1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This implements the TargetLoweringBase class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/BitVector.h" 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/StringExtras.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/ADT/Triple.h" 19 #include "llvm/ADT/Twine.h" 20 #include "llvm/Analysis/Loads.h" 21 #include "llvm/CodeGen/Analysis.h" 22 #include "llvm/CodeGen/ISDOpcodes.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineMemOperand.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/RuntimeLibcalls.h" 32 #include "llvm/CodeGen/StackMaps.h" 33 #include "llvm/CodeGen/TargetLowering.h" 34 #include "llvm/CodeGen/TargetOpcodes.h" 35 #include "llvm/CodeGen/TargetRegisterInfo.h" 36 #include "llvm/CodeGen/ValueTypes.h" 37 #include "llvm/IR/Attributes.h" 38 #include "llvm/IR/CallingConv.h" 39 #include "llvm/IR/DataLayout.h" 40 #include "llvm/IR/DerivedTypes.h" 41 #include "llvm/IR/Function.h" 42 #include "llvm/IR/GlobalValue.h" 43 #include "llvm/IR/GlobalVariable.h" 44 #include "llvm/IR/IRBuilder.h" 45 #include "llvm/IR/Module.h" 46 #include "llvm/IR/Type.h" 47 #include "llvm/Support/BranchProbability.h" 48 #include "llvm/Support/Casting.h" 49 #include "llvm/Support/CommandLine.h" 50 #include "llvm/Support/Compiler.h" 51 #include "llvm/Support/ErrorHandling.h" 52 #include "llvm/Support/MachineValueType.h" 53 #include "llvm/Support/MathExtras.h" 54 #include "llvm/Target/TargetMachine.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstddef> 58 #include <cstdint> 59 #include <cstring> 60 #include <iterator> 61 #include <string> 62 #include <tuple> 63 #include <utility> 64 65 using namespace llvm; 66 67 static cl::opt<bool> JumpIsExpensiveOverride( 68 "jump-is-expensive", cl::init(false), 69 cl::desc("Do not create extra branches to split comparison logic."), 70 cl::Hidden); 71 72 static cl::opt<unsigned> MinimumJumpTableEntries 73 ("min-jump-table-entries", cl::init(4), cl::Hidden, 74 cl::desc("Set minimum number of entries to use a jump table.")); 75 76 static cl::opt<unsigned> MaximumJumpTableSize 77 ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden, 78 cl::desc("Set maximum size of jump tables.")); 79 80 /// Minimum jump table density for normal functions. 81 static cl::opt<unsigned> 82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 83 cl::desc("Minimum density for building a jump table in " 84 "a normal function")); 85 86 /// Minimum jump table density for -Os or -Oz functions. 87 static cl::opt<unsigned> OptsizeJumpTableDensity( 88 "optsize-jump-table-density", cl::init(40), cl::Hidden, 89 cl::desc("Minimum density for building a jump table in " 90 "an optsize function")); 91 92 // FIXME: This option is only to test if the strict fp operation processed 93 // correctly by preventing mutating strict fp operation to normal fp operation 94 // during development. When the backend supports strict float operation, this 95 // option will be meaningless. 96 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation", 97 cl::desc("Don't mutate strict-float node to a legalize node"), 98 cl::init(false), cl::Hidden); 99 100 static bool darwinHasSinCos(const Triple &TT) { 101 assert(TT.isOSDarwin() && "should be called with darwin triple"); 102 // Don't bother with 32 bit x86. 103 if (TT.getArch() == Triple::x86) 104 return false; 105 // Macos < 10.9 has no sincos_stret. 106 if (TT.isMacOSX()) 107 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 108 // iOS < 7.0 has no sincos_stret. 109 if (TT.isiOS()) 110 return !TT.isOSVersionLT(7, 0); 111 // Any other darwin such as WatchOS/TvOS is new enough. 112 return true; 113 } 114 115 // Although this default value is arbitrary, it is not random. It is assumed 116 // that a condition that evaluates the same way by a higher percentage than this 117 // is best represented as control flow. Therefore, the default value N should be 118 // set such that the win from N% correct executions is greater than the loss 119 // from (100 - N)% mispredicted executions for the majority of intended targets. 120 static cl::opt<int> MinPercentageForPredictableBranch( 121 "min-predictable-branch", cl::init(99), 122 cl::desc("Minimum percentage (0-100) that a condition must be either true " 123 "or false to assume that the condition is predictable"), 124 cl::Hidden); 125 126 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 127 #define HANDLE_LIBCALL(code, name) \ 128 setLibcallName(RTLIB::code, name); 129 #include "llvm/IR/RuntimeLibcalls.def" 130 #undef HANDLE_LIBCALL 131 // Initialize calling conventions to their default. 132 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 133 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 134 135 // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf". 136 if (TT.getArch() == Triple::ppc || TT.isPPC64()) { 137 setLibcallName(RTLIB::ADD_F128, "__addkf3"); 138 setLibcallName(RTLIB::SUB_F128, "__subkf3"); 139 setLibcallName(RTLIB::MUL_F128, "__mulkf3"); 140 setLibcallName(RTLIB::DIV_F128, "__divkf3"); 141 setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2"); 142 setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2"); 143 setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2"); 144 setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2"); 145 setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi"); 146 setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi"); 147 setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi"); 148 setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi"); 149 setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf"); 150 setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf"); 151 setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf"); 152 setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf"); 153 setLibcallName(RTLIB::OEQ_F128, "__eqkf2"); 154 setLibcallName(RTLIB::UNE_F128, "__nekf2"); 155 setLibcallName(RTLIB::OGE_F128, "__gekf2"); 156 setLibcallName(RTLIB::OLT_F128, "__ltkf2"); 157 setLibcallName(RTLIB::OLE_F128, "__lekf2"); 158 setLibcallName(RTLIB::OGT_F128, "__gtkf2"); 159 setLibcallName(RTLIB::UO_F128, "__unordkf2"); 160 } 161 162 // A few names are different on particular architectures or environments. 163 if (TT.isOSDarwin()) { 164 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 165 // of the gnueabi-style __gnu_*_ieee. 166 // FIXME: What about other targets? 167 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 168 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 169 170 // Some darwins have an optimized __bzero/bzero function. 171 switch (TT.getArch()) { 172 case Triple::x86: 173 case Triple::x86_64: 174 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 175 setLibcallName(RTLIB::BZERO, "__bzero"); 176 break; 177 case Triple::aarch64: 178 case Triple::aarch64_32: 179 setLibcallName(RTLIB::BZERO, "bzero"); 180 break; 181 default: 182 break; 183 } 184 185 if (darwinHasSinCos(TT)) { 186 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 187 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 188 if (TT.isWatchABI()) { 189 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 190 CallingConv::ARM_AAPCS_VFP); 191 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 192 CallingConv::ARM_AAPCS_VFP); 193 } 194 } 195 } else { 196 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 197 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 198 } 199 200 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 201 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 202 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 203 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 204 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 205 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 206 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 207 } 208 209 if (TT.isPS4CPU()) { 210 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 211 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 212 } 213 214 if (TT.isOSOpenBSD()) { 215 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 216 } 217 } 218 219 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 220 /// UNKNOWN_LIBCALL if there is none. 221 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 222 if (OpVT == MVT::f16) { 223 if (RetVT == MVT::f32) 224 return FPEXT_F16_F32; 225 } else if (OpVT == MVT::f32) { 226 if (RetVT == MVT::f64) 227 return FPEXT_F32_F64; 228 if (RetVT == MVT::f128) 229 return FPEXT_F32_F128; 230 if (RetVT == MVT::ppcf128) 231 return FPEXT_F32_PPCF128; 232 } else if (OpVT == MVT::f64) { 233 if (RetVT == MVT::f128) 234 return FPEXT_F64_F128; 235 else if (RetVT == MVT::ppcf128) 236 return FPEXT_F64_PPCF128; 237 } else if (OpVT == MVT::f80) { 238 if (RetVT == MVT::f128) 239 return FPEXT_F80_F128; 240 } 241 242 return UNKNOWN_LIBCALL; 243 } 244 245 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 246 /// UNKNOWN_LIBCALL if there is none. 247 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 248 if (RetVT == MVT::f16) { 249 if (OpVT == MVT::f32) 250 return FPROUND_F32_F16; 251 if (OpVT == MVT::f64) 252 return FPROUND_F64_F16; 253 if (OpVT == MVT::f80) 254 return FPROUND_F80_F16; 255 if (OpVT == MVT::f128) 256 return FPROUND_F128_F16; 257 if (OpVT == MVT::ppcf128) 258 return FPROUND_PPCF128_F16; 259 } else if (RetVT == MVT::f32) { 260 if (OpVT == MVT::f64) 261 return FPROUND_F64_F32; 262 if (OpVT == MVT::f80) 263 return FPROUND_F80_F32; 264 if (OpVT == MVT::f128) 265 return FPROUND_F128_F32; 266 if (OpVT == MVT::ppcf128) 267 return FPROUND_PPCF128_F32; 268 } else if (RetVT == MVT::f64) { 269 if (OpVT == MVT::f80) 270 return FPROUND_F80_F64; 271 if (OpVT == MVT::f128) 272 return FPROUND_F128_F64; 273 if (OpVT == MVT::ppcf128) 274 return FPROUND_PPCF128_F64; 275 } else if (RetVT == MVT::f80) { 276 if (OpVT == MVT::f128) 277 return FPROUND_F128_F80; 278 } 279 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOSINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOSINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOSINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOSINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOSINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOSINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOSINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOSINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOSINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOSINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOSINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOSINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOSINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOSINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOSINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::f32) { 329 if (RetVT == MVT::i32) 330 return FPTOUINT_F32_I32; 331 if (RetVT == MVT::i64) 332 return FPTOUINT_F32_I64; 333 if (RetVT == MVT::i128) 334 return FPTOUINT_F32_I128; 335 } else if (OpVT == MVT::f64) { 336 if (RetVT == MVT::i32) 337 return FPTOUINT_F64_I32; 338 if (RetVT == MVT::i64) 339 return FPTOUINT_F64_I64; 340 if (RetVT == MVT::i128) 341 return FPTOUINT_F64_I128; 342 } else if (OpVT == MVT::f80) { 343 if (RetVT == MVT::i32) 344 return FPTOUINT_F80_I32; 345 if (RetVT == MVT::i64) 346 return FPTOUINT_F80_I64; 347 if (RetVT == MVT::i128) 348 return FPTOUINT_F80_I128; 349 } else if (OpVT == MVT::f128) { 350 if (RetVT == MVT::i32) 351 return FPTOUINT_F128_I32; 352 if (RetVT == MVT::i64) 353 return FPTOUINT_F128_I64; 354 if (RetVT == MVT::i128) 355 return FPTOUINT_F128_I128; 356 } else if (OpVT == MVT::ppcf128) { 357 if (RetVT == MVT::i32) 358 return FPTOUINT_PPCF128_I32; 359 if (RetVT == MVT::i64) 360 return FPTOUINT_PPCF128_I64; 361 if (RetVT == MVT::i128) 362 return FPTOUINT_PPCF128_I128; 363 } 364 return UNKNOWN_LIBCALL; 365 } 366 367 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 368 /// UNKNOWN_LIBCALL if there is none. 369 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 370 if (OpVT == MVT::i32) { 371 if (RetVT == MVT::f32) 372 return SINTTOFP_I32_F32; 373 if (RetVT == MVT::f64) 374 return SINTTOFP_I32_F64; 375 if (RetVT == MVT::f80) 376 return SINTTOFP_I32_F80; 377 if (RetVT == MVT::f128) 378 return SINTTOFP_I32_F128; 379 if (RetVT == MVT::ppcf128) 380 return SINTTOFP_I32_PPCF128; 381 } else if (OpVT == MVT::i64) { 382 if (RetVT == MVT::f32) 383 return SINTTOFP_I64_F32; 384 if (RetVT == MVT::f64) 385 return SINTTOFP_I64_F64; 386 if (RetVT == MVT::f80) 387 return SINTTOFP_I64_F80; 388 if (RetVT == MVT::f128) 389 return SINTTOFP_I64_F128; 390 if (RetVT == MVT::ppcf128) 391 return SINTTOFP_I64_PPCF128; 392 } else if (OpVT == MVT::i128) { 393 if (RetVT == MVT::f32) 394 return SINTTOFP_I128_F32; 395 if (RetVT == MVT::f64) 396 return SINTTOFP_I128_F64; 397 if (RetVT == MVT::f80) 398 return SINTTOFP_I128_F80; 399 if (RetVT == MVT::f128) 400 return SINTTOFP_I128_F128; 401 if (RetVT == MVT::ppcf128) 402 return SINTTOFP_I128_PPCF128; 403 } 404 return UNKNOWN_LIBCALL; 405 } 406 407 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 408 /// UNKNOWN_LIBCALL if there is none. 409 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 410 if (OpVT == MVT::i32) { 411 if (RetVT == MVT::f32) 412 return UINTTOFP_I32_F32; 413 if (RetVT == MVT::f64) 414 return UINTTOFP_I32_F64; 415 if (RetVT == MVT::f80) 416 return UINTTOFP_I32_F80; 417 if (RetVT == MVT::f128) 418 return UINTTOFP_I32_F128; 419 if (RetVT == MVT::ppcf128) 420 return UINTTOFP_I32_PPCF128; 421 } else if (OpVT == MVT::i64) { 422 if (RetVT == MVT::f32) 423 return UINTTOFP_I64_F32; 424 if (RetVT == MVT::f64) 425 return UINTTOFP_I64_F64; 426 if (RetVT == MVT::f80) 427 return UINTTOFP_I64_F80; 428 if (RetVT == MVT::f128) 429 return UINTTOFP_I64_F128; 430 if (RetVT == MVT::ppcf128) 431 return UINTTOFP_I64_PPCF128; 432 } else if (OpVT == MVT::i128) { 433 if (RetVT == MVT::f32) 434 return UINTTOFP_I128_F32; 435 if (RetVT == MVT::f64) 436 return UINTTOFP_I128_F64; 437 if (RetVT == MVT::f80) 438 return UINTTOFP_I128_F80; 439 if (RetVT == MVT::f128) 440 return UINTTOFP_I128_F128; 441 if (RetVT == MVT::ppcf128) 442 return UINTTOFP_I128_PPCF128; 443 } 444 return UNKNOWN_LIBCALL; 445 } 446 447 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 448 #define OP_TO_LIBCALL(Name, Enum) \ 449 case Name: \ 450 switch (VT.SimpleTy) { \ 451 default: \ 452 return UNKNOWN_LIBCALL; \ 453 case MVT::i8: \ 454 return Enum##_1; \ 455 case MVT::i16: \ 456 return Enum##_2; \ 457 case MVT::i32: \ 458 return Enum##_4; \ 459 case MVT::i64: \ 460 return Enum##_8; \ 461 case MVT::i128: \ 462 return Enum##_16; \ 463 } 464 465 switch (Opc) { 466 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 467 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 469 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 470 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 471 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 472 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 473 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 474 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 475 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 476 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 477 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 478 } 479 480 #undef OP_TO_LIBCALL 481 482 return UNKNOWN_LIBCALL; 483 } 484 485 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 486 switch (ElementSize) { 487 case 1: 488 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 489 case 2: 490 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 491 case 4: 492 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 493 case 8: 494 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 495 case 16: 496 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 497 default: 498 return UNKNOWN_LIBCALL; 499 } 500 } 501 502 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 503 switch (ElementSize) { 504 case 1: 505 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 506 case 2: 507 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 508 case 4: 509 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 510 case 8: 511 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 512 case 16: 513 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 514 default: 515 return UNKNOWN_LIBCALL; 516 } 517 } 518 519 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 520 switch (ElementSize) { 521 case 1: 522 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 523 case 2: 524 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 525 case 4: 526 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 527 case 8: 528 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 529 case 16: 530 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 531 default: 532 return UNKNOWN_LIBCALL; 533 } 534 } 535 536 /// InitCmpLibcallCCs - Set default comparison libcall CC. 537 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 538 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 539 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 540 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 541 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 542 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 543 CCs[RTLIB::UNE_F32] = ISD::SETNE; 544 CCs[RTLIB::UNE_F64] = ISD::SETNE; 545 CCs[RTLIB::UNE_F128] = ISD::SETNE; 546 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 547 CCs[RTLIB::OGE_F32] = ISD::SETGE; 548 CCs[RTLIB::OGE_F64] = ISD::SETGE; 549 CCs[RTLIB::OGE_F128] = ISD::SETGE; 550 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 551 CCs[RTLIB::OLT_F32] = ISD::SETLT; 552 CCs[RTLIB::OLT_F64] = ISD::SETLT; 553 CCs[RTLIB::OLT_F128] = ISD::SETLT; 554 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 555 CCs[RTLIB::OLE_F32] = ISD::SETLE; 556 CCs[RTLIB::OLE_F64] = ISD::SETLE; 557 CCs[RTLIB::OLE_F128] = ISD::SETLE; 558 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 559 CCs[RTLIB::OGT_F32] = ISD::SETGT; 560 CCs[RTLIB::OGT_F64] = ISD::SETGT; 561 CCs[RTLIB::OGT_F128] = ISD::SETGT; 562 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 563 CCs[RTLIB::UO_F32] = ISD::SETNE; 564 CCs[RTLIB::UO_F64] = ISD::SETNE; 565 CCs[RTLIB::UO_F128] = ISD::SETNE; 566 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 567 } 568 569 /// NOTE: The TargetMachine owns TLOF. 570 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 571 initActions(); 572 573 // Perform these initializations only once. 574 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 575 MaxLoadsPerMemcmp = 8; 576 MaxGluedStoresPerMemcpy = 0; 577 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 578 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 579 HasMultipleConditionRegisters = false; 580 HasExtractBitsInsn = false; 581 JumpIsExpensive = JumpIsExpensiveOverride; 582 PredictableSelectIsExpensive = false; 583 EnableExtLdPromotion = false; 584 StackPointerRegisterToSaveRestore = 0; 585 BooleanContents = UndefinedBooleanContent; 586 BooleanFloatContents = UndefinedBooleanContent; 587 BooleanVectorContents = UndefinedBooleanContent; 588 SchedPreferenceInfo = Sched::ILP; 589 GatherAllAliasesMaxDepth = 18; 590 IsStrictFPEnabled = DisableStrictNodeMutation; 591 // TODO: the default will be switched to 0 in the next commit, along 592 // with the Target-specific changes necessary. 593 MaxAtomicSizeInBitsSupported = 1024; 594 595 MinCmpXchgSizeInBits = 0; 596 SupportsUnalignedAtomics = false; 597 598 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 599 600 InitLibcalls(TM.getTargetTriple()); 601 InitCmpLibcallCCs(CmpLibcallCCs); 602 } 603 604 void TargetLoweringBase::initActions() { 605 // All operations default to being supported. 606 memset(OpActions, 0, sizeof(OpActions)); 607 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 608 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 609 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 610 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 611 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 612 std::fill(std::begin(TargetDAGCombineArray), 613 std::end(TargetDAGCombineArray), 0); 614 615 for (MVT VT : MVT::fp_valuetypes()) { 616 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 617 if (IntVT.isValid()) { 618 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 619 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 620 } 621 } 622 623 // Set default actions for various operations. 624 for (MVT VT : MVT::all_valuetypes()) { 625 // Default all indexed load / store to expand. 626 for (unsigned IM = (unsigned)ISD::PRE_INC; 627 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 628 setIndexedLoadAction(IM, VT, Expand); 629 setIndexedStoreAction(IM, VT, Expand); 630 setIndexedMaskedLoadAction(IM, VT, Expand); 631 setIndexedMaskedStoreAction(IM, VT, Expand); 632 } 633 634 // Most backends expect to see the node which just returns the value loaded. 635 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 636 637 // These operations default to expand. 638 setOperationAction(ISD::FGETSIGN, VT, Expand); 639 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 640 setOperationAction(ISD::FMINNUM, VT, Expand); 641 setOperationAction(ISD::FMAXNUM, VT, Expand); 642 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 643 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 644 setOperationAction(ISD::FMINIMUM, VT, Expand); 645 setOperationAction(ISD::FMAXIMUM, VT, Expand); 646 setOperationAction(ISD::FMAD, VT, Expand); 647 setOperationAction(ISD::SMIN, VT, Expand); 648 setOperationAction(ISD::SMAX, VT, Expand); 649 setOperationAction(ISD::UMIN, VT, Expand); 650 setOperationAction(ISD::UMAX, VT, Expand); 651 setOperationAction(ISD::ABS, VT, Expand); 652 setOperationAction(ISD::FSHL, VT, Expand); 653 setOperationAction(ISD::FSHR, VT, Expand); 654 setOperationAction(ISD::SADDSAT, VT, Expand); 655 setOperationAction(ISD::UADDSAT, VT, Expand); 656 setOperationAction(ISD::SSUBSAT, VT, Expand); 657 setOperationAction(ISD::USUBSAT, VT, Expand); 658 setOperationAction(ISD::SMULFIX, VT, Expand); 659 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 660 setOperationAction(ISD::UMULFIX, VT, Expand); 661 setOperationAction(ISD::UMULFIXSAT, VT, Expand); 662 setOperationAction(ISD::SDIVFIX, VT, Expand); 663 setOperationAction(ISD::UDIVFIX, VT, Expand); 664 665 // Overflow operations default to expand 666 setOperationAction(ISD::SADDO, VT, Expand); 667 setOperationAction(ISD::SSUBO, VT, Expand); 668 setOperationAction(ISD::UADDO, VT, Expand); 669 setOperationAction(ISD::USUBO, VT, Expand); 670 setOperationAction(ISD::SMULO, VT, Expand); 671 setOperationAction(ISD::UMULO, VT, Expand); 672 673 // ADDCARRY operations default to expand 674 setOperationAction(ISD::ADDCARRY, VT, Expand); 675 setOperationAction(ISD::SUBCARRY, VT, Expand); 676 setOperationAction(ISD::SETCCCARRY, VT, Expand); 677 678 // ADDC/ADDE/SUBC/SUBE default to expand. 679 setOperationAction(ISD::ADDC, VT, Expand); 680 setOperationAction(ISD::ADDE, VT, Expand); 681 setOperationAction(ISD::SUBC, VT, Expand); 682 setOperationAction(ISD::SUBE, VT, Expand); 683 684 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 685 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 686 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 687 688 setOperationAction(ISD::BITREVERSE, VT, Expand); 689 690 // These library functions default to expand. 691 setOperationAction(ISD::FROUND, VT, Expand); 692 setOperationAction(ISD::FPOWI, VT, Expand); 693 694 // These operations default to expand for vector types. 695 if (VT.isVector()) { 696 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 697 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); 698 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 699 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 700 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 701 setOperationAction(ISD::SPLAT_VECTOR, VT, Expand); 702 } 703 704 // Constrained floating-point operations default to expand. 705 #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ 706 setOperationAction(ISD::STRICT_##DAGN, VT, Expand); 707 #include "llvm/IR/ConstrainedOps.def" 708 709 // For most targets @llvm.get.dynamic.area.offset just returns 0. 710 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 711 712 // Vector reduction default to expand. 713 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 714 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 715 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 716 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 717 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 718 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 719 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 720 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 721 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 722 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 723 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 724 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 725 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 726 } 727 728 // Most targets ignore the @llvm.prefetch intrinsic. 729 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 730 731 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 732 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 733 734 // ConstantFP nodes default to expand. Targets can either change this to 735 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 736 // to optimize expansions for certain constants. 737 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 738 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 739 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 740 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 741 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 742 743 // These library functions default to expand. 744 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 745 setOperationAction(ISD::FCBRT, VT, Expand); 746 setOperationAction(ISD::FLOG , VT, Expand); 747 setOperationAction(ISD::FLOG2, VT, Expand); 748 setOperationAction(ISD::FLOG10, VT, Expand); 749 setOperationAction(ISD::FEXP , VT, Expand); 750 setOperationAction(ISD::FEXP2, VT, Expand); 751 setOperationAction(ISD::FFLOOR, VT, Expand); 752 setOperationAction(ISD::FNEARBYINT, VT, Expand); 753 setOperationAction(ISD::FCEIL, VT, Expand); 754 setOperationAction(ISD::FRINT, VT, Expand); 755 setOperationAction(ISD::FTRUNC, VT, Expand); 756 setOperationAction(ISD::FROUND, VT, Expand); 757 setOperationAction(ISD::LROUND, VT, Expand); 758 setOperationAction(ISD::LLROUND, VT, Expand); 759 setOperationAction(ISD::LRINT, VT, Expand); 760 setOperationAction(ISD::LLRINT, VT, Expand); 761 } 762 763 // Default ISD::TRAP to expand (which turns it into abort). 764 setOperationAction(ISD::TRAP, MVT::Other, Expand); 765 766 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 767 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 768 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 769 } 770 771 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 772 EVT) const { 773 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 774 } 775 776 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 777 bool LegalTypes) const { 778 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 779 if (LHSTy.isVector()) 780 return LHSTy; 781 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 782 : getPointerTy(DL); 783 } 784 785 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 786 assert(isTypeLegal(VT)); 787 switch (Op) { 788 default: 789 return false; 790 case ISD::SDIV: 791 case ISD::UDIV: 792 case ISD::SREM: 793 case ISD::UREM: 794 return true; 795 } 796 } 797 798 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 799 // If the command-line option was specified, ignore this request. 800 if (!JumpIsExpensiveOverride.getNumOccurrences()) 801 JumpIsExpensive = isExpensive; 802 } 803 804 TargetLoweringBase::LegalizeKind 805 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 806 // If this is a simple type, use the ComputeRegisterProp mechanism. 807 if (VT.isSimple()) { 808 MVT SVT = VT.getSimpleVT(); 809 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 810 MVT NVT = TransformToType[SVT.SimpleTy]; 811 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 812 813 assert((LA == TypeLegal || LA == TypeSoftenFloat || 814 (NVT.isVector() || 815 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 816 "Promote may not follow Expand or Promote"); 817 818 if (LA == TypeSplitVector) 819 return LegalizeKind(LA, 820 EVT::getVectorVT(Context, SVT.getVectorElementType(), 821 SVT.getVectorNumElements() / 2)); 822 if (LA == TypeScalarizeVector) 823 return LegalizeKind(LA, SVT.getVectorElementType()); 824 return LegalizeKind(LA, NVT); 825 } 826 827 // Handle Extended Scalar Types. 828 if (!VT.isVector()) { 829 assert(VT.isInteger() && "Float types must be simple"); 830 unsigned BitSize = VT.getSizeInBits(); 831 // First promote to a power-of-two size, then expand if necessary. 832 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 833 EVT NVT = VT.getRoundIntegerType(Context); 834 assert(NVT != VT && "Unable to round integer VT"); 835 LegalizeKind NextStep = getTypeConversion(Context, NVT); 836 // Avoid multi-step promotion. 837 if (NextStep.first == TypePromoteInteger) 838 return NextStep; 839 // Return rounded integer type. 840 return LegalizeKind(TypePromoteInteger, NVT); 841 } 842 843 return LegalizeKind(TypeExpandInteger, 844 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 845 } 846 847 // Handle vector types. 848 unsigned NumElts = VT.getVectorNumElements(); 849 EVT EltVT = VT.getVectorElementType(); 850 851 // Vectors with only one element are always scalarized. 852 if (NumElts == 1) 853 return LegalizeKind(TypeScalarizeVector, EltVT); 854 855 // Try to widen vector elements until the element type is a power of two and 856 // promote it to a legal type later on, for example: 857 // <3 x i8> -> <4 x i8> -> <4 x i32> 858 if (EltVT.isInteger()) { 859 // Vectors with a number of elements that is not a power of two are always 860 // widened, for example <3 x i8> -> <4 x i8>. 861 if (!VT.isPow2VectorType()) { 862 NumElts = (unsigned)NextPowerOf2(NumElts); 863 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 864 return LegalizeKind(TypeWidenVector, NVT); 865 } 866 867 // Examine the element type. 868 LegalizeKind LK = getTypeConversion(Context, EltVT); 869 870 // If type is to be expanded, split the vector. 871 // <4 x i140> -> <2 x i140> 872 if (LK.first == TypeExpandInteger) 873 return LegalizeKind(TypeSplitVector, 874 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 875 876 // Promote the integer element types until a legal vector type is found 877 // or until the element integer type is too big. If a legal type was not 878 // found, fallback to the usual mechanism of widening/splitting the 879 // vector. 880 EVT OldEltVT = EltVT; 881 while (true) { 882 // Increase the bitwidth of the element to the next pow-of-two 883 // (which is greater than 8 bits). 884 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 885 .getRoundIntegerType(Context); 886 887 // Stop trying when getting a non-simple element type. 888 // Note that vector elements may be greater than legal vector element 889 // types. Example: X86 XMM registers hold 64bit element on 32bit 890 // systems. 891 if (!EltVT.isSimple()) 892 break; 893 894 // Build a new vector type and check if it is legal. 895 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 896 // Found a legal promoted vector type. 897 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 898 return LegalizeKind(TypePromoteInteger, 899 EVT::getVectorVT(Context, EltVT, NumElts)); 900 } 901 902 // Reset the type to the unexpanded type if we did not find a legal vector 903 // type with a promoted vector element type. 904 EltVT = OldEltVT; 905 } 906 907 // Try to widen the vector until a legal type is found. 908 // If there is no wider legal type, split the vector. 909 while (true) { 910 // Round up to the next power of 2. 911 NumElts = (unsigned)NextPowerOf2(NumElts); 912 913 // If there is no simple vector type with this many elements then there 914 // cannot be a larger legal vector type. Note that this assumes that 915 // there are no skipped intermediate vector types in the simple types. 916 if (!EltVT.isSimple()) 917 break; 918 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 919 if (LargerVector == MVT()) 920 break; 921 922 // If this type is legal then widen the vector. 923 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 924 return LegalizeKind(TypeWidenVector, LargerVector); 925 } 926 927 // Widen odd vectors to next power of two. 928 if (!VT.isPow2VectorType()) { 929 EVT NVT = VT.getPow2VectorType(Context); 930 return LegalizeKind(TypeWidenVector, NVT); 931 } 932 933 // Vectors with illegal element types are expanded. 934 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 935 return LegalizeKind(TypeSplitVector, NVT); 936 } 937 938 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 939 unsigned &NumIntermediates, 940 MVT &RegisterVT, 941 TargetLoweringBase *TLI) { 942 // Figure out the right, legal destination reg to copy into. 943 unsigned NumElts = VT.getVectorNumElements(); 944 MVT EltTy = VT.getVectorElementType(); 945 946 unsigned NumVectorRegs = 1; 947 948 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 949 // could break down into LHS/RHS like LegalizeDAG does. 950 if (!isPowerOf2_32(NumElts)) { 951 NumVectorRegs = NumElts; 952 NumElts = 1; 953 } 954 955 // Divide the input until we get to a supported size. This will always 956 // end with a scalar if the target doesn't support vectors. 957 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 958 NumElts >>= 1; 959 NumVectorRegs <<= 1; 960 } 961 962 NumIntermediates = NumVectorRegs; 963 964 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 965 if (!TLI->isTypeLegal(NewVT)) 966 NewVT = EltTy; 967 IntermediateVT = NewVT; 968 969 unsigned NewVTSize = NewVT.getSizeInBits(); 970 971 // Convert sizes such as i33 to i64. 972 if (!isPowerOf2_32(NewVTSize)) 973 NewVTSize = NextPowerOf2(NewVTSize); 974 975 MVT DestVT = TLI->getRegisterType(NewVT); 976 RegisterVT = DestVT; 977 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 978 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 979 980 // Otherwise, promotion or legal types use the same number of registers as 981 // the vector decimated to the appropriate level. 982 return NumVectorRegs; 983 } 984 985 /// isLegalRC - Return true if the value types that can be represented by the 986 /// specified register class are all legal. 987 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 988 const TargetRegisterClass &RC) const { 989 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 990 if (isTypeLegal(*I)) 991 return true; 992 return false; 993 } 994 995 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 996 /// sequence of memory operands that is recognized by PrologEpilogInserter. 997 MachineBasicBlock * 998 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 999 MachineBasicBlock *MBB) const { 1000 MachineInstr *MI = &InitialMI; 1001 MachineFunction &MF = *MI->getMF(); 1002 MachineFrameInfo &MFI = MF.getFrameInfo(); 1003 1004 // We're handling multiple types of operands here: 1005 // PATCHPOINT MetaArgs - live-in, read only, direct 1006 // STATEPOINT Deopt Spill - live-through, read only, indirect 1007 // STATEPOINT Deopt Alloca - live-through, read only, direct 1008 // (We're currently conservative and mark the deopt slots read/write in 1009 // practice.) 1010 // STATEPOINT GC Spill - live-through, read/write, indirect 1011 // STATEPOINT GC Alloca - live-through, read/write, direct 1012 // The live-in vs live-through is handled already (the live through ones are 1013 // all stack slots), but we need to handle the different type of stackmap 1014 // operands and memory effects here. 1015 1016 // MI changes inside this loop as we grow operands. 1017 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1018 MachineOperand &MO = MI->getOperand(OperIdx); 1019 if (!MO.isFI()) 1020 continue; 1021 1022 // foldMemoryOperand builds a new MI after replacing a single FI operand 1023 // with the canonical set of five x86 addressing-mode operands. 1024 int FI = MO.getIndex(); 1025 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1026 1027 // Copy operands before the frame-index. 1028 for (unsigned i = 0; i < OperIdx; ++i) 1029 MIB.add(MI->getOperand(i)); 1030 // Add frame index operands recognized by stackmaps.cpp 1031 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1032 // indirect-mem-ref tag, size, #FI, offset. 1033 // Used for spills inserted by StatepointLowering. This codepath is not 1034 // used for patchpoints/stackmaps at all, for these spilling is done via 1035 // foldMemoryOperand callback only. 1036 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1037 MIB.addImm(StackMaps::IndirectMemRefOp); 1038 MIB.addImm(MFI.getObjectSize(FI)); 1039 MIB.add(MI->getOperand(OperIdx)); 1040 MIB.addImm(0); 1041 } else { 1042 // direct-mem-ref tag, #FI, offset. 1043 // Used by patchpoint, and direct alloca arguments to statepoints 1044 MIB.addImm(StackMaps::DirectMemRefOp); 1045 MIB.add(MI->getOperand(OperIdx)); 1046 MIB.addImm(0); 1047 } 1048 // Copy the operands after the frame index. 1049 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1050 MIB.add(MI->getOperand(i)); 1051 1052 // Inherit previous memory operands. 1053 MIB.cloneMemRefs(*MI); 1054 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1055 1056 // Add a new memory operand for this FI. 1057 assert(MFI.getObjectOffset(FI) != -1); 1058 1059 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1060 // PATCHPOINT should be updated to do the same. (TODO) 1061 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1062 auto Flags = MachineMemOperand::MOLoad; 1063 MachineMemOperand *MMO = MF.getMachineMemOperand( 1064 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1065 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1066 MIB->addMemOperand(MF, MMO); 1067 } 1068 1069 // Replace the instruction and update the operand index. 1070 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1071 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1072 MI->eraseFromParent(); 1073 MI = MIB; 1074 } 1075 return MBB; 1076 } 1077 1078 MachineBasicBlock * 1079 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1080 MachineBasicBlock *MBB) const { 1081 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1082 "Called emitXRayCustomEvent on the wrong MI!"); 1083 auto &MF = *MI.getMF(); 1084 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1085 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1086 MIB.add(MI.getOperand(OpIdx)); 1087 1088 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1089 MI.eraseFromParent(); 1090 return MBB; 1091 } 1092 1093 MachineBasicBlock * 1094 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1095 MachineBasicBlock *MBB) const { 1096 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1097 "Called emitXRayTypedEvent on the wrong MI!"); 1098 auto &MF = *MI.getMF(); 1099 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1100 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1101 MIB.add(MI.getOperand(OpIdx)); 1102 1103 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1104 MI.eraseFromParent(); 1105 return MBB; 1106 } 1107 1108 /// findRepresentativeClass - Return the largest legal super-reg register class 1109 /// of the register class for the specified type and its associated "cost". 1110 // This function is in TargetLowering because it uses RegClassForVT which would 1111 // need to be moved to TargetRegisterInfo and would necessitate moving 1112 // isTypeLegal over as well - a massive change that would just require 1113 // TargetLowering having a TargetRegisterInfo class member that it would use. 1114 std::pair<const TargetRegisterClass *, uint8_t> 1115 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1116 MVT VT) const { 1117 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1118 if (!RC) 1119 return std::make_pair(RC, 0); 1120 1121 // Compute the set of all super-register classes. 1122 BitVector SuperRegRC(TRI->getNumRegClasses()); 1123 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1124 SuperRegRC.setBitsInMask(RCI.getMask()); 1125 1126 // Find the first legal register class with the largest spill size. 1127 const TargetRegisterClass *BestRC = RC; 1128 for (unsigned i : SuperRegRC.set_bits()) { 1129 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1130 // We want the largest possible spill size. 1131 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1132 continue; 1133 if (!isLegalRC(*TRI, *SuperRC)) 1134 continue; 1135 BestRC = SuperRC; 1136 } 1137 return std::make_pair(BestRC, 1); 1138 } 1139 1140 /// computeRegisterProperties - Once all of the register classes are added, 1141 /// this allows us to compute derived properties we expose. 1142 void TargetLoweringBase::computeRegisterProperties( 1143 const TargetRegisterInfo *TRI) { 1144 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1145 "Too many value types for ValueTypeActions to hold!"); 1146 1147 // Everything defaults to needing one register. 1148 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1149 NumRegistersForVT[i] = 1; 1150 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1151 } 1152 // ...except isVoid, which doesn't need any registers. 1153 NumRegistersForVT[MVT::isVoid] = 0; 1154 1155 // Find the largest integer register class. 1156 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1157 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1158 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1159 1160 // Every integer value type larger than this largest register takes twice as 1161 // many registers to represent as the previous ValueType. 1162 for (unsigned ExpandedReg = LargestIntReg + 1; 1163 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1164 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1165 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1166 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1167 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1168 TypeExpandInteger); 1169 } 1170 1171 // Inspect all of the ValueType's smaller than the largest integer 1172 // register to see which ones need promotion. 1173 unsigned LegalIntReg = LargestIntReg; 1174 for (unsigned IntReg = LargestIntReg - 1; 1175 IntReg >= (unsigned)MVT::i1; --IntReg) { 1176 MVT IVT = (MVT::SimpleValueType)IntReg; 1177 if (isTypeLegal(IVT)) { 1178 LegalIntReg = IntReg; 1179 } else { 1180 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1181 (MVT::SimpleValueType)LegalIntReg; 1182 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1183 } 1184 } 1185 1186 // ppcf128 type is really two f64's. 1187 if (!isTypeLegal(MVT::ppcf128)) { 1188 if (isTypeLegal(MVT::f64)) { 1189 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1190 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1191 TransformToType[MVT::ppcf128] = MVT::f64; 1192 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1193 } else { 1194 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1195 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1196 TransformToType[MVT::ppcf128] = MVT::i128; 1197 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1198 } 1199 } 1200 1201 // Decide how to handle f128. If the target does not have native f128 support, 1202 // expand it to i128 and we will be generating soft float library calls. 1203 if (!isTypeLegal(MVT::f128)) { 1204 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1205 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1206 TransformToType[MVT::f128] = MVT::i128; 1207 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1208 } 1209 1210 // Decide how to handle f64. If the target does not have native f64 support, 1211 // expand it to i64 and we will be generating soft float library calls. 1212 if (!isTypeLegal(MVT::f64)) { 1213 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1214 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1215 TransformToType[MVT::f64] = MVT::i64; 1216 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1217 } 1218 1219 // Decide how to handle f32. If the target does not have native f32 support, 1220 // expand it to i32 and we will be generating soft float library calls. 1221 if (!isTypeLegal(MVT::f32)) { 1222 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1223 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1224 TransformToType[MVT::f32] = MVT::i32; 1225 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1226 } 1227 1228 // Decide how to handle f16. If the target does not have native f16 support, 1229 // promote it to f32, because there are no f16 library calls (except for 1230 // conversions). 1231 if (!isTypeLegal(MVT::f16)) { 1232 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1233 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1234 TransformToType[MVT::f16] = MVT::f32; 1235 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1236 } 1237 1238 // Loop over all of the vector value types to see which need transformations. 1239 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1240 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1241 MVT VT = (MVT::SimpleValueType) i; 1242 if (isTypeLegal(VT)) 1243 continue; 1244 1245 MVT EltVT = VT.getVectorElementType(); 1246 unsigned NElts = VT.getVectorNumElements(); 1247 bool IsLegalWiderType = false; 1248 bool IsScalable = VT.isScalableVector(); 1249 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1250 switch (PreferredAction) { 1251 case TypePromoteInteger: { 1252 MVT::SimpleValueType EndVT = IsScalable ? 1253 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE : 1254 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE; 1255 // Try to promote the elements of integer vectors. If no legal 1256 // promotion was found, fall through to the widen-vector method. 1257 for (unsigned nVT = i + 1; 1258 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) { 1259 MVT SVT = (MVT::SimpleValueType) nVT; 1260 // Promote vectors of integers to vectors with the same number 1261 // of elements, with a wider element type. 1262 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1263 SVT.getVectorNumElements() == NElts && 1264 SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1265 TransformToType[i] = SVT; 1266 RegisterTypeForVT[i] = SVT; 1267 NumRegistersForVT[i] = 1; 1268 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1269 IsLegalWiderType = true; 1270 break; 1271 } 1272 } 1273 if (IsLegalWiderType) 1274 break; 1275 LLVM_FALLTHROUGH; 1276 } 1277 1278 case TypeWidenVector: 1279 if (isPowerOf2_32(NElts)) { 1280 // Try to widen the vector. 1281 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1282 MVT SVT = (MVT::SimpleValueType) nVT; 1283 if (SVT.getVectorElementType() == EltVT 1284 && SVT.getVectorNumElements() > NElts 1285 && SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1286 TransformToType[i] = SVT; 1287 RegisterTypeForVT[i] = SVT; 1288 NumRegistersForVT[i] = 1; 1289 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1290 IsLegalWiderType = true; 1291 break; 1292 } 1293 } 1294 if (IsLegalWiderType) 1295 break; 1296 } else { 1297 // Only widen to the next power of 2 to keep consistency with EVT. 1298 MVT NVT = VT.getPow2VectorType(); 1299 if (isTypeLegal(NVT)) { 1300 TransformToType[i] = NVT; 1301 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1302 RegisterTypeForVT[i] = NVT; 1303 NumRegistersForVT[i] = 1; 1304 break; 1305 } 1306 } 1307 LLVM_FALLTHROUGH; 1308 1309 case TypeSplitVector: 1310 case TypeScalarizeVector: { 1311 MVT IntermediateVT; 1312 MVT RegisterVT; 1313 unsigned NumIntermediates; 1314 unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1315 NumIntermediates, RegisterVT, this); 1316 NumRegistersForVT[i] = NumRegisters; 1317 assert(NumRegistersForVT[i] == NumRegisters && 1318 "NumRegistersForVT size cannot represent NumRegisters!"); 1319 RegisterTypeForVT[i] = RegisterVT; 1320 1321 MVT NVT = VT.getPow2VectorType(); 1322 if (NVT == VT) { 1323 // Type is already a power of 2. The default action is to split. 1324 TransformToType[i] = MVT::Other; 1325 if (PreferredAction == TypeScalarizeVector) 1326 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1327 else if (PreferredAction == TypeSplitVector) 1328 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1329 else 1330 // Set type action according to the number of elements. 1331 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1332 : TypeSplitVector); 1333 } else { 1334 TransformToType[i] = NVT; 1335 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1336 } 1337 break; 1338 } 1339 default: 1340 llvm_unreachable("Unknown vector legalization action!"); 1341 } 1342 } 1343 1344 // Determine the 'representative' register class for each value type. 1345 // An representative register class is the largest (meaning one which is 1346 // not a sub-register class / subreg register class) legal register class for 1347 // a group of value types. For example, on i386, i8, i16, and i32 1348 // representative would be GR32; while on x86_64 it's GR64. 1349 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1350 const TargetRegisterClass* RRC; 1351 uint8_t Cost; 1352 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1353 RepRegClassForVT[i] = RRC; 1354 RepRegClassCostForVT[i] = Cost; 1355 } 1356 } 1357 1358 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1359 EVT VT) const { 1360 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1361 return getPointerTy(DL).SimpleTy; 1362 } 1363 1364 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1365 return MVT::i32; // return the default value 1366 } 1367 1368 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1369 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1370 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1371 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1372 /// 1373 /// This method returns the number of registers needed, and the VT for each 1374 /// register. It also returns the VT and quantity of the intermediate values 1375 /// before they are promoted/expanded. 1376 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1377 EVT &IntermediateVT, 1378 unsigned &NumIntermediates, 1379 MVT &RegisterVT) const { 1380 unsigned NumElts = VT.getVectorNumElements(); 1381 1382 // If there is a wider vector type with the same element type as this one, 1383 // or a promoted vector type that has the same number of elements which 1384 // are wider, then we should convert to that legal vector type. 1385 // This handles things like <2 x float> -> <4 x float> and 1386 // <4 x i1> -> <4 x i32>. 1387 LegalizeTypeAction TA = getTypeAction(Context, VT); 1388 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1389 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1390 if (isTypeLegal(RegisterEVT)) { 1391 IntermediateVT = RegisterEVT; 1392 RegisterVT = RegisterEVT.getSimpleVT(); 1393 NumIntermediates = 1; 1394 return 1; 1395 } 1396 } 1397 1398 // Figure out the right, legal destination reg to copy into. 1399 EVT EltTy = VT.getVectorElementType(); 1400 1401 unsigned NumVectorRegs = 1; 1402 1403 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1404 // could break down into LHS/RHS like LegalizeDAG does. 1405 if (!isPowerOf2_32(NumElts)) { 1406 NumVectorRegs = NumElts; 1407 NumElts = 1; 1408 } 1409 1410 // Divide the input until we get to a supported size. This will always 1411 // end with a scalar if the target doesn't support vectors. 1412 while (NumElts > 1 && !isTypeLegal( 1413 EVT::getVectorVT(Context, EltTy, NumElts))) { 1414 NumElts >>= 1; 1415 NumVectorRegs <<= 1; 1416 } 1417 1418 NumIntermediates = NumVectorRegs; 1419 1420 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1421 if (!isTypeLegal(NewVT)) 1422 NewVT = EltTy; 1423 IntermediateVT = NewVT; 1424 1425 MVT DestVT = getRegisterType(Context, NewVT); 1426 RegisterVT = DestVT; 1427 unsigned NewVTSize = NewVT.getSizeInBits(); 1428 1429 // Convert sizes such as i33 to i64. 1430 if (!isPowerOf2_32(NewVTSize)) 1431 NewVTSize = NextPowerOf2(NewVTSize); 1432 1433 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1434 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1435 1436 // Otherwise, promotion or legal types use the same number of registers as 1437 // the vector decimated to the appropriate level. 1438 return NumVectorRegs; 1439 } 1440 1441 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI, 1442 uint64_t NumCases, 1443 uint64_t Range, 1444 ProfileSummaryInfo *PSI, 1445 BlockFrequencyInfo *BFI) const { 1446 // FIXME: This function check the maximum table size and density, but the 1447 // minimum size is not checked. It would be nice if the minimum size is 1448 // also combined within this function. Currently, the minimum size check is 1449 // performed in findJumpTable() in SelectionDAGBuiler and 1450 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl. 1451 const bool OptForSize = 1452 SI->getParent()->getParent()->hasOptSize() || 1453 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI); 1454 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize); 1455 const unsigned MaxJumpTableSize = getMaximumJumpTableSize(); 1456 1457 // Check whether the number of cases is small enough and 1458 // the range is dense enough for a jump table. 1459 return (OptForSize || Range <= MaxJumpTableSize) && 1460 (NumCases * 100 >= Range * MinDensity); 1461 } 1462 1463 /// Get the EVTs and ArgFlags collections that represent the legalized return 1464 /// type of the given function. This does not require a DAG or a return value, 1465 /// and is suitable for use before any DAGs for the function are constructed. 1466 /// TODO: Move this out of TargetLowering.cpp. 1467 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1468 AttributeList attr, 1469 SmallVectorImpl<ISD::OutputArg> &Outs, 1470 const TargetLowering &TLI, const DataLayout &DL) { 1471 SmallVector<EVT, 4> ValueVTs; 1472 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1473 unsigned NumValues = ValueVTs.size(); 1474 if (NumValues == 0) return; 1475 1476 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1477 EVT VT = ValueVTs[j]; 1478 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1479 1480 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1481 ExtendKind = ISD::SIGN_EXTEND; 1482 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1483 ExtendKind = ISD::ZERO_EXTEND; 1484 1485 // FIXME: C calling convention requires the return type to be promoted to 1486 // at least 32-bit. But this is not necessary for non-C calling 1487 // conventions. The frontend should mark functions whose return values 1488 // require promoting with signext or zeroext attributes. 1489 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1490 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1491 if (VT.bitsLT(MinVT)) 1492 VT = MinVT; 1493 } 1494 1495 unsigned NumParts = 1496 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1497 MVT PartVT = 1498 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1499 1500 // 'inreg' on function refers to return value 1501 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1502 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1503 Flags.setInReg(); 1504 1505 // Propagate extension type if any 1506 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1507 Flags.setSExt(); 1508 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1509 Flags.setZExt(); 1510 1511 for (unsigned i = 0; i < NumParts; ++i) 1512 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1513 } 1514 } 1515 1516 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1517 /// function arguments in the caller parameter area. This is the actual 1518 /// alignment, not its logarithm. 1519 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1520 const DataLayout &DL) const { 1521 return DL.getABITypeAlignment(Ty); 1522 } 1523 1524 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1525 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1526 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1527 // Check if the specified alignment is sufficient based on the data layout. 1528 // TODO: While using the data layout works in practice, a better solution 1529 // would be to implement this check directly (make this a virtual function). 1530 // For example, the ABI alignment may change based on software platform while 1531 // this function should only be affected by hardware implementation. 1532 Type *Ty = VT.getTypeForEVT(Context); 1533 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1534 // Assume that an access that meets the ABI-specified alignment is fast. 1535 if (Fast != nullptr) 1536 *Fast = true; 1537 return true; 1538 } 1539 1540 // This is a misaligned access. 1541 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1542 } 1543 1544 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1545 LLVMContext &Context, const DataLayout &DL, EVT VT, 1546 const MachineMemOperand &MMO, bool *Fast) const { 1547 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(), 1548 MMO.getAlignment(), MMO.getFlags(), 1549 Fast); 1550 } 1551 1552 bool TargetLoweringBase::allowsMemoryAccess( 1553 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1554 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1555 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment, 1556 Flags, Fast); 1557 } 1558 1559 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1560 const DataLayout &DL, EVT VT, 1561 const MachineMemOperand &MMO, 1562 bool *Fast) const { 1563 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1564 MMO.getAlignment(), MMO.getFlags(), Fast); 1565 } 1566 1567 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1568 return BranchProbability(MinPercentageForPredictableBranch, 100); 1569 } 1570 1571 //===----------------------------------------------------------------------===// 1572 // TargetTransformInfo Helpers 1573 //===----------------------------------------------------------------------===// 1574 1575 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1576 enum InstructionOpcodes { 1577 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1578 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1579 #include "llvm/IR/Instruction.def" 1580 }; 1581 switch (static_cast<InstructionOpcodes>(Opcode)) { 1582 case Ret: return 0; 1583 case Br: return 0; 1584 case Switch: return 0; 1585 case IndirectBr: return 0; 1586 case Invoke: return 0; 1587 case CallBr: return 0; 1588 case Resume: return 0; 1589 case Unreachable: return 0; 1590 case CleanupRet: return 0; 1591 case CatchRet: return 0; 1592 case CatchPad: return 0; 1593 case CatchSwitch: return 0; 1594 case CleanupPad: return 0; 1595 case FNeg: return ISD::FNEG; 1596 case Add: return ISD::ADD; 1597 case FAdd: return ISD::FADD; 1598 case Sub: return ISD::SUB; 1599 case FSub: return ISD::FSUB; 1600 case Mul: return ISD::MUL; 1601 case FMul: return ISD::FMUL; 1602 case UDiv: return ISD::UDIV; 1603 case SDiv: return ISD::SDIV; 1604 case FDiv: return ISD::FDIV; 1605 case URem: return ISD::UREM; 1606 case SRem: return ISD::SREM; 1607 case FRem: return ISD::FREM; 1608 case Shl: return ISD::SHL; 1609 case LShr: return ISD::SRL; 1610 case AShr: return ISD::SRA; 1611 case And: return ISD::AND; 1612 case Or: return ISD::OR; 1613 case Xor: return ISD::XOR; 1614 case Alloca: return 0; 1615 case Load: return ISD::LOAD; 1616 case Store: return ISD::STORE; 1617 case GetElementPtr: return 0; 1618 case Fence: return 0; 1619 case AtomicCmpXchg: return 0; 1620 case AtomicRMW: return 0; 1621 case Trunc: return ISD::TRUNCATE; 1622 case ZExt: return ISD::ZERO_EXTEND; 1623 case SExt: return ISD::SIGN_EXTEND; 1624 case FPToUI: return ISD::FP_TO_UINT; 1625 case FPToSI: return ISD::FP_TO_SINT; 1626 case UIToFP: return ISD::UINT_TO_FP; 1627 case SIToFP: return ISD::SINT_TO_FP; 1628 case FPTrunc: return ISD::FP_ROUND; 1629 case FPExt: return ISD::FP_EXTEND; 1630 case PtrToInt: return ISD::BITCAST; 1631 case IntToPtr: return ISD::BITCAST; 1632 case BitCast: return ISD::BITCAST; 1633 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1634 case ICmp: return ISD::SETCC; 1635 case FCmp: return ISD::SETCC; 1636 case PHI: return 0; 1637 case Call: return 0; 1638 case Select: return ISD::SELECT; 1639 case UserOp1: return 0; 1640 case UserOp2: return 0; 1641 case VAArg: return 0; 1642 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1643 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1644 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1645 case ExtractValue: return ISD::MERGE_VALUES; 1646 case InsertValue: return ISD::MERGE_VALUES; 1647 case LandingPad: return 0; 1648 case Freeze: return 0; 1649 } 1650 1651 llvm_unreachable("Unknown instruction type encountered!"); 1652 } 1653 1654 std::pair<int, MVT> 1655 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1656 Type *Ty) const { 1657 LLVMContext &C = Ty->getContext(); 1658 EVT MTy = getValueType(DL, Ty); 1659 1660 int Cost = 1; 1661 // We keep legalizing the type until we find a legal kind. We assume that 1662 // the only operation that costs anything is the split. After splitting 1663 // we need to handle two types. 1664 while (true) { 1665 LegalizeKind LK = getTypeConversion(C, MTy); 1666 1667 if (LK.first == TypeLegal) 1668 return std::make_pair(Cost, MTy.getSimpleVT()); 1669 1670 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1671 Cost *= 2; 1672 1673 // Do not loop with f128 type. 1674 if (MTy == LK.second) 1675 return std::make_pair(Cost, MTy.getSimpleVT()); 1676 1677 // Keep legalizing the type. 1678 MTy = LK.second; 1679 } 1680 } 1681 1682 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1683 bool UseTLS) const { 1684 // compiler-rt provides a variable with a magic name. Targets that do not 1685 // link with compiler-rt may also provide such a variable. 1686 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1687 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1688 auto UnsafeStackPtr = 1689 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1690 1691 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1692 1693 if (!UnsafeStackPtr) { 1694 auto TLSModel = UseTLS ? 1695 GlobalValue::InitialExecTLSModel : 1696 GlobalValue::NotThreadLocal; 1697 // The global variable is not defined yet, define it ourselves. 1698 // We use the initial-exec TLS model because we do not support the 1699 // variable living anywhere other than in the main executable. 1700 UnsafeStackPtr = new GlobalVariable( 1701 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1702 UnsafeStackPtrVar, nullptr, TLSModel); 1703 } else { 1704 // The variable exists, check its type and attributes. 1705 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1706 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1707 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1708 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1709 (UseTLS ? "" : "not ") + "be thread-local"); 1710 } 1711 return UnsafeStackPtr; 1712 } 1713 1714 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1715 if (!TM.getTargetTriple().isAndroid()) 1716 return getDefaultSafeStackPointerLocation(IRB, true); 1717 1718 // Android provides a libc function to retrieve the address of the current 1719 // thread's unsafe stack pointer. 1720 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1721 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1722 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1723 StackPtrTy->getPointerTo(0)); 1724 return IRB.CreateCall(Fn); 1725 } 1726 1727 //===----------------------------------------------------------------------===// 1728 // Loop Strength Reduction hooks 1729 //===----------------------------------------------------------------------===// 1730 1731 /// isLegalAddressingMode - Return true if the addressing mode represented 1732 /// by AM is legal for this target, for a load/store of the specified type. 1733 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1734 const AddrMode &AM, Type *Ty, 1735 unsigned AS, Instruction *I) const { 1736 // The default implementation of this implements a conservative RISCy, r+r and 1737 // r+i addr mode. 1738 1739 // Allows a sign-extended 16-bit immediate field. 1740 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1741 return false; 1742 1743 // No global is ever allowed as a base. 1744 if (AM.BaseGV) 1745 return false; 1746 1747 // Only support r+r, 1748 switch (AM.Scale) { 1749 case 0: // "r+i" or just "i", depending on HasBaseReg. 1750 break; 1751 case 1: 1752 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1753 return false; 1754 // Otherwise we have r+r or r+i. 1755 break; 1756 case 2: 1757 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1758 return false; 1759 // Allow 2*r as r+r. 1760 break; 1761 default: // Don't allow n * r 1762 return false; 1763 } 1764 1765 return true; 1766 } 1767 1768 //===----------------------------------------------------------------------===// 1769 // Stack Protector 1770 //===----------------------------------------------------------------------===// 1771 1772 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1773 // so that SelectionDAG handle SSP. 1774 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1775 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1776 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1777 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1778 return M.getOrInsertGlobal("__guard_local", PtrTy); 1779 } 1780 return nullptr; 1781 } 1782 1783 // Currently only support "standard" __stack_chk_guard. 1784 // TODO: add LOAD_STACK_GUARD support. 1785 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1786 if (!M.getNamedValue("__stack_chk_guard")) 1787 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1788 GlobalVariable::ExternalLinkage, 1789 nullptr, "__stack_chk_guard"); 1790 } 1791 1792 // Currently only support "standard" __stack_chk_guard. 1793 // TODO: add LOAD_STACK_GUARD support. 1794 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1795 return M.getNamedValue("__stack_chk_guard"); 1796 } 1797 1798 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1799 return nullptr; 1800 } 1801 1802 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1803 return MinimumJumpTableEntries; 1804 } 1805 1806 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1807 MinimumJumpTableEntries = Val; 1808 } 1809 1810 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1811 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1812 } 1813 1814 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1815 return MaximumJumpTableSize; 1816 } 1817 1818 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1819 MaximumJumpTableSize = Val; 1820 } 1821 1822 //===----------------------------------------------------------------------===// 1823 // Reciprocal Estimates 1824 //===----------------------------------------------------------------------===// 1825 1826 /// Get the reciprocal estimate attribute string for a function that will 1827 /// override the target defaults. 1828 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1829 const Function &F = MF.getFunction(); 1830 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1831 } 1832 1833 /// Construct a string for the given reciprocal operation of the given type. 1834 /// This string should match the corresponding option to the front-end's 1835 /// "-mrecip" flag assuming those strings have been passed through in an 1836 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1837 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1838 std::string Name = VT.isVector() ? "vec-" : ""; 1839 1840 Name += IsSqrt ? "sqrt" : "div"; 1841 1842 // TODO: Handle "half" or other float types? 1843 if (VT.getScalarType() == MVT::f64) { 1844 Name += "d"; 1845 } else { 1846 assert(VT.getScalarType() == MVT::f32 && 1847 "Unexpected FP type for reciprocal estimate"); 1848 Name += "f"; 1849 } 1850 1851 return Name; 1852 } 1853 1854 /// Return the character position and value (a single numeric character) of a 1855 /// customized refinement operation in the input string if it exists. Return 1856 /// false if there is no customized refinement step count. 1857 static bool parseRefinementStep(StringRef In, size_t &Position, 1858 uint8_t &Value) { 1859 const char RefStepToken = ':'; 1860 Position = In.find(RefStepToken); 1861 if (Position == StringRef::npos) 1862 return false; 1863 1864 StringRef RefStepString = In.substr(Position + 1); 1865 // Allow exactly one numeric character for the additional refinement 1866 // step parameter. 1867 if (RefStepString.size() == 1) { 1868 char RefStepChar = RefStepString[0]; 1869 if (RefStepChar >= '0' && RefStepChar <= '9') { 1870 Value = RefStepChar - '0'; 1871 return true; 1872 } 1873 } 1874 report_fatal_error("Invalid refinement step for -recip."); 1875 } 1876 1877 /// For the input attribute string, return one of the ReciprocalEstimate enum 1878 /// status values (enabled, disabled, or not specified) for this operation on 1879 /// the specified data type. 1880 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1881 if (Override.empty()) 1882 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1883 1884 SmallVector<StringRef, 4> OverrideVector; 1885 Override.split(OverrideVector, ','); 1886 unsigned NumArgs = OverrideVector.size(); 1887 1888 // Check if "all", "none", or "default" was specified. 1889 if (NumArgs == 1) { 1890 // Look for an optional setting of the number of refinement steps needed 1891 // for this type of reciprocal operation. 1892 size_t RefPos; 1893 uint8_t RefSteps; 1894 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1895 // Split the string for further processing. 1896 Override = Override.substr(0, RefPos); 1897 } 1898 1899 // All reciprocal types are enabled. 1900 if (Override == "all") 1901 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1902 1903 // All reciprocal types are disabled. 1904 if (Override == "none") 1905 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1906 1907 // Target defaults for enablement are used. 1908 if (Override == "default") 1909 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1910 } 1911 1912 // The attribute string may omit the size suffix ('f'/'d'). 1913 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1914 std::string VTNameNoSize = VTName; 1915 VTNameNoSize.pop_back(); 1916 static const char DisabledPrefix = '!'; 1917 1918 for (StringRef RecipType : OverrideVector) { 1919 size_t RefPos; 1920 uint8_t RefSteps; 1921 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1922 RecipType = RecipType.substr(0, RefPos); 1923 1924 // Ignore the disablement token for string matching. 1925 bool IsDisabled = RecipType[0] == DisabledPrefix; 1926 if (IsDisabled) 1927 RecipType = RecipType.substr(1); 1928 1929 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1930 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1931 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1932 } 1933 1934 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1935 } 1936 1937 /// For the input attribute string, return the customized refinement step count 1938 /// for this operation on the specified data type. If the step count does not 1939 /// exist, return the ReciprocalEstimate enum value for unspecified. 1940 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1941 if (Override.empty()) 1942 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1943 1944 SmallVector<StringRef, 4> OverrideVector; 1945 Override.split(OverrideVector, ','); 1946 unsigned NumArgs = OverrideVector.size(); 1947 1948 // Check if "all", "default", or "none" was specified. 1949 if (NumArgs == 1) { 1950 // Look for an optional setting of the number of refinement steps needed 1951 // for this type of reciprocal operation. 1952 size_t RefPos; 1953 uint8_t RefSteps; 1954 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1955 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1956 1957 // Split the string for further processing. 1958 Override = Override.substr(0, RefPos); 1959 assert(Override != "none" && 1960 "Disabled reciprocals, but specifed refinement steps?"); 1961 1962 // If this is a general override, return the specified number of steps. 1963 if (Override == "all" || Override == "default") 1964 return RefSteps; 1965 } 1966 1967 // The attribute string may omit the size suffix ('f'/'d'). 1968 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1969 std::string VTNameNoSize = VTName; 1970 VTNameNoSize.pop_back(); 1971 1972 for (StringRef RecipType : OverrideVector) { 1973 size_t RefPos; 1974 uint8_t RefSteps; 1975 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1976 continue; 1977 1978 RecipType = RecipType.substr(0, RefPos); 1979 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1980 return RefSteps; 1981 } 1982 1983 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1984 } 1985 1986 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1987 MachineFunction &MF) const { 1988 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1989 } 1990 1991 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1992 MachineFunction &MF) const { 1993 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1994 } 1995 1996 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1997 MachineFunction &MF) const { 1998 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1999 } 2000 2001 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 2002 MachineFunction &MF) const { 2003 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 2004 } 2005 2006 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 2007 MF.getRegInfo().freezeReservedRegs(MF); 2008 } 2009 2010 MachineMemOperand::Flags 2011 TargetLoweringBase::getLoadMemOperandFlags(const LoadInst &LI, 2012 const DataLayout &DL) const { 2013 MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad; 2014 if (LI.isVolatile()) 2015 Flags |= MachineMemOperand::MOVolatile; 2016 2017 if (LI.hasMetadata(LLVMContext::MD_nontemporal)) 2018 Flags |= MachineMemOperand::MONonTemporal; 2019 2020 if (LI.hasMetadata(LLVMContext::MD_invariant_load)) 2021 Flags |= MachineMemOperand::MOInvariant; 2022 2023 if (isDereferenceablePointer(LI.getPointerOperand(), LI.getType(), DL)) 2024 Flags |= MachineMemOperand::MODereferenceable; 2025 2026 Flags |= getTargetMMOFlags(LI); 2027 return Flags; 2028 } 2029 2030 MachineMemOperand::Flags 2031 TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI, 2032 const DataLayout &DL) const { 2033 MachineMemOperand::Flags Flags = MachineMemOperand::MOStore; 2034 2035 if (SI.isVolatile()) 2036 Flags |= MachineMemOperand::MOVolatile; 2037 2038 if (SI.hasMetadata(LLVMContext::MD_nontemporal)) 2039 Flags |= MachineMemOperand::MONonTemporal; 2040 2041 // FIXME: Not preserving dereferenceable 2042 Flags |= getTargetMMOFlags(SI); 2043 return Flags; 2044 } 2045 2046 MachineMemOperand::Flags 2047 TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI, 2048 const DataLayout &DL) const { 2049 auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 2050 2051 if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(&AI)) { 2052 if (RMW->isVolatile()) 2053 Flags |= MachineMemOperand::MOVolatile; 2054 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(&AI)) { 2055 if (CmpX->isVolatile()) 2056 Flags |= MachineMemOperand::MOVolatile; 2057 } else 2058 llvm_unreachable("not an atomic instruction"); 2059 2060 // FIXME: Not preserving dereferenceable 2061 Flags |= getTargetMMOFlags(AI); 2062 return Flags; 2063 } 2064