } 3712 3713 /// Return condition code of a SET opcode. 3714 static X86::CondCode getCondFromSETOpc(unsigned Opc) { 3715 switch (Opc) { 3716 default: return X86::COND_INVALID; 3717 case X86::SETAr: case X86::SETAm: return X86::COND_A; 3718 case X86::SETAEr: case X86::SETAEm: return X86::COND_AE; 3719 case X86::SETBr: case X86::SETBm: return X86::COND_B; 3720 case X86::SETBEr: case X86::SETBEm: return X86::COND_BE; 3721 case X86::SETEr: case X86::SETEm: return X86::COND_E; 3722 case X86::SETGr: case X86::SETGm: return X86::COND_G; 3723 case X86::SETGEr: case X86::SETGEm: return X86::COND_GE; 3724 case X86::SETLr: case X86::SETLm: return X86::COND_L; 3725 case X86::SETLEr: case X86::SETLEm: return X86::COND_LE; 3726 case X86::SETNEr: case X86::SETNEm: return X86::COND_NE; 3727 case X86::SETNOr: case X86::SETNOm: return X86::COND_NO; 3728 case X86::SETNPr: case X86::SETNPm: return X86::COND_NP; 3729 case X86::SETNSr: case X86::SETNSm: return X86::COND_NS; 3730 case X86::SETOr: case X86::SETOm: return X86::COND_O; 3731 case X86::SETPr: case X86::SETPm: return X86::COND_P; 3732 case X86::SETSr: case X86::SETSm: return X86::COND_S; 3733 } 3734 } 3735 3736 /// Return condition code of a CMov opcode. 3737 X86::CondCode X86::getCondFromCMovOpc(unsigned Opc) { 3738 switch (Opc) { 3739 default: return X86::COND_INVALID; 3740 case X86::CMOVA16rm: case X86::CMOVA16rr: case X86::CMOVA32rm: 3741 case X86::CMOVA32rr: case X86::CMOVA64rm: case X86::CMOVA64rr: 3742 return X86::COND_A; 3743 case X86::CMOVAE16rm: case X86::CMOVAE16rr: case X86::CMOVAE32rm: 3744 case X86::CMOVAE32rr: case X86::CMOVAE64rm: case X86::CMOVAE64rr: 3745 return X86::COND_AE; 3746 case X86::CMOVB16rm: case X86::CMOVB16rr: case X86::CMOVB32rm: 3747 case X86::CMOVB32rr: case X86::CMOVB64rm: case X86::CMOVB64rr: 3748 return X86::COND_B; 3749 case X86::CMOVBE16rm: case X86::CMOVBE16rr: case X86::CMOVBE32rm: 3750 case X86::CMOVBE32rr: case X86::CMOVBE64rm: case X86::CMOVBE64rr: 3751 return X86::COND_BE; 3752 case X86::CMOVE16rm: case X86::CMOVE16rr: case X86::CMOVE32rm: 3753 case X86::CMOVE32rr: case X86::CMOVE64rm: case X86::CMOVE64rr: 3754 return X86::COND_E; 3755 case X86::CMOVG16rm: case X86::CMOVG16rr: case X86::CMOVG32rm: 3756 case X86::CMOVG32rr: case X86::CMOVG64rm: case X86::CMOVG64rr: 3757 return X86::COND_G; 3758 case X86::CMOVGE16rm: case X86::CMOVGE16rr: case X86::CMOVGE32rm: 3759 case X86::CMOVGE32rr: case X86::CMOVGE64rm: case X86::CMOVGE64rr: 3760 return X86::COND_GE; 3761 case X86::CMOVL16rm: case X86::CMOVL16rr: case X86::CMOVL32rm: 3762 case X86::CMOVL32rr: case X86::CMOVL64rm: case X86::CMOVL64rr: 3763 return X86::COND_L; 3764 case X86::CMOVLE16rm: case X86::CMOVLE16rr: case X86::CMOVLE32rm: 3765 case X86::CMOVLE32rr: case X86::CMOVLE64rm: case X86::CMOVLE64rr: 3766 return X86::COND_LE; 3767 case X86::CMOVNE16rm: case X86::CMOVNE16rr: case X86::CMOVNE32rm: 3768 case X86::CMOVNE32rr: case X86::CMOVNE64rm: case X86::CMOVNE64rr: 3769 return X86::COND_NE; 3770 case X86::CMOVNO16rm: case X86::CMOVNO16rr: case X86::CMOVNO32rm: 3771 case X86::CMOVNO32rr: case X86::CMOVNO64rm: case X86::CMOVNO64rr: 3772 return X86::COND_NO; 3773 case X86::CMOVNP16rm: case X86::CMOVNP16rr: case X86::CMOVNP32rm: 3774 case X86::CMOVNP32rr: case X86::CMOVNP64rm: case X86::CMOVNP64rr: 3775 return X86::COND_NP; 3776 case X86::CMOVNS16rm: case X86::CMOVNS16rr: case X86::CMOVNS32rm: 3777 case X86::CMOVNS32rr: case X86::CMOVNS64rm: case X86::CMOVNS64rr: 3778 return X86::COND_NS; 3779 case X86::CMOVO16rm: case X86::CMOVO16rr: case X86::CMOVO32rm: 3780 case X86::CMOVO32rr: case X86::CMOVO64rm: case X86::CMOVO64rr: 3781 return X86::COND_O; 3782 case X86::CMOVP16rm: case X86::CMOVP16rr: case X86::CMOVP32rm: 3783 case X86::CMOVP32rr: case X86::CMOVP64rm: case X86::CMOVP64rr: 3784 return X86::COND_P; 3785 case X86::CMOVS16rm: case X86::CMOVS16rr: case X86::CMOVS32rm: 3786 case X86::CMOVS32rr: case X86::CMOVS64rm: case X86::CMOVS64rr: 3787 return X86::COND_S; 3788 } 3789 } 3790 3791 unsigned X86::GetCondBranchFromCond(X86::CondCode CC) { 3792 switch (CC) { 3793 default: llvm_unreachable("Illegal condition code!"); 3794 case X86::COND_E: return X86::JE_1; 3795 case X86::COND_NE: return X86::JNE_1; 3796 case X86::COND_L: return X86::JL_1; 3797 case X86::COND_LE: return X86::JLE_1; 3798 case X86::COND_G: return X86::JG_1; 3799 case X86::COND_GE: return X86::JGE_1; 3800 case X86::COND_B: return X86::JB_1; 3801 case X86::COND_BE: return X86::JBE_1; 3802 case X86::COND_A: return X86::JA_1; 3803 case X86::COND_AE: return X86::JAE_1; 3804 case X86::COND_S: return X86::JS_1; 3805 case X86::COND_NS: return X86::JNS_1; 3806 case X86::COND_P: return X86::JP_1; 3807 case X86::COND_NP: return X86::JNP_1; 3808 case X86::COND_O: return X86::JO_1; 3809 case X86::COND_NO: return X86::JNO_1; 3810 } 3811 } 3812 3813 /// Return the inverse of the specified condition, 3814 /// e.g. turning COND_E to COND_NE. 3815 X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) { 3816 switch (CC) { 3817 default: llvm_unreachable("Illegal condition code!"); 3818 case X86::COND_E: return X86::COND_NE; 3819 case X86::COND_NE: return X86::COND_E; 3820 case X86::COND_L: return X86::COND_GE; 3821 case X86::COND_LE: return X86::COND_G; 3822 case X86::COND_G: return X86::COND_LE; 3823 case X86::COND_GE: return X86::COND_L; 3824 case X86::COND_B: return X86::COND_AE; 3825 case X86::COND_BE: return X86::COND_A; 3826 case X86::COND_A: return X86::COND_BE; 3827 case X86::COND_AE: return X86::COND_B; 3828 case X86::COND_S: return X86::COND_NS; 3829 case X86::COND_NS: return X86::COND_S; 3830 case X86::COND_P: return X86::COND_NP; 3831 case X86::COND_NP: return X86::COND_P; 3832 case X86::COND_O: return X86::COND_NO; 3833 case X86::COND_NO: return X86::COND_O; 3834 case X86::COND_NE_OR_P: return X86::COND_E_AND_NP; 3835 case X86::COND_E_AND_NP: return X86::COND_NE_OR_P; 3836 } 3837 } 3838 3839 /// Assuming the flags are set by MI(a,b), return the condition code if we 3840 /// modify the instructions such that flags are set by MI(b,a). 3841 static X86::CondCode getSwappedCondition(X86::CondCode CC) { 3842 switch (CC) { 3843 default: return X86::COND_INVALID; 3844 case X86::COND_E: return X86::COND_E; 3845 case X86::COND_NE: return X86::COND_NE; 3846 case X86::COND_L: return X86::COND_G; 3847 case X86::COND_LE: return X86::COND_GE; 3848 case X86::COND_G: return X86::COND_L; 3849 case X86::COND_GE: return X86::COND_LE; 3850 case X86::COND_B: return X86::COND_A; 3851 case X86::COND_BE: return X86::COND_AE; 3852 case X86::COND_A: return X86::COND_B; 3853 case X86::COND_AE: return X86::COND_BE; 3854 } 3855 } 3856 3857 /// Return a set opcode for the given condition and 3858 /// whether it has memory operand. 3859 unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) { 3860 static const uint16_t Opc[16][2] = { 3861 { X86::SETAr, X86::SETAm }, 3862 { X86::SETAEr, X86::SETAEm }, 3863 { X86::SETBr, X86::SETBm }, 3864 { X86::SETBEr, X86::SETBEm }, 3865 { X86::SETEr, X86::SETEm }, 3866 { X86::SETGr, X86::SETGm }, 3867 { X86::SETGEr, X86::SETGEm }, 3868 { X86::SETLr, X86::SETLm }, 3869 { X86::SETLEr, X86::SETLEm }, 3870 { X86::SETNEr, X86::SETNEm }, 3871 { X86::SETNOr, X86::SETNOm }, 3872 { X86::SETNPr, X86::SETNPm }, 3873 { X86::SETNSr, X86::SETNSm }, 3874 { X86::SETOr, X86::SETOm }, 3875 { X86::SETPr, X86::SETPm }, 3876 { X86::SETSr, X86::SETSm } 3877 }; 3878 3879 assert(CC <= LAST_VALID_COND && "Can only handle standard cond codes"); 3880 return Opc[CC][HasMemoryOperand ? 1 : 0]; 3881 } 3882 3883 /// Return a cmov opcode for the given condition, 3884 /// register size in bytes, and operand type. 3885 unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes, 3886 bool HasMemoryOperand) { 3887 static const uint16_t Opc[32][3] = { 3888 { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr }, 3889 { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr }, 3890 { X86::CMOVB16rr, X86::CMOVB32rr, X86::CMOVB64rr }, 3891 { X86::CMOVBE16rr, X86::CMOVBE32rr, X86::CMOVBE64rr }, 3892 { X86::CMOVE16rr, X86::CMOVE32rr, X86::CMOVE64rr }, 3893 { X86::CMOVG16rr, X86::CMOVG32rr, X86::CMOVG64rr }, 3894 { X86::CMOVGE16rr, X86::CMOVGE32rr, X86::CMOVGE64rr }, 3895 { X86::CMOVL16rr, X86::CMOVL32rr, X86::CMOVL64rr }, 3896 { X86::CMOVLE16rr, X86::CMOVLE32rr, X86::CMOVLE64rr }, 3897 { X86::CMOVNE16rr, X86::CMOVNE32rr, X86::CMOVNE64rr }, 3898 { X86::CMOVNO16rr, X86::CMOVNO32rr, X86::CMOVNO64rr }, 3899 { X86::CMOVNP16rr, X86::CMOVNP32rr, X86::CMOVNP64rr }, 3900 { X86::CMOVNS16rr, X86::CMOVNS32rr, X86::CMOVNS64rr }, 3901 { X86::CMOVO16rr, X86::CMOVO32rr, X86::CMOVO64rr }, 3902 { X86::CMOVP16rr, X86::CMOVP32rr, X86::CMOVP64rr }, 3903 { X86::CMOVS16rr, X86::CMOVS32rr, X86::CMOVS64rr }, 3904 { X86::CMOVA16rm, X86::CMOVA32rm, X86::CMOVA64rm }, 3905 { X86::CMOVAE16rm, X86::CMOVAE32rm, X86::CMOVAE64rm }, 3906 { X86::CMOVB16rm, X86::CMOVB32rm, X86::CMOVB64rm }, 3907 { X86::CMOVBE16rm, X86::CMOVBE32rm, X86::CMOVBE64rm }, 3908 { X86::CMOVE16rm, X86::CMOVE32rm, X86::CMOVE64rm }, 3909 { X86::CMOVG16rm, X86::CMOVG32rm, X86::CMOVG64rm }, 3910 { X86::CMOVGE16rm, X86::CMOVGE32rm, X86::CMOVGE64rm }, 3911 { X86::CMOVL16rm, X86::CMOVL32rm, X86::CMOVL64rm }, 3912 { X86::CMOVLE16rm, X86::CMOVLE32rm, X86::CMOVLE64rm }, 3913 { X86::CMOVNE16rm, X86::CMOVNE32rm, X86::CMOVNE64rm }, 3914 { X86::CMOVNO16rm, X86::CMOVNO32rm, X86::CMOVNO64rm }, 3915 { X86::CMOVNP16rm, X86::CMOVNP32rm, X86::CMOVNP64rm }, 3916 { X86::CMOVNS16rm, X86::CMOVNS32rm, X86::CMOVNS64rm }, 3917 { X86::CMOVO16rm, X86::CMOVO32rm, X86::CMOVO64rm }, 3918 { X86::CMOVP16rm, X86::CMOVP32rm, X86::CMOVP64rm }, 3919 { X86::CMOVS16rm, X86::CMOVS32rm, X86::CMOVS64rm } 3920 }; 3921 3922 assert(CC < 16 && "Can only handle standard cond codes"); 3923 unsigned Idx = HasMemoryOperand ? 16+CC : CC; 3924 switch(RegBytes) { 3925 default: llvm_unreachable("Illegal register size!"); 3926 case 2: return Opc[Idx][0]; 3927 case 4: return Opc[Idx][1]; 3928 case 8: return Opc[Idx][2]; 3929 } 3930 } 3931 3932 bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { 3933 if (!MI.isTerminator()) return false; 3934 3935 // Conditional branch is a special case. 3936 if (MI.isBranch() && !MI.isBarrier()) 3937 return true; 3938 if (!MI.isPredicable()) 3939 return true; 3940 return !isPredicated(MI); 3941 } 3942 3943 // Given a MBB and its TBB, find the FBB which was a fallthrough MBB (it may 3944 // not be a fallthrough MBB now due to layout changes). Return nullptr if the 3945 // fallthrough MBB cannot be identified. 3946 static MachineBasicBlock *getFallThroughMBB(MachineBasicBlock *MBB, 3947 MachineBasicBlock *TBB) { 3948 // Look for non-EHPad successors other than TBB. If we find exactly one, it 3949 // is the fallthrough MBB. If we find zero, then TBB is both the target MBB 3950 // and fallthrough MBB. If we find more than one, we cannot identify the 3951 // fallthrough MBB and should return nullptr. 3952 MachineBasicBlock *FallthroughBB = nullptr; 3953 for (auto SI = MBB->succ_begin(), SE = MBB->succ_end(); SI != SE; ++SI) { 3954 if ((*SI)->isEHPad() || (*SI == TBB && FallthroughBB)) 3955 continue; 3956 // Return a nullptr if we found more than one fallthrough successor. 3957 if (FallthroughBB && FallthroughBB != TBB) 3958 return nullptr; 3959 FallthroughBB = *SI; 3960 } 3961 return FallthroughBB; 3962 } 3963 3964 bool X86InstrInfo::AnalyzeBranchImpl( 3965 MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, 3966 SmallVectorImpl<MachineOperand> &Cond, 3967 SmallVectorImpl<MachineInstr *> &CondBranches, bool AllowModify) const { 3968 3969 // Start from the bottom of the block and work up, examining the 3970 // terminator instructions. 3971 MachineBasicBlock::iterator I = MBB.end(); 3972 MachineBasicBlock::iterator UnCondBrIter = MBB.end(); 3973 while (I != MBB.begin()) { 3974 --I; 3975 if (I->isDebugValue()) 3976 continue; 3977 3978 // Working from the bottom, when we see a non-terminator instruction, we're 3979 // done. 3980 if (!isUnpredicatedTerminator(*I)) 3981 break; 3982 3983 // A terminator that isn't a branch can't easily be handled by this 3984 // analysis. 3985 if (!I->isBranch()) 3986 return true; 3987 3988 // Handle unconditional branches. 3989 if (I->getOpcode() == X86::JMP_1) { 3990 UnCondBrIter = I; 3991 3992 if (!AllowModify) { 3993 TBB = I->getOperand(0).getMBB(); 3994 continue; 3995 } 3996 3997 // If the block has any instructions after a JMP, delete them. 3998 while (std::next(I) != MBB.end()) 3999 std::next(I)->eraseFromParent(); 4000 4001 Cond.clear(); 4002 FBB = nullptr; 4003 4004 // Delete the JMP if it's equivalent to a fall-through. 4005 if (MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 4006 TBB = nullptr; 4007 I->eraseFromParent(); 4008 I = MBB.end(); 4009 UnCondBrIter = MBB.end(); 4010 continue; 4011 } 4012 4013 // TBB is used to indicate the unconditional destination. 4014 TBB = I->getOperand(0).getMBB(); 4015 continue; 4016 } 4017 4018 // Handle conditional branches. 4019 X86::CondCode BranchCode = getCondFromBranchOpc(I->getOpcode()); 4020 if (BranchCode == X86::COND_INVALID) 4021 return true; // Can't handle indirect branch. 4022 4023 // Working from the bottom, handle the first conditional branch. 4024 if (Cond.empty()) { 4025 MachineBasicBlock *TargetBB = I->getOperand(0).getMBB(); 4026 if (AllowModify && UnCondBrIter != MBB.end() && 4027 MBB.isLayoutSuccessor(TargetBB)) { 4028 // If we can modify the code and it ends in something like: 4029 // 4030 // jCC L1 4031 // jmp L2 4032 // L1: 4033 // ... 4034 // L2: 4035 // 4036 // Then we can change this to: 4037 // 4038 // jnCC L2 4039 // L1: 4040 // ... 4041 // L2: 4042 // 4043 // Which is a bit more efficient. 4044 // We conditionally jump to the fall-through block. 4045 BranchCode = GetOppositeBranchCondition(BranchCode); 4046 unsigned JNCC = GetCondBranchFromCond(BranchCode); 4047 MachineBasicBlock::iterator OldInst = I; 4048 4049 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC)) 4050 .addMBB(UnCondBrIter->getOperand(0).getMBB()); 4051 BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1)) 4052 .addMBB(TargetBB); 4053 4054 OldInst->eraseFromParent(); 4055 UnCondBrIter->eraseFromParent(); 4056 4057 // Restart the analysis. 4058 UnCondBrIter = MBB.end(); 4059 I = MBB.end(); 4060 continue; 4061 } 4062 4063 FBB = TBB; 4064 TBB = I->getOperand(0).getMBB(); 4065 Cond.push_back(MachineOperand::CreateImm(BranchCode)); 4066 CondBranches.push_back(I); 4067 continue; 4068 } 4069 4070 // Handle subsequent conditional branches. Only handle the case where all 4071 // conditional branches branch to the same destination and their condition 4072 // opcodes fit one of the special multi-branch idioms. 4073 assert(Cond.size() == 1); 4074 assert(TBB); 4075 4076 // If the conditions are the same, we can leave them alone. 4077 X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm(); 4078 auto NewTBB = I->getOperand(0).getMBB(); 4079 if (OldBranchCode == BranchCode && TBB == NewTBB) 4080 continue; 4081 4082 // If they differ, see if they fit one of the known patterns. Theoretically, 4083 // we could handle more patterns here, but we shouldn't expect to see them 4084 // if instruction selection has done a reasonable job. 4085 if (TBB == NewTBB && 4086 ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) || 4087 (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) { 4088 BranchCode = X86::COND_NE_OR_P; 4089 } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) || 4090 (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) { 4091 if (NewTBB != (FBB ? FBB : getFallThroughMBB(&MBB, TBB))) 4092 return true; 4093 4094 // X86::COND_E_AND_NP usually has two different branch destinations. 4095 // 4096 // JP B1 4097 // JE B2 4098 // JMP B1 4099 // B1: 4100 // B2: 4101 // 4102 // Here this condition branches to B2 only if NP && E. It has another 4103 // equivalent form: 4104 // 4105 // JNE B1 4106 // JNP B2 4107 // JMP B1 4108 // B1: 4109 // B2: 4110 // 4111 // Similarly it branches to B2 only if E && NP. That is why this condition 4112 // is named with COND_E_AND_NP. 4113 BranchCode = X86::COND_E_AND_NP; 4114 } else 4115 return true; 4116 4117 // Update the MachineOperand. 4118 Cond[0].setImm(BranchCode); 4119 CondBranches.push_back(I); 4120 } 4121 4122 return false; 4123 } 4124 4125 bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, 4126 MachineBasicBlock *&TBB, 4127 MachineBasicBlock *&FBB, 4128 SmallVectorImpl<MachineOperand> &Cond, 4129 bool AllowModify) const { 4130 SmallVector<MachineInstr *, 4> CondBranches; 4131 return AnalyzeBranchImpl(MBB, TBB, FBB, Cond, CondBranches, AllowModify); 4132 } 4133 4134 bool X86InstrInfo::AnalyzeBranchPredicate(MachineBasicBlock &MBB, 4135 MachineBranchPredicate &MBP, 4136 bool AllowModify) const { 4137 using namespace std::placeholders; 4138 4139 SmallVector<MachineOperand, 4> Cond; 4140 SmallVector<MachineInstr *, 4> CondBranches; 4141 if (AnalyzeBranchImpl(MBB, MBP.TrueDest, MBP.FalseDest, Cond, CondBranches, 4142 AllowModify)) 4143 return true; 4144 4145 if (Cond.size() != 1) 4146 return true; 4147 4148 assert(MBP.TrueDest && "expected!"); 4149 4150 if (!MBP.FalseDest) 4151 MBP.FalseDest = MBB.getNextNode(); 4152 4153 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4154 4155 MachineInstr *ConditionDef = nullptr; 4156 bool SingleUseCondition = true; 4157 4158 for (auto I = std::next(MBB.rbegin()), E = MBB.rend(); I != E; ++I) { 4159 if (I->modifiesRegister(X86::EFLAGS, TRI)) { 4160 ConditionDef = &*I; 4161 break; 4162 } 4163 4164 if (I->readsRegister(X86::EFLAGS, TRI)) 4165 SingleUseCondition = false; 4166 } 4167 4168 if (!ConditionDef) 4169 return true; 4170 4171 if (SingleUseCondition) { 4172 for (auto *Succ : MBB.successors()) 4173 if (Succ->isLiveIn(X86::EFLAGS)) 4174 SingleUseCondition = false; 4175 } 4176 4177 MBP.ConditionDef = ConditionDef; 4178 MBP.SingleUseCondition = SingleUseCondition; 4179 4180 // Currently we only recognize the simple pattern: 4181 // 4182 // test %reg, %reg 4183 // je %label 4184 // 4185 const unsigned TestOpcode = 4186 Subtarget.is64Bit() ? X86::TEST64rr : X86::TEST32rr; 4187 4188 if (ConditionDef->getOpcode() == TestOpcode && 4189 ConditionDef->getNumOperands() == 3 && 4190 ConditionDef->getOperand(0).isIdenticalTo(ConditionDef->getOperand(1)) && 4191 (Cond[0].getImm() == X86::COND_NE || Cond[0].getImm() == X86::COND_E)) { 4192 MBP.LHS = ConditionDef->getOperand(0); 4193 MBP.RHS = MachineOperand::CreateImm(0); 4194 MBP.Predicate = Cond[0].getImm() == X86::COND_NE 4195 ? MachineBranchPredicate::PRED_NE 4196 : MachineBranchPredicate::PRED_EQ; 4197 return false; 4198 } 4199 4200 return true; 4201 } 4202 4203 unsigned X86InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 4204 MachineBasicBlock::iterator I = MBB.end(); 4205 unsigned Count = 0; 4206 4207 while (I != MBB.begin()) { 4208 --I; 4209 if (I->isDebugValue()) 4210 continue; 4211 if (I->getOpcode() != X86::JMP_1 && 4212 getCondFromBranchOpc(I->getOpcode()) == X86::COND_INVALID) 4213 break; 4214 // Remove the branch. 4215 I->eraseFromParent(); 4216 I = MBB.end(); 4217 ++Count; 4218 } 4219 4220 return Count; 4221 } 4222 4223 unsigned X86InstrInfo::InsertBranch(MachineBasicBlock &MBB, 4224 MachineBasicBlock *TBB, 4225 MachineBasicBlock *FBB, 4226 ArrayRef<MachineOperand> Cond, 4227 const DebugLoc &DL) const { 4228 // Shouldn't be a fall through. 4229 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 4230 assert((Cond.size() == 1 || Cond.size() == 0) && 4231 "X86 branch conditions have one component!"); 4232 4233 if (Cond.empty()) { 4234 // Unconditional branch? 4235 assert(!FBB && "Unconditional branch with multiple successors!"); 4236 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(TBB); 4237 return 1; 4238 } 4239 4240 // If FBB is null, it is implied to be a fall-through block. 4241 bool FallThru = FBB == nullptr; 4242 4243 // Conditional branch. 4244 unsigned Count = 0; 4245 X86::CondCode CC = (X86::CondCode)Cond[0].getImm(); 4246 switch (CC) { 4247 case X86::COND_NE_OR_P: 4248 // Synthesize NE_OR_P with two branches. 4249 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB); 4250 ++Count; 4251 BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB); 4252 ++Count; 4253 break; 4254 case X86::COND_E_AND_NP: 4255 // Use the next block of MBB as FBB if it is null. 4256 if (FBB == nullptr) { 4257 FBB = getFallThroughMBB(&MBB, TBB); 4258 assert(FBB && "MBB cannot be the last block in function when the false " 4259 "body is a fall-through."); 4260 } 4261 // Synthesize COND_E_AND_NP with two branches. 4262 BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB); 4263 ++Count; 4264 BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB); 4265 ++Count; 4266 break; 4267 default: { 4268 unsigned Opc = GetCondBranchFromCond(CC); 4269 BuildMI(&MBB, DL, get(Opc)).addMBB(TBB); 4270 ++Count; 4271 } 4272 } 4273 if (!FallThru) { 4274 // Two-way Conditional branch. Insert the second branch. 4275 BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB); 4276 ++Count; 4277 } 4278 return Count; 4279 } 4280 4281 bool X86InstrInfo:: 4282 canInsertSelect(const MachineBasicBlock &MBB, 4283 ArrayRef<MachineOperand> Cond, 4284 unsigned TrueReg, unsigned FalseReg, 4285 int &CondCycles, int &TrueCycles, int &FalseCycles) const { 4286 // Not all subtargets have cmov instructions. 4287 if (!Subtarget.hasCMov()) 4288 return false; 4289 if (Cond.size() != 1) 4290 return false; 4291 // We cannot do the composite conditions, at least not in SSA form. 4292 if ((X86::CondCode)Cond[0].getImm() > X86::COND_S) 4293 return false; 4294 4295 // Check register classes. 4296 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 4297 const TargetRegisterClass *RC = 4298 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 4299 if (!RC) 4300 return false; 4301 4302 // We have cmov instructions for 16, 32, and 64 bit general purpose registers. 4303 if (X86::GR16RegClass.hasSubClassEq(RC) || 4304 X86::GR32RegClass.hasSubClassEq(RC) || 4305 X86::GR64RegClass.hasSubClassEq(RC)) { 4306 // This latency applies to Pentium M, Merom, Wolfdale, Nehalem, and Sandy 4307 // Bridge. Probably Ivy Bridge as well. 4308 CondCycles = 2; 4309 TrueCycles = 2; 4310 FalseCycles = 2; 4311 return true; 4312 } 4313 4314 // Can't do vectors. 4315 return false; 4316 } 4317 4318 void X86InstrInfo::insertSelect(MachineBasicBlock &MBB, 4319 MachineBasicBlock::iterator I, 4320 const DebugLoc &DL, unsigned DstReg, 4321 ArrayRef<MachineOperand> Cond, unsigned TrueReg, 4322 unsigned FalseReg) const { 4323 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 4324 assert(Cond.size() == 1 && "Invalid Cond array"); 4325 unsigned Opc = getCMovFromCond((X86::CondCode)Cond[0].getImm(), 4326 MRI.getRegClass(DstReg)->getSize(), 4327 false /*HasMemoryOperand*/); 4328 BuildMI(MBB, I, DL, get(Opc), DstReg).addReg(FalseReg).addReg(TrueReg); 4329 } 4330 4331 /// Test if the given register is a physical h register. 4332 static bool isHReg(unsigned Reg) { 4333 return X86::GR8_ABCD_HRegClass.contains(Reg); 4334 } 4335 4336 // Try and copy between VR128/VR64 and GR64 registers. 4337 static unsigned CopyToFromAsymmetricReg(unsigned DestReg, unsigned SrcReg, 4338 const X86Subtarget &Subtarget) { 4339 4340 // SrcReg(VR128) -> DestReg(GR64) 4341 // SrcReg(VR64) -> DestReg(GR64) 4342 // SrcReg(GR64) -> DestReg(VR128) 4343 // SrcReg(GR64) -> DestReg(VR64) 4344 4345 bool HasAVX = Subtarget.hasAVX(); 4346 bool HasAVX512 = Subtarget.hasAVX512(); 4347 if (X86::GR64RegClass.contains(DestReg)) { 4348 if (X86::VR128XRegClass.contains(SrcReg)) 4349 // Copy from a VR128 register to a GR64 register. 4350 return HasAVX512 ? X86::VMOVPQIto64Zrr: (HasAVX ? X86::VMOVPQIto64rr : 4351 X86::MOVPQIto64rr); 4352 if (X86::VR64RegClass.contains(SrcReg)) 4353 // Copy from a VR64 register to a GR64 register. 4354 return X86::MMX_MOVD64from64rr; 4355 } else if (X86::GR64RegClass.contains(SrcReg)) { 4356 // Copy from a GR64 register to a VR128 register. 4357 if (X86::VR128XRegClass.contains(DestReg)) 4358 return HasAVX512 ? X86::VMOV64toPQIZrr: (HasAVX ? X86::VMOV64toPQIrr : 4359 X86::MOV64toPQIrr); 4360 // Copy from a GR64 register to a VR64 register. 4361 if (X86::VR64RegClass.contains(DestReg)) 4362 return X86::MMX_MOVD64to64rr; 4363 } 4364 4365 // SrcReg(FR32) -> DestReg(GR32) 4366 // SrcReg(GR32) -> DestReg(FR32) 4367 4368 if (X86::GR32RegClass.contains(DestReg) && X86::FR32XRegClass.contains(SrcReg)) 4369 // Copy from a FR32 register to a GR32 register. 4370 return HasAVX512 ? X86::VMOVSS2DIZrr : (HasAVX ? X86::VMOVSS2DIrr : X86::MOVSS2DIrr); 4371 4372 if (X86::FR32XRegClass.contains(DestReg) && X86::GR32RegClass.contains(SrcReg)) 4373 // Copy from a GR32 register to a FR32 register. 4374 return HasAVX512 ? X86::VMOVDI2SSZrr : (HasAVX ? X86::VMOVDI2SSrr : X86::MOVDI2SSrr); 4375 return 0; 4376 } 4377 4378 static bool isMaskRegClass(const TargetRegisterClass *RC) { 4379 // All KMASK RegClasses hold the same k registers, can be tested against anyone. 4380 return X86::VK16RegClass.hasSubClassEq(RC); 4381 } 4382 4383 static bool MaskRegClassContains(unsigned Reg) { 4384 // All KMASK RegClasses hold the same k registers, can be tested against anyone. 4385 return X86::VK16RegClass.contains(Reg); 4386 } 4387 4388 static bool GRRegClassContains(unsigned Reg) { 4389 return X86::GR64RegClass.contains(Reg) || 4390 X86::GR32RegClass.contains(Reg) || 4391 X86::GR16RegClass.contains(Reg) || 4392 X86::GR8RegClass.contains(Reg); 4393 } 4394 static 4395 unsigned copyPhysRegOpcode_AVX512_DQ(unsigned& DestReg, unsigned& SrcReg) { 4396 if (MaskRegClassContains(SrcReg) && X86::GR8RegClass.contains(DestReg)) { 4397 DestReg = getX86SubSuperRegister(DestReg, 32); 4398 return X86::KMOVBrk; 4399 } 4400 if (MaskRegClassContains(DestReg) && X86::GR8RegClass.contains(SrcReg)) { 4401 SrcReg = getX86SubSuperRegister(SrcReg, 32); 4402 return X86::KMOVBkr; 4403 } 4404 return 0; 4405 } 4406 4407 static 4408 unsigned copyPhysRegOpcode_AVX512_BW(unsigned& DestReg, unsigned& SrcReg) { 4409 if (MaskRegClassContains(SrcReg) && MaskRegClassContains(DestReg)) 4410 return X86::KMOVQkk; 4411 if (MaskRegClassContains(SrcReg) && X86::GR32RegClass.contains(DestReg)) 4412 return X86::KMOVDrk; 4413 if (MaskRegClassContains(SrcReg) && X86::GR64RegClass.contains(DestReg)) 4414 return X86::KMOVQrk; 4415 if (MaskRegClassContains(DestReg) && X86::GR32RegClass.contains(SrcReg)) 4416 return X86::KMOVDkr; 4417 if (MaskRegClassContains(DestReg) && X86::GR64RegClass.contains(SrcReg)) 4418 return X86::KMOVQkr; 4419 return 0; 4420 } 4421 4422 static 4423 unsigned copyPhysRegOpcode_AVX512(unsigned& DestReg, unsigned& SrcReg, 4424 const X86Subtarget &Subtarget) 4425 { 4426 if (Subtarget.hasDQI()) 4427 if (auto Opc = copyPhysRegOpcode_AVX512_DQ(DestReg, SrcReg)) 4428 return Opc; 4429 if (Subtarget.hasBWI()) 4430 if (auto Opc = copyPhysRegOpcode_AVX512_BW(DestReg, SrcReg)) 4431 return Opc; 4432 if (X86::VR128XRegClass.contains(DestReg, SrcReg) || 4433 X86::VR256XRegClass.contains(DestReg, SrcReg) || 4434 X86::VR512RegClass.contains(DestReg, SrcReg)) { 4435 DestReg = get512BitSuperRegister(DestReg); 4436 SrcReg = get512BitSuperRegister(SrcReg); 4437 return X86::VMOVAPSZrr; 4438 } 4439 if (MaskRegClassContains(DestReg) && MaskRegClassContains(SrcReg)) 4440 return X86::KMOVWkk; 4441 if (MaskRegClassContains(DestReg) && GRRegClassContains(SrcReg)) { 4442 SrcReg = getX86SubSuperRegister(SrcReg, 32); 4443 return X86::KMOVWkr; 4444 } 4445 if (GRRegClassContains(DestReg) && MaskRegClassContains(SrcReg)) { 4446 DestReg = getX86SubSuperRegister(DestReg, 32); 4447 return X86::KMOVWrk; 4448 } 4449 return 0; 4450 } 4451 4452 void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB, 4453 MachineBasicBlock::iterator MI, 4454 const DebugLoc &DL, unsigned DestReg, 4455 unsigned SrcReg, bool KillSrc) const { 4456 // First deal with the normal symmetric copies. 4457 bool HasAVX = Subtarget.hasAVX(); 4458 bool HasAVX512 = Subtarget.hasAVX512(); 4459 unsigned Opc = 0; 4460 if (X86::GR64RegClass.contains(DestReg, SrcReg)) 4461 Opc = X86::MOV64rr; 4462 else if (X86::GR32RegClass.contains(DestReg, SrcReg)) 4463 Opc = X86::MOV32rr; 4464 else if (X86::GR16RegClass.contains(DestReg, SrcReg)) 4465 Opc = X86::MOV16rr; 4466 else if (X86::GR8RegClass.contains(DestReg, SrcReg)) { 4467 // Copying to or from a physical H register on x86-64 requires a NOREX 4468 // move. Otherwise use a normal move. 4469 if ((isHReg(DestReg) || isHReg(SrcReg)) && 4470 Subtarget.is64Bit()) { 4471 Opc = X86::MOV8rr_NOREX; 4472 // Both operands must be encodable without an REX prefix. 4473 assert(X86::GR8_NOREXRegClass.contains(SrcReg, DestReg) && 4474 "8-bit H register can not be copied outside GR8_NOREX"); 4475 } else 4476 Opc = X86::MOV8rr; 4477 } 4478 else if (X86::VR64RegClass.contains(DestReg, SrcReg)) 4479 Opc = X86::MMX_MOVQ64rr; 4480 else if (HasAVX512) 4481 Opc = copyPhysRegOpcode_AVX512(DestReg, SrcReg, Subtarget); 4482 else if (X86::VR128RegClass.contains(DestReg, SrcReg)) 4483 Opc = HasAVX ? X86::VMOVAPSrr : X86::MOVAPSrr; 4484 else if (X86::VR256RegClass.contains(DestReg, SrcReg)) 4485 Opc = X86::VMOVAPSYrr; 4486 if (!Opc) 4487 Opc = CopyToFromAsymmetricReg(DestReg, SrcReg, Subtarget); 4488 4489 if (Opc) { 4490 BuildMI(MBB, MI, DL, get(Opc), DestReg) 4491 .addReg(SrcReg, getKillRegState(KillSrc)); 4492 return; 4493 } 4494 4495 bool FromEFLAGS = SrcReg == X86::EFLAGS; 4496 bool ToEFLAGS = DestReg == X86::EFLAGS; 4497 int Reg = FromEFLAGS ? DestReg : SrcReg; 4498 bool is32 = X86::GR32RegClass.contains(Reg); 4499 bool is64 = X86::GR64RegClass.contains(Reg); 4500 4501 if ((FromEFLAGS || ToEFLAGS) && (is32 || is64)) { 4502 int Mov = is64 ? X86::MOV64rr : X86::MOV32rr; 4503 int Push = is64 ? X86::PUSH64r : X86::PUSH32r; 4504 int PushF = is64 ? X86::PUSHF64 : X86::PUSHF32; 4505 int Pop = is64 ? X86::POP64r : X86::POP32r; 4506 int PopF = is64 ? X86::POPF64 : X86::POPF32; 4507 int AX = is64 ? X86::RAX : X86::EAX; 4508 4509 if (!Subtarget.hasLAHFSAHF()) { 4510 assert(Subtarget.is64Bit() && 4511 "Not having LAHF/SAHF only happens on 64-bit."); 4512 // Moving EFLAGS to / from another register requires a push and a pop. 4513 // Notice that we have to adjust the stack if we don't want to clobber the 4514 // first frame index. See X86FrameLowering.cpp - usesTheStack. 4515 if (FromEFLAGS) { 4516 BuildMI(MBB, MI, DL, get(PushF)); 4517 BuildMI(MBB, MI, DL, get(Pop), DestReg); 4518 } 4519 if (ToEFLAGS) { 4520 BuildMI(MBB, MI, DL, get(Push)) 4521 .addReg(SrcReg, getKillRegState(KillSrc)); 4522 BuildMI(MBB, MI, DL, get(PopF)); 4523 } 4524 return; 4525 } 4526 4527 // The flags need to be saved, but saving EFLAGS with PUSHF/POPF is 4528 // inefficient. Instead: 4529 // - Save the overflow flag OF into AL using SETO, and restore it using a 4530 // signed 8-bit addition of AL and INT8_MAX. 4531 // - Save/restore the bottom 8 EFLAGS bits (CF, PF, AF, ZF, SF) to/from AH 4532 // using LAHF/SAHF. 4533 // - When RAX/EAX is live and isn't the destination register, make sure it 4534 // isn't clobbered by PUSH/POP'ing it before and after saving/restoring 4535 // the flags. 4536 // This approach is ~2.25x faster than using PUSHF/POPF. 4537 // 4538 // This is still somewhat inefficient because we don't know which flags are 4539 // actually live inside EFLAGS. Were we able to do a single SETcc instead of 4540 // SETO+LAHF / ADDB+SAHF the code could be 1.02x faster. 4541 // 4542 // PUSHF/POPF is also potentially incorrect because it affects other flags 4543 // such as TF/IF/DF, which LLVM doesn't model. 4544 // 4545 // Notice that we have to adjust the stack if we don't want to clobber the 4546 // first frame index. 4547 // See X86ISelLowering.cpp - X86::hasCopyImplyingStackAdjustment. 4548 4549 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4550 MachineBasicBlock::LivenessQueryResult LQR = 4551 MBB.computeRegisterLiveness(TRI, AX, MI); 4552 // We do not want to save and restore AX if we do not have to. 4553 // Moreover, if we do so whereas AX is dead, we would need to set 4554 // an undef flag on the use of AX, otherwise the verifier will 4555 // complain that we read an undef value. 4556 // We do not want to change the behavior of the machine verifier 4557 // as this is usually wrong to read an undef value. 4558 if (MachineBasicBlock::LQR_Unknown == LQR) { 4559 LivePhysRegs LPR(TRI); 4560 LPR.addLiveOuts(MBB); 4561 MachineBasicBlock::iterator I = MBB.end(); 4562 while (I != MI) { 4563 --I; 4564 LPR.stepBackward(*I); 4565 } 4566 // AX contains the top most register in the aliasing hierarchy. 4567 // It may not be live, but one of its aliases may be. 4568 for (MCRegAliasIterator AI(AX, TRI, true); 4569 AI.isValid() && LQR != MachineBasicBlock::LQR_Live; ++AI) 4570 LQR = LPR.contains(*AI) ? MachineBasicBlock::LQR_Live 4571 : MachineBasicBlock::LQR_Dead; 4572 } 4573 bool AXDead = (Reg == AX) || (MachineBasicBlock::LQR_Dead == LQR); 4574 if (!AXDead) 4575 BuildMI(MBB, MI, DL, get(Push)).addReg(AX, getKillRegState(true)); 4576 if (FromEFLAGS) { 4577 BuildMI(MBB, MI, DL, get(X86::SETOr), X86::AL); 4578 BuildMI(MBB, MI, DL, get(X86::LAHF)); 4579 BuildMI(MBB, MI, DL, get(Mov), Reg).addReg(AX); 4580 } 4581 if (ToEFLAGS) { 4582 BuildMI(MBB, MI, DL, get(Mov), AX).addReg(Reg, getKillRegState(KillSrc)); 4583 BuildMI(MBB, MI, DL, get(X86::ADD8ri), X86::AL) 4584 .addReg(X86::AL) 4585 .addImm(INT8_MAX); 4586 BuildMI(MBB, MI, DL, get(X86::SAHF)); 4587 } 4588 if (!AXDead) 4589 BuildMI(MBB, MI, DL, get(Pop), AX); 4590 return; 4591 } 4592 4593 DEBUG(dbgs() << "Cannot copy " << RI.getName(SrcReg) 4594 << " to " << RI.getName(DestReg) << '\n'); 4595 llvm_unreachable("Cannot emit physreg copy instruction"); 4596 } 4597 4598 static unsigned getLoadStoreMaskRegOpcode(const TargetRegisterClass *RC, 4599 bool load) { 4600 switch (RC->getSize()) { 4601 default: 4602 llvm_unreachable("Unknown spill size"); 4603 case 2: 4604 return load ? X86::KMOVWkm : X86::KMOVWmk; 4605 case 4: 4606 return load ? X86::KMOVDkm : X86::KMOVDmk; 4607 case 8: 4608 return load ? X86::KMOVQkm : X86::KMOVQmk; 4609 } 4610 } 4611 4612 static unsigned getLoadStoreRegOpcode(unsigned Reg, 4613 const TargetRegisterClass *RC, 4614 bool isStackAligned, 4615 const X86Subtarget &STI, 4616 bool load) { 4617 if (STI.hasAVX512()) { 4618 if (isMaskRegClass(RC)) 4619 return getLoadStoreMaskRegOpcode(RC, load); 4620 if (RC->getSize() == 4 && X86::FR32XRegClass.hasSubClassEq(RC)) 4621 return load ? X86::VMOVSSZrm : X86::VMOVSSZmr; 4622 if (RC->getSize() == 8 && X86::FR64XRegClass.hasSubClassEq(RC)) 4623 return load ? X86::VMOVSDZrm : X86::VMOVSDZmr; 4624 if (X86::VR512RegClass.hasSubClassEq(RC)) 4625 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr; 4626 } 4627 4628 bool HasAVX = STI.hasAVX(); 4629 switch (RC->getSize()) { 4630 default: 4631 llvm_unreachable("Unknown spill size"); 4632 case 1: 4633 assert(X86::GR8RegClass.hasSubClassEq(RC) && "Unknown 1-byte regclass"); 4634 if (STI.is64Bit()) 4635 // Copying to or from a physical H register on x86-64 requires a NOREX 4636 // move. Otherwise use a normal move. 4637 if (isHReg(Reg) || X86::GR8_ABCD_HRegClass.hasSubClassEq(RC)) 4638 return load ? X86::MOV8rm_NOREX : X86::MOV8mr_NOREX; 4639 return load ? X86::MOV8rm : X86::MOV8mr; 4640 case 2: 4641 assert(X86::GR16RegClass.hasSubClassEq(RC) && "Unknown 2-byte regclass"); 4642 return load ? X86::MOV16rm : X86::MOV16mr; 4643 case 4: 4644 if (X86::GR32RegClass.hasSubClassEq(RC)) 4645 return load ? X86::MOV32rm : X86::MOV32mr; 4646 if (X86::FR32RegClass.hasSubClassEq(RC)) 4647 return load ? 4648 (HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) : 4649 (HasAVX ? X86::VMOVSSmr : X86::MOVSSmr); 4650 if (X86::RFP32RegClass.hasSubClassEq(RC)) 4651 return load ? X86::LD_Fp32m : X86::ST_Fp32m; 4652 llvm_unreachable("Unknown 4-byte regclass"); 4653 case 8: 4654 if (X86::GR64RegClass.hasSubClassEq(RC)) 4655 return load ? X86::MOV64rm : X86::MOV64mr; 4656 if (X86::FR64RegClass.hasSubClassEq(RC)) 4657 return load ? 4658 (HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) : 4659 (HasAVX ? X86::VMOVSDmr : X86::MOVSDmr); 4660 if (X86::VR64RegClass.hasSubClassEq(RC)) 4661 return load ? X86::MMX_MOVQ64rm : X86::MMX_MOVQ64mr; 4662 if (X86::RFP64RegClass.hasSubClassEq(RC)) 4663 return load ? X86::LD_Fp64m : X86::ST_Fp64m; 4664 llvm_unreachable("Unknown 8-byte regclass"); 4665 case 10: 4666 assert(X86::RFP80RegClass.hasSubClassEq(RC) && "Unknown 10-byte regclass"); 4667 return load ? X86::LD_Fp80m : X86::ST_FpP80m; 4668 case 16: { 4669 assert((X86::VR128RegClass.hasSubClassEq(RC) || 4670 X86::VR128XRegClass.hasSubClassEq(RC))&& "Unknown 16-byte regclass"); 4671 // If stack is realigned we can use aligned stores. 4672 if (X86::VR128RegClass.hasSubClassEq(RC)) { 4673 if (isStackAligned) 4674 return load ? (HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) 4675 : (HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr); 4676 else 4677 return load ? (HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) 4678 : (HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr); 4679 } 4680 assert(STI.hasVLX() && "Using extended register requires VLX"); 4681 if (isStackAligned) 4682 return load ? X86::VMOVAPSZ128rm : X86::VMOVAPSZ128mr; 4683 else 4684 return load ? X86::VMOVUPSZ128rm : X86::VMOVUPSZ128mr; 4685 } 4686 case 32: 4687 assert((X86::VR256RegClass.hasSubClassEq(RC) || 4688 X86::VR256XRegClass.hasSubClassEq(RC)) && "Unknown 32-byte regclass"); 4689 // If stack is realigned we can use aligned stores. 4690 if (X86::VR256RegClass.hasSubClassEq(RC)) { 4691 if (isStackAligned) 4692 return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr; 4693 else 4694 return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr; 4695 } 4696 assert(STI.hasVLX() && "Using extended register requires VLX"); 4697 if (isStackAligned) 4698 return load ? X86::VMOVAPSZ256rm : X86::VMOVAPSZ256mr; 4699 else 4700 return load ? X86::VMOVUPSZ256rm : X86::VMOVUPSZ256mr; 4701 case 64: 4702 assert(X86::VR512RegClass.hasSubClassEq(RC) && "Unknown 64-byte regclass"); 4703 assert(STI.hasVLX() && "Using 512-bit register requires AVX512"); 4704 if (isStackAligned) 4705 return load ? X86::VMOVAPSZrm : X86::VMOVAPSZmr; 4706 else 4707 return load ? X86::VMOVUPSZrm : X86::VMOVUPSZmr; 4708 } 4709 } 4710 4711 bool X86InstrInfo::getMemOpBaseRegImmOfs(MachineInstr *MemOp, unsigned &BaseReg, 4712 int64_t &Offset, 4713 const TargetRegisterInfo *TRI) const { 4714 const MCInstrDesc &Desc = MemOp->getDesc(); 4715 int MemRefBegin = X86II::getMemoryOperandNo(Desc.TSFlags); 4716 if (MemRefBegin < 0) 4717 return false; 4718 4719 MemRefBegin += X86II::getOperandBias(Desc); 4720 4721 MachineOperand &BaseMO = MemOp->getOperand(MemRefBegin + X86::AddrBaseReg); 4722 if (!BaseMO.isReg()) // Can be an MO_FrameIndex 4723 return false; 4724 4725 BaseReg = BaseMO.getReg(); 4726 if (MemOp->getOperand(MemRefBegin + X86::AddrScaleAmt).getImm() != 1) 4727 return false; 4728 4729 if (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() != 4730 X86::NoRegister) 4731 return false; 4732 4733 const MachineOperand &DispMO = MemOp->getOperand(MemRefBegin + X86::AddrDisp); 4734 4735 // Displacement can be symbolic 4736 if (!DispMO.isImm()) 4737 return false; 4738 4739 Offset = DispMO.getImm(); 4740 4741 return (MemOp->getOperand(MemRefBegin + X86::AddrIndexReg).getReg() == 4742 X86::NoRegister); 4743 } 4744 4745 static unsigned getStoreRegOpcode(unsigned SrcReg, 4746 const TargetRegisterClass *RC, 4747 bool isStackAligned, 4748 const X86Subtarget &STI) { 4749 return getLoadStoreRegOpcode(SrcReg, RC, isStackAligned, STI, false); 4750 } 4751 4752 4753 static unsigned getLoadRegOpcode(unsigned DestReg, 4754 const TargetRegisterClass *RC, 4755 bool isStackAligned, 4756 const X86Subtarget &STI) { 4757 return getLoadStoreRegOpcode(DestReg, RC, isStackAligned, STI, true); 4758 } 4759 4760 void X86InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 4761 MachineBasicBlock::iterator MI, 4762 unsigned SrcReg, bool isKill, int FrameIdx, 4763 const TargetRegisterClass *RC, 4764 const TargetRegisterInfo *TRI) const { 4765 const MachineFunction &MF = *MBB.getParent(); 4766 assert(MF.getFrameInfo()->getObjectSize(FrameIdx) >= RC->getSize() && 4767 "Stack slot too small for store"); 4768 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16); 4769 bool isAligned = 4770 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) || 4771 RI.canRealignStack(MF); 4772 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget); 4773 DebugLoc DL = MBB.findDebugLoc(MI); 4774 addFrameReference(BuildMI(MBB, MI, DL, get(Opc)), FrameIdx) 4775 .addReg(SrcReg, getKillRegState(isKill)); 4776 } 4777 4778 void X86InstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg, 4779 bool isKill, 4780 SmallVectorImpl<MachineOperand> &Addr, 4781 const TargetRegisterClass *RC, 4782 MachineInstr::mmo_iterator MMOBegin, 4783 MachineInstr::mmo_iterator MMOEnd, 4784 SmallVectorImpl<MachineInstr*> &NewMIs) const { 4785 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16); 4786 bool isAligned = MMOBegin != MMOEnd && 4787 (*MMOBegin)->getAlignment() >= Alignment; 4788 unsigned Opc = getStoreRegOpcode(SrcReg, RC, isAligned, Subtarget); 4789 DebugLoc DL; 4790 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc)); 4791 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 4792 MIB.addOperand(Addr[i]); 4793 MIB.addReg(SrcReg, getKillRegState(isKill)); 4794 (*MIB).setMemRefs(MMOBegin, MMOEnd); 4795 NewMIs.push_back(MIB); 4796 } 4797 4798 4799 void X86InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 4800 MachineBasicBlock::iterator MI, 4801 unsigned DestReg, int FrameIdx, 4802 const TargetRegisterClass *RC, 4803 const TargetRegisterInfo *TRI) const { 4804 const MachineFunction &MF = *MBB.getParent(); 4805 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16); 4806 bool isAligned = 4807 (Subtarget.getFrameLowering()->getStackAlignment() >= Alignment) || 4808 RI.canRealignStack(MF); 4809 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget); 4810 DebugLoc DL = MBB.findDebugLoc(MI); 4811 addFrameReference(BuildMI(MBB, MI, DL, get(Opc), DestReg), FrameIdx); 4812 } 4813 4814 void X86InstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg, 4815 SmallVectorImpl<MachineOperand> &Addr, 4816 const TargetRegisterClass *RC, 4817 MachineInstr::mmo_iterator MMOBegin, 4818 MachineInstr::mmo_iterator MMOEnd, 4819 SmallVectorImpl<MachineInstr*> &NewMIs) const { 4820 unsigned Alignment = std::max<uint32_t>(RC->getSize(), 16); 4821 bool isAligned = MMOBegin != MMOEnd && 4822 (*MMOBegin)->getAlignment() >= Alignment; 4823 unsigned Opc = getLoadRegOpcode(DestReg, RC, isAligned, Subtarget); 4824 DebugLoc DL; 4825 MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), DestReg); 4826 for (unsigned i = 0, e = Addr.size(); i != e; ++i) 4827 MIB.addOperand(Addr[i]); 4828 (*MIB).setMemRefs(MMOBegin, MMOEnd); 4829 NewMIs.push_back(MIB); 4830 } 4831 4832 bool X86InstrInfo:: 4833 analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2, 4834 int &CmpMask, int &CmpValue) const { 4835 switch (MI->getOpcode()) { 4836 default: break; 4837 case X86::CMP64ri32: 4838 case X86::CMP64ri8: 4839 case X86::CMP32ri: 4840 case X86::CMP32ri8: 4841 case X86::CMP16ri: 4842 case X86::CMP16ri8: 4843 case X86::CMP8ri: 4844 SrcReg = MI->getOperand(0).getReg(); 4845 SrcReg2 = 0; 4846 CmpMask = ~0; 4847 CmpValue = MI->getOperand(1).getImm(); 4848 return true; 4849 // A SUB can be used to perform comparison. 4850 case X86::SUB64rm: 4851 case X86::SUB32rm: 4852 case X86::SUB16rm: 4853 case X86::SUB8rm: 4854 SrcReg = MI->getOperand(1).getReg(); 4855 SrcReg2 = 0; 4856 CmpMask = ~0; 4857 CmpValue = 0; 4858 return true; 4859 case X86::SUB64rr: 4860 case X86::SUB32rr: 4861 case X86::SUB16rr: 4862 case X86::SUB8rr: 4863 SrcReg = MI->getOperand(1).getReg(); 4864 SrcReg2 = MI->getOperand(2).getReg(); 4865 CmpMask = ~0; 4866 CmpValue = 0; 4867 return true; 4868 case X86::SUB64ri32: 4869 case X86::SUB64ri8: 4870 case X86::SUB32ri: 4871 case X86::SUB32ri8: 4872 case X86::SUB16ri: 4873 case X86::SUB16ri8: 4874 case X86::SUB8ri: 4875 SrcReg = MI->getOperand(1).getReg(); 4876 SrcReg2 = 0; 4877 CmpMask = ~0; 4878 CmpValue = MI->getOperand(2).getImm(); 4879 return true; 4880 case X86::CMP64rr: 4881 case X86::CMP32rr: 4882 case X86::CMP16rr: 4883 case X86::CMP8rr: 4884 SrcReg = MI->getOperand(0).getReg(); 4885 SrcReg2 = MI->getOperand(1).getReg(); 4886 CmpMask = ~0; 4887 CmpValue = 0; 4888 return true; 4889 case X86::TEST8rr: 4890 case X86::TEST16rr: 4891 case X86::TEST32rr: 4892 case X86::TEST64rr: 4893 SrcReg = MI->getOperand(0).getReg(); 4894 if (MI->getOperand(1).getReg() != SrcReg) return false; 4895 // Compare against zero. 4896 SrcReg2 = 0; 4897 CmpMask = ~0; 4898 CmpValue = 0; 4899 return true; 4900 } 4901 return false; 4902 } 4903 4904 /// Check whether the first instruction, whose only 4905 /// purpose is to update flags, can be made redundant. 4906 /// CMPrr can be made redundant by SUBrr if the operands are the same. 4907 /// This function can be extended later on. 4908 /// SrcReg, SrcRegs: register operands for FlagI. 4909 /// ImmValue: immediate for FlagI if it takes an immediate. 4910 inline static bool isRedundantFlagInstr(MachineInstr *FlagI, unsigned SrcReg, 4911 unsigned SrcReg2, int ImmValue, 4912 MachineInstr *OI) { 4913 if (((FlagI->getOpcode() == X86::CMP64rr && 4914 OI->getOpcode() == X86::SUB64rr) || 4915 (FlagI->getOpcode() == X86::CMP32rr && 4916 OI->getOpcode() == X86::SUB32rr)|| 4917 (FlagI->getOpcode() == X86::CMP16rr && 4918 OI->getOpcode() == X86::SUB16rr)|| 4919 (FlagI->getOpcode() == X86::CMP8rr && 4920 OI->getOpcode() == X86::SUB8rr)) && 4921 ((OI->getOperand(1).getReg() == SrcReg && 4922 OI->getOperand(2).getReg() == SrcReg2) || 4923 (OI->getOperand(1).getReg() == SrcReg2 && 4924 OI->getOperand(2).getReg() == SrcReg))) 4925 return true; 4926 4927 if (((FlagI->getOpcode() == X86::CMP64ri32 && 4928 OI->getOpcode() == X86::SUB64ri32) || 4929 (FlagI->getOpcode() == X86::CMP64ri8 && 4930 OI->getOpcode() == X86::SUB64ri8) || 4931 (FlagI->getOpcode() == X86::CMP32ri && 4932 OI->getOpcode() == X86::SUB32ri) || 4933 (FlagI->getOpcode() == X86::CMP32ri8 && 4934 OI->getOpcode() == X86::SUB32ri8) || 4935 (FlagI->getOpcode() == X86::CMP16ri && 4936 OI->getOpcode() == X86::SUB16ri) || 4937 (FlagI->getOpcode() == X86::CMP16ri8 && 4938 OI->getOpcode() == X86::SUB16ri8) || 4939 (FlagI->getOpcode() == X86::CMP8ri && 4940 OI->getOpcode() == X86::SUB8ri)) && 4941 OI->getOperand(1).getReg() == SrcReg && 4942 OI->getOperand(2).getImm() == ImmValue) 4943 return true; 4944 return false; 4945 } 4946 4947 /// Check whether the definition can be converted 4948 /// to remove a comparison against zero. 4949 inline static bool isDefConvertible(MachineInstr *MI) { 4950 switch (MI->getOpcode()) { 4951 default: return false; 4952 4953 // The shift instructions only modify ZF if their shift count is non-zero. 4954 // N.B.: The processor truncates the shift count depending on the encoding. 4955 case X86::SAR8ri: case X86::SAR16ri: case X86::SAR32ri:case X86::SAR64ri: 4956 case X86::SHR8ri: case X86::SHR16ri: case X86::SHR32ri:case X86::SHR64ri: 4957 return getTruncatedShiftCount(MI, 2) != 0; 4958 4959 // Some left shift instructions can be turned into LEA instructions but only 4960 // if their flags aren't used. Avoid transforming such instructions. 4961 case X86::SHL8ri: case X86::SHL16ri: case X86::SHL32ri:case X86::SHL64ri:{ 4962 unsigned ShAmt = getTruncatedShiftCount(MI, 2); 4963 if (isTruncatedShiftCountForLEA(ShAmt)) return false; 4964 return ShAmt != 0; 4965 } 4966 4967 case X86::SHRD16rri8:case X86::SHRD32rri8:case X86::SHRD64rri8: 4968 case X86::SHLD16rri8:case X86::SHLD32rri8:case X86::SHLD64rri8: 4969 return getTruncatedShiftCount(MI, 3) != 0; 4970 4971 case X86::SUB64ri32: case X86::SUB64ri8: case X86::SUB32ri: 4972 case X86::SUB32ri8: case X86::SUB16ri: case X86::SUB16ri8: 4973 case X86::SUB8ri: case X86::SUB64rr: case X86::SUB32rr: 4974 case X86::SUB16rr: case X86::SUB8rr: case X86::SUB64rm: 4975 case X86::SUB32rm: case X86::SUB16rm: case X86::SUB8rm: 4976 case X86::DEC64r: case X86::DEC32r: case X86::DEC16r: case X86::DEC8r: 4977 case X86::ADD64ri32: case X86::ADD64ri8: case X86::ADD32ri: 4978 case X86::ADD32ri8: case X86::ADD16ri: case X86::ADD16ri8: 4979 case X86::ADD8ri: case X86::ADD64rr: case X86::ADD32rr: 4980 case X86::ADD16rr: case X86::ADD8rr: case X86::ADD64rm: 4981 case X86::ADD32rm: case X86::ADD16rm: case X86::ADD8rm: 4982 case X86::INC64r: case X86::INC32r: case X86::INC16r: case X86::INC8r: 4983 case X86::AND64ri32: case X86::AND64ri8: case X86::AND32ri: 4984 case X86::AND32ri8: case X86::AND16ri: case X86::AND16ri8: 4985 case X86::AND8ri: case X86::AND64rr: case X86::AND32rr: 4986 case X86::AND16rr: case X86::AND8rr: case X86::AND64rm: 4987 case X86::AND32rm: case X86::AND16rm: case X86::AND8rm: 4988 case X86::XOR64ri32: case X86::XOR64ri8: case X86::XOR32ri: 4989 case X86::XOR32ri8: case X86::XOR16ri: case X86::XOR16ri8: 4990 case X86::XOR8ri: case X86::XOR64rr: case X86::XOR32rr: 4991 case X86::XOR16rr: case X86::XOR8rr: case X86::XOR64rm: 4992 case X86::XOR32rm: case X86::XOR16rm: case X86::XOR8rm: 4993 case X86::OR64ri32: case X86::OR64ri8: case X86::OR32ri: 4994 case X86::OR32ri8: case X86::OR16ri: case X86::OR16ri8: 4995 case X86::OR8ri: case X86::OR64rr: case X86::OR32rr: 4996 case X86::OR16rr: case X86::OR8rr: case X86::OR64rm: 4997 case X86::OR32rm: case X86::OR16rm: case X86::OR8rm: 4998 case X86::NEG8r: case X86::NEG16r: case X86::NEG32r: case X86::NEG64r: 4999 case X86::SAR8r1: case X86::SAR16r1: case X86::SAR32r1:case X86::SAR64r1: 5000 case X86::SHR8r1: case X86::SHR16r1: case X86::SHR32r1:case X86::SHR64r1: 5001 case X86::SHL8r1: case X86::SHL16r1: case X86::SHL32r1:case X86::SHL64r1: 5002 case X86::ADC32ri: case X86::ADC32ri8: 5003 case X86::ADC32rr: case X86::ADC64ri32: 5004 case X86::ADC64ri8: case X86::ADC64rr: 5005 case X86::SBB32ri: case X86::SBB32ri8: 5006 case X86::SBB32rr: case X86::SBB64ri32: 5007 case X86::SBB64ri8: case X86::SBB64rr: 5008 case X86::ANDN32rr: case X86::ANDN32rm: 5009 case X86::ANDN64rr: case X86::ANDN64rm: 5010 case X86::BEXTR32rr: case X86::BEXTR64rr: 5011 case X86::BEXTR32rm: case X86::BEXTR64rm: 5012 case X86::BLSI32rr: case X86::BLSI32rm: 5013 case X86::BLSI64rr: case X86::BLSI64rm: 5014 case X86::BLSMSK32rr:case X86::BLSMSK32rm: 5015 case X86::BLSMSK64rr:case X86::BLSMSK64rm: 5016 case X86::BLSR32rr: case X86::BLSR32rm: 5017 case X86::BLSR64rr: case X86::BLSR64rm: 5018 case X86::BZHI32rr: case X86::BZHI32rm: 5019 case X86::BZHI64rr: case X86::BZHI64rm: 5020 case X86::LZCNT16rr: case X86::LZCNT16rm: 5021 case X86::LZCNT32rr: case X86::LZCNT32rm: 5022 case X86::LZCNT64rr: case X86::LZCNT64rm: 5023 case X86::POPCNT16rr:case X86::POPCNT16rm: 5024 case X86::POPCNT32rr:case X86::POPCNT32rm: 5025 case X86::POPCNT64rr:case X86::POPCNT64rm: 5026 case X86::TZCNT16rr: case X86::TZCNT16rm: 5027 case X86::TZCNT32rr: case X86::TZCNT32rm: 5028 case X86::TZCNT64rr: case X86::TZCNT64rm: 5029 return true; 5030 } 5031 } 5032 5033 /// Check whether the use can be converted to remove a comparison against zero. 5034 static X86::CondCode isUseDefConvertible(MachineInstr *MI) { 5035 switch (MI->getOpcode()) { 5036 default: return X86::COND_INVALID; 5037 case X86::LZCNT16rr: case X86::LZCNT16rm: 5038 case X86::LZCNT32rr: case X86::LZCNT32rm: 5039 case X86::LZCNT64rr: case X86::LZCNT64rm: 5040 return X86::COND_B; 5041 case X86::POPCNT16rr:case X86::POPCNT16rm: 5042 case X86::POPCNT32rr:case X86::POPCNT32rm: 5043 case X86::POPCNT64rr:case X86::POPCNT64rm: 5044 return X86::COND_E; 5045 case X86::TZCNT16rr: case X86::TZCNT16rm: 5046 case X86::TZCNT32rr: case X86::TZCNT32rm: 5047 case X86::TZCNT64rr: case X86::TZCNT64rm: 5048 return X86::COND_B; 5049 } 5050 } 5051 5052 /// Check if there exists an earlier instruction that 5053 /// operates on the same source operands and sets flags in the same way as 5054 /// Compare; remove Compare if possible. 5055 bool X86InstrInfo:: 5056 optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2, 5057 int CmpMask, int CmpValue, 5058 const MachineRegisterInfo *MRI) const { 5059 // Check whether we can replace SUB with CMP. 5060 unsigned NewOpcode = 0; 5061 switch (CmpInstr->getOpcode()) { 5062 default: break; 5063 case X86::SUB64ri32: 5064 case X86::SUB64ri8: 5065 case X86::SUB32ri: 5066 case X86::SUB32ri8: 5067 case X86::SUB16ri: 5068 case X86::SUB16ri8: 5069 case X86::SUB8ri: 5070 case X86::SUB64rm: 5071 case X86::SUB32rm: 5072 case X86::SUB16rm: 5073 case X86::SUB8rm: 5074 case X86::SUB64rr: 5075 case X86::SUB32rr: 5076 case X86::SUB16rr: 5077 case X86::SUB8rr: { 5078 if (!MRI->use_nodbg_empty(CmpInstr->getOperand(0).getReg())) 5079 return false; 5080 // There is no use of the destination register, we can replace SUB with CMP. 5081 switch (CmpInstr->getOpcode()) { 5082 default: llvm_unreachable("Unreachable!"); 5083 case X86::SUB64rm: NewOpcode = X86::CMP64rm; break; 5084 case X86::SUB32rm: NewOpcode = X86::CMP32rm; break; 5085 case X86::SUB16rm: NewOpcode = X86::CMP16rm; break; 5086 case X86::SUB8rm: NewOpcode = X86::CMP8rm; break; 5087 case X86::SUB64rr: NewOpcode = X86::CMP64rr; break; 5088 case X86::SUB32rr: NewOpcode = X86::CMP32rr; break; 5089 case X86::SUB16rr: NewOpcode = X86::CMP16rr; break; 5090 case X86::SUB8rr: NewOpcode = X86::CMP8rr; break; 5091 case X86::SUB64ri32: NewOpcode = X86::CMP64ri32; break; 5092 case X86::SUB64ri8: NewOpcode = X86::CMP64ri8; break; 5093 case X86::SUB32ri: NewOpcode = X86::CMP32ri; break; 5094 case X86::SUB32ri8: NewOpcode = X86::CMP32ri8; break; 5095 case X86::SUB16ri: NewOpcode = X86::CMP16ri; break; 5096 case X86::SUB16ri8: NewOpcode = X86::CMP16ri8; break; 5097 case X86::SUB8ri: NewOpcode = X86::CMP8ri; break; 5098 } 5099 CmpInstr->setDesc(get(NewOpcode)); 5100 CmpInstr->RemoveOperand(0); 5101 // Fall through to optimize Cmp if Cmp is CMPrr or CMPri. 5102 if (NewOpcode == X86::CMP64rm || NewOpcode == X86::CMP32rm || 5103 NewOpcode == X86::CMP16rm || NewOpcode == X86::CMP8rm) 5104 return false; 5105 } 5106 } 5107 5108 // Get the unique definition of SrcReg. 5109 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 5110 if (!MI) return false; 5111 5112 // CmpInstr is the first instruction of the BB. 5113 MachineBasicBlock::iterator I = CmpInstr, Def = MI; 5114 5115 // If we are comparing against zero, check whether we can use MI to update 5116 // EFLAGS. If MI is not in the same BB as CmpInstr, do not optimize. 5117 bool IsCmpZero = (SrcReg2 == 0 && CmpValue == 0); 5118 if (IsCmpZero && MI->getParent() != CmpInstr->getParent()) 5119 return false; 5120 5121 // If we have a use of the source register between the def and our compare 5122 // instruction we can eliminate the compare iff the use sets EFLAGS in the 5123 // right way. 5124 bool ShouldUpdateCC = false; 5125 X86::CondCode NewCC = X86::COND_INVALID; 5126 if (IsCmpZero && !isDefConvertible(MI)) { 5127 // Scan forward from the use until we hit the use we're looking for or the 5128 // compare instruction. 5129 for (MachineBasicBlock::iterator J = MI;; ++J) { 5130 // Do we have a convertible instruction? 5131 NewCC = isUseDefConvertible(J); 5132 if (NewCC != X86::COND_INVALID && J->getOperand(1).isReg() && 5133 J->getOperand(1).getReg() == SrcReg) { 5134 assert(J->definesRegister(X86::EFLAGS) && "Must be an EFLAGS def!"); 5135 ShouldUpdateCC = true; // Update CC later on. 5136 // This is not a def of SrcReg, but still a def of EFLAGS. Keep going 5137 // with the new def. 5138 MI = Def = J; 5139 break; 5140 } 5141 5142 if (J == I) 5143 return false; 5144 } 5145 } 5146 5147 // We are searching for an earlier instruction that can make CmpInstr 5148 // redundant and that instruction will be saved in Sub. 5149 MachineInstr *Sub = nullptr; 5150 const TargetRegisterInfo *TRI = &getRegisterInfo(); 5151 5152 // We iterate backward, starting from the instruction before CmpInstr and 5153 // stop when reaching the definition of a source register or done with the BB. 5154 // RI points to the instruction before CmpInstr. 5155 // If the definition is in this basic block, RE points to the definition; 5156 // otherwise, RE is the rend of the basic block. 5157 MachineBasicBlock::reverse_iterator 5158 RI = MachineBasicBlock::reverse_iterator(I), 5159 RE = CmpInstr->getParent() == MI->getParent() ? 5160 MachineBasicBlock::reverse_iterator(++Def) /* points to MI */ : 5161 CmpInstr->getParent()->rend(); 5162 MachineInstr *Movr0Inst = nullptr; 5163 for (; RI != RE; ++RI) { 5164 MachineInstr *Instr = &*RI; 5165 // Check whether CmpInstr can be made redundant by the current instruction. 5166 if (!IsCmpZero && 5167 isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, Instr)) { 5168 Sub = Instr; 5169 break; 5170 } 5171 5172 if (Instr->modifiesRegister(X86::EFLAGS, TRI) || 5173 Instr->readsRegister(X86::EFLAGS, TRI)) { 5174 // This instruction modifies or uses EFLAGS. 5175 5176 // MOV32r0 etc. are implemented with xor which clobbers condition code. 5177 // They are safe to move up, if the definition to EFLAGS is dead and 5178 // earlier instructions do not read or write EFLAGS. 5179 if (!Movr0Inst && Instr->getOpcode() == X86::MOV32r0 && 5180 Instr->registerDefIsDead(X86::EFLAGS, TRI)) { 5181 Movr0Inst = Instr; 5182 continue; 5183 } 5184 5185 // We can't remove CmpInstr. 5186 return false; 5187 } 5188 } 5189 5190 // Return false if no candidates exist. 5191 if (!IsCmpZero && !Sub) 5192 return false; 5193 5194 bool IsSwapped = (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 5195 Sub->getOperand(2).getReg() == SrcReg); 5196 5197 // Scan forward from the instruction after CmpInstr for uses of EFLAGS. 5198 // It is safe to remove CmpInstr if EFLAGS is redefined or killed. 5199 // If we are done with the basic block, we need to check whether EFLAGS is 5200 // live-out. 5201 bool IsSafe = false; 5202 SmallVector<std::pair<MachineInstr*, unsigned /*NewOpc*/>, 4> OpsToUpdate; 5203 MachineBasicBlock::iterator E = CmpInstr->getParent()->end(); 5204 for (++I; I != E; ++I) { 5205 const MachineInstr &Instr = *I; 5206 bool ModifyEFLAGS = Instr.modifiesRegister(X86::EFLAGS, TRI); 5207 bool UseEFLAGS = Instr.readsRegister(X86::EFLAGS, TRI); 5208 // We should check the usage if this instruction uses and updates EFLAGS. 5209 if (!UseEFLAGS && ModifyEFLAGS) { 5210 // It is safe to remove CmpInstr if EFLAGS is updated again. 5211 IsSafe = true; 5212 break; 5213 } 5214 if (!UseEFLAGS && !ModifyEFLAGS) 5215 continue; 5216 5217 // EFLAGS is used by this instruction. 5218 X86::CondCode OldCC = X86::COND_INVALID; 5219 bool OpcIsSET = false; 5220 if (IsCmpZero || IsSwapped) { 5221 // We decode the condition code from opcode. 5222 if (Instr.isBranch()) 5223 OldCC = getCondFromBranchOpc(Instr.getOpcode()); 5224 else { 5225 OldCC = getCondFromSETOpc(Instr.getOpcode()); 5226 if (OldCC != X86::COND_INVALID) 5227 OpcIsSET = true; 5228 else 5229 OldCC = X86::getCondFromCMovOpc(Instr.getOpcode()); 5230 } 5231 if (OldCC == X86::COND_INVALID) return false; 5232 } 5233 if (IsCmpZero) { 5234 switch (OldCC) { 5235 default: break; 5236 case X86::COND_A: case X86::COND_AE: 5237 case X86::COND_B: case X86::COND_BE: 5238 case X86::COND_G: case X86::COND_GE: 5239 case X86::COND_L: case X86::COND_LE: 5240 case X86::COND_O: case X86::COND_NO: 5241 // CF and OF are used, we can't perform this optimization. 5242 return false; 5243 } 5244 5245 // If we're updating the condition code check if we have to reverse the 5246 // condition. 5247 if (ShouldUpdateCC) 5248 switch (OldCC) { 5249 default: 5250 return false; 5251 case X86::COND_E: 5252 break; 5253 case X86::COND_NE: 5254 NewCC = GetOppositeBranchCondition(NewCC); 5255 break; 5256 } 5257 } else if (IsSwapped) { 5258 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code needs 5259 // to be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 5260 // We swap the condition code and synthesize the new opcode. 5261 NewCC = getSwappedCondition(OldCC); 5262 if (NewCC == X86::COND_INVALID) return false; 5263 } 5264 5265 if ((ShouldUpdateCC || IsSwapped) && NewCC != OldCC) { 5266 // Synthesize the new opcode. 5267 bool HasMemoryOperand = Instr.hasOneMemOperand(); 5268 unsigned NewOpc; 5269 if (Instr.isBranch()) 5270 NewOpc = GetCondBranchFromCond(NewCC); 5271 else if(OpcIsSET) 5272 NewOpc = getSETFromCond(NewCC, HasMemoryOperand); 5273 else { 5274 unsigned DstReg = Instr.getOperand(0).getReg(); 5275 NewOpc = getCMovFromCond(NewCC, MRI->getRegClass(DstReg)->getSize(), 5276 HasMemoryOperand); 5277 } 5278 5279 // Push the MachineInstr to OpsToUpdate. 5280 // If it is safe to remove CmpInstr, the condition code of these 5281 // instructions will be modified. 5282 OpsToUpdate.push_back(std::make_pair(&*I, NewOpc)); 5283 } 5284 if (ModifyEFLAGS || Instr.killsRegister(X86::EFLAGS, TRI)) { 5285 // It is safe to remove CmpInstr if EFLAGS is updated again or killed. 5286 IsSafe = true; 5287 break; 5288 } 5289 } 5290 5291 // If EFLAGS is not killed nor re-defined, we should check whether it is 5292 // live-out. If it is live-out, do not optimize. 5293 if ((IsCmpZero || IsSwapped) && !IsSafe) { 5294 MachineBasicBlock *MBB = CmpInstr->getParent(); 5295 for (MachineBasicBlock *Successor : MBB->successors()) 5296 if (Successor->isLiveIn(X86::EFLAGS)) 5297 return false; 5298 } 5299 5300 // The instruction to be updated is either Sub or MI. 5301 Sub = IsCmpZero ? MI : Sub; 5302 // Move Movr0Inst to the appropriate place before Sub. 5303 if (Movr0Inst) { 5304 // Look backwards until we find a def that doesn't use the current EFLAGS. 5305 Def = Sub; 5306 MachineBasicBlock::reverse_iterator 5307 InsertI = MachineBasicBlock::reverse_iterator(++Def), 5308 InsertE = Sub->getParent()->rend(); 5309 for (; InsertI != InsertE; ++InsertI) { 5310 MachineInstr *Instr = &*InsertI; 5311 if (!Instr->readsRegister(X86::EFLAGS, TRI) && 5312 Instr->modifiesRegister(X86::EFLAGS, TRI)) { 5313 Sub->getParent()->remove(Movr0Inst); 5314 Instr->getParent()->insert(MachineBasicBlock::iterator(Instr), 5315 Movr0Inst); 5316 break; 5317 } 5318 } 5319 if (InsertI == InsertE) 5320 return false; 5321 } 5322 5323 // Make sure Sub instruction defines EFLAGS and mark the def live. 5324 unsigned i = 0, e = Sub->getNumOperands(); 5325 for (; i != e; ++i) { 5326 MachineOperand &MO = Sub->getOperand(i); 5327 if (MO.isReg() && MO.isDef() && MO.getReg() == X86::EFLAGS) { 5328 MO.setIsDead(false); 5329 break; 5330 } 5331 } 5332 assert(i != e && "Unable to locate a def EFLAGS operand"); 5333 5334 CmpInstr->eraseFromParent(); 5335 5336 // Modify the condition code of instructions in OpsToUpdate. 5337 for (auto &Op : OpsToUpdate) 5338 Op.first->setDesc(get(Op.second)); 5339 return true; 5340 } 5341 5342 /// Try to remove the load by folding it to a register 5343 /// operand at the use. We fold the load instructions if load defines a virtual 5344 /// register, the virtual register is used once in the same BB, and the 5345 /// instructions in-between do not load or store, and have no side effects. 5346 MachineInstr *X86InstrInfo::optimizeLoadInstr(MachineInstr *MI, 5347 const MachineRegisterInfo *MRI, 5348 unsigned &FoldAsLoadDefReg, 5349 MachineInstr *&DefMI) const { 5350 if (FoldAsLoadDefReg == 0) 5351 return nullptr; 5352 // To be conservative, if there exists another load, clear the load candidate. 5353 if (MI->mayLoad()) { 5354 FoldAsLoadDefReg = 0; 5355 return nullptr; 5356 } 5357 5358 // Check whether we can move DefMI here. 5359 DefMI = MRI->getVRegDef(FoldAsLoadDefReg); 5360 assert(DefMI); 5361 bool SawStore = false; 5362 if (!DefMI->isSafeToMove(nullptr, SawStore)) 5363 return nullptr; 5364 5365 // Collect information about virtual register operands of MI. 5366 unsigned SrcOperandId = 0; 5367 bool FoundSrcOperand = false; 5368 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { 5369 MachineOperand &MO = MI->getOperand(i); 5370 if (!MO.isReg()) 5371 continue; 5372 unsigned Reg = MO.getReg(); 5373 if (Reg != FoldAsLoadDefReg) 5374 continue; 5375 // Do not fold if we have a subreg use or a def or multiple uses. 5376 if (MO.getSubReg() || MO.isDef() || FoundSrcOperand) 5377 return nullptr; 5378 5379 SrcOperandId = i; 5380 FoundSrcOperand = true; 5381 } 5382 if (!FoundSrcOperand) 5383 return nullptr; 5384 5385 // Check whether we can fold the def into SrcOperandId. 5386 if (MachineInstr *FoldMI = foldMemoryOperand(MI, SrcOperandId, DefMI)) { 5387 FoldAsLoadDefReg = 0; 5388 return FoldMI; 5389 } 5390 5391 return nullptr; 5392 } 5393 5394 /// Expand a single-def pseudo instruction to a two-addr 5395 /// instruction with two undef reads of the register being defined. 5396 /// This is used for mapping: 5397 /// %xmm4 = V_SET0 5398 /// to: 5399 /// %xmm4 = PXORrr %xmm4<undef>, %xmm4<undef> 5400 /// 5401 static bool Expand2AddrUndef(MachineInstrBuilder &MIB, 5402 const MCInstrDesc &Desc) { 5403 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction."); 5404 unsigned Reg = MIB->getOperand(0).getReg(); 5405 MIB->setDesc(Desc); 5406 5407 // MachineInstr::addOperand() will insert explicit operands before any 5408 // implicit operands. 5409 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 5410 // But we don't trust that. 5411 assert(MIB->getOperand(1).getReg() == Reg && 5412 MIB->getOperand(2).getReg() == Reg && "Misplaced operand"); 5413 return true; 5414 } 5415 5416 /// Expand a single-def pseudo instruction to a two-addr 5417 /// instruction with two %k0 reads. 5418 /// This is used for mapping: 5419 /// %k4 = K_SET1 5420 /// to: 5421 /// %k4 = KXNORrr %k0, %k0 5422 static bool Expand2AddrKreg(MachineInstrBuilder &MIB, 5423 const MCInstrDesc &Desc, unsigned Reg) { 5424 assert(Desc.getNumOperands() == 3 && "Expected two-addr instruction."); 5425 MIB->setDesc(Desc); 5426 MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 5427 return true; 5428 } 5429 5430 static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII, 5431 bool MinusOne) { 5432 MachineBasicBlock &MBB = *MIB->getParent(); 5433 DebugLoc DL = MIB->getDebugLoc(); 5434 unsigned Reg = MIB->getOperand(0).getReg(); 5435 5436 // Insert the XOR. 5437 BuildMI(MBB, MIB.getInstr(), DL, TII.get(X86::XOR32rr), Reg) 5438 .addReg(Reg, RegState::Undef) 5439 .addReg(Reg, RegState::Undef); 5440 5441 // Turn the pseudo into an INC or DEC. 5442 MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r)); 5443 MIB.addReg(Reg); 5444 5445 return true; 5446 } 5447 5448 bool X86InstrInfo::ExpandMOVImmSExti8(MachineInstrBuilder &MIB) const { 5449 MachineBasicBlock &MBB = *MIB->getParent(); 5450 DebugLoc DL = MIB->getDebugLoc(); 5451 int64_t Imm = MIB->getOperand(1).getImm(); 5452 assert(Imm != 0 && "Using push/pop for 0 is not efficient."); 5453 MachineBasicBlock::iterator I = MIB.getInstr(); 5454 5455 int StackAdjustment; 5456 5457 if (Subtarget.is64Bit()) { 5458 assert(MIB->getOpcode() == X86::MOV64ImmSExti8 || 5459 MIB->getOpcode() == X86::MOV32ImmSExti8); 5460 5461 // Can't use push/pop lowering if the function might write to the red zone. 5462 X86MachineFunctionInfo *X86FI = 5463 MBB.getParent()->getInfo<X86MachineFunctionInfo>(); 5464 if (X86FI->getUsesRedZone()) { 5465 MIB->setDesc(get(MIB->getOpcode() == X86::MOV32ImmSExti8 ? X86::MOV32ri 5466 : X86::MOV64ri)); 5467 return true; 5468 } 5469 5470 // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and 5471 // widen the register if necessary. 5472 StackAdjustment = 8; 5473 BuildMI(MBB, I, DL, get(X86::PUSH64i8)).addImm(Imm); 5474 MIB->setDesc(get(X86::POP64r)); 5475 MIB->getOperand(0) 5476 .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64)); 5477 } else { 5478 assert(MIB->getOpcode() == X86::MOV32ImmSExti8); 5479 StackAdjustment = 4; 5480 BuildMI(MBB, I, DL, get(X86::PUSH32i8)).addImm(Imm); 5481 MIB->setDesc(get(X86::POP32r)); 5482 } 5483 5484 // Build CFI if necessary. 5485 MachineFunction &MF = *MBB.getParent(); 5486 const X86FrameLowering *TFL = Subtarget.getFrameLowering(); 5487 bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI(); 5488 bool NeedsDwarfCFI = 5489 !IsWin64Prologue && 5490 (MF.getMMI().hasDebugInfo() || MF.getFunction()->needsUnwindTableEntry()); 5491 bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI; 5492 if (EmitCFI) { 5493 TFL->BuildCFI(MBB, I, DL, 5494 MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment)); 5495 TFL->BuildCFI(MBB, std::next(I), DL, 5496 MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment)); 5497 } 5498 5499 return true; 5500 } 5501 5502 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different 5503 // code sequence is needed for other targets. 5504 static void expandLoadStackGuard(MachineInstrBuilder &MIB, 5505 const TargetInstrInfo &TII) { 5506 MachineBasicBlock &MBB = *MIB->getParent(); 5507 DebugLoc DL = MIB->getDebugLoc(); 5508 unsigned Reg = MIB->getOperand(0).getReg(); 5509 const GlobalValue *GV = 5510 cast<GlobalValue>((*MIB->memoperands_begin())->getValue()); 5511 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant; 5512 MachineMemOperand *MMO = MBB.getParent()->getMachineMemOperand( 5513 MachinePointerInfo::getGOT(*MBB.getParent()), Flag, 8, 8); 5514 MachineBasicBlock::iterator I = MIB.getInstr(); 5515 5516 BuildMI(MBB, I, DL, TII.get(X86::MOV64rm), Reg).addReg(X86::RIP).addImm(1) 5517 .addReg(0).addGlobalAddress(GV, 0, X86II::MO_GOTPCREL).addReg(0) 5518 .addMemOperand(MMO); 5519 MIB->setDebugLoc(DL); 5520 MIB->setDesc(TII.get(X86::MOV64rm)); 5521 MIB.addReg(Reg, RegState::Kill).addImm(1).addReg(0).addImm(0).addReg(0); 5522 } 5523 5524 bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const { 5525 bool HasAVX = Subtarget.hasAVX(); 5526 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 5527 switch (MI->getOpcode()) { 5528 case X86::MOV32r0: 5529 return Expand2AddrUndef(MIB, get(X86::XOR32rr)); 5530 case X86::MOV32r1: 5531 return expandMOV32r1(MIB, *this, /*MinusOne=*/ false); 5532 case X86::MOV32r_1: 5533 return expandMOV32r1(MIB, *this, /*MinusOne=*/ true); 5534 case X86::MOV32ImmSExti8: 5535 case X86::MOV64ImmSExti8: 5536 return ExpandMOVImmSExti8(MIB); 5537 case X86::SETB_C8r: 5538 return Expand2AddrUndef(MIB, get(X86::SBB8rr)); 5539 case X86::SETB_C16r: 5540 return Expand2AddrUndef(MIB, get(X86::SBB16rr)); 5541 case X86::SETB_C32r: 5542 return Expand2AddrUndef(MIB, get(X86::SBB32rr)); 5543 case X86::SETB_C64r: 5544 return Expand2AddrUndef(MIB, get(X86::SBB64rr)); 5545 case X86::V_SET0: 5546 case X86::FsFLD0SS: 5547 case X86::FsFLD0SD: 5548 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr)); 5549 case X86::AVX_SET0: 5550 assert(HasAVX && "AVX not supported"); 5551 return Expand2AddrUndef(MIB, get(X86::VXORPSYrr)); 5552 case X86::AVX512_128_SET0: 5553 return Expand2AddrUndef(MIB, get(X86::VPXORDZ128rr)); 5554 case X86::AVX512_256_SET0: 5555 return Expand2AddrUndef(MIB, get(X86::VPXORDZ256rr)); 5556 case X86::AVX512_512_SET0: 5557 return Expand2AddrUndef(MIB, get(X86::VPXORDZrr)); 5558 case X86::V_SETALLONES: 5559 return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr)); 5560 case X86::AVX2_SETALLONES: 5561 return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr)); 5562 case X86::TEST8ri_NOREX: 5563 MI->setDesc(get(X86::TEST8ri)); 5564 return true; 5565 case X86::MOV32ri64: 5566 MI->setDesc(get(X86::MOV32ri)); 5567 return true; 5568 5569 // KNL does not recognize dependency-breaking idioms for mask registers, 5570 // so kxnor %k1, %k1, %k2 has a RAW dependence on %k1. 5571 // Using %k0 as the undef input register is a performance heuristic based 5572 // on the assumption that %k0 is used less frequently than the other mask 5573 // registers, since it is not usable as a write mask. 5574 // FIXME: A more advanced approach would be to choose the best input mask 5575 // register based on context. 5576 case X86::KSET0B: 5577 case X86::KSET0W: return Expand2AddrKreg(MIB, get(X86::KXORWrr), X86::K0); 5578 case X86::KSET0D: return Expand2AddrKreg(MIB, get(X86::KXORDrr), X86::K0); 5579 case X86::KSET0Q: return Expand2AddrKreg(MIB, get(X86::KXORQrr), X86::K0); 5580 case X86::KSET1B: 5581 case X86::KSET1W: return Expand2AddrKreg(MIB, get(X86::KXNORWrr), X86::K0); 5582 case X86::KSET1D: return Expand2AddrKreg(MIB, get(X86::KXNORDrr), X86::K0); 5583 case X86::KSET1Q: return Expand2AddrKreg(MIB, get(X86::KXNORQrr), X86::K0); 5584 case TargetOpcode::LOAD_STACK_GUARD: 5585 expandLoadStackGuard(MIB, *this); 5586 return true; 5587 } 5588 return false; 5589 } 5590 5591 static void addOperands(MachineInstrBuilder &MIB, ArrayRef<MachineOperand> MOs, 5592 int PtrOffset = 0) { 5593 unsigned NumAddrOps = MOs.size(); 5594 5595 if (NumAddrOps < 4) { 5596 // FrameIndex only - add an immediate offset (whether its zero or not). 5597 for (unsigned i = 0; i != NumAddrOps; ++i) 5598 MIB.addOperand(MOs[i]); 5599 addOffset(MIB, PtrOffset); 5600 } else { 5601 // General Memory Addressing - we need to add any offset to an existing 5602 // offset. 5603 assert(MOs.size() == 5 && "Unexpected memory operand list length"); 5604 for (unsigned i = 0; i != NumAddrOps; ++i) { 5605 const MachineOperand &MO = MOs[i]; 5606 if (i == 3 && PtrOffset != 0) { 5607 MIB.addDisp(MO, PtrOffset); 5608 } else { 5609 MIB.addOperand(MO); 5610 } 5611 } 5612 } 5613 } 5614 5615 static MachineInstr *FuseTwoAddrInst(MachineFunction &MF, unsigned Opcode, 5616 ArrayRef<MachineOperand> MOs, 5617 MachineBasicBlock::iterator InsertPt, 5618 MachineInstr *MI, 5619 const TargetInstrInfo &TII) { 5620 // Create the base instruction with the memory operand as the first part. 5621 // Omit the implicit operands, something BuildMI can't do. 5622 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), 5623 MI->getDebugLoc(), true); 5624 MachineInstrBuilder MIB(MF, NewMI); 5625 addOperands(MIB, MOs); 5626 5627 // Loop over the rest of the ri operands, converting them over. 5628 unsigned NumOps = MI->getDesc().getNumOperands()-2; 5629 for (unsigned i = 0; i != NumOps; ++i) { 5630 MachineOperand &MO = MI->getOperand(i+2); 5631 MIB.addOperand(MO); 5632 } 5633 for (unsigned i = NumOps+2, e = MI->getNumOperands(); i != e; ++i) { 5634 MachineOperand &MO = MI->getOperand(i); 5635 MIB.addOperand(MO); 5636 } 5637 5638 MachineBasicBlock *MBB = InsertPt->getParent(); 5639 MBB->insert(InsertPt, NewMI); 5640 5641 return MIB; 5642 } 5643 5644 static MachineInstr *FuseInst(MachineFunction &MF, unsigned Opcode, 5645 unsigned OpNo, ArrayRef<MachineOperand> MOs, 5646 MachineBasicBlock::iterator InsertPt, 5647 MachineInstr *MI, const TargetInstrInfo &TII, 5648 int PtrOffset = 0) { 5649 // Omit the implicit operands, something BuildMI can't do. 5650 MachineInstr *NewMI = MF.CreateMachineInstr(TII.get(Opcode), 5651 MI->getDebugLoc(), true); 5652 MachineInstrBuilder MIB(MF, NewMI); 5653 5654 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 5655 MachineOperand &MO = MI->getOperand(i); 5656 if (i == OpNo) { 5657 assert(MO.isReg() && "Expected to fold into reg operand!"); 5658 addOperands(MIB, MOs, PtrOffset); 5659 } else { 5660 MIB.addOperand(MO); 5661 } 5662 } 5663 5664 MachineBasicBlock *MBB = InsertPt->getParent(); 5665 MBB->insert(InsertPt, NewMI); 5666 5667 return MIB; 5668 } 5669 5670 static MachineInstr *MakeM0Inst(const TargetInstrInfo &TII, unsigned Opcode, 5671 ArrayRef<MachineOperand> MOs, 5672 MachineBasicBlock::iterator InsertPt, 5673 MachineInstr *MI) { 5674 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 5675 MI->getDebugLoc(), TII.get(Opcode)); 5676 addOperands(MIB, MOs); 5677 return MIB.addImm(0); 5678 } 5679 5680 MachineInstr *X86InstrInfo::foldMemoryOperandCustom( 5681 MachineFunction &MF, MachineInstr *MI, unsigned OpNum, 5682 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt, 5683 unsigned Size, unsigned Align) const { 5684 switch (MI->getOpcode()) { 5685 case X86::INSERTPSrr: 5686 case X86::VINSERTPSrr: 5687 // Attempt to convert the load of inserted vector into a fold load 5688 // of a single float. 5689 if (OpNum == 2) { 5690 unsigned Imm = MI->getOperand(MI->getNumOperands() - 1).getImm(); 5691 unsigned ZMask = Imm & 15; 5692 unsigned DstIdx = (Imm >> 4) & 3; 5693 unsigned SrcIdx = (Imm >> 6) & 3; 5694 5695 unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize(); 5696 if (Size <= RCSize && 4 <= Align) { 5697 int PtrOffset = SrcIdx * 4; 5698 unsigned NewImm = (DstIdx << 4) | ZMask; 5699 unsigned NewOpCode = 5700 (MI->getOpcode() == X86::VINSERTPSrr ? X86::VINSERTPSrm 5701 : X86::INSERTPSrm); 5702 MachineInstr *NewMI = 5703 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, PtrOffset); 5704 NewMI->getOperand(NewMI->getNumOperands() - 1).setImm(NewImm); 5705 return NewMI; 5706 } 5707 } 5708 break; 5709 case X86::MOVHLPSrr: 5710 case X86::VMOVHLPSrr: 5711 // Move the upper 64-bits of the second operand to the lower 64-bits. 5712 // To fold the load, adjust the pointer to the upper and use (V)MOVLPS. 5713 // TODO: In most cases AVX doesn't have a 8-byte alignment requirement. 5714 if (OpNum == 2) { 5715 unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize(); 5716 if (Size <= RCSize && 8 <= Align) { 5717 unsigned NewOpCode = 5718 (MI->getOpcode() == X86::VMOVHLPSrr ? X86::VMOVLPSrm 5719 : X86::MOVLPSrm); 5720 MachineInstr *NewMI = 5721 FuseInst(MF, NewOpCode, OpNum, MOs, InsertPt, MI, *this, 8); 5722 return NewMI; 5723 } 5724 } 5725 break; 5726 }; 5727 5728 return nullptr; 5729 } 5730 5731 MachineInstr *X86InstrInfo::foldMemoryOperandImpl( 5732 MachineFunction &MF, MachineInstr *MI, unsigned OpNum, 5733 ArrayRef<MachineOperand> MOs, MachineBasicBlock::iterator InsertPt, 5734 unsigned Size, unsigned Align, bool AllowCommute) const { 5735 const DenseMap<unsigned, 5736 std::pair<uint16_t, uint16_t> > *OpcodeTablePtr = nullptr; 5737 bool isCallRegIndirect = Subtarget.callRegIndirect(); 5738 bool isTwoAddrFold = false; 5739 5740 // For CPUs that favor the register form of a call or push, 5741 // do not fold loads into calls or pushes, unless optimizing for size 5742 // aggressively. 5743 if (isCallRegIndirect && !MF.getFunction()->optForMinSize() && 5744 (MI->getOpcode() == X86::CALL32r || MI->getOpcode() == X86::CALL64r || 5745 MI->getOpcode() == X86::PUSH16r || MI->getOpcode() == X86::PUSH32r || 5746 MI->getOpcode() == X86::PUSH64r)) 5747 return nullptr; 5748 5749 unsigned NumOps = MI->getDesc().getNumOperands(); 5750 bool isTwoAddr = NumOps > 1 && 5751 MI->getDesc().getOperandConstraint(1, MCOI::TIED_TO) != -1; 5752 5753 // FIXME: AsmPrinter doesn't know how to handle 5754 // X86II::MO_GOT_ABSOLUTE_ADDRESS after folding. 5755 if (MI->getOpcode() == X86::ADD32ri && 5756 MI->getOperand(2).getTargetFlags() == X86II::MO_GOT_ABSOLUTE_ADDRESS) 5757 return nullptr; 5758 5759 MachineInstr *NewMI = nullptr; 5760 5761 // Attempt to fold any custom cases we have. 5762 if (MachineInstr *CustomMI = 5763 foldMemoryOperandCustom(MF, MI, OpNum, MOs, InsertPt, Size, Align)) 5764 return CustomMI; 5765 5766 // Folding a memory location into the two-address part of a two-address 5767 // instruction is different than folding it other places. It requires 5768 // replacing the *two* registers with the memory location. 5769 if (isTwoAddr && NumOps >= 2 && OpNum < 2 && 5770 MI->getOperand(0).isReg() && 5771 MI->getOperand(1).isReg() && 5772 MI->getOperand(0).getReg() == MI->getOperand(1).getReg()) { 5773 OpcodeTablePtr = &RegOp2MemOpTable2Addr; 5774 isTwoAddrFold = true; 5775 } else if (OpNum == 0) { 5776 if (MI->getOpcode() == X86::MOV32r0) { 5777 NewMI = MakeM0Inst(*this, X86::MOV32mi, MOs, InsertPt, MI); 5778 if (NewMI) 5779 return NewMI; 5780 } 5781 5782 OpcodeTablePtr = &RegOp2MemOpTable0; 5783 } else if (OpNum == 1) { 5784 OpcodeTablePtr = &RegOp2MemOpTable1; 5785 } else if (OpNum == 2) { 5786 OpcodeTablePtr = &RegOp2MemOpTable2; 5787 } else if (OpNum == 3) { 5788 OpcodeTablePtr = &RegOp2MemOpTable3; 5789 } else if (OpNum == 4) { 5790 OpcodeTablePtr = &RegOp2MemOpTable4; 5791 } 5792 5793 // If table selected... 5794 if (OpcodeTablePtr) { 5795 // Find the Opcode to fuse 5796 auto I = OpcodeTablePtr->find(MI->getOpcode()); 5797 if (I != OpcodeTablePtr->end()) { 5798 unsigned Opcode = I->second.first; 5799 unsigned MinAlign = (I->second.second & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT; 5800 if (Align < MinAlign) 5801 return nullptr; 5802 bool NarrowToMOV32rm = false; 5803 if (Size) { 5804 unsigned RCSize = getRegClass(MI->getDesc(), OpNum, &RI, MF)->getSize(); 5805 if (Size < RCSize) { 5806 // Check if it's safe to fold the load. If the size of the object is 5807 // narrower than the load width, then it's not. 5808 if (Opcode != X86::MOV64rm || RCSize != 8 || Size != 4) 5809 return nullptr; 5810 // If this is a 64-bit load, but the spill slot is 32, then we can do 5811 // a 32-bit load which is implicitly zero-extended. This likely is 5812 // due to live interval analysis remat'ing a load from stack slot. 5813 if (MI->getOperand(0).getSubReg() || MI->getOperand(1).getSubReg()) 5814 return nullptr; 5815 Opcode = X86::MOV32rm; 5816 NarrowToMOV32rm = true; 5817 } 5818 } 5819 5820 if (isTwoAddrFold) 5821 NewMI = FuseTwoAddrInst(MF, Opcode, MOs, InsertPt, MI, *this); 5822 else 5823 NewMI = FuseInst(MF, Opcode, OpNum, MOs, InsertPt, MI, *this); 5824 5825 if (NarrowToMOV32rm) { 5826 // If this is the special case where we use a MOV32rm to load a 32-bit 5827 // value and zero-extend the top bits. Change the destination register 5828 // to a 32-bit one. 5829 unsigned DstReg = NewMI->getOperand(0).getReg(); 5830 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) 5831 NewMI->getOperand(0).setReg(RI.getSubReg(DstReg, X86::sub_32bit)); 5832 else 5833 NewMI->getOperand(0).setSubReg(X86::sub_32bit); 5834 } 5835 return NewMI; 5836 } 5837 } 5838 5839 // If the instruction and target operand are commutable, commute the 5840 // instruction and try again. 5841 if (AllowCommute) { 5842 unsigned CommuteOpIdx1 = OpNum, CommuteOpIdx2 = CommuteAnyOperandIndex; 5843 if (findCommutedOpIndices(MI, CommuteOpIdx1, CommuteOpIdx2)) { 5844 bool HasDef = MI->getDesc().getNumDefs(); 5845 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0; 5846 unsigned Reg1 = MI->getOperand(CommuteOpIdx1).getReg(); 5847 unsigned Reg2 = MI->getOperand(CommuteOpIdx2).getReg(); 5848 bool Tied1 = 5849 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx1, MCOI::TIED_TO); 5850 bool Tied2 = 5851 0 == MI->getDesc().getOperandConstraint(CommuteOpIdx2, MCOI::TIED_TO); 5852 5853 // If either of the commutable operands are tied to the destination 5854 // then we can not commute + fold. 5855 if ((HasDef && Reg0 == Reg1 && Tied1) || 5856 (HasDef && Reg0 == Reg2 && Tied2)) 5857 return nullptr; 5858 5859 MachineInstr *CommutedMI = 5860 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2); 5861 if (!CommutedMI) { 5862 // Unable to commute. 5863 return nullptr; 5864 } 5865 if (CommutedMI != MI) { 5866 // New instruction. We can't fold from this. 5867 CommutedMI->eraseFromParent(); 5868 return nullptr; 5869 } 5870 5871 // Attempt to fold with the commuted version of the instruction. 5872 NewMI = foldMemoryOperandImpl(MF, MI, CommuteOpIdx2, MOs, InsertPt, 5873 Size, Align, /*AllowCommute=*/false); 5874 if (NewMI) 5875 return NewMI; 5876 5877 // Folding failed again - undo the commute before returning. 5878 MachineInstr *UncommutedMI = 5879 commuteInstruction(MI, false, CommuteOpIdx1, CommuteOpIdx2); 5880 if (!UncommutedMI) { 5881 // Unable to commute. 5882 return nullptr; 5883 } 5884 if (UncommutedMI != MI) { 5885 // New instruction. It doesn't need to be kept. 5886 UncommutedMI->eraseFromParent(); 5887 return nullptr; 5888 } 5889 5890 // Return here to prevent duplicate fuse failure report. 5891 return nullptr; 5892 } 5893 } 5894 5895 // No fusion 5896 if (PrintFailedFusing && !MI->isCopy()) 5897 dbgs() << "We failed to fuse operand " << OpNum << " in " << *MI; 5898 return nullptr; 5899 } 5900 5901 /// Return true for all instructions that only update 5902 /// the first 32 or 64-bits of the destination register and leave the rest 5903 /// unmodified. This can be used to avoid folding loads if the instructions 5904 /// only update part of the destination register, and the non-updated part is 5905 /// not needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these 5906 /// instructions breaks the partial register dependency and it can improve 5907 /// performance. e.g.: 5908 /// 5909 /// movss (%rdi), %xmm0 5910 /// cvtss2sd %xmm0, %xmm0 5911 /// 5912 /// Instead of 5913 /// cvtss2sd (%rdi), %xmm0 5914 /// 5915 /// FIXME: This should be turned into a TSFlags. 5916 /// 5917 static bool hasPartialRegUpdate(unsigned Opcode) { 5918 switch (Opcode) { 5919 case X86::CVTSI2SSrr: 5920 case X86::CVTSI2SSrm: 5921 case X86::CVTSI2SS64rr: 5922 case X86::CVTSI2SS64rm: 5923 case X86::CVTSI2SDrr: 5924 case X86::CVTSI2SDrm: 5925 case X86::CVTSI2SD64rr: 5926 case X86::CVTSI2SD64rm: 5927 case X86::CVTSD2SSrr: 5928 case X86::CVTSD2SSrm: 5929 case X86::Int_CVTSD2SSrr: 5930 case X86::Int_CVTSD2SSrm: 5931 case X86::CVTSS2SDrr: 5932 case X86::CVTSS2SDrm: 5933 case X86::Int_CVTSS2SDrr: 5934 case X86::Int_CVTSS2SDrm: 5935 case X86::MOVHPDrm: 5936 case X86::MOVHPSrm: 5937 case X86::MOVLPDrm: 5938 case X86::MOVLPSrm: 5939 case X86::RCPSSr: 5940 case X86::RCPSSm: 5941 case X86::RCPSSr_Int: 5942 case X86::RCPSSm_Int: 5943 case X86::ROUNDSDr: 5944 case X86::ROUNDSDm: 5945 case X86::ROUNDSDr_Int: 5946 case X86::ROUNDSSr: 5947 case X86::ROUNDSSm: 5948 case X86::ROUNDSSr_Int: 5949 case X86::RSQRTSSr: 5950 case X86::RSQRTSSm: 5951 case X86::RSQRTSSr_Int: 5952 case X86::RSQRTSSm_Int: 5953 case X86::SQRTSSr: 5954 case X86::SQRTSSm: 5955 case X86::SQRTSSr_Int: 5956 case X86::SQRTSSm_Int: 5957 case X86::SQRTSDr: 5958 case X86::SQRTSDm: 5959 case X86::SQRTSDr_Int: 5960 case X86::SQRTSDm_Int: 5961 return true; 5962 } 5963 5964 return false; 5965 } 5966 5967 /// Inform the ExeDepsFix pass how many idle 5968 /// instructions we would like before a partial register update. 5969 unsigned X86InstrInfo:: 5970 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum, 5971 const TargetRegisterInfo *TRI) const { 5972 if (OpNum != 0 || !hasPartialRegUpdate(MI->getOpcode())) 5973 return 0; 5974 5975 // If MI is marked as reading Reg, the partial register update is wanted. 5976 const MachineOperand &MO = MI->getOperand(0); 5977 unsigned Reg = MO.getReg(); 5978 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 5979 if (MO.readsReg() || MI->readsVirtualRegister(Reg)) 5980 return 0; 5981 } else { 5982 if (MI->readsRegister(Reg, TRI)) 5983 return 0; 5984 } 5985 5986 // If any instructions in the clearance range are reading Reg, insert a 5987 // dependency breaking instruction, which is inexpensive and is likely to 5988 // be hidden in other instruction's cycles. 5989 return PartialRegUpdateClearance; 5990 } 5991 5992 // Return true for any instruction the copies the high bits of the first source 5993 // operand into the unused high bits of the destination operand. 5994 static bool hasUndefRegUpdate(unsigned Opcode) { 5995 switch (Opcode) { 5996 case X86::VCVTSI2SSrr: 5997 case X86::VCVTSI2SSrm: 5998 case X86::Int_VCVTSI2SSrr: 5999 case X86::Int_VCVTSI2SSrm: 6000 case X86::VCVTSI2SS64rr: 6001 case X86::VCVTSI2SS64rm: 6002 case X86::Int_VCVTSI2SS64rr: 6003 case X86::Int_VCVTSI2SS64rm: 6004 case X86::VCVTSI2SDrr: 6005 case X86::VCVTSI2SDrm: 6006 case X86::Int_VCVTSI2SDrr: 6007 case X86::Int_VCVTSI2SDrm: 6008 case X86::VCVTSI2SD64rr: 6009 case X86::VCVTSI2SD64rm: 6010 case X86::Int_VCVTSI2SD64rr: 6011 case X86::Int_VCVTSI2SD64rm: 6012 case X86::VCVTSD2SSrr: 6013 case X86::VCVTSD2SSrm: 6014 case X86::Int_VCVTSD2SSrr: 6015 case X86::Int_VCVTSD2SSrm: 6016 case X86::VCVTSS2SDrr: 6017 case X86::VCVTSS2SDrm: 6018 case X86::Int_VCVTSS2SDrr: 6019 case X86::Int_VCVTSS2SDrm: 6020 case X86::VRCPSSr: 6021 case X86::VRCPSSm: 6022 case X86::VRCPSSm_Int: 6023 case X86::VROUNDSDr: 6024 case X86::VROUNDSDm: 6025 case X86::VROUNDSDr_Int: 6026 case X86::VROUNDSSr: 6027 case X86::VROUNDSSm: 6028 case X86::VROUNDSSr_Int: 6029 case X86::VRSQRTSSr: 6030 case X86::VRSQRTSSm: 6031 case X86::VRSQRTSSm_Int: 6032 case X86::VSQRTSSr: 6033 case X86::VSQRTSSm: 6034 case X86::VSQRTSSm_Int: 6035 case X86::VSQRTSDr: 6036 case X86::VSQRTSDm: 6037 case X86::VSQRTSDm_Int: 6038 // AVX-512 6039 case X86::VCVTSD2SSZrr: 6040 case X86::VCVTSD2SSZrm: 6041 case X86::VCVTSS2SDZrr: 6042 case X86::VCVTSS2SDZrm: 6043 return true; 6044 } 6045 6046 return false; 6047 } 6048 6049 /// Inform the ExeDepsFix pass how many idle instructions we would like before 6050 /// certain undef register reads. 6051 /// 6052 /// This catches the VCVTSI2SD family of instructions: 6053 /// 6054 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14 6055 /// 6056 /// We should to be careful *not* to catch VXOR idioms which are presumably 6057 /// handled specially in the pipeline: 6058 /// 6059 /// vxorps %xmm1<undef>, %xmm1<undef>, %xmm1 6060 /// 6061 /// Like getPartialRegUpdateClearance, this makes a strong assumption that the 6062 /// high bits that are passed-through are not live. 6063 unsigned X86InstrInfo:: 6064 getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum, 6065 const TargetRegisterInfo *TRI) const { 6066 if (!hasUndefRegUpdate(MI->getOpcode())) 6067 return 0; 6068 6069 // Set the OpNum parameter to the first source operand. 6070 OpNum = 1; 6071 6072 const MachineOperand &MO = MI->getOperand(OpNum); 6073 if (MO.isUndef() && TargetRegisterInfo::isPhysicalRegister(MO.getReg())) { 6074 return UndefRegClearance; 6075 } 6076 return 0; 6077 } 6078 6079 void X86InstrInfo:: 6080 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum, 6081 const TargetRegisterInfo *TRI) const { 6082 unsigned Reg = MI->getOperand(OpNum).getReg(); 6083 // If MI kills this register, the false dependence is already broken. 6084 if (MI->killsRegister(Reg, TRI)) 6085 return; 6086 6087 if (X86::VR128RegClass.contains(Reg)) { 6088 // These instructions are all floating point domain, so xorps is the best 6089 // choice. 6090 unsigned Opc = Subtarget.hasAVX() ? X86::VXORPSrr : X86::XORPSrr; 6091 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(Opc), Reg) 6092 .addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef); 6093 MI->addRegisterKilled(Reg, TRI, true); 6094 } else if (X86::VR256RegClass.contains(Reg)) { 6095 // Use vxorps to clear the full ymm register. 6096 // It wants to read and write the xmm sub-register. 6097 unsigned XReg = TRI->getSubReg(Reg, X86::sub_xmm); 6098 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), get(X86::VXORPSrr), XReg) 6099 .addReg(XReg, RegState::Undef).addReg(XReg, RegState::Undef) 6100 .addReg(Reg, RegState::ImplicitDefine); 6101 MI->addRegisterKilled(Reg, TRI, true); 6102 } 6103 } 6104 6105 MachineInstr *X86InstrInfo::foldMemoryOperandImpl( 6106 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops, 6107 MachineBasicBlock::iterator InsertPt, int FrameIndex, 6108 LiveIntervals *LIS) const { 6109 // Check switch flag 6110 if (NoFusing) 6111 return nullptr; 6112 6113 // Unless optimizing for size, don't fold to avoid partial 6114 // register update stalls 6115 if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI->getOpcode())) 6116 return nullptr; 6117 6118 const MachineFrameInfo *MFI = MF.getFrameInfo(); 6119 unsigned Size = MFI->getObjectSize(FrameIndex); 6120 unsigned Alignment = MFI->getObjectAlignment(FrameIndex); 6121 // If the function stack isn't realigned we don't want to fold instructions 6122 // that need increased alignment. 6123 if (!RI.needsStackRealignment(MF)) 6124 Alignment = 6125 std::min(Alignment, Subtarget.getFrameLowering()->getStackAlignment()); 6126 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 6127 unsigned NewOpc = 0; 6128 unsigned RCSize = 0; 6129 switch (MI->getOpcode()) { 6130 default: return nullptr; 6131 case X86::TEST8rr: NewOpc = X86::CMP8ri; RCSize = 1; break; 6132 case X86::TEST16rr: NewOpc = X86::CMP16ri8; RCSize = 2; break; 6133 case X86::TEST32rr: NewOpc = X86::CMP32ri8; RCSize = 4; break; 6134 case X86::TEST64rr: NewOpc = X86::CMP64ri8; RCSize = 8; break; 6135 } 6136 // Check if it's safe to fold the load. If the size of the object is 6137 // narrower than the load width, then it's not. 6138 if (Size < RCSize) 6139 return nullptr; 6140 // Change to CMPXXri r, 0 first. 6141 MI->setDesc(get(NewOpc)); 6142 MI->getOperand(1).ChangeToImmediate(0); 6143 } else if (Ops.size() != 1) 6144 return nullptr; 6145 6146 return foldMemoryOperandImpl(MF, MI, Ops[0], 6147 MachineOperand::CreateFI(FrameIndex), InsertPt, 6148 Size, Alignment, /*AllowCommute=*/true); 6149 } 6150 6151 /// Check if \p LoadMI is a partial register load that we can't fold into \p MI 6152 /// because the latter uses contents that wouldn't be defined in the folded 6153 /// version. For instance, this transformation isn't legal: 6154 /// movss (%rdi), %xmm0 6155 /// addps %xmm0, %xmm0 6156 /// -> 6157 /// addps (%rdi), %xmm0 6158 /// 6159 /// But this one is: 6160 /// movss (%rdi), %xmm0 6161 /// addss %xmm0, %xmm0 6162 /// -> 6163 /// addss (%rdi), %xmm0 6164 /// 6165 static bool isNonFoldablePartialRegisterLoad(const MachineInstr &LoadMI, 6166 const MachineInstr &UserMI, 6167 const MachineFunction &MF) { 6168 unsigned Opc = LoadMI.getOpcode(); 6169 unsigned UserOpc = UserMI.getOpcode(); 6170 unsigned RegSize = 6171 MF.getRegInfo().getRegClass(LoadMI.getOperand(0).getReg())->getSize(); 6172 6173 if ((Opc == X86::MOVSSrm || Opc == X86::VMOVSSrm) && RegSize > 4) { 6174 // These instructions only load 32 bits, we can't fold them if the 6175 // destination register is wider than 32 bits (4 bytes), and its user 6176 // instruction isn't scalar (SS). 6177 switch (UserOpc) { 6178 case X86::ADDSSrr_Int: case X86::VADDSSrr_Int: 6179 case X86::DIVSSrr_Int: case X86::VDIVSSrr_Int: 6180 case X86::MULSSrr_Int: case X86::VMULSSrr_Int: 6181 case X86::SUBSSrr_Int: case X86::VSUBSSrr_Int: 6182 case X86::VFMADDSSr132r_Int: case X86::VFNMADDSSr132r_Int: 6183 case X86::VFMADDSSr213r_Int: case X86::VFNMADDSSr213r_Int: 6184 case X86::VFMADDSSr231r_Int: case X86::VFNMADDSSr231r_Int: 6185 case X86::VFMSUBSSr132r_Int: case X86::VFNMSUBSSr132r_Int: 6186 case X86::VFMSUBSSr213r_Int: case X86::VFNMSUBSSr213r_Int: 6187 case X86::VFMSUBSSr231r_Int: case X86::VFNMSUBSSr231r_Int: 6188 return false; 6189 default: 6190 return true; 6191 } 6192 } 6193 6194 if ((Opc == X86::MOVSDrm || Opc == X86::VMOVSDrm) && RegSize > 8) { 6195 // These instructions only load 64 bits, we can't fold them if the 6196 // destination register is wider than 64 bits (8 bytes), and its user 6197 // instruction isn't scalar (SD). 6198 switch (UserOpc) { 6199 case X86::ADDSDrr_Int: case X86::VADDSDrr_Int: 6200 case X86::DIVSDrr_Int: case X86::VDIVSDrr_Int: 6201 case X86::MULSDrr_Int: case X86::VMULSDrr_Int: 6202 case X86::SUBSDrr_Int: case X86::VSUBSDrr_Int: 6203 case X86::VFMADDSDr132r_Int: case X86::VFNMADDSDr132r_Int: 6204 case X86::VFMADDSDr213r_Int: case X86::VFNMADDSDr213r_Int: 6205 case X86::VFMADDSDr231r_Int: case X86::VFNMADDSDr231r_Int: 6206 case X86::VFMSUBSDr132r_Int: case X86::VFNMSUBSDr132r_Int: 6207 case X86::VFMSUBSDr213r_Int: case X86::VFNMSUBSDr213r_Int: 6208 case X86::VFMSUBSDr231r_Int: case X86::VFNMSUBSDr231r_Int: 6209 return false; 6210 default: 6211 return true; 6212 } 6213 } 6214 6215 return false; 6216 } 6217 6218 MachineInstr *X86InstrInfo::foldMemoryOperandImpl( 6219 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops, 6220 MachineBasicBlock::iterator InsertPt, MachineInstr *LoadMI, 6221 LiveIntervals *LIS) const { 6222 // If loading from a FrameIndex, fold directly from the FrameIndex. 6223 unsigned NumOps = LoadMI->getDesc().getNumOperands(); 6224 int FrameIndex; 6225 if (isLoadFromStackSlot(LoadMI, FrameIndex)) { 6226 if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF)) 6227 return nullptr; 6228 return foldMemoryOperandImpl(MF, MI, Ops, InsertPt, FrameIndex, LIS); 6229 } 6230 6231 // Check switch flag 6232 if (NoFusing) return nullptr; 6233 6234 // Avoid partial register update stalls unless optimizing for size. 6235 if (!MF.getFunction()->optForSize() && hasPartialRegUpdate(MI->getOpcode())) 6236 return nullptr; 6237 6238 // Determine the alignment of the load. 6239 unsigned Alignment = 0; 6240 if (LoadMI->hasOneMemOperand()) 6241 Alignment = (*LoadMI->memoperands_begin())->getAlignment(); 6242 else 6243 switch (LoadMI->getOpcode()) { 6244 case X86::AVX2_SETALLONES: 6245 case X86::AVX_SET0: 6246 Alignment = 32; 6247 break; 6248 case X86::V_SET0: 6249 case X86::V_SETALLONES: 6250 Alignment = 16; 6251 break; 6252 case X86::FsFLD0SD: 6253 Alignment = 8; 6254 break; 6255 case X86::FsFLD0SS: 6256 Alignment = 4; 6257 break; 6258 default: 6259 return nullptr; 6260 } 6261 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 6262 unsigned NewOpc = 0; 6263 switch (MI->getOpcode()) { 6264 default: return nullptr; 6265 case X86::TEST8rr: NewOpc = X86::CMP8ri; break; 6266 case X86::TEST16rr: NewOpc = X86::CMP16ri8; break; 6267 case X86::TEST32rr: NewOpc = X86::CMP32ri8; break; 6268 case X86::TEST64rr: NewOpc = X86::CMP64ri8; break; 6269 } 6270 // Change to CMPXXri r, 0 first. 6271 MI->setDesc(get(NewOpc)); 6272 MI->getOperand(1).ChangeToImmediate(0); 6273 } else if (Ops.size() != 1) 6274 return nullptr; 6275 6276 // Make sure the subregisters match. 6277 // Otherwise we risk changing the size of the load. 6278 if (LoadMI->getOperand(0).getSubReg() != MI->getOperand(Ops[0]).getSubReg()) 6279 return nullptr; 6280 6281 SmallVector<MachineOperand,X86::AddrNumOperands> MOs; 6282 switch (LoadMI->getOpcode()) { 6283 case X86::V_SET0: 6284 case X86::V_SETALLONES: 6285 case X86::AVX2_SETALLONES: 6286 case X86::AVX_SET0: 6287 case X86::FsFLD0SD: 6288 case X86::FsFLD0SS: { 6289 // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure. 6290 // Create a constant-pool entry and operands to load from it. 6291 6292 // Medium and large mode can't fold loads this way. 6293 if (MF.getTarget().getCodeModel() != CodeModel::Small && 6294 MF.getTarget().getCodeModel() != CodeModel::Kernel) 6295 return nullptr; 6296 6297 // x86-32 PIC requires a PIC base register for constant pools. 6298 unsigned PICBase = 0; 6299 if (MF.getTarget().isPositionIndependent()) { 6300 if (Subtarget.is64Bit()) 6301 PICBase = X86::RIP; 6302 else 6303 // FIXME: PICBase = getGlobalBaseReg(&MF); 6304 // This doesn't work for several reasons. 6305 // 1. GlobalBaseReg may have been spilled. 6306 // 2. It may not be live at MI. 6307 return nullptr; 6308 } 6309 6310 // Create a constant-pool entry. 6311 MachineConstantPool &MCP = *MF.getConstantPool(); 6312 Type *Ty; 6313 unsigned Opc = LoadMI->getOpcode(); 6314 if (Opc == X86::FsFLD0SS) 6315 Ty = Type::getFloatTy(MF.getFunction()->getContext()); 6316 else if (Opc == X86::FsFLD0SD) 6317 Ty = Type::getDoubleTy(MF.getFunction()->getContext()); 6318 else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0) 6319 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8); 6320 else 6321 Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4); 6322 6323 bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES); 6324 const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) : 6325 Constant::getNullValue(Ty); 6326 unsigned CPI = MCP.getConstantPoolIndex(C, Alignment); 6327 6328 // Create operands to load from the constant pool entry. 6329 MOs.push_back(MachineOperand::CreateReg(PICBase, false)); 6330 MOs.push_back(MachineOperand::CreateImm(1)); 6331 MOs.push_back(MachineOperand::CreateReg(0, false)); 6332 MOs.push_back(MachineOperand::CreateCPI(CPI, 0)); 6333 MOs.push_back(MachineOperand::CreateReg(0, false)); 6334 break; 6335 } 6336 default: { 6337 if (isNonFoldablePartialRegisterLoad(*LoadMI, *MI, MF)) 6338 return nullptr; 6339 6340 // Folding a normal load. Just copy the load's address operands. 6341 MOs.append(LoadMI->operands_begin() + NumOps - X86::AddrNumOperands, 6342 LoadMI->operands_begin() + NumOps); 6343 break; 6344 } 6345 } 6346 return foldMemoryOperandImpl(MF, MI, Ops[0], MOs, InsertPt, 6347 /*Size=*/0, Alignment, /*AllowCommute=*/true); 6348 } 6349 6350 bool X86InstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI, 6351 unsigned Reg, bool UnfoldLoad, bool UnfoldStore, 6352 SmallVectorImpl<MachineInstr*> &NewMIs) const { 6353 auto I = MemOp2RegOpTable.find(MI->getOpcode()); 6354 if (I == MemOp2RegOpTable.end()) 6355 return false; 6356 unsigned Opc = I->second.first; 6357 unsigned Index = I->second.second & TB_INDEX_MASK; 6358 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 6359 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 6360 if (UnfoldLoad && !FoldedLoad) 6361 return false; 6362 UnfoldLoad &= FoldedLoad; 6363 if (UnfoldStore && !FoldedStore) 6364 return false; 6365 UnfoldStore &= FoldedStore; 6366 6367 const MCInstrDesc &MCID = get(Opc); 6368 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 6369 // TODO: Check if 32-byte or greater accesses are slow too? 6370 if (!MI->hasOneMemOperand() && 6371 RC == &X86::VR128RegClass && 6372 Subtarget.isUnalignedMem16Slow()) 6373 // Without memoperands, loadRegFromAddr and storeRegToStackSlot will 6374 // conservatively assume the address is unaligned. That's bad for 6375 // performance. 6376 return false; 6377 SmallVector<MachineOperand, X86::AddrNumOperands> AddrOps; 6378 SmallVector<MachineOperand,2> BeforeOps; 6379 SmallVector<MachineOperand,2> AfterOps; 6380 SmallVector<MachineOperand,4> ImpOps; 6381 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 6382 MachineOperand &Op = MI->getOperand(i); 6383 if (i >= Index && i < Index + X86::AddrNumOperands) 6384 AddrOps.push_back(Op); 6385 else if (Op.isReg() && Op.isImplicit()) 6386 ImpOps.push_back(Op); 6387 else if (i < Index) 6388 BeforeOps.push_back(Op); 6389 else if (i > Index) 6390 AfterOps.push_back(Op); 6391 } 6392 6393 // Emit the load instruction. 6394 if (UnfoldLoad) { 6395 std::pair<MachineInstr::mmo_iterator, 6396 MachineInstr::mmo_iterator> MMOs = 6397 MF.extractLoadMemRefs(MI->memoperands_begin(), 6398 MI->memoperands_end()); 6399 loadRegFromAddr(MF, Reg, AddrOps, RC, MMOs.first, MMOs.second, NewMIs); 6400 if (UnfoldStore) { 6401 // Address operands cannot be marked isKill. 6402 for (unsigned i = 1; i != 1 + X86::AddrNumOperands; ++i) { 6403 MachineOperand &MO = NewMIs[0]->getOperand(i); 6404 if (MO.isReg()) 6405 MO.setIsKill(false); 6406 } 6407 } 6408 } 6409 6410 // Emit the data processing instruction. 6411 MachineInstr *DataMI = MF.CreateMachineInstr(MCID, MI->getDebugLoc(), true); 6412 MachineInstrBuilder MIB(MF, DataMI); 6413 6414 if (FoldedStore) 6415 MIB.addReg(Reg, RegState::Define); 6416 for (MachineOperand &BeforeOp : BeforeOps) 6417 MIB.addOperand(BeforeOp); 6418 if (FoldedLoad) 6419 MIB.addReg(Reg); 6420 for (MachineOperand &AfterOp : AfterOps) 6421 MIB.addOperand(AfterOp); 6422 for (MachineOperand &ImpOp : ImpOps) { 6423 MIB.addReg(ImpOp.getReg(), 6424 getDefRegState(ImpOp.isDef()) | 6425 RegState::Implicit | 6426 getKillRegState(ImpOp.isKill()) | 6427 getDeadRegState(ImpOp.isDead()) | 6428 getUndefRegState(ImpOp.isUndef())); 6429 } 6430 // Change CMP32ri r, 0 back to TEST32rr r, r, etc. 6431 switch (DataMI->getOpcode()) { 6432 default: break; 6433 case X86::CMP64ri32: 6434 case X86::CMP64ri8: 6435 case X86::CMP32ri: 6436 case X86::CMP32ri8: 6437 case X86::CMP16ri: 6438 case X86::CMP16ri8: 6439 case X86::CMP8ri: { 6440 MachineOperand &MO0 = DataMI->getOperand(0); 6441 MachineOperand &MO1 = DataMI->getOperand(1); 6442 if (MO1.getImm() == 0) { 6443 unsigned NewOpc; 6444 switch (DataMI->getOpcode()) { 6445 default: llvm_unreachable("Unreachable!"); 6446 case X86::CMP64ri8: 6447 case X86::CMP64ri32: NewOpc = X86::TEST64rr; break; 6448 case X86::CMP32ri8: 6449 case X86::CMP32ri: NewOpc = X86::TEST32rr; break; 6450 case X86::CMP16ri8: 6451 case X86::CMP16ri: NewOpc = X86::TEST16rr; break; 6452 case X86::CMP8ri: NewOpc = X86::TEST8rr; break; 6453 } 6454 DataMI->setDesc(get(NewOpc)); 6455 MO1.ChangeToRegister(MO0.getReg(), false); 6456 } 6457 } 6458 } 6459 NewMIs.push_back(DataMI); 6460 6461 // Emit the store instruction. 6462 if (UnfoldStore) { 6463 const TargetRegisterClass *DstRC = getRegClass(MCID, 0, &RI, MF); 6464 std::pair<MachineInstr::mmo_iterator, 6465 MachineInstr::mmo_iterator> MMOs = 6466 MF.extractStoreMemRefs(MI->memoperands_begin(), 6467 MI->memoperands_end()); 6468 storeRegToAddr(MF, Reg, true, AddrOps, DstRC, MMOs.first, MMOs.second, NewMIs); 6469 } 6470 6471 return true; 6472 } 6473 6474 bool 6475 X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, 6476 SmallVectorImpl<SDNode*> &NewNodes) const { 6477 if (!N->isMachineOpcode()) 6478 return false; 6479 6480 auto I = MemOp2RegOpTable.find(N->getMachineOpcode()); 6481 if (I == MemOp2RegOpTable.end()) 6482 return false; 6483 unsigned Opc = I->second.first; 6484 unsigned Index = I->second.second & TB_INDEX_MASK; 6485 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 6486 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 6487 const MCInstrDesc &MCID = get(Opc); 6488 MachineFunction &MF = DAG.getMachineFunction(); 6489 const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF); 6490 unsigned NumDefs = MCID.NumDefs; 6491 std::vector<SDValue> AddrOps; 6492 std::vector<SDValue> BeforeOps; 6493 std::vector<SDValue> AfterOps; 6494 SDLoc dl(N); 6495 unsigned NumOps = N->getNumOperands(); 6496 for (unsigned i = 0; i != NumOps-1; ++i) { 6497 SDValue Op = N->getOperand(i); 6498 if (i >= Index-NumDefs && i < Index-NumDefs + X86::AddrNumOperands) 6499 AddrOps.push_back(Op); 6500 else if (i < Index-NumDefs) 6501 BeforeOps.push_back(Op); 6502 else if (i > Index-NumDefs) 6503 AfterOps.push_back(Op); 6504 } 6505 SDValue Chain = N->getOperand(NumOps-1); 6506 AddrOps.push_back(Chain); 6507 6508 // Emit the load instruction. 6509 SDNode *Load = nullptr; 6510 if (FoldedLoad) { 6511 EVT VT = *RC->vt_begin(); 6512 std::pair<MachineInstr::mmo_iterator, 6513 MachineInstr::mmo_iterator> MMOs = 6514 MF.extractLoadMemRefs(cast<MachineSDNode>(N)->memoperands_begin(), 6515 cast<MachineSDNode>(N)->memoperands_end()); 6516 if (!(*MMOs.first) && 6517 RC == &X86::VR128RegClass && 6518 Subtarget.isUnalignedMem16Slow()) 6519 // Do not introduce a slow unaligned load. 6520 return false; 6521 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte 6522 // memory access is slow above. 6523 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 6524 bool isAligned = (*MMOs.first) && 6525 (*MMOs.first)->getAlignment() >= Alignment; 6526 Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl, 6527 VT, MVT::Other, AddrOps); 6528 NewNodes.push_back(Load); 6529 6530 // Preserve memory reference information. 6531 cast<MachineSDNode>(Load)->setMemRefs(MMOs.first, MMOs.second); 6532 } 6533 6534 // Emit the data processing instruction. 6535 std::vector<EVT> VTs; 6536 const TargetRegisterClass *DstRC = nullptr; 6537 if (MCID.getNumDefs() > 0) { 6538 DstRC = getRegClass(MCID, 0, &RI, MF); 6539 VTs.push_back(*DstRC->vt_begin()); 6540 } 6541 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) { 6542 EVT VT = N->getValueType(i); 6543 if (VT != MVT::Other && i >= (unsigned)MCID.getNumDefs()) 6544 VTs.push_back(VT); 6545 } 6546 if (Load) 6547 BeforeOps.push_back(SDValue(Load, 0)); 6548 BeforeOps.insert(BeforeOps.end(), AfterOps.begin(), AfterOps.end()); 6549 SDNode *NewNode= DAG.getMachineNode(Opc, dl, VTs, BeforeOps); 6550 NewNodes.push_back(NewNode); 6551 6552 // Emit the store instruction. 6553 if (FoldedStore) { 6554 AddrOps.pop_back(); 6555 AddrOps.push_back(SDValue(NewNode, 0)); 6556 AddrOps.push_back(Chain); 6557 std::pair<MachineInstr::mmo_iterator, 6558 MachineInstr::mmo_iterator> MMOs = 6559 MF.extractStoreMemRefs(cast<MachineSDNode>(N)->memoperands_begin(), 6560 cast<MachineSDNode>(N)->memoperands_end()); 6561 if (!(*MMOs.first) && 6562 RC == &X86::VR128RegClass && 6563 Subtarget.isUnalignedMem16Slow()) 6564 // Do not introduce a slow unaligned store. 6565 return false; 6566 // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte 6567 // memory access is slow above. 6568 unsigned Alignment = RC->getSize() == 32 ? 32 : 16; 6569 bool isAligned = (*MMOs.first) && 6570 (*MMOs.first)->getAlignment() >= Alignment; 6571 SDNode *Store = 6572 DAG.getMachineNode(getStoreRegOpcode(0, DstRC, isAligned, Subtarget), 6573 dl, MVT::Other, AddrOps); 6574 NewNodes.push_back(Store); 6575 6576 // Preserve memory reference information. 6577 cast<MachineSDNode>(Store)->setMemRefs(MMOs.first, MMOs.second); 6578 } 6579 6580 return true; 6581 } 6582 6583 unsigned X86InstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc, 6584 bool UnfoldLoad, bool UnfoldStore, 6585 unsigned *LoadRegIndex) const { 6586 auto I = MemOp2RegOpTable.find(Opc); 6587 if (I == MemOp2RegOpTable.end()) 6588 return 0; 6589 bool FoldedLoad = I->second.second & TB_FOLDED_LOAD; 6590 bool FoldedStore = I->second.second & TB_FOLDED_STORE; 6591 if (UnfoldLoad && !FoldedLoad) 6592 return 0; 6593 if (UnfoldStore && !FoldedStore) 6594 return 0; 6595 if (LoadRegIndex) 6596 *LoadRegIndex = I->second.second & TB_INDEX_MASK; 6597 return I->second.first; 6598 } 6599 6600 bool 6601 X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 6602 int64_t &Offset1, int64_t &Offset2) const { 6603 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode()) 6604 return false; 6605 unsigned Opc1 = Load1->getMachineOpcode(); 6606 unsigned Opc2 = Load2->getMachineOpcode(); 6607 switch (Opc1) { 6608 default: return false; 6609 case X86::MOV8rm: 6610 case X86::MOV16rm: 6611 case X86::MOV32rm: 6612 case X86::MOV64rm: 6613 case X86::LD_Fp32m: 6614 case X86::LD_Fp64m: 6615 case X86::LD_Fp80m: 6616 case X86::MOVSSrm: 6617 case X86::MOVSDrm: 6618 case X86::MMX_MOVD64rm: 6619 case X86::MMX_MOVQ64rm: 6620 case X86::FsMOVAPSrm: 6621 case X86::FsMOVAPDrm: 6622 case X86::MOVAPSrm: 6623 case X86::MOVUPSrm: 6624 case X86::MOVAPDrm: 6625 case X86::MOVDQArm: 6626 case X86::MOVDQUrm: 6627 // AVX load instructions 6628 case X86::VMOVSSrm: 6629 case X86::VMOVSDrm: 6630 case X86::FsVMOVAPSrm: 6631 case X86::FsVMOVAPDrm: 6632 case X86::VMOVAPSrm: 6633 case X86::VMOVUPSrm: 6634 case X86::VMOVAPDrm: 6635 case X86::VMOVDQArm: 6636 case X86::VMOVDQUrm: 6637 case X86::VMOVAPSYrm: 6638 case X86::VMOVUPSYrm: 6639 case X86::VMOVAPDYrm: 6640 case X86::VMOVDQAYrm: 6641 case X86::VMOVDQUYrm: 6642 break; 6643 } 6644 switch (Opc2) { 6645 default: return false; 6646 case X86::MOV8rm: 6647 case X86::MOV16rm: 6648 case X86::MOV32rm: 6649 case X86::MOV64rm: 6650 case X86::LD_Fp32m: 6651 case X86::LD_Fp64m: 6652 case X86::LD_Fp80m: 6653 case X86::MOVSSrm: 6654 case X86::MOVSDrm: 6655 case X86::MMX_MOVD64rm: 6656 case X86::MMX_MOVQ64rm: 6657 case X86::FsMOVAPSrm: 6658 case X86::FsMOVAPDrm: 6659 case X86::MOVAPSrm: 6660 case X86::MOVUPSrm: 6661 case X86::MOVAPDrm: 6662 case X86::MOVDQArm: 6663 case X86::MOVDQUrm: 6664 // AVX load instructions 6665 case X86::VMOVSSrm: 6666 case X86::VMOVSDrm: 6667 case X86::FsVMOVAPSrm: 6668 case X86::FsVMOVAPDrm: 6669 case X86::VMOVAPSrm: 6670 case X86::VMOVUPSrm: 6671 case X86::VMOVAPDrm: 6672 case X86::VMOVDQArm: 6673 case X86::VMOVDQUrm: 6674 case X86::VMOVAPSYrm: 6675 case X86::VMOVUPSYrm: 6676 case X86::VMOVAPDYrm: 6677 case X86::VMOVDQAYrm: 6678 case X86::VMOVDQUYrm: 6679 break; 6680 } 6681 6682 // Check if chain operands and base addresses match. 6683 if (Load1->getOperand(0) != Load2->getOperand(0) || 6684 Load1->getOperand(5) != Load2->getOperand(5)) 6685 return false; 6686 // Segment operands should match as well. 6687 if (Load1->getOperand(4) != Load2->getOperand(4)) 6688 return false; 6689 // Scale should be 1, Index should be Reg0. 6690 if (Load1->getOperand(1) == Load2->getOperand(1) && 6691 Load1->getOperand(2) == Load2->getOperand(2)) { 6692 if (cast<ConstantSDNode>(Load1->getOperand(1))->getZExtValue() != 1) 6693 return false; 6694 6695 // Now let's examine the displacements. 6696 if (isa<ConstantSDNode>(Load1->getOperand(3)) && 6697 isa<ConstantSDNode>(Load2->getOperand(3))) { 6698 Offset1 = cast<ConstantSDNode>(Load1->getOperand(3))->getSExtValue(); 6699 Offset2 = cast<ConstantSDNode>(Load2->getOperand(3))->getSExtValue(); 6700 return true; 6701 } 6702 } 6703 return false; 6704 } 6705 6706 bool X86InstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 6707 int64_t Offset1, int64_t Offset2, 6708 unsigned NumLoads) const { 6709 assert(Offset2 > Offset1); 6710 if ((Offset2 - Offset1) / 8 > 64) 6711 return false; 6712 6713 unsigned Opc1 = Load1->getMachineOpcode(); 6714 unsigned Opc2 = Load2->getMachineOpcode(); 6715 if (Opc1 != Opc2) 6716 return false; // FIXME: overly conservative? 6717 6718 switch (Opc1) { 6719 default: break; 6720 case X86::LD_Fp32m: 6721 case X86::LD_Fp64m: 6722 case X86::LD_Fp80m: 6723 case X86::MMX_MOVD64rm: 6724 case X86::MMX_MOVQ64rm: 6725 return false; 6726 } 6727 6728 EVT VT = Load1->getValueType(0); 6729 switch (VT.getSimpleVT().SimpleTy) { 6730 default: 6731 // XMM registers. In 64-bit mode we can be a bit more aggressive since we 6732 // have 16 of them to play with. 6733 if (Subtarget.is64Bit()) { 6734 if (NumLoads >= 3) 6735 return false; 6736 } else if (NumLoads) { 6737 return false; 6738 } 6739 break; 6740 case MVT::i8: 6741 case MVT::i16: 6742 case MVT::i32: 6743 case MVT::i64: 6744 case MVT::f32: 6745 case MVT::f64: 6746 if (NumLoads) 6747 return false; 6748 break; 6749 } 6750 6751 return true; 6752 } 6753 6754 bool X86InstrInfo::shouldScheduleAdjacent(MachineInstr* First, 6755 MachineInstr *Second) const { 6756 // Check if this processor supports macro-fusion. Since this is a minor 6757 // heuristic, we haven't specifically reserved a feature. hasAVX is a decent 6758 // proxy for SandyBridge+. 6759 if (!Subtarget.hasAVX()) 6760 return false; 6761 6762 enum { 6763 FuseTest, 6764 FuseCmp, 6765 FuseInc 6766 } FuseKind; 6767 6768 switch(Second->getOpcode()) { 6769 default: 6770 return false; 6771 case X86::JE_1: 6772 case X86::JNE_1: 6773 case X86::JL_1: 6774 case X86::JLE_1: 6775 case X86::JG_1: 6776 case X86::JGE_1: 6777 FuseKind = FuseInc; 6778 break; 6779 case X86::JB_1: 6780 case X86::JBE_1: 6781 case X86::JA_1: 6782 case X86::JAE_1: 6783 FuseKind = FuseCmp; 6784 break; 6785 case X86::JS_1: 6786 case X86::JNS_1: 6787 case X86::JP_1: 6788 case X86::JNP_1: 6789 case X86::JO_1: 6790 case X86::JNO_1: 6791 FuseKind = FuseTest; 6792 break; 6793 } 6794 switch (First->getOpcode()) { 6795 default: 6796 return false; 6797 case X86::TEST8rr: 6798 case X86::TEST16rr: 6799 case X86::TEST32rr: 6800 case X86::TEST64rr: 6801 case X86::TEST8ri: 6802 case X86::TEST16ri: 6803 case X86::TEST32ri: 6804 case X86::TEST32i32: 6805 case X86::TEST64i32: 6806 case X86::TEST64ri32: 6807 case X86::TEST8rm: 6808 case X86::TEST16rm: 6809 case X86::TEST32rm: 6810 case X86::TEST64rm: 6811 case X86::TEST8ri_NOREX: 6812 case X86::AND16i16: 6813 case X86::AND16ri: 6814 case X86::AND16ri8: 6815 case X86::AND16rm: 6816 case X86::AND16rr: 6817 case X86::AND32i32: 6818 case X86::AND32ri: 6819 case X86::AND32ri8: 6820 case X86::AND32rm: 6821 case X86::AND32rr: 6822 case X86::AND64i32: 6823 case X86::AND64ri32: 6824 case X86::AND64ri8: 6825 case X86::AND64rm: 6826 case X86::AND64rr: 6827 case X86::AND8i8: 6828 case X86::AND8ri: 6829 case X86::AND8rm: 6830 case X86::AND8rr: 6831 return true; 6832 case X86::CMP16i16: 6833 case X86::CMP16ri: 6834 case X86::CMP16ri8: 6835 case X86::CMP16rm: 6836 case X86::CMP16rr: 6837 case X86::CMP32i32: 6838 case X86::CMP32ri: 6839 case X86::CMP32ri8: 6840 case X86::CMP32rm: 6841 case X86::CMP32rr: 6842 case X86::CMP64i32: 6843 case X86::CMP64ri32: 6844 case X86::CMP64ri8: 6845 case X86::CMP64rm: 6846 case X86::CMP64rr: 6847 case X86::CMP8i8: 6848 case X86::CMP8ri: 6849 case X86::CMP8rm: 6850 case X86::CMP8rr: 6851 case X86::ADD16i16: 6852 case X86::ADD16ri: 6853 case X86::ADD16ri8: 6854 case X86::ADD16ri8_DB: 6855 case X86::ADD16ri_DB: 6856 case X86::ADD16rm: 6857 case X86::ADD16rr: 6858 case X86::ADD16rr_DB: 6859 case X86::ADD32i32: 6860 case X86::ADD32ri: 6861 case X86::ADD32ri8: 6862 case X86::ADD32ri8_DB: 6863 case X86::ADD32ri_DB: 6864 case X86::ADD32rm: 6865 case X86::ADD32rr: 6866 case X86::ADD32rr_DB: 6867 case X86::ADD64i32: 6868 case X86::ADD64ri32: 6869 case X86::ADD64ri32_DB: 6870 case X86::ADD64ri8: 6871 case X86::ADD64ri8_DB: 6872 case X86::ADD64rm: 6873 case X86::ADD64rr: 6874 case X86::ADD64rr_DB: 6875 case X86::ADD8i8: 6876 case X86::ADD8mi: 6877 case X86::ADD8mr: 6878 case X86::ADD8ri: 6879 case X86::ADD8rm: 6880 case X86::ADD8rr: 6881 case X86::SUB16i16: 6882 case X86::SUB16ri: 6883 case X86::SUB16ri8: 6884 case X86::SUB16rm: 6885 case X86::SUB16rr: 6886 case X86::SUB32i32: 6887 case X86::SUB32ri: 6888 case X86::SUB32ri8: 6889 case X86::SUB32rm: 6890 case X86::SUB32rr: 6891 case X86::SUB64i32: 6892 case X86::SUB64ri32: 6893 case X86::SUB64ri8: 6894 case X86::SUB64rm: 6895 case X86::SUB64rr: 6896 case X86::SUB8i8: 6897 case X86::SUB8ri: 6898 case X86::SUB8rm: 6899 case X86::SUB8rr: 6900 return FuseKind == FuseCmp || FuseKind == FuseInc; 6901 case X86::INC16r: 6902 case X86::INC32r: 6903 case X86::INC64r: 6904 case X86::INC8r: 6905 case X86::DEC16r: 6906 case X86::DEC32r: 6907 case X86::DEC64r: 6908 case X86::DEC8r: 6909 return FuseKind == FuseInc; 6910 } 6911 } 6912 6913 bool X86InstrInfo:: 6914 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 6915 assert(Cond.size() == 1 && "Invalid X86 branch condition!"); 6916 X86::CondCode CC = static_cast<X86::CondCode>(Cond[0].getImm()); 6917 Cond[0].setImm(GetOppositeBranchCondition(CC)); 6918 return false; 6919 } 6920 6921 bool X86InstrInfo:: 6922 isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { 6923 // FIXME: Return false for x87 stack register classes for now. We can't 6924 // allow any loads of these registers before FpGet_ST0_80. 6925 return !(RC == &X86::CCRRegClass || RC == &X86::RFP32RegClass || 6926 RC == &X86::RFP64RegClass || RC == &X86::RFP80RegClass); 6927 } 6928 6929 /// Return a virtual register initialized with the 6930 /// the global base register value. Output instructions required to 6931 /// initialize the register in the function entry block, if necessary. 6932 /// 6933 /// TODO: Eliminate this and move the code to X86MachineFunctionInfo. 6934 /// 6935 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const { 6936 assert(!Subtarget.is64Bit() && 6937 "X86-64 PIC uses RIP relative addressing"); 6938 6939 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); 6940 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 6941 if (GlobalBaseReg != 0) 6942 return GlobalBaseReg; 6943 6944 // Create the register. The code to initialize it is inserted 6945 // later, by the CGBR pass (below). 6946 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 6947 GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); 6948 X86FI->setGlobalBaseReg(GlobalBaseReg); 6949 return GlobalBaseReg; 6950 } 6951 6952 // These are the replaceable SSE instructions. Some of these have Int variants 6953 // that we don't include here. We don't want to replace instructions selected 6954 // by intrinsics. 6955 static const uint16_t ReplaceableInstrs[][3] = { 6956 //PackedSingle PackedDouble PackedInt 6957 { X86::MOVAPSmr, X86::MOVAPDmr, X86::MOVDQAmr }, 6958 { X86::MOVAPSrm, X86::MOVAPDrm, X86::MOVDQArm }, 6959 { X86::MOVAPSrr, X86::MOVAPDrr, X86::MOVDQArr }, 6960 { X86::MOVUPSmr, X86::MOVUPDmr, X86::MOVDQUmr }, 6961 { X86::MOVUPSrm, X86::MOVUPDrm, X86::MOVDQUrm }, 6962 { X86::MOVLPSmr, X86::MOVLPDmr, X86::MOVPQI2QImr }, 6963 { X86::MOVNTPSmr, X86::MOVNTPDmr, X86::MOVNTDQmr }, 6964 { X86::ANDNPSrm, X86::ANDNPDrm, X86::PANDNrm }, 6965 { X86::ANDNPSrr, X86::ANDNPDrr, X86::PANDNrr }, 6966 { X86::ANDPSrm, X86::ANDPDrm, X86::PANDrm }, 6967 { X86::ANDPSrr, X86::ANDPDrr, X86::PANDrr }, 6968 { X86::ORPSrm, X86::ORPDrm, X86::PORrm }, 6969 { X86::ORPSrr, X86::ORPDrr, X86::PORrr }, 6970 { X86::XORPSrm, X86::XORPDrm, X86::PXORrm }, 6971 { X86::XORPSrr, X86::XORPDrr, X86::PXORrr }, 6972 // AVX 128-bit support 6973 { X86::VMOVAPSmr, X86::VMOVAPDmr, X86::VMOVDQAmr }, 6974 { X86::VMOVAPSrm, X86::VMOVAPDrm, X86::VMOVDQArm }, 6975 { X86::VMOVAPSrr, X86::VMOVAPDrr, X86::VMOVDQArr }, 6976 { X86::VMOVUPSmr, X86::VMOVUPDmr, X86::VMOVDQUmr }, 6977 { X86::VMOVUPSrm, X86::VMOVUPDrm, X86::VMOVDQUrm }, 6978 { X86::VMOVLPSmr, X86::VMOVLPDmr, X86::VMOVPQI2QImr }, 6979 { X86::VMOVNTPSmr, X86::VMOVNTPDmr, X86::VMOVNTDQmr }, 6980 { X86::VANDNPSrm, X86::VANDNPDrm, X86::VPANDNrm }, 6981 { X86::VANDNPSrr, X86::VANDNPDrr, X86::VPANDNrr }, 6982 { X86::VANDPSrm, X86::VANDPDrm, X86::VPANDrm }, 6983 { X86::VANDPSrr, X86::VANDPDrr, X86::VPANDrr }, 6984 { X86::VORPSrm, X86::VORPDrm, X86::VPORrm }, 6985 { X86::VORPSrr, X86::VORPDrr, X86::VPORrr }, 6986 { X86::VXORPSrm, X86::VXORPDrm, X86::VPXORrm }, 6987 { X86::VXORPSrr, X86::VXORPDrr, X86::VPXORrr }, 6988 // AVX 256-bit support 6989 { X86::VMOVAPSYmr, X86::VMOVAPDYmr, X86::VMOVDQAYmr }, 6990 { X86::VMOVAPSYrm, X86::VMOVAPDYrm, X86::VMOVDQAYrm }, 6991 { X86::VMOVAPSYrr, X86::VMOVAPDYrr, X86::VMOVDQAYrr }, 6992 { X86::VMOVUPSYmr, X86::VMOVUPDYmr, X86::VMOVDQUYmr }, 6993 { X86::VMOVUPSYrm, X86::VMOVUPDYrm, X86::VMOVDQUYrm }, 6994 { X86::VMOVNTPSYmr, X86::VMOVNTPDYmr, X86::VMOVNTDQYmr } 6995 }; 6996 6997 static const uint16_t ReplaceableInstrsAVX2[][3] = { 6998 //PackedSingle PackedDouble PackedInt 6999 { X86::VANDNPSYrm, X86::VANDNPDYrm, X86::VPANDNYrm }, 7000 { X86::VANDNPSYrr, X86::VANDNPDYrr, X86::VPANDNYrr }, 7001 { X86::VANDPSYrm, X86::VANDPDYrm, X86::VPANDYrm }, 7002 { X86::VANDPSYrr, X86::VANDPDYrr, X86::VPANDYrr }, 7003 { X86::VORPSYrm, X86::VORPDYrm, X86::VPORYrm }, 7004 { X86::VORPSYrr, X86::VORPDYrr, X86::VPORYrr }, 7005 { X86::VXORPSYrm, X86::VXORPDYrm, X86::VPXORYrm }, 7006 { X86::VXORPSYrr, X86::VXORPDYrr, X86::VPXORYrr }, 7007 { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr }, 7008 { X86::VEXTRACTF128rr, X86::VEXTRACTF128rr, X86::VEXTRACTI128rr }, 7009 { X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm }, 7010 { X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr }, 7011 { X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm }, 7012 { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }, 7013 { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm}, 7014 { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr}, 7015 { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr}, 7016 { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm}, 7017 { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr}, 7018 { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm} 7019 }; 7020 7021 // FIXME: Some shuffle and unpack instructions have equivalents in different 7022 // domains, but they require a bit more work than just switching opcodes. 7023 7024 static const uint16_t *lookup(unsigned opcode, unsigned domain) { 7025 for (const uint16_t (&Row)[3] : ReplaceableInstrs) 7026 if (Row[domain-1] == opcode) 7027 return Row; 7028 return nullptr; 7029 } 7030 7031 static const uint16_t *lookupAVX2(unsigned opcode, unsigned domain) { 7032 for (const uint16_t (&Row)[3] : ReplaceableInstrsAVX2) 7033 if (Row[domain-1] == opcode) 7034 return Row; 7035 return nullptr; 7036 } 7037 7038 std::pair<uint16_t, uint16_t> 7039 X86InstrInfo::getExecutionDomain(const MachineInstr *MI) const { 7040 uint16_t domain = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 7041 bool hasAVX2 = Subtarget.hasAVX2(); 7042 uint16_t validDomains = 0; 7043 if (domain && lookup(MI->getOpcode(), domain)) 7044 validDomains = 0xe; 7045 else if (domain && lookupAVX2(MI->getOpcode(), domain)) 7046 validDomains = hasAVX2 ? 0xe : 0x6; 7047 return std::make_pair(domain, validDomains); 7048 } 7049 7050 void X86InstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const { 7051 assert(Domain>0 && Domain<4 && "Invalid execution domain"); 7052 uint16_t dom = (MI->getDesc().TSFlags >> X86II::SSEDomainShift) & 3; 7053 assert(dom && "Not an SSE instruction"); 7054 const uint16_t *table = lookup(MI->getOpcode(), dom); 7055 if (!table) { // try the other table 7056 assert((Subtarget.hasAVX2() || Domain < 3) && 7057 "256-bit vector operations only available in AVX2"); 7058 table = lookupAVX2(MI->getOpcode(), dom); 7059 } 7060 assert(table && "Cannot change domain"); 7061 MI->setDesc(get(table[Domain-1])); 7062 } 7063 7064 /// Return the noop instruction to use for a noop. 7065 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { 7066 NopInst.setOpcode(X86::NOOP); 7067 } 7068 7069 // This code must remain in sync with getJumpInstrTableEntryBound in this class! 7070 // In particular, getJumpInstrTableEntryBound must always return an upper bound 7071 // on the encoding lengths of the instructions generated by 7072 // getUnconditionalBranch and getTrap. 7073 void X86InstrInfo::getUnconditionalBranch( 7074 MCInst &Branch, const MCSymbolRefExpr *BranchTarget) const { 7075 Branch.setOpcode(X86::JMP_1); 7076 Branch.addOperand(MCOperand::createExpr(BranchTarget)); 7077 } 7078 7079 // This code must remain in sync with getJumpInstrTableEntryBound in this class! 7080 // In particular, getJumpInstrTableEntryBound must always return an upper bound 7081 // on the encoding lengths of the instructions generated by 7082 // getUnconditionalBranch and getTrap. 7083 void X86InstrInfo::getTrap(MCInst &MI) const { 7084 MI.setOpcode(X86::TRAP); 7085 } 7086 7087 // See getTrap and getUnconditionalBranch for conditions on the value returned 7088 // by this function. 7089 unsigned X86InstrInfo::getJumpInstrTableEntryBound() const { 7090 // 5 bytes suffice: JMP_4 Symbol@PLT is uses 1 byte (E9) for the JMP_4 and 4 7091 // bytes for the symbol offset. And TRAP is ud2, which is two bytes (0F 0B). 7092 return 5; 7093 } 7094 7095 bool X86InstrInfo::isHighLatencyDef(int opc) const { 7096 switch (opc) { 7097 default: return false; 7098 case X86::DIVSDrm: 7099 case X86::DIVSDrm_Int: 7100 case X86::DIVSDrr: 7101 case X86::DIVSDrr_Int: 7102 case X86::DIVSSrm: 7103 case X86::DIVSSrm_Int: 7104 case X86::DIVSSrr: 7105 case X86::DIVSSrr_Int: 7106 case X86::SQRTPDm: 7107 case X86::SQRTPDr: 7108 case X86::SQRTPSm: 7109 case X86::SQRTPSr: 7110 case X86::SQRTSDm: 7111 case X86::SQRTSDm_Int: 7112 case X86::SQRTSDr: 7113 case X86::SQRTSDr_Int: 7114 case X86::SQRTSSm: 7115 case X86::SQRTSSm_Int: 7116 case X86::SQRTSSr: 7117 case X86::SQRTSSr_Int: 7118 // AVX instructions with high latency 7119 case X86::VDIVSDrm: 7120 case X86::VDIVSDrm_Int: 7121 case X86::VDIVSDrr: 7122 case X86::VDIVSDrr_Int: 7123 case X86::VDIVSSrm: 7124 case X86::VDIVSSrm_Int: 7125 case X86::VDIVSSrr: 7126 case X86::VDIVSSrr_Int: 7127 case X86::VSQRTPDm: 7128 case X86::VSQRTPDr: 7129 case X86::VSQRTPSm: 7130 case X86::VSQRTPSr: 7131 case X86::VSQRTSDm: 7132 case X86::VSQRTSDm_Int: 7133 case X86::VSQRTSDr: 7134 case X86::VSQRTSSm: 7135 case X86::VSQRTSSm_Int: 7136 case X86::VSQRTSSr: 7137 case X86::VSQRTPDZm: 7138 case X86::VSQRTPDZr: 7139 case X86::VSQRTPSZm: 7140 case X86::VSQRTPSZr: 7141 case X86::VSQRTSDZm: 7142 case X86::VSQRTSDZm_Int: 7143 case X86::VSQRTSDZr: 7144 case X86::VSQRTSSZm_Int: 7145 case X86::VSQRTSSZr: 7146 case X86::VSQRTSSZm: 7147 case X86::VDIVSDZrm: 7148 case X86::VDIVSDZrr: 7149 case X86::VDIVSSZrm: 7150 case X86::VDIVSSZrr: 7151 7152 case X86::VGATHERQPSZrm: 7153 case X86::VGATHERQPDZrm: 7154 case X86::VGATHERDPDZrm: 7155 case X86::VGATHERDPSZrm: 7156 case X86::VPGATHERQDZrm: 7157 case X86::VPGATHERQQZrm: 7158 case X86::VPGATHERDDZrm: 7159 case X86::VPGATHERDQZrm: 7160 case X86::VSCATTERQPDZmr: 7161 case X86::VSCATTERQPSZmr: 7162 case X86::VSCATTERDPDZmr: 7163 case X86::VSCATTERDPSZmr: 7164 case X86::VPSCATTERQDZmr: 7165 case X86::VPSCATTERQQZmr: 7166 case X86::VPSCATTERDDZmr: 7167 case X86::VPSCATTERDQZmr: 7168 return true; 7169 } 7170 } 7171 7172 bool X86InstrInfo:: 7173 hasHighOperandLatency(const TargetSchedModel &SchedModel, 7174 const MachineRegisterInfo *MRI, 7175 const MachineInstr *DefMI, unsigned DefIdx, 7176 const MachineInstr *UseMI, unsigned UseIdx) const { 7177 return isHighLatencyDef(DefMI->getOpcode()); 7178 } 7179 7180 bool X86InstrInfo::hasReassociableOperands(const MachineInstr &Inst, 7181 const MachineBasicBlock *MBB) const { 7182 assert((Inst.getNumOperands() == 3 || Inst.getNumOperands() == 4) && 7183 "Reassociation needs binary operators"); 7184 7185 // Integer binary math/logic instructions have a third source operand: 7186 // the EFLAGS register. That operand must be both defined here and never 7187 // used; ie, it must be dead. If the EFLAGS operand is live, then we can 7188 // not change anything because rearranging the operands could affect other 7189 // instructions that depend on the exact status flags (zero, sign, etc.) 7190 // that are set by using these particular operands with this operation. 7191 if (Inst.getNumOperands() == 4) { 7192 assert(Inst.getOperand(3).isReg() && 7193 Inst.getOperand(3).getReg() == X86::EFLAGS && 7194 "Unexpected operand in reassociable instruction"); 7195 if (!Inst.getOperand(3).isDead()) 7196 return false; 7197 } 7198 7199 return TargetInstrInfo::hasReassociableOperands(Inst, MBB); 7200 } 7201 7202 // TODO: There are many more machine instruction opcodes to match: 7203 // 1. Other data types (integer, vectors) 7204 // 2. Other math / logic operations (xor, or) 7205 // 3. Other forms of the same operation (intrinsics and other variants) 7206 bool X86InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst) const { 7207 switch (Inst.getOpcode()) { 7208 case X86::AND8rr: 7209 case X86::AND16rr: 7210 case X86::AND32rr: 7211 case X86::AND64rr: 7212 case X86::OR8rr: 7213 case X86::OR16rr: 7214 case X86::OR32rr: 7215 case X86::OR64rr: 7216 case X86::XOR8rr: 7217 case X86::XOR16rr: 7218 case X86::XOR32rr: 7219 case X86::XOR64rr: 7220 case X86::IMUL16rr: 7221 case X86::IMUL32rr: 7222 case X86::IMUL64rr: 7223 case X86::PANDrr: 7224 case X86::PORrr: 7225 case X86::PXORrr: 7226 case X86::VPANDrr: 7227 case X86::VPANDYrr: 7228 case X86::VPORrr: 7229 case X86::VPORYrr: 7230 case X86::VPXORrr: 7231 case X86::VPXORYrr: 7232 // Normal min/max instructions are not commutative because of NaN and signed 7233 // zero semantics, but these are. Thus, there's no need to check for global 7234 // relaxed math; the instructions themselves have the properties we need. 7235 case X86::MAXCPDrr: 7236 case X86::MAXCPSrr: 7237 case X86::MAXCSDrr: 7238 case X86::MAXCSSrr: 7239 case X86::MINCPDrr: 7240 case X86::MINCPSrr: 7241 case X86::MINCSDrr: 7242 case X86::MINCSSrr: 7243 case X86::VMAXCPDrr: 7244 case X86::VMAXCPSrr: 7245 case X86::VMAXCPDYrr: 7246 case X86::VMAXCPSYrr: 7247 case X86::VMAXCSDrr: 7248 case X86::VMAXCSSrr: 7249 case X86::VMINCPDrr: 7250 case X86::VMINCPSrr: 7251 case X86::VMINCPDYrr: 7252 case X86::VMINCPSYrr: 7253 case X86::VMINCSDrr: 7254 case X86::VMINCSSrr: 7255 return true; 7256 case X86::ADDPDrr: 7257 case X86::ADDPSrr: 7258 case X86::ADDSDrr: 7259 case X86::ADDSSrr: 7260 case X86::MULPDrr: 7261 case X86::MULPSrr: 7262 case X86::MULSDrr: 7263 case X86::MULSSrr: 7264 case X86::VADDPDrr: 7265 case X86::VADDPSrr: 7266 case X86::VADDPDYrr: 7267 case X86::VADDPSYrr: 7268 case X86::VADDSDrr: 7269 case X86::VADDSSrr: 7270 case X86::VMULPDrr: 7271 case X86::VMULPSrr: 7272 case X86::VMULPDYrr: 7273 case X86::VMULPSYrr: 7274 case X86::VMULSDrr: 7275 case X86::VMULSSrr: 7276 return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath; 7277 default: 7278 return false; 7279 } 7280 } 7281 7282 /// This is an architecture-specific helper function of reassociateOps. 7283 /// Set special operand attributes for new instructions after reassociation. 7284 void X86InstrInfo::setSpecialOperandAttr(MachineInstr &OldMI1, 7285 MachineInstr &OldMI2, 7286 MachineInstr &NewMI1, 7287 MachineInstr &NewMI2) const { 7288 // Integer instructions define an implicit EFLAGS source register operand as 7289 // the third source (fourth total) operand. 7290 if (OldMI1.getNumOperands() != 4 || OldMI2.getNumOperands() != 4) 7291 return; 7292 7293 assert(NewMI1.getNumOperands() == 4 && NewMI2.getNumOperands() == 4 && 7294 "Unexpected instruction type for reassociation"); 7295 7296 MachineOperand &OldOp1 = OldMI1.getOperand(3); 7297 MachineOperand &OldOp2 = OldMI2.getOperand(3); 7298 MachineOperand &NewOp1 = NewMI1.getOperand(3); 7299 MachineOperand &NewOp2 = NewMI2.getOperand(3); 7300 7301 assert(OldOp1.isReg() && OldOp1.getReg() == X86::EFLAGS && OldOp1.isDead() && 7302 "Must have dead EFLAGS operand in reassociable instruction"); 7303 assert(OldOp2.isReg() && OldOp2.getReg() == X86::EFLAGS && OldOp2.isDead() && 7304 "Must have dead EFLAGS operand in reassociable instruction"); 7305 7306 (void)OldOp1; 7307 (void)OldOp2; 7308 7309 assert(NewOp1.isReg() && NewOp1.getReg() == X86::EFLAGS && 7310 "Unexpected operand in reassociable instruction"); 7311 assert(NewOp2.isReg() && NewOp2.getReg() == X86::EFLAGS && 7312 "Unexpected operand in reassociable instruction"); 7313 7314 // Mark the new EFLAGS operands as dead to be helpful to subsequent iterations 7315 // of this pass or other passes. The EFLAGS operands must be dead in these new 7316 // instructions because the EFLAGS operands in the original instructions must 7317 // be dead in order for reassociation to occur. 7318 NewOp1.setIsDead(); 7319 NewOp2.setIsDead(); 7320 } 7321 7322 std::pair<unsigned, unsigned> 7323 X86InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 7324 return std::make_pair(TF, 0u); 7325 } 7326 7327 ArrayRef<std::pair<unsigned, const char *>> 7328 X86InstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 7329 using namespace X86II; 7330 static const std::pair<unsigned, const char *> TargetFlags[] = { 7331 {MO_GOT_ABSOLUTE_ADDRESS, "x86-got-absolute-address"}, 7332 {MO_PIC_BASE_OFFSET, "x86-pic-base-offset"}, 7333 {MO_GOT, "x86-got"}, 7334 {MO_GOTOFF, "x86-gotoff"}, 7335 {MO_GOTPCREL, "x86-gotpcrel"}, 7336 {MO_PLT, "x86-plt"}, 7337 {MO_TLSGD, "x86-tlsgd"}, 7338 {MO_TLSLD, "x86-tlsld"}, 7339 {MO_TLSLDM, "x86-tlsldm"}, 7340 {MO_GOTTPOFF, "x86-gottpoff"}, 7341 {MO_INDNTPOFF, "x86-indntpoff"}, 7342 {MO_TPOFF, "x86-tpoff"}, 7343 {MO_DTPOFF, "x86-dtpoff"}, 7344 {MO_NTPOFF, "x86-ntpoff"}, 7345 {MO_GOTNTPOFF, "x86-gotntpoff"}, 7346 {MO_DLLIMPORT, "x86-dllimport"}, 7347 {MO_DARWIN_NONLAZY, "x86-darwin-nonlazy"}, 7348 {MO_DARWIN_NONLAZY_PIC_BASE, "x86-darwin-nonlazy-pic-base"}, 7349 {MO_TLVP, "x86-tlvp"}, 7350 {MO_TLVP_PIC_BASE, "x86-tlvp-pic-base"}, 7351 {MO_SECREL, "x86-secrel"}}; 7352 return makeArrayRef(TargetFlags); 7353 } 7354 7355 namespace { 7356 /// Create Global Base Reg pass. This initializes the PIC 7357 /// global base register for x86-32. 7358 struct CGBR : public MachineFunctionPass { 7359 static char ID; 7360 CGBR() : MachineFunctionPass(ID) {} 7361 7362 bool runOnMachineFunction(MachineFunction &MF) override { 7363 const X86TargetMachine *TM = 7364 static_cast<const X86TargetMachine *>(&MF.getTarget()); 7365 const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>(); 7366 7367 // Don't do anything if this is 64-bit as 64-bit PIC 7368 // uses RIP relative addressing. 7369 if (STI.is64Bit()) 7370 return false; 7371 7372 // Only emit a global base reg in PIC mode. 7373 if (!TM->isPositionIndependent()) 7374 return false; 7375 7376 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); 7377 unsigned GlobalBaseReg = X86FI->getGlobalBaseReg(); 7378 7379 // If we didn't need a GlobalBaseReg, don't insert code. 7380 if (GlobalBaseReg == 0) 7381 return false; 7382 7383 // Insert the set of GlobalBaseReg into the first MBB of the function 7384 MachineBasicBlock &FirstMBB = MF.front(); 7385 MachineBasicBlock::iterator MBBI = FirstMBB.begin(); 7386 DebugLoc DL = FirstMBB.findDebugLoc(MBBI); 7387 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 7388 const X86InstrInfo *TII = STI.getInstrInfo(); 7389 7390 unsigned PC; 7391 if (STI.isPICStyleGOT()) 7392 PC = RegInfo.createVirtualRegister(&X86::GR32RegClass); 7393 else 7394 PC = GlobalBaseReg; 7395 7396 // Operand of MovePCtoStack is completely ignored by asm printer. It's 7397 // only used in JIT code emission as displacement to pc. 7398 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0); 7399 7400 // If we're using vanilla 'GOT' PIC style, we should use relative addressing 7401 // not to pc, but to _GLOBAL_OFFSET_TABLE_ external. 7402 if (STI.isPICStyleGOT()) { 7403 // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register 7404 BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg) 7405 .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_", 7406 X86II::MO_GOT_ABSOLUTE_ADDRESS); 7407 } 7408 7409 return true; 7410 } 7411 7412 const char *getPassName() const override { 7413 return "X86 PIC Global Base Reg Initialization"; 7414 } 7415 7416 void getAnalysisUsage(AnalysisUsage &AU) const override { 7417 AU.setPreservesCFG(); 7418 MachineFunctionPass::getAnalysisUsage(AU); 7419 } 7420 }; 7421 } 7422 7423 char CGBR::ID = 0; 7424 FunctionPass* 7425 llvm::createX86GlobalBaseRegPass() { return new CGBR(); } 7426 7427 namespace { 7428 struct LDTLSCleanup : public MachineFunctionPass { 7429 static char ID; 7430 LDTLSCleanup() : MachineFunctionPass(ID) {} 7431 7432 bool runOnMachineFunction(MachineFunction &MF) override { 7433 if (skipFunction(*MF.getFunction())) 7434 return false; 7435 7436 X86MachineFunctionInfo *MFI = MF.getInfo<X86MachineFunctionInfo>(); 7437 if (MFI->getNumLocalDynamicTLSAccesses() < 2) { 7438 // No point folding accesses if there isn't at least two. 7439 return false; 7440 } 7441 7442 MachineDominatorTree *DT = &getAnalysis<MachineDominatorTree>(); 7443 return VisitNode(DT->getRootNode(), 0); 7444 } 7445 7446 // Visit the dominator subtree rooted at Node in pre-order. 7447 // If TLSBaseAddrReg is non-null, then use that to replace any 7448 // TLS_base_addr instructions. Otherwise, create the register 7449 // when the first such instruction is seen, and then use it 7450 // as we encounter more instructions. 7451 bool VisitNode(MachineDomTreeNode *Node, unsigned TLSBaseAddrReg) { 7452 MachineBasicBlock *BB = Node->getBlock(); 7453 bool Changed = false; 7454 7455 // Traverse the current block. 7456 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; 7457 ++I) { 7458 switch (I->getOpcode()) { 7459 case X86::TLS_base_addr32: 7460 case X86::TLS_base_addr64: 7461 if (TLSBaseAddrReg) 7462 I = ReplaceTLSBaseAddrCall(I, TLSBaseAddrReg); 7463 else 7464 I = SetRegister(I, &TLSBaseAddrReg); 7465 Changed = true; 7466 break; 7467 default: 7468 break; 7469 } 7470 } 7471 7472 // Visit the children of this block in the dominator tree. 7473 for (MachineDomTreeNode::iterator I = Node->begin(), E = Node->end(); 7474 I != E; ++I) { 7475 Changed |= VisitNode(*I, TLSBaseAddrReg); 7476 } 7477 7478 return Changed; 7479 } 7480 7481 // Replace the TLS_base_addr instruction I with a copy from 7482 // TLSBaseAddrReg, returning the new instruction. 7483 MachineInstr *ReplaceTLSBaseAddrCall(MachineInstr *I, 7484 unsigned TLSBaseAddrReg) { 7485 MachineFunction *MF = I->getParent()->getParent(); 7486 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>(); 7487 const bool is64Bit = STI.is64Bit(); 7488 const X86InstrInfo *TII = STI.getInstrInfo(); 7489 7490 // Insert a Copy from TLSBaseAddrReg to RAX/EAX. 7491 MachineInstr *Copy = BuildMI(*I->getParent(), I, I->getDebugLoc(), 7492 TII->get(TargetOpcode::COPY), 7493 is64Bit ? X86::RAX : X86::EAX) 7494 .addReg(TLSBaseAddrReg); 7495 7496 // Erase the TLS_base_addr instruction. 7497 I->eraseFromParent(); 7498 7499 return Copy; 7500 } 7501 7502 // Create a virtal register in *TLSBaseAddrReg, and populate it by 7503 // inserting a copy instruction after I. Returns the new instruction. 7504 MachineInstr *SetRegister(MachineInstr *I, unsigned *TLSBaseAddrReg) { 7505 MachineFunction *MF = I->getParent()->getParent(); 7506 const X86Subtarget &STI = MF->getSubtarget<X86Subtarget>(); 7507 const bool is64Bit = STI.is64Bit(); 7508 const X86InstrInfo *TII = STI.getInstrInfo(); 7509 7510 // Create a virtual register for the TLS base address. 7511 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 7512 *TLSBaseAddrReg = RegInfo.createVirtualRegister(is64Bit 7513 ? &X86::GR64RegClass 7514 : &X86::GR32RegClass); 7515 7516 // Insert a copy from RAX/EAX to TLSBaseAddrReg. 7517 MachineInstr *Next = I->getNextNode(); 7518 MachineInstr *Copy = BuildMI(*I->getParent(), Next, I->getDebugLoc(), 7519 TII->get(TargetOpcode::COPY), 7520 *TLSBaseAddrReg) 7521 .addReg(is64Bit ? X86::RAX : X86::EAX); 7522 7523 return Copy; 7524 } 7525 7526 const char *getPassName() const override { 7527 return "Local Dynamic TLS Access Clean-up"; 7528 } 7529 7530 void getAnalysisUsage(AnalysisUsage &AU) const override { 7531 AU.setPreservesCFG(); 7532 AU.addRequired<MachineDominatorTree>(); 7533 MachineFunctionPass::getAnalysisUsage(AU); 7534 } 7535 }; 7536 } 7537 7538 char LDTLSCleanup::ID = 0; 7539 FunctionPass* 7540 llvm::createCleanupLocalDynamicTLSPass() { return new LDTLSCleanup(); } 7541