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/CodeGen/Analysis.h" 21 #include "llvm/CodeGen/ISDOpcodes.h" 22 #include "llvm/CodeGen/MachineBasicBlock.h" 23 #include "llvm/CodeGen/MachineFrameInfo.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/RuntimeLibcalls.h" 31 #include "llvm/CodeGen/StackMaps.h" 32 #include "llvm/CodeGen/TargetLowering.h" 33 #include "llvm/CodeGen/TargetOpcodes.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/ValueTypes.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/CallingConv.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DerivedTypes.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/IR/GlobalValue.h" 42 #include "llvm/IR/GlobalVariable.h" 43 #include "llvm/IR/IRBuilder.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/IR/Type.h" 46 #include "llvm/Support/BranchProbability.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Compiler.h" 50 #include "llvm/Support/ErrorHandling.h" 51 #include "llvm/Support/MachineValueType.h" 52 #include "llvm/Support/MathExtras.h" 53 #include "llvm/Target/TargetMachine.h" 54 #include <algorithm> 55 #include <cassert> 56 #include <cstddef> 57 #include <cstdint> 58 #include <cstring> 59 #include <iterator> 60 #include <string> 61 #include <tuple> 62 #include <utility> 63 64 using namespace llvm; 65 66 static cl::opt<bool> JumpIsExpensiveOverride( 67 "jump-is-expensive", cl::init(false), 68 cl::desc("Do not create extra branches to split comparison logic."), 69 cl::Hidden); 70 71 static cl::opt<unsigned> MinimumJumpTableEntries 72 ("min-jump-table-entries", cl::init(4), cl::Hidden, 73 cl::desc("Set minimum number of entries to use a jump table.")); 74 75 static cl::opt<unsigned> MaximumJumpTableSize 76 ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden, 77 cl::desc("Set maximum size of jump tables.")); 78 79 /// Minimum jump table density for normal functions. 80 static cl::opt<unsigned> 81 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 82 cl::desc("Minimum density for building a jump table in " 83 "a normal function")); 84 85 /// Minimum jump table density for -Os or -Oz functions. 86 static cl::opt<unsigned> OptsizeJumpTableDensity( 87 "optsize-jump-table-density", cl::init(40), cl::Hidden, 88 cl::desc("Minimum density for building a jump table in " 89 "an optsize function")); 90 91 static bool darwinHasSinCos(const Triple &TT) { 92 assert(TT.isOSDarwin() && "should be called with darwin triple"); 93 // Don't bother with 32 bit x86. 94 if (TT.getArch() == Triple::x86) 95 return false; 96 // Macos < 10.9 has no sincos_stret. 97 if (TT.isMacOSX()) 98 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 99 // iOS < 7.0 has no sincos_stret. 100 if (TT.isiOS()) 101 return !TT.isOSVersionLT(7, 0); 102 // Any other darwin such as WatchOS/TvOS is new enough. 103 return true; 104 } 105 106 // Although this default value is arbitrary, it is not random. It is assumed 107 // that a condition that evaluates the same way by a higher percentage than this 108 // is best represented as control flow. Therefore, the default value N should be 109 // set such that the win from N% correct executions is greater than the loss 110 // from (100 - N)% mispredicted executions for the majority of intended targets. 111 static cl::opt<int> MinPercentageForPredictableBranch( 112 "min-predictable-branch", cl::init(99), 113 cl::desc("Minimum percentage (0-100) that a condition must be either true " 114 "or false to assume that the condition is predictable"), 115 cl::Hidden); 116 117 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 118 #define HANDLE_LIBCALL(code, name) \ 119 setLibcallName(RTLIB::code, name); 120 #include "llvm/IR/RuntimeLibcalls.def" 121 #undef HANDLE_LIBCALL 122 // Initialize calling conventions to their default. 123 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 124 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 125 126 // A few names are different on particular architectures or environments. 127 if (TT.isOSDarwin()) { 128 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 129 // of the gnueabi-style __gnu_*_ieee. 130 // FIXME: What about other targets? 131 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 132 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 133 134 // Some darwins have an optimized __bzero/bzero function. 135 switch (TT.getArch()) { 136 case Triple::x86: 137 case Triple::x86_64: 138 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 139 setLibcallName(RTLIB::BZERO, "__bzero"); 140 break; 141 case Triple::aarch64: 142 setLibcallName(RTLIB::BZERO, "bzero"); 143 break; 144 default: 145 break; 146 } 147 148 if (darwinHasSinCos(TT)) { 149 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 150 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 151 if (TT.isWatchABI()) { 152 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 153 CallingConv::ARM_AAPCS_VFP); 154 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 155 CallingConv::ARM_AAPCS_VFP); 156 } 157 } 158 } else { 159 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 160 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 161 } 162 163 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 164 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 165 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 166 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 167 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 168 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 169 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 170 } 171 172 if (TT.isOSOpenBSD()) { 173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 174 } 175 } 176 177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 178 /// UNKNOWN_LIBCALL if there is none. 179 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 180 if (OpVT == MVT::f16) { 181 if (RetVT == MVT::f32) 182 return FPEXT_F16_F32; 183 } else if (OpVT == MVT::f32) { 184 if (RetVT == MVT::f64) 185 return FPEXT_F32_F64; 186 if (RetVT == MVT::f128) 187 return FPEXT_F32_F128; 188 if (RetVT == MVT::ppcf128) 189 return FPEXT_F32_PPCF128; 190 } else if (OpVT == MVT::f64) { 191 if (RetVT == MVT::f128) 192 return FPEXT_F64_F128; 193 else if (RetVT == MVT::ppcf128) 194 return FPEXT_F64_PPCF128; 195 } else if (OpVT == MVT::f80) { 196 if (RetVT == MVT::f128) 197 return FPEXT_F80_F128; 198 } 199 200 return UNKNOWN_LIBCALL; 201 } 202 203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 204 /// UNKNOWN_LIBCALL if there is none. 205 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 206 if (RetVT == MVT::f16) { 207 if (OpVT == MVT::f32) 208 return FPROUND_F32_F16; 209 if (OpVT == MVT::f64) 210 return FPROUND_F64_F16; 211 if (OpVT == MVT::f80) 212 return FPROUND_F80_F16; 213 if (OpVT == MVT::f128) 214 return FPROUND_F128_F16; 215 if (OpVT == MVT::ppcf128) 216 return FPROUND_PPCF128_F16; 217 } else if (RetVT == MVT::f32) { 218 if (OpVT == MVT::f64) 219 return FPROUND_F64_F32; 220 if (OpVT == MVT::f80) 221 return FPROUND_F80_F32; 222 if (OpVT == MVT::f128) 223 return FPROUND_F128_F32; 224 if (OpVT == MVT::ppcf128) 225 return FPROUND_PPCF128_F32; 226 } else if (RetVT == MVT::f64) { 227 if (OpVT == MVT::f80) 228 return FPROUND_F80_F64; 229 if (OpVT == MVT::f128) 230 return FPROUND_F128_F64; 231 if (OpVT == MVT::ppcf128) 232 return FPROUND_PPCF128_F64; 233 } else if (RetVT == MVT::f80) { 234 if (OpVT == MVT::f128) 235 return FPROUND_F128_F80; 236 } 237 238 return UNKNOWN_LIBCALL; 239 } 240 241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 242 /// UNKNOWN_LIBCALL if there is none. 243 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 244 if (OpVT == MVT::f32) { 245 if (RetVT == MVT::i32) 246 return FPTOSINT_F32_I32; 247 if (RetVT == MVT::i64) 248 return FPTOSINT_F32_I64; 249 if (RetVT == MVT::i128) 250 return FPTOSINT_F32_I128; 251 } else if (OpVT == MVT::f64) { 252 if (RetVT == MVT::i32) 253 return FPTOSINT_F64_I32; 254 if (RetVT == MVT::i64) 255 return FPTOSINT_F64_I64; 256 if (RetVT == MVT::i128) 257 return FPTOSINT_F64_I128; 258 } else if (OpVT == MVT::f80) { 259 if (RetVT == MVT::i32) 260 return FPTOSINT_F80_I32; 261 if (RetVT == MVT::i64) 262 return FPTOSINT_F80_I64; 263 if (RetVT == MVT::i128) 264 return FPTOSINT_F80_I128; 265 } else if (OpVT == MVT::f128) { 266 if (RetVT == MVT::i32) 267 return FPTOSINT_F128_I32; 268 if (RetVT == MVT::i64) 269 return FPTOSINT_F128_I64; 270 if (RetVT == MVT::i128) 271 return FPTOSINT_F128_I128; 272 } else if (OpVT == MVT::ppcf128) { 273 if (RetVT == MVT::i32) 274 return FPTOSINT_PPCF128_I32; 275 if (RetVT == MVT::i64) 276 return FPTOSINT_PPCF128_I64; 277 if (RetVT == MVT::i128) 278 return FPTOSINT_PPCF128_I128; 279 } 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOUINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOUINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOUINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOUINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOUINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOUINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOUINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOUINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOUINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOUINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOUINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOUINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOUINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOUINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOUINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::i32) { 329 if (RetVT == MVT::f32) 330 return SINTTOFP_I32_F32; 331 if (RetVT == MVT::f64) 332 return SINTTOFP_I32_F64; 333 if (RetVT == MVT::f80) 334 return SINTTOFP_I32_F80; 335 if (RetVT == MVT::f128) 336 return SINTTOFP_I32_F128; 337 if (RetVT == MVT::ppcf128) 338 return SINTTOFP_I32_PPCF128; 339 } else if (OpVT == MVT::i64) { 340 if (RetVT == MVT::f32) 341 return SINTTOFP_I64_F32; 342 if (RetVT == MVT::f64) 343 return SINTTOFP_I64_F64; 344 if (RetVT == MVT::f80) 345 return SINTTOFP_I64_F80; 346 if (RetVT == MVT::f128) 347 return SINTTOFP_I64_F128; 348 if (RetVT == MVT::ppcf128) 349 return SINTTOFP_I64_PPCF128; 350 } else if (OpVT == MVT::i128) { 351 if (RetVT == MVT::f32) 352 return SINTTOFP_I128_F32; 353 if (RetVT == MVT::f64) 354 return SINTTOFP_I128_F64; 355 if (RetVT == MVT::f80) 356 return SINTTOFP_I128_F80; 357 if (RetVT == MVT::f128) 358 return SINTTOFP_I128_F128; 359 if (RetVT == MVT::ppcf128) 360 return SINTTOFP_I128_PPCF128; 361 } 362 return UNKNOWN_LIBCALL; 363 } 364 365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 366 /// UNKNOWN_LIBCALL if there is none. 367 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 368 if (OpVT == MVT::i32) { 369 if (RetVT == MVT::f32) 370 return UINTTOFP_I32_F32; 371 if (RetVT == MVT::f64) 372 return UINTTOFP_I32_F64; 373 if (RetVT == MVT::f80) 374 return UINTTOFP_I32_F80; 375 if (RetVT == MVT::f128) 376 return UINTTOFP_I32_F128; 377 if (RetVT == MVT::ppcf128) 378 return UINTTOFP_I32_PPCF128; 379 } else if (OpVT == MVT::i64) { 380 if (RetVT == MVT::f32) 381 return UINTTOFP_I64_F32; 382 if (RetVT == MVT::f64) 383 return UINTTOFP_I64_F64; 384 if (RetVT == MVT::f80) 385 return UINTTOFP_I64_F80; 386 if (RetVT == MVT::f128) 387 return UINTTOFP_I64_F128; 388 if (RetVT == MVT::ppcf128) 389 return UINTTOFP_I64_PPCF128; 390 } else if (OpVT == MVT::i128) { 391 if (RetVT == MVT::f32) 392 return UINTTOFP_I128_F32; 393 if (RetVT == MVT::f64) 394 return UINTTOFP_I128_F64; 395 if (RetVT == MVT::f80) 396 return UINTTOFP_I128_F80; 397 if (RetVT == MVT::f128) 398 return UINTTOFP_I128_F128; 399 if (RetVT == MVT::ppcf128) 400 return UINTTOFP_I128_PPCF128; 401 } 402 return UNKNOWN_LIBCALL; 403 } 404 405 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 406 #define OP_TO_LIBCALL(Name, Enum) \ 407 case Name: \ 408 switch (VT.SimpleTy) { \ 409 default: \ 410 return UNKNOWN_LIBCALL; \ 411 case MVT::i8: \ 412 return Enum##_1; \ 413 case MVT::i16: \ 414 return Enum##_2; \ 415 case MVT::i32: \ 416 return Enum##_4; \ 417 case MVT::i64: \ 418 return Enum##_8; \ 419 case MVT::i128: \ 420 return Enum##_16; \ 421 } 422 423 switch (Opc) { 424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 436 } 437 438 #undef OP_TO_LIBCALL 439 440 return UNKNOWN_LIBCALL; 441 } 442 443 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 444 switch (ElementSize) { 445 case 1: 446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 447 case 2: 448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 449 case 4: 450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 451 case 8: 452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 453 case 16: 454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 455 default: 456 return UNKNOWN_LIBCALL; 457 } 458 } 459 460 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 461 switch (ElementSize) { 462 case 1: 463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 464 case 2: 465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 466 case 4: 467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 468 case 8: 469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 470 case 16: 471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 472 default: 473 return UNKNOWN_LIBCALL; 474 } 475 } 476 477 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 /// InitCmpLibcallCCs - Set default comparison libcall CC. 495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 500 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 501 CCs[RTLIB::UNE_F32] = ISD::SETNE; 502 CCs[RTLIB::UNE_F64] = ISD::SETNE; 503 CCs[RTLIB::UNE_F128] = ISD::SETNE; 504 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 505 CCs[RTLIB::OGE_F32] = ISD::SETGE; 506 CCs[RTLIB::OGE_F64] = ISD::SETGE; 507 CCs[RTLIB::OGE_F128] = ISD::SETGE; 508 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 509 CCs[RTLIB::OLT_F32] = ISD::SETLT; 510 CCs[RTLIB::OLT_F64] = ISD::SETLT; 511 CCs[RTLIB::OLT_F128] = ISD::SETLT; 512 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 513 CCs[RTLIB::OLE_F32] = ISD::SETLE; 514 CCs[RTLIB::OLE_F64] = ISD::SETLE; 515 CCs[RTLIB::OLE_F128] = ISD::SETLE; 516 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 517 CCs[RTLIB::OGT_F32] = ISD::SETGT; 518 CCs[RTLIB::OGT_F64] = ISD::SETGT; 519 CCs[RTLIB::OGT_F128] = ISD::SETGT; 520 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 521 CCs[RTLIB::UO_F32] = ISD::SETNE; 522 CCs[RTLIB::UO_F64] = ISD::SETNE; 523 CCs[RTLIB::UO_F128] = ISD::SETNE; 524 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 525 CCs[RTLIB::O_F32] = ISD::SETEQ; 526 CCs[RTLIB::O_F64] = ISD::SETEQ; 527 CCs[RTLIB::O_F128] = ISD::SETEQ; 528 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 529 } 530 531 /// NOTE: The TargetMachine owns TLOF. 532 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 533 initActions(); 534 535 // Perform these initializations only once. 536 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 537 MaxLoadsPerMemcmp = 8; 538 MaxGluedStoresPerMemcpy = 0; 539 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 540 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 541 UseUnderscoreSetJmp = false; 542 UseUnderscoreLongJmp = false; 543 HasMultipleConditionRegisters = false; 544 HasExtractBitsInsn = false; 545 JumpIsExpensive = JumpIsExpensiveOverride; 546 PredictableSelectIsExpensive = false; 547 EnableExtLdPromotion = false; 548 StackPointerRegisterToSaveRestore = 0; 549 BooleanContents = UndefinedBooleanContent; 550 BooleanFloatContents = UndefinedBooleanContent; 551 BooleanVectorContents = UndefinedBooleanContent; 552 SchedPreferenceInfo = Sched::ILP; 553 JumpBufSize = 0; 554 JumpBufAlignment = 0; 555 MinFunctionAlignment = 0; 556 PrefFunctionAlignment = 0; 557 PrefLoopAlignment = 0; 558 GatherAllAliasesMaxDepth = 18; 559 MinStackArgumentAlignment = 1; 560 // TODO: the default will be switched to 0 in the next commit, along 561 // with the Target-specific changes necessary. 562 MaxAtomicSizeInBitsSupported = 1024; 563 564 MinCmpXchgSizeInBits = 0; 565 SupportsUnalignedAtomics = false; 566 567 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 568 569 InitLibcalls(TM.getTargetTriple()); 570 InitCmpLibcallCCs(CmpLibcallCCs); 571 } 572 573 void TargetLoweringBase::initActions() { 574 // All operations default to being supported. 575 memset(OpActions, 0, sizeof(OpActions)); 576 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 577 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 578 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 579 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 580 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 581 std::fill(std::begin(TargetDAGCombineArray), 582 std::end(TargetDAGCombineArray), 0); 583 584 for (MVT VT : MVT::fp_valuetypes()) { 585 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 586 if (IntVT.isValid()) { 587 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 588 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 589 } 590 } 591 592 // Set default actions for various operations. 593 for (MVT VT : MVT::all_valuetypes()) { 594 // Default all indexed load / store to expand. 595 for (unsigned IM = (unsigned)ISD::PRE_INC; 596 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 597 setIndexedLoadAction(IM, VT, Expand); 598 setIndexedStoreAction(IM, VT, Expand); 599 } 600 601 // Most backends expect to see the node which just returns the value loaded. 602 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 603 604 // These operations default to expand. 605 setOperationAction(ISD::FGETSIGN, VT, Expand); 606 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 607 setOperationAction(ISD::FMINNUM, VT, Expand); 608 setOperationAction(ISD::FMAXNUM, VT, Expand); 609 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 610 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 611 setOperationAction(ISD::FMINIMUM, VT, Expand); 612 setOperationAction(ISD::FMAXIMUM, VT, Expand); 613 setOperationAction(ISD::FMAD, VT, Expand); 614 setOperationAction(ISD::SMIN, VT, Expand); 615 setOperationAction(ISD::SMAX, VT, Expand); 616 setOperationAction(ISD::UMIN, VT, Expand); 617 setOperationAction(ISD::UMAX, VT, Expand); 618 setOperationAction(ISD::ABS, VT, Expand); 619 setOperationAction(ISD::FSHL, VT, Expand); 620 setOperationAction(ISD::FSHR, VT, Expand); 621 setOperationAction(ISD::SADDSAT, VT, Expand); 622 setOperationAction(ISD::UADDSAT, VT, Expand); 623 setOperationAction(ISD::SSUBSAT, VT, Expand); 624 setOperationAction(ISD::USUBSAT, VT, Expand); 625 setOperationAction(ISD::SMULFIX, VT, Expand); 626 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 627 setOperationAction(ISD::UMULFIX, VT, Expand); 628 629 // Overflow operations default to expand 630 setOperationAction(ISD::SADDO, VT, Expand); 631 setOperationAction(ISD::SSUBO, VT, Expand); 632 setOperationAction(ISD::UADDO, VT, Expand); 633 setOperationAction(ISD::USUBO, VT, Expand); 634 setOperationAction(ISD::SMULO, VT, Expand); 635 setOperationAction(ISD::UMULO, VT, Expand); 636 637 // ADDCARRY operations default to expand 638 setOperationAction(ISD::ADDCARRY, VT, Expand); 639 setOperationAction(ISD::SUBCARRY, VT, Expand); 640 setOperationAction(ISD::SETCCCARRY, VT, Expand); 641 642 // ADDC/ADDE/SUBC/SUBE default to expand. 643 setOperationAction(ISD::ADDC, VT, Expand); 644 setOperationAction(ISD::ADDE, VT, Expand); 645 setOperationAction(ISD::SUBC, VT, Expand); 646 setOperationAction(ISD::SUBE, VT, Expand); 647 648 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 649 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 650 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 651 652 setOperationAction(ISD::BITREVERSE, VT, Expand); 653 654 // These library functions default to expand. 655 setOperationAction(ISD::FROUND, VT, Expand); 656 setOperationAction(ISD::FPOWI, VT, Expand); 657 658 // These operations default to expand for vector types. 659 if (VT.isVector()) { 660 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 661 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 662 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 663 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 664 } 665 666 // For most targets @llvm.get.dynamic.area.offset just returns 0. 667 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 668 669 // Vector reduction default to expand. 670 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 671 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 672 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 673 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 674 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 675 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 676 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 677 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 678 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 679 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 680 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 681 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 682 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 683 } 684 685 // Most targets ignore the @llvm.prefetch intrinsic. 686 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 687 688 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 689 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 690 691 // ConstantFP nodes default to expand. Targets can either change this to 692 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 693 // to optimize expansions for certain constants. 694 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 695 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 696 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 697 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 698 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 699 700 // These library functions default to expand. 701 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 702 setOperationAction(ISD::FCBRT, VT, Expand); 703 setOperationAction(ISD::FLOG , VT, Expand); 704 setOperationAction(ISD::FLOG2, VT, Expand); 705 setOperationAction(ISD::FLOG10, VT, Expand); 706 setOperationAction(ISD::FEXP , VT, Expand); 707 setOperationAction(ISD::FEXP2, VT, Expand); 708 setOperationAction(ISD::FFLOOR, VT, Expand); 709 setOperationAction(ISD::FNEARBYINT, VT, Expand); 710 setOperationAction(ISD::FCEIL, VT, Expand); 711 setOperationAction(ISD::FRINT, VT, Expand); 712 setOperationAction(ISD::FTRUNC, VT, Expand); 713 setOperationAction(ISD::FROUND, VT, Expand); 714 setOperationAction(ISD::LROUND, VT, Expand); 715 setOperationAction(ISD::LLROUND, VT, Expand); 716 setOperationAction(ISD::LRINT, VT, Expand); 717 setOperationAction(ISD::LLRINT, VT, Expand); 718 } 719 720 // Default ISD::TRAP to expand (which turns it into abort). 721 setOperationAction(ISD::TRAP, MVT::Other, Expand); 722 723 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 724 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 725 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 726 } 727 728 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 729 EVT) const { 730 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 731 } 732 733 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 734 bool LegalTypes) const { 735 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 736 if (LHSTy.isVector()) 737 return LHSTy; 738 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 739 : getPointerTy(DL); 740 } 741 742 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 743 assert(isTypeLegal(VT)); 744 switch (Op) { 745 default: 746 return false; 747 case ISD::SDIV: 748 case ISD::UDIV: 749 case ISD::SREM: 750 case ISD::UREM: 751 return true; 752 } 753 } 754 755 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 756 // If the command-line option was specified, ignore this request. 757 if (!JumpIsExpensiveOverride.getNumOccurrences()) 758 JumpIsExpensive = isExpensive; 759 } 760 761 TargetLoweringBase::LegalizeKind 762 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 763 // If this is a simple type, use the ComputeRegisterProp mechanism. 764 if (VT.isSimple()) { 765 MVT SVT = VT.getSimpleVT(); 766 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 767 MVT NVT = TransformToType[SVT.SimpleTy]; 768 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 769 770 assert((LA == TypeLegal || LA == TypeSoftenFloat || 771 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 772 "Promote may not follow Expand or Promote"); 773 774 if (LA == TypeSplitVector) 775 return LegalizeKind(LA, 776 EVT::getVectorVT(Context, SVT.getVectorElementType(), 777 SVT.getVectorNumElements() / 2)); 778 if (LA == TypeScalarizeVector) 779 return LegalizeKind(LA, SVT.getVectorElementType()); 780 return LegalizeKind(LA, NVT); 781 } 782 783 // Handle Extended Scalar Types. 784 if (!VT.isVector()) { 785 assert(VT.isInteger() && "Float types must be simple"); 786 unsigned BitSize = VT.getSizeInBits(); 787 // First promote to a power-of-two size, then expand if necessary. 788 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 789 EVT NVT = VT.getRoundIntegerType(Context); 790 assert(NVT != VT && "Unable to round integer VT"); 791 LegalizeKind NextStep = getTypeConversion(Context, NVT); 792 // Avoid multi-step promotion. 793 if (NextStep.first == TypePromoteInteger) 794 return NextStep; 795 // Return rounded integer type. 796 return LegalizeKind(TypePromoteInteger, NVT); 797 } 798 799 return LegalizeKind(TypeExpandInteger, 800 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 801 } 802 803 // Handle vector types. 804 unsigned NumElts = VT.getVectorNumElements(); 805 EVT EltVT = VT.getVectorElementType(); 806 807 // Vectors with only one element are always scalarized. 808 if (NumElts == 1) 809 return LegalizeKind(TypeScalarizeVector, EltVT); 810 811 // Try to widen vector elements until the element type is a power of two and 812 // promote it to a legal type later on, for example: 813 // <3 x i8> -> <4 x i8> -> <4 x i32> 814 if (EltVT.isInteger()) { 815 // Vectors with a number of elements that is not a power of two are always 816 // widened, for example <3 x i8> -> <4 x i8>. 817 if (!VT.isPow2VectorType()) { 818 NumElts = (unsigned)NextPowerOf2(NumElts); 819 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 820 return LegalizeKind(TypeWidenVector, NVT); 821 } 822 823 // Examine the element type. 824 LegalizeKind LK = getTypeConversion(Context, EltVT); 825 826 // If type is to be expanded, split the vector. 827 // <4 x i140> -> <2 x i140> 828 if (LK.first == TypeExpandInteger) 829 return LegalizeKind(TypeSplitVector, 830 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 831 832 // Promote the integer element types until a legal vector type is found 833 // or until the element integer type is too big. If a legal type was not 834 // found, fallback to the usual mechanism of widening/splitting the 835 // vector. 836 EVT OldEltVT = EltVT; 837 while (true) { 838 // Increase the bitwidth of the element to the next pow-of-two 839 // (which is greater than 8 bits). 840 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 841 .getRoundIntegerType(Context); 842 843 // Stop trying when getting a non-simple element type. 844 // Note that vector elements may be greater than legal vector element 845 // types. Example: X86 XMM registers hold 64bit element on 32bit 846 // systems. 847 if (!EltVT.isSimple()) 848 break; 849 850 // Build a new vector type and check if it is legal. 851 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 852 // Found a legal promoted vector type. 853 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 854 return LegalizeKind(TypePromoteInteger, 855 EVT::getVectorVT(Context, EltVT, NumElts)); 856 } 857 858 // Reset the type to the unexpanded type if we did not find a legal vector 859 // type with a promoted vector element type. 860 EltVT = OldEltVT; 861 } 862 863 // Try to widen the vector until a legal type is found. 864 // If there is no wider legal type, split the vector. 865 while (true) { 866 // Round up to the next power of 2. 867 NumElts = (unsigned)NextPowerOf2(NumElts); 868 869 // If there is no simple vector type with this many elements then there 870 // cannot be a larger legal vector type. Note that this assumes that 871 // there are no skipped intermediate vector types in the simple types. 872 if (!EltVT.isSimple()) 873 break; 874 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 875 if (LargerVector == MVT()) 876 break; 877 878 // If this type is legal then widen the vector. 879 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 880 return LegalizeKind(TypeWidenVector, LargerVector); 881 } 882 883 // Widen odd vectors to next power of two. 884 if (!VT.isPow2VectorType()) { 885 EVT NVT = VT.getPow2VectorType(Context); 886 return LegalizeKind(TypeWidenVector, NVT); 887 } 888 889 // Vectors with illegal element types are expanded. 890 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 891 return LegalizeKind(TypeSplitVector, NVT); 892 } 893 894 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 895 unsigned &NumIntermediates, 896 MVT &RegisterVT, 897 TargetLoweringBase *TLI) { 898 // Figure out the right, legal destination reg to copy into. 899 unsigned NumElts = VT.getVectorNumElements(); 900 MVT EltTy = VT.getVectorElementType(); 901 902 unsigned NumVectorRegs = 1; 903 904 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 905 // could break down into LHS/RHS like LegalizeDAG does. 906 if (!isPowerOf2_32(NumElts)) { 907 NumVectorRegs = NumElts; 908 NumElts = 1; 909 } 910 911 // Divide the input until we get to a supported size. This will always 912 // end with a scalar if the target doesn't support vectors. 913 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 914 NumElts >>= 1; 915 NumVectorRegs <<= 1; 916 } 917 918 NumIntermediates = NumVectorRegs; 919 920 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 921 if (!TLI->isTypeLegal(NewVT)) 922 NewVT = EltTy; 923 IntermediateVT = NewVT; 924 925 unsigned NewVTSize = NewVT.getSizeInBits(); 926 927 // Convert sizes such as i33 to i64. 928 if (!isPowerOf2_32(NewVTSize)) 929 NewVTSize = NextPowerOf2(NewVTSize); 930 931 MVT DestVT = TLI->getRegisterType(NewVT); 932 RegisterVT = DestVT; 933 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 934 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 935 936 // Otherwise, promotion or legal types use the same number of registers as 937 // the vector decimated to the appropriate level. 938 return NumVectorRegs; 939 } 940 941 /// isLegalRC - Return true if the value types that can be represented by the 942 /// specified register class are all legal. 943 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 944 const TargetRegisterClass &RC) const { 945 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 946 if (isTypeLegal(*I)) 947 return true; 948 return false; 949 } 950 951 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 952 /// sequence of memory operands that is recognized by PrologEpilogInserter. 953 MachineBasicBlock * 954 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 955 MachineBasicBlock *MBB) const { 956 MachineInstr *MI = &InitialMI; 957 MachineFunction &MF = *MI->getMF(); 958 MachineFrameInfo &MFI = MF.getFrameInfo(); 959 960 // We're handling multiple types of operands here: 961 // PATCHPOINT MetaArgs - live-in, read only, direct 962 // STATEPOINT Deopt Spill - live-through, read only, indirect 963 // STATEPOINT Deopt Alloca - live-through, read only, direct 964 // (We're currently conservative and mark the deopt slots read/write in 965 // practice.) 966 // STATEPOINT GC Spill - live-through, read/write, indirect 967 // STATEPOINT GC Alloca - live-through, read/write, direct 968 // The live-in vs live-through is handled already (the live through ones are 969 // all stack slots), but we need to handle the different type of stackmap 970 // operands and memory effects here. 971 972 // MI changes inside this loop as we grow operands. 973 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 974 MachineOperand &MO = MI->getOperand(OperIdx); 975 if (!MO.isFI()) 976 continue; 977 978 // foldMemoryOperand builds a new MI after replacing a single FI operand 979 // with the canonical set of five x86 addressing-mode operands. 980 int FI = MO.getIndex(); 981 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 982 983 // Copy operands before the frame-index. 984 for (unsigned i = 0; i < OperIdx; ++i) 985 MIB.add(MI->getOperand(i)); 986 // Add frame index operands recognized by stackmaps.cpp 987 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 988 // indirect-mem-ref tag, size, #FI, offset. 989 // Used for spills inserted by StatepointLowering. This codepath is not 990 // used for patchpoints/stackmaps at all, for these spilling is done via 991 // foldMemoryOperand callback only. 992 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 993 MIB.addImm(StackMaps::IndirectMemRefOp); 994 MIB.addImm(MFI.getObjectSize(FI)); 995 MIB.add(MI->getOperand(OperIdx)); 996 MIB.addImm(0); 997 } else { 998 // direct-mem-ref tag, #FI, offset. 999 // Used by patchpoint, and direct alloca arguments to statepoints 1000 MIB.addImm(StackMaps::DirectMemRefOp); 1001 MIB.add(MI->getOperand(OperIdx)); 1002 MIB.addImm(0); 1003 } 1004 // Copy the operands after the frame index. 1005 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1006 MIB.add(MI->getOperand(i)); 1007 1008 // Inherit previous memory operands. 1009 MIB.cloneMemRefs(*MI); 1010 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1011 1012 // Add a new memory operand for this FI. 1013 assert(MFI.getObjectOffset(FI) != -1); 1014 1015 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1016 // PATCHPOINT should be updated to do the same. (TODO) 1017 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1018 auto Flags = MachineMemOperand::MOLoad; 1019 MachineMemOperand *MMO = MF.getMachineMemOperand( 1020 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1021 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1022 MIB->addMemOperand(MF, MMO); 1023 } 1024 1025 // Replace the instruction and update the operand index. 1026 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1027 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1028 MI->eraseFromParent(); 1029 MI = MIB; 1030 } 1031 return MBB; 1032 } 1033 1034 MachineBasicBlock * 1035 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1036 MachineBasicBlock *MBB) const { 1037 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1038 "Called emitXRayCustomEvent on the wrong MI!"); 1039 auto &MF = *MI.getMF(); 1040 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1041 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1042 MIB.add(MI.getOperand(OpIdx)); 1043 1044 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1045 MI.eraseFromParent(); 1046 return MBB; 1047 } 1048 1049 MachineBasicBlock * 1050 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1051 MachineBasicBlock *MBB) const { 1052 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1053 "Called emitXRayTypedEvent on the wrong MI!"); 1054 auto &MF = *MI.getMF(); 1055 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1056 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1057 MIB.add(MI.getOperand(OpIdx)); 1058 1059 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1060 MI.eraseFromParent(); 1061 return MBB; 1062 } 1063 1064 /// findRepresentativeClass - Return the largest legal super-reg register class 1065 /// of the register class for the specified type and its associated "cost". 1066 // This function is in TargetLowering because it uses RegClassForVT which would 1067 // need to be moved to TargetRegisterInfo and would necessitate moving 1068 // isTypeLegal over as well - a massive change that would just require 1069 // TargetLowering having a TargetRegisterInfo class member that it would use. 1070 std::pair<const TargetRegisterClass *, uint8_t> 1071 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1072 MVT VT) const { 1073 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1074 if (!RC) 1075 return std::make_pair(RC, 0); 1076 1077 // Compute the set of all super-register classes. 1078 BitVector SuperRegRC(TRI->getNumRegClasses()); 1079 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1080 SuperRegRC.setBitsInMask(RCI.getMask()); 1081 1082 // Find the first legal register class with the largest spill size. 1083 const TargetRegisterClass *BestRC = RC; 1084 for (unsigned i : SuperRegRC.set_bits()) { 1085 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1086 // We want the largest possible spill size. 1087 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1088 continue; 1089 if (!isLegalRC(*TRI, *SuperRC)) 1090 continue; 1091 BestRC = SuperRC; 1092 } 1093 return std::make_pair(BestRC, 1); 1094 } 1095 1096 /// computeRegisterProperties - Once all of the register classes are added, 1097 /// this allows us to compute derived properties we expose. 1098 void TargetLoweringBase::computeRegisterProperties( 1099 const TargetRegisterInfo *TRI) { 1100 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1101 "Too many value types for ValueTypeActions to hold!"); 1102 1103 // Everything defaults to needing one register. 1104 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1105 NumRegistersForVT[i] = 1; 1106 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1107 } 1108 // ...except isVoid, which doesn't need any registers. 1109 NumRegistersForVT[MVT::isVoid] = 0; 1110 1111 // Find the largest integer register class. 1112 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1113 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1114 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1115 1116 // Every integer value type larger than this largest register takes twice as 1117 // many registers to represent as the previous ValueType. 1118 for (unsigned ExpandedReg = LargestIntReg + 1; 1119 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1120 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1121 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1122 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1123 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1124 TypeExpandInteger); 1125 } 1126 1127 // Inspect all of the ValueType's smaller than the largest integer 1128 // register to see which ones need promotion. 1129 unsigned LegalIntReg = LargestIntReg; 1130 for (unsigned IntReg = LargestIntReg - 1; 1131 IntReg >= (unsigned)MVT::i1; --IntReg) { 1132 MVT IVT = (MVT::SimpleValueType)IntReg; 1133 if (isTypeLegal(IVT)) { 1134 LegalIntReg = IntReg; 1135 } else { 1136 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1137 (MVT::SimpleValueType)LegalIntReg; 1138 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1139 } 1140 } 1141 1142 // ppcf128 type is really two f64's. 1143 if (!isTypeLegal(MVT::ppcf128)) { 1144 if (isTypeLegal(MVT::f64)) { 1145 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1146 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1147 TransformToType[MVT::ppcf128] = MVT::f64; 1148 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1149 } else { 1150 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1151 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1152 TransformToType[MVT::ppcf128] = MVT::i128; 1153 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1154 } 1155 } 1156 1157 // Decide how to handle f128. If the target does not have native f128 support, 1158 // expand it to i128 and we will be generating soft float library calls. 1159 if (!isTypeLegal(MVT::f128)) { 1160 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1161 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1162 TransformToType[MVT::f128] = MVT::i128; 1163 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1164 } 1165 1166 // Decide how to handle f64. If the target does not have native f64 support, 1167 // expand it to i64 and we will be generating soft float library calls. 1168 if (!isTypeLegal(MVT::f64)) { 1169 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1170 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1171 TransformToType[MVT::f64] = MVT::i64; 1172 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1173 } 1174 1175 // Decide how to handle f32. If the target does not have native f32 support, 1176 // expand it to i32 and we will be generating soft float library calls. 1177 if (!isTypeLegal(MVT::f32)) { 1178 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1179 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1180 TransformToType[MVT::f32] = MVT::i32; 1181 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1182 } 1183 1184 // Decide how to handle f16. If the target does not have native f16 support, 1185 // promote it to f32, because there are no f16 library calls (except for 1186 // conversions). 1187 if (!isTypeLegal(MVT::f16)) { 1188 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1189 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1190 TransformToType[MVT::f16] = MVT::f32; 1191 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1192 } 1193 1194 // Loop over all of the vector value types to see which need transformations. 1195 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1196 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1197 MVT VT = (MVT::SimpleValueType) i; 1198 if (isTypeLegal(VT)) 1199 continue; 1200 1201 MVT EltVT = VT.getVectorElementType(); 1202 unsigned NElts = VT.getVectorNumElements(); 1203 bool IsLegalWiderType = false; 1204 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1205 switch (PreferredAction) { 1206 case TypePromoteInteger: 1207 // Try to promote the elements of integer vectors. If no legal 1208 // promotion was found, fall through to the widen-vector method. 1209 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1210 MVT SVT = (MVT::SimpleValueType) nVT; 1211 // Promote vectors of integers to vectors with the same number 1212 // of elements, with a wider element type. 1213 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1214 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1215 TransformToType[i] = SVT; 1216 RegisterTypeForVT[i] = SVT; 1217 NumRegistersForVT[i] = 1; 1218 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1219 IsLegalWiderType = true; 1220 break; 1221 } 1222 } 1223 if (IsLegalWiderType) 1224 break; 1225 LLVM_FALLTHROUGH; 1226 1227 case TypeWidenVector: 1228 // Try to widen the vector. 1229 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1230 MVT SVT = (MVT::SimpleValueType) nVT; 1231 if (SVT.getVectorElementType() == EltVT 1232 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1233 TransformToType[i] = SVT; 1234 RegisterTypeForVT[i] = SVT; 1235 NumRegistersForVT[i] = 1; 1236 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1237 IsLegalWiderType = true; 1238 break; 1239 } 1240 } 1241 if (IsLegalWiderType) 1242 break; 1243 LLVM_FALLTHROUGH; 1244 1245 case TypeSplitVector: 1246 case TypeScalarizeVector: { 1247 MVT IntermediateVT; 1248 MVT RegisterVT; 1249 unsigned NumIntermediates; 1250 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1251 NumIntermediates, RegisterVT, this); 1252 RegisterTypeForVT[i] = RegisterVT; 1253 1254 MVT NVT = VT.getPow2VectorType(); 1255 if (NVT == VT) { 1256 // Type is already a power of 2. The default action is to split. 1257 TransformToType[i] = MVT::Other; 1258 if (PreferredAction == TypeScalarizeVector) 1259 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1260 else if (PreferredAction == TypeSplitVector) 1261 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1262 else 1263 // Set type action according to the number of elements. 1264 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1265 : TypeSplitVector); 1266 } else { 1267 TransformToType[i] = NVT; 1268 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1269 } 1270 break; 1271 } 1272 default: 1273 llvm_unreachable("Unknown vector legalization action!"); 1274 } 1275 } 1276 1277 // Determine the 'representative' register class for each value type. 1278 // An representative register class is the largest (meaning one which is 1279 // not a sub-register class / subreg register class) legal register class for 1280 // a group of value types. For example, on i386, i8, i16, and i32 1281 // representative would be GR32; while on x86_64 it's GR64. 1282 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1283 const TargetRegisterClass* RRC; 1284 uint8_t Cost; 1285 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1286 RepRegClassForVT[i] = RRC; 1287 RepRegClassCostForVT[i] = Cost; 1288 } 1289 } 1290 1291 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1292 EVT VT) const { 1293 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1294 return getPointerTy(DL).SimpleTy; 1295 } 1296 1297 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1298 return MVT::i32; // return the default value 1299 } 1300 1301 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1302 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1303 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1304 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1305 /// 1306 /// This method returns the number of registers needed, and the VT for each 1307 /// register. It also returns the VT and quantity of the intermediate values 1308 /// before they are promoted/expanded. 1309 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1310 EVT &IntermediateVT, 1311 unsigned &NumIntermediates, 1312 MVT &RegisterVT) const { 1313 unsigned NumElts = VT.getVectorNumElements(); 1314 1315 // If there is a wider vector type with the same element type as this one, 1316 // or a promoted vector type that has the same number of elements which 1317 // are wider, then we should convert to that legal vector type. 1318 // This handles things like <2 x float> -> <4 x float> and 1319 // <4 x i1> -> <4 x i32>. 1320 LegalizeTypeAction TA = getTypeAction(Context, VT); 1321 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1322 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1323 if (isTypeLegal(RegisterEVT)) { 1324 IntermediateVT = RegisterEVT; 1325 RegisterVT = RegisterEVT.getSimpleVT(); 1326 NumIntermediates = 1; 1327 return 1; 1328 } 1329 } 1330 1331 // Figure out the right, legal destination reg to copy into. 1332 EVT EltTy = VT.getVectorElementType(); 1333 1334 unsigned NumVectorRegs = 1; 1335 1336 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1337 // could break down into LHS/RHS like LegalizeDAG does. 1338 if (!isPowerOf2_32(NumElts)) { 1339 NumVectorRegs = NumElts; 1340 NumElts = 1; 1341 } 1342 1343 // Divide the input until we get to a supported size. This will always 1344 // end with a scalar if the target doesn't support vectors. 1345 while (NumElts > 1 && !isTypeLegal( 1346 EVT::getVectorVT(Context, EltTy, NumElts))) { 1347 NumElts >>= 1; 1348 NumVectorRegs <<= 1; 1349 } 1350 1351 NumIntermediates = NumVectorRegs; 1352 1353 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1354 if (!isTypeLegal(NewVT)) 1355 NewVT = EltTy; 1356 IntermediateVT = NewVT; 1357 1358 MVT DestVT = getRegisterType(Context, NewVT); 1359 RegisterVT = DestVT; 1360 unsigned NewVTSize = NewVT.getSizeInBits(); 1361 1362 // Convert sizes such as i33 to i64. 1363 if (!isPowerOf2_32(NewVTSize)) 1364 NewVTSize = NextPowerOf2(NewVTSize); 1365 1366 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1367 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1368 1369 // Otherwise, promotion or legal types use the same number of registers as 1370 // the vector decimated to the appropriate level. 1371 return NumVectorRegs; 1372 } 1373 1374 /// Get the EVTs and ArgFlags collections that represent the legalized return 1375 /// type of the given function. This does not require a DAG or a return value, 1376 /// and is suitable for use before any DAGs for the function are constructed. 1377 /// TODO: Move this out of TargetLowering.cpp. 1378 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1379 AttributeList attr, 1380 SmallVectorImpl<ISD::OutputArg> &Outs, 1381 const TargetLowering &TLI, const DataLayout &DL) { 1382 SmallVector<EVT, 4> ValueVTs; 1383 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1384 unsigned NumValues = ValueVTs.size(); 1385 if (NumValues == 0) return; 1386 1387 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1388 EVT VT = ValueVTs[j]; 1389 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1390 1391 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1392 ExtendKind = ISD::SIGN_EXTEND; 1393 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1394 ExtendKind = ISD::ZERO_EXTEND; 1395 1396 // FIXME: C calling convention requires the return type to be promoted to 1397 // at least 32-bit. But this is not necessary for non-C calling 1398 // conventions. The frontend should mark functions whose return values 1399 // require promoting with signext or zeroext attributes. 1400 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1401 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1402 if (VT.bitsLT(MinVT)) 1403 VT = MinVT; 1404 } 1405 1406 unsigned NumParts = 1407 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1408 MVT PartVT = 1409 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1410 1411 // 'inreg' on function refers to return value 1412 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1413 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1414 Flags.setInReg(); 1415 1416 // Propagate extension type if any 1417 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1418 Flags.setSExt(); 1419 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1420 Flags.setZExt(); 1421 1422 for (unsigned i = 0; i < NumParts; ++i) 1423 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1424 } 1425 } 1426 1427 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1428 /// function arguments in the caller parameter area. This is the actual 1429 /// alignment, not its logarithm. 1430 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1431 const DataLayout &DL) const { 1432 return DL.getABITypeAlignment(Ty); 1433 } 1434 1435 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1436 const DataLayout &DL, EVT VT, 1437 unsigned AddrSpace, 1438 unsigned Alignment, 1439 bool *Fast) const { 1440 // Check if the specified alignment is sufficient based on the data layout. 1441 // TODO: While using the data layout works in practice, a better solution 1442 // would be to implement this check directly (make this a virtual function). 1443 // For example, the ABI alignment may change based on software platform while 1444 // this function should only be affected by hardware implementation. 1445 Type *Ty = VT.getTypeForEVT(Context); 1446 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1447 // Assume that an access that meets the ABI-specified alignment is fast. 1448 if (Fast != nullptr) 1449 *Fast = true; 1450 return true; 1451 } 1452 1453 // This is a misaligned access. 1454 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1455 } 1456 1457 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1458 return BranchProbability(MinPercentageForPredictableBranch, 100); 1459 } 1460 1461 //===----------------------------------------------------------------------===// 1462 // TargetTransformInfo Helpers 1463 //===----------------------------------------------------------------------===// 1464 1465 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1466 enum InstructionOpcodes { 1467 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1468 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1469 #include "llvm/IR/Instruction.def" 1470 }; 1471 switch (static_cast<InstructionOpcodes>(Opcode)) { 1472 case Ret: return 0; 1473 case Br: return 0; 1474 case Switch: return 0; 1475 case IndirectBr: return 0; 1476 case Invoke: return 0; 1477 case CallBr: return 0; 1478 case Resume: return 0; 1479 case Unreachable: return 0; 1480 case CleanupRet: return 0; 1481 case CatchRet: return 0; 1482 case CatchPad: return 0; 1483 case CatchSwitch: return 0; 1484 case CleanupPad: return 0; 1485 case FNeg: return ISD::FNEG; 1486 case Add: return ISD::ADD; 1487 case FAdd: return ISD::FADD; 1488 case Sub: return ISD::SUB; 1489 case FSub: return ISD::FSUB; 1490 case Mul: return ISD::MUL; 1491 case FMul: return ISD::FMUL; 1492 case UDiv: return ISD::UDIV; 1493 case SDiv: return ISD::SDIV; 1494 case FDiv: return ISD::FDIV; 1495 case URem: return ISD::UREM; 1496 case SRem: return ISD::SREM; 1497 case FRem: return ISD::FREM; 1498 case Shl: return ISD::SHL; 1499 case LShr: return ISD::SRL; 1500 case AShr: return ISD::SRA; 1501 case And: return ISD::AND; 1502 case Or: return ISD::OR; 1503 case Xor: return ISD::XOR; 1504 case Alloca: return 0; 1505 case Load: return ISD::LOAD; 1506 case Store: return ISD::STORE; 1507 case GetElementPtr: return 0; 1508 case Fence: return 0; 1509 case AtomicCmpXchg: return 0; 1510 case AtomicRMW: return 0; 1511 case Trunc: return ISD::TRUNCATE; 1512 case ZExt: return ISD::ZERO_EXTEND; 1513 case SExt: return ISD::SIGN_EXTEND; 1514 case FPToUI: return ISD::FP_TO_UINT; 1515 case FPToSI: return ISD::FP_TO_SINT; 1516 case UIToFP: return ISD::UINT_TO_FP; 1517 case SIToFP: return ISD::SINT_TO_FP; 1518 case FPTrunc: return ISD::FP_ROUND; 1519 case FPExt: return ISD::FP_EXTEND; 1520 case PtrToInt: return ISD::BITCAST; 1521 case IntToPtr: return ISD::BITCAST; 1522 case BitCast: return ISD::BITCAST; 1523 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1524 case ICmp: return ISD::SETCC; 1525 case FCmp: return ISD::SETCC; 1526 case PHI: return 0; 1527 case Call: return 0; 1528 case Select: return ISD::SELECT; 1529 case UserOp1: return 0; 1530 case UserOp2: return 0; 1531 case VAArg: return 0; 1532 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1533 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1534 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1535 case ExtractValue: return ISD::MERGE_VALUES; 1536 case InsertValue: return ISD::MERGE_VALUES; 1537 case LandingPad: return 0; 1538 } 1539 1540 llvm_unreachable("Unknown instruction type encountered!"); 1541 } 1542 1543 std::pair<int, MVT> 1544 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1545 Type *Ty) const { 1546 LLVMContext &C = Ty->getContext(); 1547 EVT MTy = getValueType(DL, Ty); 1548 1549 int Cost = 1; 1550 // We keep legalizing the type until we find a legal kind. We assume that 1551 // the only operation that costs anything is the split. After splitting 1552 // we need to handle two types. 1553 while (true) { 1554 LegalizeKind LK = getTypeConversion(C, MTy); 1555 1556 if (LK.first == TypeLegal) 1557 return std::make_pair(Cost, MTy.getSimpleVT()); 1558 1559 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1560 Cost *= 2; 1561 1562 // Do not loop with f128 type. 1563 if (MTy == LK.second) 1564 return std::make_pair(Cost, MTy.getSimpleVT()); 1565 1566 // Keep legalizing the type. 1567 MTy = LK.second; 1568 } 1569 } 1570 1571 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1572 bool UseTLS) const { 1573 // compiler-rt provides a variable with a magic name. Targets that do not 1574 // link with compiler-rt may also provide such a variable. 1575 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1576 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1577 auto UnsafeStackPtr = 1578 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1579 1580 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1581 1582 if (!UnsafeStackPtr) { 1583 auto TLSModel = UseTLS ? 1584 GlobalValue::InitialExecTLSModel : 1585 GlobalValue::NotThreadLocal; 1586 // The global variable is not defined yet, define it ourselves. 1587 // We use the initial-exec TLS model because we do not support the 1588 // variable living anywhere other than in the main executable. 1589 UnsafeStackPtr = new GlobalVariable( 1590 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1591 UnsafeStackPtrVar, nullptr, TLSModel); 1592 } else { 1593 // The variable exists, check its type and attributes. 1594 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1595 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1596 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1597 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1598 (UseTLS ? "" : "not ") + "be thread-local"); 1599 } 1600 return UnsafeStackPtr; 1601 } 1602 1603 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1604 if (!TM.getTargetTriple().isAndroid()) 1605 return getDefaultSafeStackPointerLocation(IRB, true); 1606 1607 // Android provides a libc function to retrieve the address of the current 1608 // thread's unsafe stack pointer. 1609 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1610 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1611 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1612 StackPtrTy->getPointerTo(0)); 1613 return IRB.CreateCall(Fn); 1614 } 1615 1616 //===----------------------------------------------------------------------===// 1617 // Loop Strength Reduction hooks 1618 //===----------------------------------------------------------------------===// 1619 1620 /// isLegalAddressingMode - Return true if the addressing mode represented 1621 /// by AM is legal for this target, for a load/store of the specified type. 1622 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1623 const AddrMode &AM, Type *Ty, 1624 unsigned AS, Instruction *I) const { 1625 // The default implementation of this implements a conservative RISCy, r+r and 1626 // r+i addr mode. 1627 1628 // Allows a sign-extended 16-bit immediate field. 1629 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1630 return false; 1631 1632 // No global is ever allowed as a base. 1633 if (AM.BaseGV) 1634 return false; 1635 1636 // Only support r+r, 1637 switch (AM.Scale) { 1638 case 0: // "r+i" or just "i", depending on HasBaseReg. 1639 break; 1640 case 1: 1641 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1642 return false; 1643 // Otherwise we have r+r or r+i. 1644 break; 1645 case 2: 1646 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1647 return false; 1648 // Allow 2*r as r+r. 1649 break; 1650 default: // Don't allow n * r 1651 return false; 1652 } 1653 1654 return true; 1655 } 1656 1657 //===----------------------------------------------------------------------===// 1658 // Stack Protector 1659 //===----------------------------------------------------------------------===// 1660 1661 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1662 // so that SelectionDAG handle SSP. 1663 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1664 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1665 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1666 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1667 return M.getOrInsertGlobal("__guard_local", PtrTy); 1668 } 1669 return nullptr; 1670 } 1671 1672 // Currently only support "standard" __stack_chk_guard. 1673 // TODO: add LOAD_STACK_GUARD support. 1674 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1675 if (!M.getNamedValue("__stack_chk_guard")) 1676 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1677 GlobalVariable::ExternalLinkage, 1678 nullptr, "__stack_chk_guard"); 1679 } 1680 1681 // Currently only support "standard" __stack_chk_guard. 1682 // TODO: add LOAD_STACK_GUARD support. 1683 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1684 return M.getNamedValue("__stack_chk_guard"); 1685 } 1686 1687 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1688 return nullptr; 1689 } 1690 1691 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1692 return MinimumJumpTableEntries; 1693 } 1694 1695 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1696 MinimumJumpTableEntries = Val; 1697 } 1698 1699 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1700 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1701 } 1702 1703 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1704 return MaximumJumpTableSize; 1705 } 1706 1707 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1708 MaximumJumpTableSize = Val; 1709 } 1710 1711 //===----------------------------------------------------------------------===// 1712 // Reciprocal Estimates 1713 //===----------------------------------------------------------------------===// 1714 1715 /// Get the reciprocal estimate attribute string for a function that will 1716 /// override the target defaults. 1717 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1718 const Function &F = MF.getFunction(); 1719 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1720 } 1721 1722 /// Construct a string for the given reciprocal operation of the given type. 1723 /// This string should match the corresponding option to the front-end's 1724 /// "-mrecip" flag assuming those strings have been passed through in an 1725 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1726 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1727 std::string Name = VT.isVector() ? "vec-" : ""; 1728 1729 Name += IsSqrt ? "sqrt" : "div"; 1730 1731 // TODO: Handle "half" or other float types? 1732 if (VT.getScalarType() == MVT::f64) { 1733 Name += "d"; 1734 } else { 1735 assert(VT.getScalarType() == MVT::f32 && 1736 "Unexpected FP type for reciprocal estimate"); 1737 Name += "f"; 1738 } 1739 1740 return Name; 1741 } 1742 1743 /// Return the character position and value (a single numeric character) of a 1744 /// customized refinement operation in the input string if it exists. Return 1745 /// false if there is no customized refinement step count. 1746 static bool parseRefinementStep(StringRef In, size_t &Position, 1747 uint8_t &Value) { 1748 const char RefStepToken = ':'; 1749 Position = In.find(RefStepToken); 1750 if (Position == StringRef::npos) 1751 return false; 1752 1753 StringRef RefStepString = In.substr(Position + 1); 1754 // Allow exactly one numeric character for the additional refinement 1755 // step parameter. 1756 if (RefStepString.size() == 1) { 1757 char RefStepChar = RefStepString[0]; 1758 if (RefStepChar >= '0' && RefStepChar <= '9') { 1759 Value = RefStepChar - '0'; 1760 return true; 1761 } 1762 } 1763 report_fatal_error("Invalid refinement step for -recip."); 1764 } 1765 1766 /// For the input attribute string, return one of the ReciprocalEstimate enum 1767 /// status values (enabled, disabled, or not specified) for this operation on 1768 /// the specified data type. 1769 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1770 if (Override.empty()) 1771 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1772 1773 SmallVector<StringRef, 4> OverrideVector; 1774 Override.split(OverrideVector, ','); 1775 unsigned NumArgs = OverrideVector.size(); 1776 1777 // Check if "all", "none", or "default" was specified. 1778 if (NumArgs == 1) { 1779 // Look for an optional setting of the number of refinement steps needed 1780 // for this type of reciprocal operation. 1781 size_t RefPos; 1782 uint8_t RefSteps; 1783 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1784 // Split the string for further processing. 1785 Override = Override.substr(0, RefPos); 1786 } 1787 1788 // All reciprocal types are enabled. 1789 if (Override == "all") 1790 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1791 1792 // All reciprocal types are disabled. 1793 if (Override == "none") 1794 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1795 1796 // Target defaults for enablement are used. 1797 if (Override == "default") 1798 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1799 } 1800 1801 // The attribute string may omit the size suffix ('f'/'d'). 1802 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1803 std::string VTNameNoSize = VTName; 1804 VTNameNoSize.pop_back(); 1805 static const char DisabledPrefix = '!'; 1806 1807 for (StringRef RecipType : OverrideVector) { 1808 size_t RefPos; 1809 uint8_t RefSteps; 1810 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1811 RecipType = RecipType.substr(0, RefPos); 1812 1813 // Ignore the disablement token for string matching. 1814 bool IsDisabled = RecipType[0] == DisabledPrefix; 1815 if (IsDisabled) 1816 RecipType = RecipType.substr(1); 1817 1818 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1819 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1820 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1821 } 1822 1823 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1824 } 1825 1826 /// For the input attribute string, return the customized refinement step count 1827 /// for this operation on the specified data type. If the step count does not 1828 /// exist, return the ReciprocalEstimate enum value for unspecified. 1829 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1830 if (Override.empty()) 1831 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1832 1833 SmallVector<StringRef, 4> OverrideVector; 1834 Override.split(OverrideVector, ','); 1835 unsigned NumArgs = OverrideVector.size(); 1836 1837 // Check if "all", "default", or "none" was specified. 1838 if (NumArgs == 1) { 1839 // Look for an optional setting of the number of refinement steps needed 1840 // for this type of reciprocal operation. 1841 size_t RefPos; 1842 uint8_t RefSteps; 1843 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1844 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1845 1846 // Split the string for further processing. 1847 Override = Override.substr(0, RefPos); 1848 assert(Override != "none" && 1849 "Disabled reciprocals, but specifed refinement steps?"); 1850 1851 // If this is a general override, return the specified number of steps. 1852 if (Override == "all" || Override == "default") 1853 return RefSteps; 1854 } 1855 1856 // The attribute string may omit the size suffix ('f'/'d'). 1857 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1858 std::string VTNameNoSize = VTName; 1859 VTNameNoSize.pop_back(); 1860 1861 for (StringRef RecipType : OverrideVector) { 1862 size_t RefPos; 1863 uint8_t RefSteps; 1864 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1865 continue; 1866 1867 RecipType = RecipType.substr(0, RefPos); 1868 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1869 return RefSteps; 1870 } 1871 1872 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1873 } 1874 1875 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1876 MachineFunction &MF) const { 1877 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1878 } 1879 1880 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1881 MachineFunction &MF) const { 1882 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1883 } 1884 1885 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1886 MachineFunction &MF) const { 1887 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1888 } 1889 1890 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1891 MachineFunction &MF) const { 1892 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1893 } 1894 1895 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1896 MF.getRegInfo().freezeReservedRegs(MF); 1897 } 1898