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 // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf". 127 if (TT.getArch() == Triple::ppc || TT.isPPC64()) { 128 setLibcallName(RTLIB::ADD_F128, "__addkf3"); 129 setLibcallName(RTLIB::SUB_F128, "__subkf3"); 130 setLibcallName(RTLIB::MUL_F128, "__mulkf3"); 131 setLibcallName(RTLIB::DIV_F128, "__divkf3"); 132 setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2"); 133 setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2"); 134 setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2"); 135 setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2"); 136 setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi"); 137 setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi"); 138 setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi"); 139 setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi"); 140 setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf"); 141 setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf"); 142 setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf"); 143 setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf"); 144 setLibcallName(RTLIB::OEQ_F128, "__eqkf2"); 145 setLibcallName(RTLIB::UNE_F128, "__nekf2"); 146 setLibcallName(RTLIB::OGE_F128, "__gekf2"); 147 setLibcallName(RTLIB::OLT_F128, "__ltkf2"); 148 setLibcallName(RTLIB::OLE_F128, "__lekf2"); 149 setLibcallName(RTLIB::OGT_F128, "__gtkf2"); 150 setLibcallName(RTLIB::UO_F128, "__unordkf2"); 151 setLibcallName(RTLIB::O_F128, "__unordkf2"); 152 } 153 154 // A few names are different on particular architectures or environments. 155 if (TT.isOSDarwin()) { 156 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 157 // of the gnueabi-style __gnu_*_ieee. 158 // FIXME: What about other targets? 159 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 160 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 161 162 // Some darwins have an optimized __bzero/bzero function. 163 switch (TT.getArch()) { 164 case Triple::x86: 165 case Triple::x86_64: 166 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 167 setLibcallName(RTLIB::BZERO, "__bzero"); 168 break; 169 case Triple::aarch64: 170 setLibcallName(RTLIB::BZERO, "bzero"); 171 break; 172 default: 173 break; 174 } 175 176 if (darwinHasSinCos(TT)) { 177 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 178 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 179 if (TT.isWatchABI()) { 180 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 181 CallingConv::ARM_AAPCS_VFP); 182 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 183 CallingConv::ARM_AAPCS_VFP); 184 } 185 } 186 } else { 187 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 188 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 189 } 190 191 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 192 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 193 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 194 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 195 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 196 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 197 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 198 } 199 200 if (TT.isPS4CPU()) { 201 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 202 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 203 } 204 205 if (TT.isOSOpenBSD()) { 206 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 207 } 208 } 209 210 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 211 /// UNKNOWN_LIBCALL if there is none. 212 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 213 if (OpVT == MVT::f16) { 214 if (RetVT == MVT::f32) 215 return FPEXT_F16_F32; 216 } else if (OpVT == MVT::f32) { 217 if (RetVT == MVT::f64) 218 return FPEXT_F32_F64; 219 if (RetVT == MVT::f128) 220 return FPEXT_F32_F128; 221 if (RetVT == MVT::ppcf128) 222 return FPEXT_F32_PPCF128; 223 } else if (OpVT == MVT::f64) { 224 if (RetVT == MVT::f128) 225 return FPEXT_F64_F128; 226 else if (RetVT == MVT::ppcf128) 227 return FPEXT_F64_PPCF128; 228 } else if (OpVT == MVT::f80) { 229 if (RetVT == MVT::f128) 230 return FPEXT_F80_F128; 231 } 232 233 return UNKNOWN_LIBCALL; 234 } 235 236 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 237 /// UNKNOWN_LIBCALL if there is none. 238 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 239 if (RetVT == MVT::f16) { 240 if (OpVT == MVT::f32) 241 return FPROUND_F32_F16; 242 if (OpVT == MVT::f64) 243 return FPROUND_F64_F16; 244 if (OpVT == MVT::f80) 245 return FPROUND_F80_F16; 246 if (OpVT == MVT::f128) 247 return FPROUND_F128_F16; 248 if (OpVT == MVT::ppcf128) 249 return FPROUND_PPCF128_F16; 250 } else if (RetVT == MVT::f32) { 251 if (OpVT == MVT::f64) 252 return FPROUND_F64_F32; 253 if (OpVT == MVT::f80) 254 return FPROUND_F80_F32; 255 if (OpVT == MVT::f128) 256 return FPROUND_F128_F32; 257 if (OpVT == MVT::ppcf128) 258 return FPROUND_PPCF128_F32; 259 } else if (RetVT == MVT::f64) { 260 if (OpVT == MVT::f80) 261 return FPROUND_F80_F64; 262 if (OpVT == MVT::f128) 263 return FPROUND_F128_F64; 264 if (OpVT == MVT::ppcf128) 265 return FPROUND_PPCF128_F64; 266 } else if (RetVT == MVT::f80) { 267 if (OpVT == MVT::f128) 268 return FPROUND_F128_F80; 269 } 270 271 return UNKNOWN_LIBCALL; 272 } 273 274 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 275 /// UNKNOWN_LIBCALL if there is none. 276 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 277 if (OpVT == MVT::f32) { 278 if (RetVT == MVT::i32) 279 return FPTOSINT_F32_I32; 280 if (RetVT == MVT::i64) 281 return FPTOSINT_F32_I64; 282 if (RetVT == MVT::i128) 283 return FPTOSINT_F32_I128; 284 } else if (OpVT == MVT::f64) { 285 if (RetVT == MVT::i32) 286 return FPTOSINT_F64_I32; 287 if (RetVT == MVT::i64) 288 return FPTOSINT_F64_I64; 289 if (RetVT == MVT::i128) 290 return FPTOSINT_F64_I128; 291 } else if (OpVT == MVT::f80) { 292 if (RetVT == MVT::i32) 293 return FPTOSINT_F80_I32; 294 if (RetVT == MVT::i64) 295 return FPTOSINT_F80_I64; 296 if (RetVT == MVT::i128) 297 return FPTOSINT_F80_I128; 298 } else if (OpVT == MVT::f128) { 299 if (RetVT == MVT::i32) 300 return FPTOSINT_F128_I32; 301 if (RetVT == MVT::i64) 302 return FPTOSINT_F128_I64; 303 if (RetVT == MVT::i128) 304 return FPTOSINT_F128_I128; 305 } else if (OpVT == MVT::ppcf128) { 306 if (RetVT == MVT::i32) 307 return FPTOSINT_PPCF128_I32; 308 if (RetVT == MVT::i64) 309 return FPTOSINT_PPCF128_I64; 310 if (RetVT == MVT::i128) 311 return FPTOSINT_PPCF128_I128; 312 } 313 return UNKNOWN_LIBCALL; 314 } 315 316 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 317 /// UNKNOWN_LIBCALL if there is none. 318 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 319 if (OpVT == MVT::f32) { 320 if (RetVT == MVT::i32) 321 return FPTOUINT_F32_I32; 322 if (RetVT == MVT::i64) 323 return FPTOUINT_F32_I64; 324 if (RetVT == MVT::i128) 325 return FPTOUINT_F32_I128; 326 } else if (OpVT == MVT::f64) { 327 if (RetVT == MVT::i32) 328 return FPTOUINT_F64_I32; 329 if (RetVT == MVT::i64) 330 return FPTOUINT_F64_I64; 331 if (RetVT == MVT::i128) 332 return FPTOUINT_F64_I128; 333 } else if (OpVT == MVT::f80) { 334 if (RetVT == MVT::i32) 335 return FPTOUINT_F80_I32; 336 if (RetVT == MVT::i64) 337 return FPTOUINT_F80_I64; 338 if (RetVT == MVT::i128) 339 return FPTOUINT_F80_I128; 340 } else if (OpVT == MVT::f128) { 341 if (RetVT == MVT::i32) 342 return FPTOUINT_F128_I32; 343 if (RetVT == MVT::i64) 344 return FPTOUINT_F128_I64; 345 if (RetVT == MVT::i128) 346 return FPTOUINT_F128_I128; 347 } else if (OpVT == MVT::ppcf128) { 348 if (RetVT == MVT::i32) 349 return FPTOUINT_PPCF128_I32; 350 if (RetVT == MVT::i64) 351 return FPTOUINT_PPCF128_I64; 352 if (RetVT == MVT::i128) 353 return FPTOUINT_PPCF128_I128; 354 } 355 return UNKNOWN_LIBCALL; 356 } 357 358 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 359 /// UNKNOWN_LIBCALL if there is none. 360 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 361 if (OpVT == MVT::i32) { 362 if (RetVT == MVT::f32) 363 return SINTTOFP_I32_F32; 364 if (RetVT == MVT::f64) 365 return SINTTOFP_I32_F64; 366 if (RetVT == MVT::f80) 367 return SINTTOFP_I32_F80; 368 if (RetVT == MVT::f128) 369 return SINTTOFP_I32_F128; 370 if (RetVT == MVT::ppcf128) 371 return SINTTOFP_I32_PPCF128; 372 } else if (OpVT == MVT::i64) { 373 if (RetVT == MVT::f32) 374 return SINTTOFP_I64_F32; 375 if (RetVT == MVT::f64) 376 return SINTTOFP_I64_F64; 377 if (RetVT == MVT::f80) 378 return SINTTOFP_I64_F80; 379 if (RetVT == MVT::f128) 380 return SINTTOFP_I64_F128; 381 if (RetVT == MVT::ppcf128) 382 return SINTTOFP_I64_PPCF128; 383 } else if (OpVT == MVT::i128) { 384 if (RetVT == MVT::f32) 385 return SINTTOFP_I128_F32; 386 if (RetVT == MVT::f64) 387 return SINTTOFP_I128_F64; 388 if (RetVT == MVT::f80) 389 return SINTTOFP_I128_F80; 390 if (RetVT == MVT::f128) 391 return SINTTOFP_I128_F128; 392 if (RetVT == MVT::ppcf128) 393 return SINTTOFP_I128_PPCF128; 394 } 395 return UNKNOWN_LIBCALL; 396 } 397 398 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 399 /// UNKNOWN_LIBCALL if there is none. 400 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 401 if (OpVT == MVT::i32) { 402 if (RetVT == MVT::f32) 403 return UINTTOFP_I32_F32; 404 if (RetVT == MVT::f64) 405 return UINTTOFP_I32_F64; 406 if (RetVT == MVT::f80) 407 return UINTTOFP_I32_F80; 408 if (RetVT == MVT::f128) 409 return UINTTOFP_I32_F128; 410 if (RetVT == MVT::ppcf128) 411 return UINTTOFP_I32_PPCF128; 412 } else if (OpVT == MVT::i64) { 413 if (RetVT == MVT::f32) 414 return UINTTOFP_I64_F32; 415 if (RetVT == MVT::f64) 416 return UINTTOFP_I64_F64; 417 if (RetVT == MVT::f80) 418 return UINTTOFP_I64_F80; 419 if (RetVT == MVT::f128) 420 return UINTTOFP_I64_F128; 421 if (RetVT == MVT::ppcf128) 422 return UINTTOFP_I64_PPCF128; 423 } else if (OpVT == MVT::i128) { 424 if (RetVT == MVT::f32) 425 return UINTTOFP_I128_F32; 426 if (RetVT == MVT::f64) 427 return UINTTOFP_I128_F64; 428 if (RetVT == MVT::f80) 429 return UINTTOFP_I128_F80; 430 if (RetVT == MVT::f128) 431 return UINTTOFP_I128_F128; 432 if (RetVT == MVT::ppcf128) 433 return UINTTOFP_I128_PPCF128; 434 } 435 return UNKNOWN_LIBCALL; 436 } 437 438 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 439 #define OP_TO_LIBCALL(Name, Enum) \ 440 case Name: \ 441 switch (VT.SimpleTy) { \ 442 default: \ 443 return UNKNOWN_LIBCALL; \ 444 case MVT::i8: \ 445 return Enum##_1; \ 446 case MVT::i16: \ 447 return Enum##_2; \ 448 case MVT::i32: \ 449 return Enum##_4; \ 450 case MVT::i64: \ 451 return Enum##_8; \ 452 case MVT::i128: \ 453 return Enum##_16; \ 454 } 455 456 switch (Opc) { 457 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 458 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 459 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 460 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 461 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 462 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 463 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 464 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 465 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 466 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 467 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 469 } 470 471 #undef OP_TO_LIBCALL 472 473 return UNKNOWN_LIBCALL; 474 } 475 476 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 477 switch (ElementSize) { 478 case 1: 479 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 480 case 2: 481 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 482 case 4: 483 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 484 case 8: 485 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 486 case 16: 487 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 488 default: 489 return UNKNOWN_LIBCALL; 490 } 491 } 492 493 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 494 switch (ElementSize) { 495 case 1: 496 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 497 case 2: 498 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 499 case 4: 500 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 501 case 8: 502 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 503 case 16: 504 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 505 default: 506 return UNKNOWN_LIBCALL; 507 } 508 } 509 510 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 511 switch (ElementSize) { 512 case 1: 513 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 514 case 2: 515 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 516 case 4: 517 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 518 case 8: 519 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 520 case 16: 521 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 522 default: 523 return UNKNOWN_LIBCALL; 524 } 525 } 526 527 /// InitCmpLibcallCCs - Set default comparison libcall CC. 528 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 529 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 530 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 531 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 532 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 533 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 534 CCs[RTLIB::UNE_F32] = ISD::SETNE; 535 CCs[RTLIB::UNE_F64] = ISD::SETNE; 536 CCs[RTLIB::UNE_F128] = ISD::SETNE; 537 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 538 CCs[RTLIB::OGE_F32] = ISD::SETGE; 539 CCs[RTLIB::OGE_F64] = ISD::SETGE; 540 CCs[RTLIB::OGE_F128] = ISD::SETGE; 541 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 542 CCs[RTLIB::OLT_F32] = ISD::SETLT; 543 CCs[RTLIB::OLT_F64] = ISD::SETLT; 544 CCs[RTLIB::OLT_F128] = ISD::SETLT; 545 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 546 CCs[RTLIB::OLE_F32] = ISD::SETLE; 547 CCs[RTLIB::OLE_F64] = ISD::SETLE; 548 CCs[RTLIB::OLE_F128] = ISD::SETLE; 549 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 550 CCs[RTLIB::OGT_F32] = ISD::SETGT; 551 CCs[RTLIB::OGT_F64] = ISD::SETGT; 552 CCs[RTLIB::OGT_F128] = ISD::SETGT; 553 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 554 CCs[RTLIB::UO_F32] = ISD::SETNE; 555 CCs[RTLIB::UO_F64] = ISD::SETNE; 556 CCs[RTLIB::UO_F128] = ISD::SETNE; 557 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 558 CCs[RTLIB::O_F32] = ISD::SETEQ; 559 CCs[RTLIB::O_F64] = ISD::SETEQ; 560 CCs[RTLIB::O_F128] = ISD::SETEQ; 561 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 562 } 563 564 /// NOTE: The TargetMachine owns TLOF. 565 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 566 initActions(); 567 568 // Perform these initializations only once. 569 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 570 MaxLoadsPerMemcmp = 8; 571 MaxGluedStoresPerMemcpy = 0; 572 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 573 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 574 UseUnderscoreSetJmp = false; 575 UseUnderscoreLongJmp = false; 576 HasMultipleConditionRegisters = false; 577 HasExtractBitsInsn = false; 578 JumpIsExpensive = JumpIsExpensiveOverride; 579 PredictableSelectIsExpensive = false; 580 EnableExtLdPromotion = false; 581 StackPointerRegisterToSaveRestore = 0; 582 BooleanContents = UndefinedBooleanContent; 583 BooleanFloatContents = UndefinedBooleanContent; 584 BooleanVectorContents = UndefinedBooleanContent; 585 SchedPreferenceInfo = Sched::ILP; 586 GatherAllAliasesMaxDepth = 18; 587 // TODO: the default will be switched to 0 in the next commit, along 588 // with the Target-specific changes necessary. 589 MaxAtomicSizeInBitsSupported = 1024; 590 591 MinCmpXchgSizeInBits = 0; 592 SupportsUnalignedAtomics = false; 593 594 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 595 596 InitLibcalls(TM.getTargetTriple()); 597 InitCmpLibcallCCs(CmpLibcallCCs); 598 } 599 600 void TargetLoweringBase::initActions() { 601 // All operations default to being supported. 602 memset(OpActions, 0, sizeof(OpActions)); 603 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 604 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 605 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 606 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 607 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 608 std::fill(std::begin(TargetDAGCombineArray), 609 std::end(TargetDAGCombineArray), 0); 610 611 for (MVT VT : MVT::fp_valuetypes()) { 612 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 613 if (IntVT.isValid()) { 614 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 615 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 616 } 617 } 618 619 // Set default actions for various operations. 620 for (MVT VT : MVT::all_valuetypes()) { 621 // Default all indexed load / store to expand. 622 for (unsigned IM = (unsigned)ISD::PRE_INC; 623 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 624 setIndexedLoadAction(IM, VT, Expand); 625 setIndexedStoreAction(IM, VT, Expand); 626 } 627 628 // Most backends expect to see the node which just returns the value loaded. 629 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 630 631 // These operations default to expand. 632 setOperationAction(ISD::FGETSIGN, VT, Expand); 633 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 634 setOperationAction(ISD::FMINNUM, VT, Expand); 635 setOperationAction(ISD::FMAXNUM, VT, Expand); 636 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 637 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 638 setOperationAction(ISD::FMINIMUM, VT, Expand); 639 setOperationAction(ISD::FMAXIMUM, VT, Expand); 640 setOperationAction(ISD::FMAD, VT, Expand); 641 setOperationAction(ISD::SMIN, VT, Expand); 642 setOperationAction(ISD::SMAX, VT, Expand); 643 setOperationAction(ISD::UMIN, VT, Expand); 644 setOperationAction(ISD::UMAX, VT, Expand); 645 setOperationAction(ISD::ABS, VT, Expand); 646 setOperationAction(ISD::FSHL, VT, Expand); 647 setOperationAction(ISD::FSHR, VT, Expand); 648 setOperationAction(ISD::SADDSAT, VT, Expand); 649 setOperationAction(ISD::UADDSAT, VT, Expand); 650 setOperationAction(ISD::SSUBSAT, VT, Expand); 651 setOperationAction(ISD::USUBSAT, VT, Expand); 652 setOperationAction(ISD::SMULFIX, VT, Expand); 653 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 654 setOperationAction(ISD::UMULFIX, VT, Expand); 655 setOperationAction(ISD::UMULFIXSAT, VT, Expand); 656 657 // Overflow operations default to expand 658 setOperationAction(ISD::SADDO, VT, Expand); 659 setOperationAction(ISD::SSUBO, VT, Expand); 660 setOperationAction(ISD::UADDO, VT, Expand); 661 setOperationAction(ISD::USUBO, VT, Expand); 662 setOperationAction(ISD::SMULO, VT, Expand); 663 setOperationAction(ISD::UMULO, VT, Expand); 664 665 // ADDCARRY operations default to expand 666 setOperationAction(ISD::ADDCARRY, VT, Expand); 667 setOperationAction(ISD::SUBCARRY, VT, Expand); 668 setOperationAction(ISD::SETCCCARRY, VT, Expand); 669 670 // ADDC/ADDE/SUBC/SUBE default to expand. 671 setOperationAction(ISD::ADDC, VT, Expand); 672 setOperationAction(ISD::ADDE, VT, Expand); 673 setOperationAction(ISD::SUBC, VT, Expand); 674 setOperationAction(ISD::SUBE, VT, Expand); 675 676 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 677 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 678 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 679 680 setOperationAction(ISD::BITREVERSE, VT, Expand); 681 682 // These library functions default to expand. 683 setOperationAction(ISD::FROUND, VT, Expand); 684 setOperationAction(ISD::FPOWI, VT, Expand); 685 686 // These operations default to expand for vector types. 687 if (VT.isVector()) { 688 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 689 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 690 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 691 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 692 } 693 694 // Constrained floating-point operations default to expand. 695 setOperationAction(ISD::STRICT_FADD, VT, Expand); 696 setOperationAction(ISD::STRICT_FSUB, VT, Expand); 697 setOperationAction(ISD::STRICT_FMUL, VT, Expand); 698 setOperationAction(ISD::STRICT_FDIV, VT, Expand); 699 setOperationAction(ISD::STRICT_FREM, VT, Expand); 700 setOperationAction(ISD::STRICT_FMA, VT, Expand); 701 setOperationAction(ISD::STRICT_FSQRT, VT, Expand); 702 setOperationAction(ISD::STRICT_FPOW, VT, Expand); 703 setOperationAction(ISD::STRICT_FPOWI, VT, Expand); 704 setOperationAction(ISD::STRICT_FSIN, VT, Expand); 705 setOperationAction(ISD::STRICT_FCOS, VT, Expand); 706 setOperationAction(ISD::STRICT_FEXP, VT, Expand); 707 setOperationAction(ISD::STRICT_FEXP2, VT, Expand); 708 setOperationAction(ISD::STRICT_FLOG, VT, Expand); 709 setOperationAction(ISD::STRICT_FLOG10, VT, Expand); 710 setOperationAction(ISD::STRICT_FLOG2, VT, Expand); 711 setOperationAction(ISD::STRICT_FRINT, VT, Expand); 712 setOperationAction(ISD::STRICT_FNEARBYINT, VT, Expand); 713 setOperationAction(ISD::STRICT_FCEIL, VT, Expand); 714 setOperationAction(ISD::STRICT_FFLOOR, VT, Expand); 715 setOperationAction(ISD::STRICT_FROUND, VT, Expand); 716 setOperationAction(ISD::STRICT_FTRUNC, VT, Expand); 717 setOperationAction(ISD::STRICT_FMAXNUM, VT, Expand); 718 setOperationAction(ISD::STRICT_FMINNUM, VT, Expand); 719 setOperationAction(ISD::STRICT_FP_ROUND, VT, Expand); 720 setOperationAction(ISD::STRICT_FP_EXTEND, VT, Expand); 721 setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Expand); 722 setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Expand); 723 724 // For most targets @llvm.get.dynamic.area.offset just returns 0. 725 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 726 727 // Vector reduction default to expand. 728 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 729 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 730 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 731 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 732 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 733 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 734 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 735 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 736 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 737 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 738 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 739 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 740 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 741 } 742 743 // Most targets ignore the @llvm.prefetch intrinsic. 744 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 745 746 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 747 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 748 749 // ConstantFP nodes default to expand. Targets can either change this to 750 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 751 // to optimize expansions for certain constants. 752 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 753 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 754 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 755 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 756 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 757 758 // These library functions default to expand. 759 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 760 setOperationAction(ISD::FCBRT, VT, Expand); 761 setOperationAction(ISD::FLOG , VT, Expand); 762 setOperationAction(ISD::FLOG2, VT, Expand); 763 setOperationAction(ISD::FLOG10, VT, Expand); 764 setOperationAction(ISD::FEXP , VT, Expand); 765 setOperationAction(ISD::FEXP2, VT, Expand); 766 setOperationAction(ISD::FFLOOR, VT, Expand); 767 setOperationAction(ISD::FNEARBYINT, VT, Expand); 768 setOperationAction(ISD::FCEIL, VT, Expand); 769 setOperationAction(ISD::FRINT, VT, Expand); 770 setOperationAction(ISD::FTRUNC, VT, Expand); 771 setOperationAction(ISD::FROUND, VT, Expand); 772 setOperationAction(ISD::LROUND, VT, Expand); 773 setOperationAction(ISD::LLROUND, VT, Expand); 774 setOperationAction(ISD::LRINT, VT, Expand); 775 setOperationAction(ISD::LLRINT, VT, Expand); 776 } 777 778 // Default ISD::TRAP to expand (which turns it into abort). 779 setOperationAction(ISD::TRAP, MVT::Other, Expand); 780 781 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 782 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 783 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 784 } 785 786 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 787 EVT) const { 788 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 789 } 790 791 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 792 bool LegalTypes) const { 793 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 794 if (LHSTy.isVector()) 795 return LHSTy; 796 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 797 : getPointerTy(DL); 798 } 799 800 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 801 assert(isTypeLegal(VT)); 802 switch (Op) { 803 default: 804 return false; 805 case ISD::SDIV: 806 case ISD::UDIV: 807 case ISD::SREM: 808 case ISD::UREM: 809 return true; 810 } 811 } 812 813 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 814 // If the command-line option was specified, ignore this request. 815 if (!JumpIsExpensiveOverride.getNumOccurrences()) 816 JumpIsExpensive = isExpensive; 817 } 818 819 TargetLoweringBase::LegalizeKind 820 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 821 // If this is a simple type, use the ComputeRegisterProp mechanism. 822 if (VT.isSimple()) { 823 MVT SVT = VT.getSimpleVT(); 824 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 825 MVT NVT = TransformToType[SVT.SimpleTy]; 826 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 827 828 assert((LA == TypeLegal || LA == TypeSoftenFloat || 829 (NVT.isVector() || 830 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 831 "Promote may not follow Expand or Promote"); 832 833 if (LA == TypeSplitVector) 834 return LegalizeKind(LA, 835 EVT::getVectorVT(Context, SVT.getVectorElementType(), 836 SVT.getVectorNumElements() / 2)); 837 if (LA == TypeScalarizeVector) 838 return LegalizeKind(LA, SVT.getVectorElementType()); 839 return LegalizeKind(LA, NVT); 840 } 841 842 // Handle Extended Scalar Types. 843 if (!VT.isVector()) { 844 assert(VT.isInteger() && "Float types must be simple"); 845 unsigned BitSize = VT.getSizeInBits(); 846 // First promote to a power-of-two size, then expand if necessary. 847 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 848 EVT NVT = VT.getRoundIntegerType(Context); 849 assert(NVT != VT && "Unable to round integer VT"); 850 LegalizeKind NextStep = getTypeConversion(Context, NVT); 851 // Avoid multi-step promotion. 852 if (NextStep.first == TypePromoteInteger) 853 return NextStep; 854 // Return rounded integer type. 855 return LegalizeKind(TypePromoteInteger, NVT); 856 } 857 858 return LegalizeKind(TypeExpandInteger, 859 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 860 } 861 862 // Handle vector types. 863 unsigned NumElts = VT.getVectorNumElements(); 864 EVT EltVT = VT.getVectorElementType(); 865 866 // Vectors with only one element are always scalarized. 867 if (NumElts == 1) 868 return LegalizeKind(TypeScalarizeVector, EltVT); 869 870 // Try to widen vector elements until the element type is a power of two and 871 // promote it to a legal type later on, for example: 872 // <3 x i8> -> <4 x i8> -> <4 x i32> 873 if (EltVT.isInteger()) { 874 // Vectors with a number of elements that is not a power of two are always 875 // widened, for example <3 x i8> -> <4 x i8>. 876 if (!VT.isPow2VectorType()) { 877 NumElts = (unsigned)NextPowerOf2(NumElts); 878 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 879 return LegalizeKind(TypeWidenVector, NVT); 880 } 881 882 // Examine the element type. 883 LegalizeKind LK = getTypeConversion(Context, EltVT); 884 885 // If type is to be expanded, split the vector. 886 // <4 x i140> -> <2 x i140> 887 if (LK.first == TypeExpandInteger) 888 return LegalizeKind(TypeSplitVector, 889 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 890 891 // Promote the integer element types until a legal vector type is found 892 // or until the element integer type is too big. If a legal type was not 893 // found, fallback to the usual mechanism of widening/splitting the 894 // vector. 895 EVT OldEltVT = EltVT; 896 while (true) { 897 // Increase the bitwidth of the element to the next pow-of-two 898 // (which is greater than 8 bits). 899 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 900 .getRoundIntegerType(Context); 901 902 // Stop trying when getting a non-simple element type. 903 // Note that vector elements may be greater than legal vector element 904 // types. Example: X86 XMM registers hold 64bit element on 32bit 905 // systems. 906 if (!EltVT.isSimple()) 907 break; 908 909 // Build a new vector type and check if it is legal. 910 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 911 // Found a legal promoted vector type. 912 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 913 return LegalizeKind(TypePromoteInteger, 914 EVT::getVectorVT(Context, EltVT, NumElts)); 915 } 916 917 // Reset the type to the unexpanded type if we did not find a legal vector 918 // type with a promoted vector element type. 919 EltVT = OldEltVT; 920 } 921 922 // Try to widen the vector until a legal type is found. 923 // If there is no wider legal type, split the vector. 924 while (true) { 925 // Round up to the next power of 2. 926 NumElts = (unsigned)NextPowerOf2(NumElts); 927 928 // If there is no simple vector type with this many elements then there 929 // cannot be a larger legal vector type. Note that this assumes that 930 // there are no skipped intermediate vector types in the simple types. 931 if (!EltVT.isSimple()) 932 break; 933 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 934 if (LargerVector == MVT()) 935 break; 936 937 // If this type is legal then widen the vector. 938 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 939 return LegalizeKind(TypeWidenVector, LargerVector); 940 } 941 942 // Widen odd vectors to next power of two. 943 if (!VT.isPow2VectorType()) { 944 EVT NVT = VT.getPow2VectorType(Context); 945 return LegalizeKind(TypeWidenVector, NVT); 946 } 947 948 // Vectors with illegal element types are expanded. 949 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 950 return LegalizeKind(TypeSplitVector, NVT); 951 } 952 953 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 954 unsigned &NumIntermediates, 955 MVT &RegisterVT, 956 TargetLoweringBase *TLI) { 957 // Figure out the right, legal destination reg to copy into. 958 unsigned NumElts = VT.getVectorNumElements(); 959 MVT EltTy = VT.getVectorElementType(); 960 961 unsigned NumVectorRegs = 1; 962 963 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 964 // could break down into LHS/RHS like LegalizeDAG does. 965 if (!isPowerOf2_32(NumElts)) { 966 NumVectorRegs = NumElts; 967 NumElts = 1; 968 } 969 970 // Divide the input until we get to a supported size. This will always 971 // end with a scalar if the target doesn't support vectors. 972 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 973 NumElts >>= 1; 974 NumVectorRegs <<= 1; 975 } 976 977 NumIntermediates = NumVectorRegs; 978 979 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 980 if (!TLI->isTypeLegal(NewVT)) 981 NewVT = EltTy; 982 IntermediateVT = NewVT; 983 984 unsigned NewVTSize = NewVT.getSizeInBits(); 985 986 // Convert sizes such as i33 to i64. 987 if (!isPowerOf2_32(NewVTSize)) 988 NewVTSize = NextPowerOf2(NewVTSize); 989 990 MVT DestVT = TLI->getRegisterType(NewVT); 991 RegisterVT = DestVT; 992 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 993 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 994 995 // Otherwise, promotion or legal types use the same number of registers as 996 // the vector decimated to the appropriate level. 997 return NumVectorRegs; 998 } 999 1000 /// isLegalRC - Return true if the value types that can be represented by the 1001 /// specified register class are all legal. 1002 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 1003 const TargetRegisterClass &RC) const { 1004 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 1005 if (isTypeLegal(*I)) 1006 return true; 1007 return false; 1008 } 1009 1010 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1011 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1012 MachineBasicBlock * 1013 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 1014 MachineBasicBlock *MBB) const { 1015 MachineInstr *MI = &InitialMI; 1016 MachineFunction &MF = *MI->getMF(); 1017 MachineFrameInfo &MFI = MF.getFrameInfo(); 1018 1019 // We're handling multiple types of operands here: 1020 // PATCHPOINT MetaArgs - live-in, read only, direct 1021 // STATEPOINT Deopt Spill - live-through, read only, indirect 1022 // STATEPOINT Deopt Alloca - live-through, read only, direct 1023 // (We're currently conservative and mark the deopt slots read/write in 1024 // practice.) 1025 // STATEPOINT GC Spill - live-through, read/write, indirect 1026 // STATEPOINT GC Alloca - live-through, read/write, direct 1027 // The live-in vs live-through is handled already (the live through ones are 1028 // all stack slots), but we need to handle the different type of stackmap 1029 // operands and memory effects here. 1030 1031 // MI changes inside this loop as we grow operands. 1032 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1033 MachineOperand &MO = MI->getOperand(OperIdx); 1034 if (!MO.isFI()) 1035 continue; 1036 1037 // foldMemoryOperand builds a new MI after replacing a single FI operand 1038 // with the canonical set of five x86 addressing-mode operands. 1039 int FI = MO.getIndex(); 1040 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1041 1042 // Copy operands before the frame-index. 1043 for (unsigned i = 0; i < OperIdx; ++i) 1044 MIB.add(MI->getOperand(i)); 1045 // Add frame index operands recognized by stackmaps.cpp 1046 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1047 // indirect-mem-ref tag, size, #FI, offset. 1048 // Used for spills inserted by StatepointLowering. This codepath is not 1049 // used for patchpoints/stackmaps at all, for these spilling is done via 1050 // foldMemoryOperand callback only. 1051 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1052 MIB.addImm(StackMaps::IndirectMemRefOp); 1053 MIB.addImm(MFI.getObjectSize(FI)); 1054 MIB.add(MI->getOperand(OperIdx)); 1055 MIB.addImm(0); 1056 } else { 1057 // direct-mem-ref tag, #FI, offset. 1058 // Used by patchpoint, and direct alloca arguments to statepoints 1059 MIB.addImm(StackMaps::DirectMemRefOp); 1060 MIB.add(MI->getOperand(OperIdx)); 1061 MIB.addImm(0); 1062 } 1063 // Copy the operands after the frame index. 1064 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1065 MIB.add(MI->getOperand(i)); 1066 1067 // Inherit previous memory operands. 1068 MIB.cloneMemRefs(*MI); 1069 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1070 1071 // Add a new memory operand for this FI. 1072 assert(MFI.getObjectOffset(FI) != -1); 1073 1074 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1075 // PATCHPOINT should be updated to do the same. (TODO) 1076 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1077 auto Flags = MachineMemOperand::MOLoad; 1078 MachineMemOperand *MMO = MF.getMachineMemOperand( 1079 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1080 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1081 MIB->addMemOperand(MF, MMO); 1082 } 1083 1084 // Replace the instruction and update the operand index. 1085 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1086 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1087 MI->eraseFromParent(); 1088 MI = MIB; 1089 } 1090 return MBB; 1091 } 1092 1093 MachineBasicBlock * 1094 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1095 MachineBasicBlock *MBB) const { 1096 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1097 "Called emitXRayCustomEvent 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 MachineBasicBlock * 1109 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1110 MachineBasicBlock *MBB) const { 1111 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1112 "Called emitXRayTypedEvent on the wrong MI!"); 1113 auto &MF = *MI.getMF(); 1114 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1115 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1116 MIB.add(MI.getOperand(OpIdx)); 1117 1118 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1119 MI.eraseFromParent(); 1120 return MBB; 1121 } 1122 1123 /// findRepresentativeClass - Return the largest legal super-reg register class 1124 /// of the register class for the specified type and its associated "cost". 1125 // This function is in TargetLowering because it uses RegClassForVT which would 1126 // need to be moved to TargetRegisterInfo and would necessitate moving 1127 // isTypeLegal over as well - a massive change that would just require 1128 // TargetLowering having a TargetRegisterInfo class member that it would use. 1129 std::pair<const TargetRegisterClass *, uint8_t> 1130 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1131 MVT VT) const { 1132 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1133 if (!RC) 1134 return std::make_pair(RC, 0); 1135 1136 // Compute the set of all super-register classes. 1137 BitVector SuperRegRC(TRI->getNumRegClasses()); 1138 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1139 SuperRegRC.setBitsInMask(RCI.getMask()); 1140 1141 // Find the first legal register class with the largest spill size. 1142 const TargetRegisterClass *BestRC = RC; 1143 for (unsigned i : SuperRegRC.set_bits()) { 1144 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1145 // We want the largest possible spill size. 1146 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1147 continue; 1148 if (!isLegalRC(*TRI, *SuperRC)) 1149 continue; 1150 BestRC = SuperRC; 1151 } 1152 return std::make_pair(BestRC, 1); 1153 } 1154 1155 /// computeRegisterProperties - Once all of the register classes are added, 1156 /// this allows us to compute derived properties we expose. 1157 void TargetLoweringBase::computeRegisterProperties( 1158 const TargetRegisterInfo *TRI) { 1159 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1160 "Too many value types for ValueTypeActions to hold!"); 1161 1162 // Everything defaults to needing one register. 1163 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1164 NumRegistersForVT[i] = 1; 1165 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1166 } 1167 // ...except isVoid, which doesn't need any registers. 1168 NumRegistersForVT[MVT::isVoid] = 0; 1169 1170 // Find the largest integer register class. 1171 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1172 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1173 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1174 1175 // Every integer value type larger than this largest register takes twice as 1176 // many registers to represent as the previous ValueType. 1177 for (unsigned ExpandedReg = LargestIntReg + 1; 1178 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1179 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1180 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1181 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1182 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1183 TypeExpandInteger); 1184 } 1185 1186 // Inspect all of the ValueType's smaller than the largest integer 1187 // register to see which ones need promotion. 1188 unsigned LegalIntReg = LargestIntReg; 1189 for (unsigned IntReg = LargestIntReg - 1; 1190 IntReg >= (unsigned)MVT::i1; --IntReg) { 1191 MVT IVT = (MVT::SimpleValueType)IntReg; 1192 if (isTypeLegal(IVT)) { 1193 LegalIntReg = IntReg; 1194 } else { 1195 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1196 (MVT::SimpleValueType)LegalIntReg; 1197 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1198 } 1199 } 1200 1201 // ppcf128 type is really two f64's. 1202 if (!isTypeLegal(MVT::ppcf128)) { 1203 if (isTypeLegal(MVT::f64)) { 1204 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1205 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1206 TransformToType[MVT::ppcf128] = MVT::f64; 1207 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1208 } else { 1209 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1210 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1211 TransformToType[MVT::ppcf128] = MVT::i128; 1212 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1213 } 1214 } 1215 1216 // Decide how to handle f128. If the target does not have native f128 support, 1217 // expand it to i128 and we will be generating soft float library calls. 1218 if (!isTypeLegal(MVT::f128)) { 1219 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1220 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1221 TransformToType[MVT::f128] = MVT::i128; 1222 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1223 } 1224 1225 // Decide how to handle f64. If the target does not have native f64 support, 1226 // expand it to i64 and we will be generating soft float library calls. 1227 if (!isTypeLegal(MVT::f64)) { 1228 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1229 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1230 TransformToType[MVT::f64] = MVT::i64; 1231 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1232 } 1233 1234 // Decide how to handle f32. If the target does not have native f32 support, 1235 // expand it to i32 and we will be generating soft float library calls. 1236 if (!isTypeLegal(MVT::f32)) { 1237 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1238 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1239 TransformToType[MVT::f32] = MVT::i32; 1240 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1241 } 1242 1243 // Decide how to handle f16. If the target does not have native f16 support, 1244 // promote it to f32, because there are no f16 library calls (except for 1245 // conversions). 1246 if (!isTypeLegal(MVT::f16)) { 1247 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1248 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1249 TransformToType[MVT::f16] = MVT::f32; 1250 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1251 } 1252 1253 // Loop over all of the vector value types to see which need transformations. 1254 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1255 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1256 MVT VT = (MVT::SimpleValueType) i; 1257 if (isTypeLegal(VT)) 1258 continue; 1259 1260 MVT EltVT = VT.getVectorElementType(); 1261 unsigned NElts = VT.getVectorNumElements(); 1262 bool IsLegalWiderType = false; 1263 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1264 switch (PreferredAction) { 1265 case TypePromoteInteger: 1266 // Try to promote the elements of integer vectors. If no legal 1267 // promotion was found, fall through to the widen-vector method. 1268 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1269 MVT SVT = (MVT::SimpleValueType) nVT; 1270 // Promote vectors of integers to vectors with the same number 1271 // of elements, with a wider element type. 1272 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1273 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1274 TransformToType[i] = SVT; 1275 RegisterTypeForVT[i] = SVT; 1276 NumRegistersForVT[i] = 1; 1277 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1278 IsLegalWiderType = true; 1279 break; 1280 } 1281 } 1282 if (IsLegalWiderType) 1283 break; 1284 LLVM_FALLTHROUGH; 1285 1286 case TypeWidenVector: 1287 if (isPowerOf2_32(NElts)) { 1288 // Try to widen the vector. 1289 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1290 MVT SVT = (MVT::SimpleValueType) nVT; 1291 if (SVT.getVectorElementType() == EltVT 1292 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1293 TransformToType[i] = SVT; 1294 RegisterTypeForVT[i] = SVT; 1295 NumRegistersForVT[i] = 1; 1296 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1297 IsLegalWiderType = true; 1298 break; 1299 } 1300 } 1301 if (IsLegalWiderType) 1302 break; 1303 } else { 1304 // Only widen to the next power of 2 to keep consistency with EVT. 1305 MVT NVT = VT.getPow2VectorType(); 1306 if (isTypeLegal(NVT)) { 1307 TransformToType[i] = NVT; 1308 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1309 RegisterTypeForVT[i] = NVT; 1310 NumRegistersForVT[i] = 1; 1311 break; 1312 } 1313 } 1314 LLVM_FALLTHROUGH; 1315 1316 case TypeSplitVector: 1317 case TypeScalarizeVector: { 1318 MVT IntermediateVT; 1319 MVT RegisterVT; 1320 unsigned NumIntermediates; 1321 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1322 NumIntermediates, RegisterVT, this); 1323 RegisterTypeForVT[i] = RegisterVT; 1324 1325 MVT NVT = VT.getPow2VectorType(); 1326 if (NVT == VT) { 1327 // Type is already a power of 2. The default action is to split. 1328 TransformToType[i] = MVT::Other; 1329 if (PreferredAction == TypeScalarizeVector) 1330 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1331 else if (PreferredAction == TypeSplitVector) 1332 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1333 else 1334 // Set type action according to the number of elements. 1335 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1336 : TypeSplitVector); 1337 } else { 1338 TransformToType[i] = NVT; 1339 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1340 } 1341 break; 1342 } 1343 default: 1344 llvm_unreachable("Unknown vector legalization action!"); 1345 } 1346 } 1347 1348 // Determine the 'representative' register class for each value type. 1349 // An representative register class is the largest (meaning one which is 1350 // not a sub-register class / subreg register class) legal register class for 1351 // a group of value types. For example, on i386, i8, i16, and i32 1352 // representative would be GR32; while on x86_64 it's GR64. 1353 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1354 const TargetRegisterClass* RRC; 1355 uint8_t Cost; 1356 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1357 RepRegClassForVT[i] = RRC; 1358 RepRegClassCostForVT[i] = Cost; 1359 } 1360 } 1361 1362 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1363 EVT VT) const { 1364 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1365 return getPointerTy(DL).SimpleTy; 1366 } 1367 1368 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1369 return MVT::i32; // return the default value 1370 } 1371 1372 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1373 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1374 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1375 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1376 /// 1377 /// This method returns the number of registers needed, and the VT for each 1378 /// register. It also returns the VT and quantity of the intermediate values 1379 /// before they are promoted/expanded. 1380 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1381 EVT &IntermediateVT, 1382 unsigned &NumIntermediates, 1383 MVT &RegisterVT) const { 1384 unsigned NumElts = VT.getVectorNumElements(); 1385 1386 // If there is a wider vector type with the same element type as this one, 1387 // or a promoted vector type that has the same number of elements which 1388 // are wider, then we should convert to that legal vector type. 1389 // This handles things like <2 x float> -> <4 x float> and 1390 // <4 x i1> -> <4 x i32>. 1391 LegalizeTypeAction TA = getTypeAction(Context, VT); 1392 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1393 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1394 if (isTypeLegal(RegisterEVT)) { 1395 IntermediateVT = RegisterEVT; 1396 RegisterVT = RegisterEVT.getSimpleVT(); 1397 NumIntermediates = 1; 1398 return 1; 1399 } 1400 } 1401 1402 // Figure out the right, legal destination reg to copy into. 1403 EVT EltTy = VT.getVectorElementType(); 1404 1405 unsigned NumVectorRegs = 1; 1406 1407 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1408 // could break down into LHS/RHS like LegalizeDAG does. 1409 if (!isPowerOf2_32(NumElts)) { 1410 NumVectorRegs = NumElts; 1411 NumElts = 1; 1412 } 1413 1414 // Divide the input until we get to a supported size. This will always 1415 // end with a scalar if the target doesn't support vectors. 1416 while (NumElts > 1 && !isTypeLegal( 1417 EVT::getVectorVT(Context, EltTy, NumElts))) { 1418 NumElts >>= 1; 1419 NumVectorRegs <<= 1; 1420 } 1421 1422 NumIntermediates = NumVectorRegs; 1423 1424 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1425 if (!isTypeLegal(NewVT)) 1426 NewVT = EltTy; 1427 IntermediateVT = NewVT; 1428 1429 MVT DestVT = getRegisterType(Context, NewVT); 1430 RegisterVT = DestVT; 1431 unsigned NewVTSize = NewVT.getSizeInBits(); 1432 1433 // Convert sizes such as i33 to i64. 1434 if (!isPowerOf2_32(NewVTSize)) 1435 NewVTSize = NextPowerOf2(NewVTSize); 1436 1437 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1438 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1439 1440 // Otherwise, promotion or legal types use the same number of registers as 1441 // the vector decimated to the appropriate level. 1442 return NumVectorRegs; 1443 } 1444 1445 /// Get the EVTs and ArgFlags collections that represent the legalized return 1446 /// type of the given function. This does not require a DAG or a return value, 1447 /// and is suitable for use before any DAGs for the function are constructed. 1448 /// TODO: Move this out of TargetLowering.cpp. 1449 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1450 AttributeList attr, 1451 SmallVectorImpl<ISD::OutputArg> &Outs, 1452 const TargetLowering &TLI, const DataLayout &DL) { 1453 SmallVector<EVT, 4> ValueVTs; 1454 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1455 unsigned NumValues = ValueVTs.size(); 1456 if (NumValues == 0) return; 1457 1458 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1459 EVT VT = ValueVTs[j]; 1460 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1461 1462 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1463 ExtendKind = ISD::SIGN_EXTEND; 1464 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1465 ExtendKind = ISD::ZERO_EXTEND; 1466 1467 // FIXME: C calling convention requires the return type to be promoted to 1468 // at least 32-bit. But this is not necessary for non-C calling 1469 // conventions. The frontend should mark functions whose return values 1470 // require promoting with signext or zeroext attributes. 1471 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1472 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1473 if (VT.bitsLT(MinVT)) 1474 VT = MinVT; 1475 } 1476 1477 unsigned NumParts = 1478 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1479 MVT PartVT = 1480 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1481 1482 // 'inreg' on function refers to return value 1483 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1484 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1485 Flags.setInReg(); 1486 1487 // Propagate extension type if any 1488 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1489 Flags.setSExt(); 1490 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1491 Flags.setZExt(); 1492 1493 for (unsigned i = 0; i < NumParts; ++i) 1494 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1495 } 1496 } 1497 1498 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1499 /// function arguments in the caller parameter area. This is the actual 1500 /// alignment, not its logarithm. 1501 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1502 const DataLayout &DL) const { 1503 return DL.getABITypeAlignment(Ty); 1504 } 1505 1506 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1507 const DataLayout &DL, EVT VT, 1508 unsigned AddrSpace, 1509 unsigned Alignment, 1510 MachineMemOperand::Flags Flags, 1511 bool *Fast) const { 1512 // Check if the specified alignment is sufficient based on the data layout. 1513 // TODO: While using the data layout works in practice, a better solution 1514 // would be to implement this check directly (make this a virtual function). 1515 // For example, the ABI alignment may change based on software platform while 1516 // this function should only be affected by hardware implementation. 1517 Type *Ty = VT.getTypeForEVT(Context); 1518 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1519 // Assume that an access that meets the ABI-specified alignment is fast. 1520 if (Fast != nullptr) 1521 *Fast = true; 1522 return true; 1523 } 1524 1525 // This is a misaligned access. 1526 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1527 } 1528 1529 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1530 const DataLayout &DL, EVT VT, 1531 const MachineMemOperand &MMO, 1532 bool *Fast) const { 1533 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1534 MMO.getAlignment(), MMO.getFlags(), Fast); 1535 } 1536 1537 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1538 return BranchProbability(MinPercentageForPredictableBranch, 100); 1539 } 1540 1541 //===----------------------------------------------------------------------===// 1542 // TargetTransformInfo Helpers 1543 //===----------------------------------------------------------------------===// 1544 1545 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1546 enum InstructionOpcodes { 1547 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1548 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1549 #include "llvm/IR/Instruction.def" 1550 }; 1551 switch (static_cast<InstructionOpcodes>(Opcode)) { 1552 case Ret: return 0; 1553 case Br: return 0; 1554 case Switch: return 0; 1555 case IndirectBr: return 0; 1556 case Invoke: return 0; 1557 case CallBr: return 0; 1558 case Resume: return 0; 1559 case Unreachable: return 0; 1560 case CleanupRet: return 0; 1561 case CatchRet: return 0; 1562 case CatchPad: return 0; 1563 case CatchSwitch: return 0; 1564 case CleanupPad: return 0; 1565 case FNeg: return ISD::FNEG; 1566 case Add: return ISD::ADD; 1567 case FAdd: return ISD::FADD; 1568 case Sub: return ISD::SUB; 1569 case FSub: return ISD::FSUB; 1570 case Mul: return ISD::MUL; 1571 case FMul: return ISD::FMUL; 1572 case UDiv: return ISD::UDIV; 1573 case SDiv: return ISD::SDIV; 1574 case FDiv: return ISD::FDIV; 1575 case URem: return ISD::UREM; 1576 case SRem: return ISD::SREM; 1577 case FRem: return ISD::FREM; 1578 case Shl: return ISD::SHL; 1579 case LShr: return ISD::SRL; 1580 case AShr: return ISD::SRA; 1581 case And: return ISD::AND; 1582 case Or: return ISD::OR; 1583 case Xor: return ISD::XOR; 1584 case Alloca: return 0; 1585 case Load: return ISD::LOAD; 1586 case Store: return ISD::STORE; 1587 case GetElementPtr: return 0; 1588 case Fence: return 0; 1589 case AtomicCmpXchg: return 0; 1590 case AtomicRMW: return 0; 1591 case Trunc: return ISD::TRUNCATE; 1592 case ZExt: return ISD::ZERO_EXTEND; 1593 case SExt: return ISD::SIGN_EXTEND; 1594 case FPToUI: return ISD::FP_TO_UINT; 1595 case FPToSI: return ISD::FP_TO_SINT; 1596 case UIToFP: return ISD::UINT_TO_FP; 1597 case SIToFP: return ISD::SINT_TO_FP; 1598 case FPTrunc: return ISD::FP_ROUND; 1599 case FPExt: return ISD::FP_EXTEND; 1600 case PtrToInt: return ISD::BITCAST; 1601 case IntToPtr: return ISD::BITCAST; 1602 case BitCast: return ISD::BITCAST; 1603 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1604 case ICmp: return ISD::SETCC; 1605 case FCmp: return ISD::SETCC; 1606 case PHI: return 0; 1607 case Call: return 0; 1608 case Select: return ISD::SELECT; 1609 case UserOp1: return 0; 1610 case UserOp2: return 0; 1611 case VAArg: return 0; 1612 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1613 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1614 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1615 case ExtractValue: return ISD::MERGE_VALUES; 1616 case InsertValue: return ISD::MERGE_VALUES; 1617 case LandingPad: return 0; 1618 } 1619 1620 llvm_unreachable("Unknown instruction type encountered!"); 1621 } 1622 1623 std::pair<int, MVT> 1624 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1625 Type *Ty) const { 1626 LLVMContext &C = Ty->getContext(); 1627 EVT MTy = getValueType(DL, Ty); 1628 1629 int Cost = 1; 1630 // We keep legalizing the type until we find a legal kind. We assume that 1631 // the only operation that costs anything is the split. After splitting 1632 // we need to handle two types. 1633 while (true) { 1634 LegalizeKind LK = getTypeConversion(C, MTy); 1635 1636 if (LK.first == TypeLegal) 1637 return std::make_pair(Cost, MTy.getSimpleVT()); 1638 1639 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1640 Cost *= 2; 1641 1642 // Do not loop with f128 type. 1643 if (MTy == LK.second) 1644 return std::make_pair(Cost, MTy.getSimpleVT()); 1645 1646 // Keep legalizing the type. 1647 MTy = LK.second; 1648 } 1649 } 1650 1651 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1652 bool UseTLS) const { 1653 // compiler-rt provides a variable with a magic name. Targets that do not 1654 // link with compiler-rt may also provide such a variable. 1655 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1656 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1657 auto UnsafeStackPtr = 1658 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1659 1660 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1661 1662 if (!UnsafeStackPtr) { 1663 auto TLSModel = UseTLS ? 1664 GlobalValue::InitialExecTLSModel : 1665 GlobalValue::NotThreadLocal; 1666 // The global variable is not defined yet, define it ourselves. 1667 // We use the initial-exec TLS model because we do not support the 1668 // variable living anywhere other than in the main executable. 1669 UnsafeStackPtr = new GlobalVariable( 1670 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1671 UnsafeStackPtrVar, nullptr, TLSModel); 1672 } else { 1673 // The variable exists, check its type and attributes. 1674 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1675 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1676 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1677 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1678 (UseTLS ? "" : "not ") + "be thread-local"); 1679 } 1680 return UnsafeStackPtr; 1681 } 1682 1683 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1684 if (!TM.getTargetTriple().isAndroid()) 1685 return getDefaultSafeStackPointerLocation(IRB, true); 1686 1687 // Android provides a libc function to retrieve the address of the current 1688 // thread's unsafe stack pointer. 1689 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1690 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1691 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1692 StackPtrTy->getPointerTo(0)); 1693 return IRB.CreateCall(Fn); 1694 } 1695 1696 //===----------------------------------------------------------------------===// 1697 // Loop Strength Reduction hooks 1698 //===----------------------------------------------------------------------===// 1699 1700 /// isLegalAddressingMode - Return true if the addressing mode represented 1701 /// by AM is legal for this target, for a load/store of the specified type. 1702 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1703 const AddrMode &AM, Type *Ty, 1704 unsigned AS, Instruction *I) const { 1705 // The default implementation of this implements a conservative RISCy, r+r and 1706 // r+i addr mode. 1707 1708 // Allows a sign-extended 16-bit immediate field. 1709 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1710 return false; 1711 1712 // No global is ever allowed as a base. 1713 if (AM.BaseGV) 1714 return false; 1715 1716 // Only support r+r, 1717 switch (AM.Scale) { 1718 case 0: // "r+i" or just "i", depending on HasBaseReg. 1719 break; 1720 case 1: 1721 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1722 return false; 1723 // Otherwise we have r+r or r+i. 1724 break; 1725 case 2: 1726 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1727 return false; 1728 // Allow 2*r as r+r. 1729 break; 1730 default: // Don't allow n * r 1731 return false; 1732 } 1733 1734 return true; 1735 } 1736 1737 //===----------------------------------------------------------------------===// 1738 // Stack Protector 1739 //===----------------------------------------------------------------------===// 1740 1741 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1742 // so that SelectionDAG handle SSP. 1743 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1744 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1745 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1746 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1747 return M.getOrInsertGlobal("__guard_local", PtrTy); 1748 } 1749 return nullptr; 1750 } 1751 1752 // Currently only support "standard" __stack_chk_guard. 1753 // TODO: add LOAD_STACK_GUARD support. 1754 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1755 if (!M.getNamedValue("__stack_chk_guard")) 1756 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1757 GlobalVariable::ExternalLinkage, 1758 nullptr, "__stack_chk_guard"); 1759 } 1760 1761 // Currently only support "standard" __stack_chk_guard. 1762 // TODO: add LOAD_STACK_GUARD support. 1763 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1764 return M.getNamedValue("__stack_chk_guard"); 1765 } 1766 1767 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1768 return nullptr; 1769 } 1770 1771 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1772 return MinimumJumpTableEntries; 1773 } 1774 1775 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1776 MinimumJumpTableEntries = Val; 1777 } 1778 1779 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1780 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1781 } 1782 1783 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1784 return MaximumJumpTableSize; 1785 } 1786 1787 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1788 MaximumJumpTableSize = Val; 1789 } 1790 1791 //===----------------------------------------------------------------------===// 1792 // Reciprocal Estimates 1793 //===----------------------------------------------------------------------===// 1794 1795 /// Get the reciprocal estimate attribute string for a function that will 1796 /// override the target defaults. 1797 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1798 const Function &F = MF.getFunction(); 1799 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1800 } 1801 1802 /// Construct a string for the given reciprocal operation of the given type. 1803 /// This string should match the corresponding option to the front-end's 1804 /// "-mrecip" flag assuming those strings have been passed through in an 1805 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1806 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1807 std::string Name = VT.isVector() ? "vec-" : ""; 1808 1809 Name += IsSqrt ? "sqrt" : "div"; 1810 1811 // TODO: Handle "half" or other float types? 1812 if (VT.getScalarType() == MVT::f64) { 1813 Name += "d"; 1814 } else { 1815 assert(VT.getScalarType() == MVT::f32 && 1816 "Unexpected FP type for reciprocal estimate"); 1817 Name += "f"; 1818 } 1819 1820 return Name; 1821 } 1822 1823 /// Return the character position and value (a single numeric character) of a 1824 /// customized refinement operation in the input string if it exists. Return 1825 /// false if there is no customized refinement step count. 1826 static bool parseRefinementStep(StringRef In, size_t &Position, 1827 uint8_t &Value) { 1828 const char RefStepToken = ':'; 1829 Position = In.find(RefStepToken); 1830 if (Position == StringRef::npos) 1831 return false; 1832 1833 StringRef RefStepString = In.substr(Position + 1); 1834 // Allow exactly one numeric character for the additional refinement 1835 // step parameter. 1836 if (RefStepString.size() == 1) { 1837 char RefStepChar = RefStepString[0]; 1838 if (RefStepChar >= '0' && RefStepChar <= '9') { 1839 Value = RefStepChar - '0'; 1840 return true; 1841 } 1842 } 1843 report_fatal_error("Invalid refinement step for -recip."); 1844 } 1845 1846 /// For the input attribute string, return one of the ReciprocalEstimate enum 1847 /// status values (enabled, disabled, or not specified) for this operation on 1848 /// the specified data type. 1849 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1850 if (Override.empty()) 1851 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1852 1853 SmallVector<StringRef, 4> OverrideVector; 1854 Override.split(OverrideVector, ','); 1855 unsigned NumArgs = OverrideVector.size(); 1856 1857 // Check if "all", "none", or "default" was specified. 1858 if (NumArgs == 1) { 1859 // Look for an optional setting of the number of refinement steps needed 1860 // for this type of reciprocal operation. 1861 size_t RefPos; 1862 uint8_t RefSteps; 1863 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1864 // Split the string for further processing. 1865 Override = Override.substr(0, RefPos); 1866 } 1867 1868 // All reciprocal types are enabled. 1869 if (Override == "all") 1870 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1871 1872 // All reciprocal types are disabled. 1873 if (Override == "none") 1874 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1875 1876 // Target defaults for enablement are used. 1877 if (Override == "default") 1878 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1879 } 1880 1881 // The attribute string may omit the size suffix ('f'/'d'). 1882 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1883 std::string VTNameNoSize = VTName; 1884 VTNameNoSize.pop_back(); 1885 static const char DisabledPrefix = '!'; 1886 1887 for (StringRef RecipType : OverrideVector) { 1888 size_t RefPos; 1889 uint8_t RefSteps; 1890 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1891 RecipType = RecipType.substr(0, RefPos); 1892 1893 // Ignore the disablement token for string matching. 1894 bool IsDisabled = RecipType[0] == DisabledPrefix; 1895 if (IsDisabled) 1896 RecipType = RecipType.substr(1); 1897 1898 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1899 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1900 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1901 } 1902 1903 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1904 } 1905 1906 /// For the input attribute string, return the customized refinement step count 1907 /// for this operation on the specified data type. If the step count does not 1908 /// exist, return the ReciprocalEstimate enum value for unspecified. 1909 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1910 if (Override.empty()) 1911 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1912 1913 SmallVector<StringRef, 4> OverrideVector; 1914 Override.split(OverrideVector, ','); 1915 unsigned NumArgs = OverrideVector.size(); 1916 1917 // Check if "all", "default", or "none" was specified. 1918 if (NumArgs == 1) { 1919 // Look for an optional setting of the number of refinement steps needed 1920 // for this type of reciprocal operation. 1921 size_t RefPos; 1922 uint8_t RefSteps; 1923 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1924 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1925 1926 // Split the string for further processing. 1927 Override = Override.substr(0, RefPos); 1928 assert(Override != "none" && 1929 "Disabled reciprocals, but specifed refinement steps?"); 1930 1931 // If this is a general override, return the specified number of steps. 1932 if (Override == "all" || Override == "default") 1933 return RefSteps; 1934 } 1935 1936 // The attribute string may omit the size suffix ('f'/'d'). 1937 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1938 std::string VTNameNoSize = VTName; 1939 VTNameNoSize.pop_back(); 1940 1941 for (StringRef RecipType : OverrideVector) { 1942 size_t RefPos; 1943 uint8_t RefSteps; 1944 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1945 continue; 1946 1947 RecipType = RecipType.substr(0, RefPos); 1948 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1949 return RefSteps; 1950 } 1951 1952 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1953 } 1954 1955 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1956 MachineFunction &MF) const { 1957 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1958 } 1959 1960 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1961 MachineFunction &MF) const { 1962 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1963 } 1964 1965 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1966 MachineFunction &MF) const { 1967 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1968 } 1969 1970 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1971 MachineFunction &MF) const { 1972 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1973 } 1974 1975 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1976 MF.getRegInfo().freezeReservedRegs(MF); 1977 } 1978