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 case Triple::aarch64_32: 171 setLibcallName(RTLIB::BZERO, "bzero"); 172 break; 173 default: 174 break; 175 } 176 177 if (darwinHasSinCos(TT)) { 178 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 179 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 180 if (TT.isWatchABI()) { 181 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 182 CallingConv::ARM_AAPCS_VFP); 183 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 184 CallingConv::ARM_AAPCS_VFP); 185 } 186 } 187 } else { 188 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 189 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 190 } 191 192 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 193 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 194 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 195 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 196 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 197 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 198 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 199 } 200 201 if (TT.isPS4CPU()) { 202 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 203 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 204 } 205 206 if (TT.isOSOpenBSD()) { 207 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 208 } 209 } 210 211 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 212 /// UNKNOWN_LIBCALL if there is none. 213 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 214 if (OpVT == MVT::f16) { 215 if (RetVT == MVT::f32) 216 return FPEXT_F16_F32; 217 } else if (OpVT == MVT::f32) { 218 if (RetVT == MVT::f64) 219 return FPEXT_F32_F64; 220 if (RetVT == MVT::f128) 221 return FPEXT_F32_F128; 222 if (RetVT == MVT::ppcf128) 223 return FPEXT_F32_PPCF128; 224 } else if (OpVT == MVT::f64) { 225 if (RetVT == MVT::f128) 226 return FPEXT_F64_F128; 227 else if (RetVT == MVT::ppcf128) 228 return FPEXT_F64_PPCF128; 229 } else if (OpVT == MVT::f80) { 230 if (RetVT == MVT::f128) 231 return FPEXT_F80_F128; 232 } 233 234 return UNKNOWN_LIBCALL; 235 } 236 237 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 238 /// UNKNOWN_LIBCALL if there is none. 239 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 240 if (RetVT == MVT::f16) { 241 if (OpVT == MVT::f32) 242 return FPROUND_F32_F16; 243 if (OpVT == MVT::f64) 244 return FPROUND_F64_F16; 245 if (OpVT == MVT::f80) 246 return FPROUND_F80_F16; 247 if (OpVT == MVT::f128) 248 return FPROUND_F128_F16; 249 if (OpVT == MVT::ppcf128) 250 return FPROUND_PPCF128_F16; 251 } else if (RetVT == MVT::f32) { 252 if (OpVT == MVT::f64) 253 return FPROUND_F64_F32; 254 if (OpVT == MVT::f80) 255 return FPROUND_F80_F32; 256 if (OpVT == MVT::f128) 257 return FPROUND_F128_F32; 258 if (OpVT == MVT::ppcf128) 259 return FPROUND_PPCF128_F32; 260 } else if (RetVT == MVT::f64) { 261 if (OpVT == MVT::f80) 262 return FPROUND_F80_F64; 263 if (OpVT == MVT::f128) 264 return FPROUND_F128_F64; 265 if (OpVT == MVT::ppcf128) 266 return FPROUND_PPCF128_F64; 267 } else if (RetVT == MVT::f80) { 268 if (OpVT == MVT::f128) 269 return FPROUND_F128_F80; 270 } 271 272 return UNKNOWN_LIBCALL; 273 } 274 275 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 276 /// UNKNOWN_LIBCALL if there is none. 277 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 278 if (OpVT == MVT::f32) { 279 if (RetVT == MVT::i32) 280 return FPTOSINT_F32_I32; 281 if (RetVT == MVT::i64) 282 return FPTOSINT_F32_I64; 283 if (RetVT == MVT::i128) 284 return FPTOSINT_F32_I128; 285 } else if (OpVT == MVT::f64) { 286 if (RetVT == MVT::i32) 287 return FPTOSINT_F64_I32; 288 if (RetVT == MVT::i64) 289 return FPTOSINT_F64_I64; 290 if (RetVT == MVT::i128) 291 return FPTOSINT_F64_I128; 292 } else if (OpVT == MVT::f80) { 293 if (RetVT == MVT::i32) 294 return FPTOSINT_F80_I32; 295 if (RetVT == MVT::i64) 296 return FPTOSINT_F80_I64; 297 if (RetVT == MVT::i128) 298 return FPTOSINT_F80_I128; 299 } else if (OpVT == MVT::f128) { 300 if (RetVT == MVT::i32) 301 return FPTOSINT_F128_I32; 302 if (RetVT == MVT::i64) 303 return FPTOSINT_F128_I64; 304 if (RetVT == MVT::i128) 305 return FPTOSINT_F128_I128; 306 } else if (OpVT == MVT::ppcf128) { 307 if (RetVT == MVT::i32) 308 return FPTOSINT_PPCF128_I32; 309 if (RetVT == MVT::i64) 310 return FPTOSINT_PPCF128_I64; 311 if (RetVT == MVT::i128) 312 return FPTOSINT_PPCF128_I128; 313 } 314 return UNKNOWN_LIBCALL; 315 } 316 317 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 318 /// UNKNOWN_LIBCALL if there is none. 319 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 320 if (OpVT == MVT::f32) { 321 if (RetVT == MVT::i32) 322 return FPTOUINT_F32_I32; 323 if (RetVT == MVT::i64) 324 return FPTOUINT_F32_I64; 325 if (RetVT == MVT::i128) 326 return FPTOUINT_F32_I128; 327 } else if (OpVT == MVT::f64) { 328 if (RetVT == MVT::i32) 329 return FPTOUINT_F64_I32; 330 if (RetVT == MVT::i64) 331 return FPTOUINT_F64_I64; 332 if (RetVT == MVT::i128) 333 return FPTOUINT_F64_I128; 334 } else if (OpVT == MVT::f80) { 335 if (RetVT == MVT::i32) 336 return FPTOUINT_F80_I32; 337 if (RetVT == MVT::i64) 338 return FPTOUINT_F80_I64; 339 if (RetVT == MVT::i128) 340 return FPTOUINT_F80_I128; 341 } else if (OpVT == MVT::f128) { 342 if (RetVT == MVT::i32) 343 return FPTOUINT_F128_I32; 344 if (RetVT == MVT::i64) 345 return FPTOUINT_F128_I64; 346 if (RetVT == MVT::i128) 347 return FPTOUINT_F128_I128; 348 } else if (OpVT == MVT::ppcf128) { 349 if (RetVT == MVT::i32) 350 return FPTOUINT_PPCF128_I32; 351 if (RetVT == MVT::i64) 352 return FPTOUINT_PPCF128_I64; 353 if (RetVT == MVT::i128) 354 return FPTOUINT_PPCF128_I128; 355 } 356 return UNKNOWN_LIBCALL; 357 } 358 359 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 360 /// UNKNOWN_LIBCALL if there is none. 361 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 362 if (OpVT == MVT::i32) { 363 if (RetVT == MVT::f32) 364 return SINTTOFP_I32_F32; 365 if (RetVT == MVT::f64) 366 return SINTTOFP_I32_F64; 367 if (RetVT == MVT::f80) 368 return SINTTOFP_I32_F80; 369 if (RetVT == MVT::f128) 370 return SINTTOFP_I32_F128; 371 if (RetVT == MVT::ppcf128) 372 return SINTTOFP_I32_PPCF128; 373 } else if (OpVT == MVT::i64) { 374 if (RetVT == MVT::f32) 375 return SINTTOFP_I64_F32; 376 if (RetVT == MVT::f64) 377 return SINTTOFP_I64_F64; 378 if (RetVT == MVT::f80) 379 return SINTTOFP_I64_F80; 380 if (RetVT == MVT::f128) 381 return SINTTOFP_I64_F128; 382 if (RetVT == MVT::ppcf128) 383 return SINTTOFP_I64_PPCF128; 384 } else if (OpVT == MVT::i128) { 385 if (RetVT == MVT::f32) 386 return SINTTOFP_I128_F32; 387 if (RetVT == MVT::f64) 388 return SINTTOFP_I128_F64; 389 if (RetVT == MVT::f80) 390 return SINTTOFP_I128_F80; 391 if (RetVT == MVT::f128) 392 return SINTTOFP_I128_F128; 393 if (RetVT == MVT::ppcf128) 394 return SINTTOFP_I128_PPCF128; 395 } 396 return UNKNOWN_LIBCALL; 397 } 398 399 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 400 /// UNKNOWN_LIBCALL if there is none. 401 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 402 if (OpVT == MVT::i32) { 403 if (RetVT == MVT::f32) 404 return UINTTOFP_I32_F32; 405 if (RetVT == MVT::f64) 406 return UINTTOFP_I32_F64; 407 if (RetVT == MVT::f80) 408 return UINTTOFP_I32_F80; 409 if (RetVT == MVT::f128) 410 return UINTTOFP_I32_F128; 411 if (RetVT == MVT::ppcf128) 412 return UINTTOFP_I32_PPCF128; 413 } else if (OpVT == MVT::i64) { 414 if (RetVT == MVT::f32) 415 return UINTTOFP_I64_F32; 416 if (RetVT == MVT::f64) 417 return UINTTOFP_I64_F64; 418 if (RetVT == MVT::f80) 419 return UINTTOFP_I64_F80; 420 if (RetVT == MVT::f128) 421 return UINTTOFP_I64_F128; 422 if (RetVT == MVT::ppcf128) 423 return UINTTOFP_I64_PPCF128; 424 } else if (OpVT == MVT::i128) { 425 if (RetVT == MVT::f32) 426 return UINTTOFP_I128_F32; 427 if (RetVT == MVT::f64) 428 return UINTTOFP_I128_F64; 429 if (RetVT == MVT::f80) 430 return UINTTOFP_I128_F80; 431 if (RetVT == MVT::f128) 432 return UINTTOFP_I128_F128; 433 if (RetVT == MVT::ppcf128) 434 return UINTTOFP_I128_PPCF128; 435 } 436 return UNKNOWN_LIBCALL; 437 } 438 439 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 440 #define OP_TO_LIBCALL(Name, Enum) \ 441 case Name: \ 442 switch (VT.SimpleTy) { \ 443 default: \ 444 return UNKNOWN_LIBCALL; \ 445 case MVT::i8: \ 446 return Enum##_1; \ 447 case MVT::i16: \ 448 return Enum##_2; \ 449 case MVT::i32: \ 450 return Enum##_4; \ 451 case MVT::i64: \ 452 return Enum##_8; \ 453 case MVT::i128: \ 454 return Enum##_16; \ 455 } 456 457 switch (Opc) { 458 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 459 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 460 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 461 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 462 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 463 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 464 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 465 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 466 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 467 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 469 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 470 } 471 472 #undef OP_TO_LIBCALL 473 474 return UNKNOWN_LIBCALL; 475 } 476 477 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 495 switch (ElementSize) { 496 case 1: 497 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 498 case 2: 499 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 500 case 4: 501 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 502 case 8: 503 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 504 case 16: 505 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 506 default: 507 return UNKNOWN_LIBCALL; 508 } 509 } 510 511 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 512 switch (ElementSize) { 513 case 1: 514 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 515 case 2: 516 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 517 case 4: 518 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 519 case 8: 520 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 521 case 16: 522 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 523 default: 524 return UNKNOWN_LIBCALL; 525 } 526 } 527 528 /// InitCmpLibcallCCs - Set default comparison libcall CC. 529 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 530 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 531 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 532 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 533 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 534 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 535 CCs[RTLIB::UNE_F32] = ISD::SETNE; 536 CCs[RTLIB::UNE_F64] = ISD::SETNE; 537 CCs[RTLIB::UNE_F128] = ISD::SETNE; 538 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 539 CCs[RTLIB::OGE_F32] = ISD::SETGE; 540 CCs[RTLIB::OGE_F64] = ISD::SETGE; 541 CCs[RTLIB::OGE_F128] = ISD::SETGE; 542 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 543 CCs[RTLIB::OLT_F32] = ISD::SETLT; 544 CCs[RTLIB::OLT_F64] = ISD::SETLT; 545 CCs[RTLIB::OLT_F128] = ISD::SETLT; 546 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 547 CCs[RTLIB::OLE_F32] = ISD::SETLE; 548 CCs[RTLIB::OLE_F64] = ISD::SETLE; 549 CCs[RTLIB::OLE_F128] = ISD::SETLE; 550 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 551 CCs[RTLIB::OGT_F32] = ISD::SETGT; 552 CCs[RTLIB::OGT_F64] = ISD::SETGT; 553 CCs[RTLIB::OGT_F128] = ISD::SETGT; 554 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 555 CCs[RTLIB::UO_F32] = ISD::SETNE; 556 CCs[RTLIB::UO_F64] = ISD::SETNE; 557 CCs[RTLIB::UO_F128] = ISD::SETNE; 558 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 559 CCs[RTLIB::O_F32] = ISD::SETEQ; 560 CCs[RTLIB::O_F64] = ISD::SETEQ; 561 CCs[RTLIB::O_F128] = ISD::SETEQ; 562 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 563 } 564 565 /// NOTE: The TargetMachine owns TLOF. 566 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 567 initActions(); 568 569 // Perform these initializations only once. 570 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 571 MaxLoadsPerMemcmp = 8; 572 MaxGluedStoresPerMemcpy = 0; 573 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 574 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 575 UseUnderscoreSetJmp = false; 576 UseUnderscoreLongJmp = false; 577 HasMultipleConditionRegisters = false; 578 HasExtractBitsInsn = false; 579 JumpIsExpensive = JumpIsExpensiveOverride; 580 PredictableSelectIsExpensive = false; 581 EnableExtLdPromotion = false; 582 StackPointerRegisterToSaveRestore = 0; 583 BooleanContents = UndefinedBooleanContent; 584 BooleanFloatContents = UndefinedBooleanContent; 585 BooleanVectorContents = UndefinedBooleanContent; 586 SchedPreferenceInfo = Sched::ILP; 587 GatherAllAliasesMaxDepth = 18; 588 // TODO: the default will be switched to 0 in the next commit, along 589 // with the Target-specific changes necessary. 590 MaxAtomicSizeInBitsSupported = 1024; 591 592 MinCmpXchgSizeInBits = 0; 593 SupportsUnalignedAtomics = false; 594 595 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 596 597 InitLibcalls(TM.getTargetTriple()); 598 InitCmpLibcallCCs(CmpLibcallCCs); 599 } 600 601 void TargetLoweringBase::initActions() { 602 // All operations default to being supported. 603 memset(OpActions, 0, sizeof(OpActions)); 604 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 605 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 606 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 607 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 608 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 609 std::fill(std::begin(TargetDAGCombineArray), 610 std::end(TargetDAGCombineArray), 0); 611 612 for (MVT VT : MVT::fp_valuetypes()) { 613 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 614 if (IntVT.isValid()) { 615 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 616 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 617 } 618 } 619 620 // Set default actions for various operations. 621 for (MVT VT : MVT::all_valuetypes()) { 622 // Default all indexed load / store to expand. 623 for (unsigned IM = (unsigned)ISD::PRE_INC; 624 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 625 setIndexedLoadAction(IM, VT, Expand); 626 setIndexedStoreAction(IM, VT, Expand); 627 } 628 629 // Most backends expect to see the node which just returns the value loaded. 630 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 631 632 // These operations default to expand. 633 setOperationAction(ISD::FGETSIGN, VT, Expand); 634 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 635 setOperationAction(ISD::FMINNUM, VT, Expand); 636 setOperationAction(ISD::FMAXNUM, VT, Expand); 637 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 638 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 639 setOperationAction(ISD::FMINIMUM, VT, Expand); 640 setOperationAction(ISD::FMAXIMUM, VT, Expand); 641 setOperationAction(ISD::FMAD, VT, Expand); 642 setOperationAction(ISD::SMIN, VT, Expand); 643 setOperationAction(ISD::SMAX, VT, Expand); 644 setOperationAction(ISD::UMIN, VT, Expand); 645 setOperationAction(ISD::UMAX, VT, Expand); 646 setOperationAction(ISD::ABS, VT, Expand); 647 setOperationAction(ISD::FSHL, VT, Expand); 648 setOperationAction(ISD::FSHR, VT, Expand); 649 setOperationAction(ISD::SADDSAT, VT, Expand); 650 setOperationAction(ISD::UADDSAT, VT, Expand); 651 setOperationAction(ISD::SSUBSAT, VT, Expand); 652 setOperationAction(ISD::USUBSAT, VT, Expand); 653 setOperationAction(ISD::SMULFIX, VT, Expand); 654 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 655 setOperationAction(ISD::UMULFIX, VT, Expand); 656 setOperationAction(ISD::UMULFIXSAT, VT, Expand); 657 658 // Overflow operations default to expand 659 setOperationAction(ISD::SADDO, VT, Expand); 660 setOperationAction(ISD::SSUBO, VT, Expand); 661 setOperationAction(ISD::UADDO, VT, Expand); 662 setOperationAction(ISD::USUBO, VT, Expand); 663 setOperationAction(ISD::SMULO, VT, Expand); 664 setOperationAction(ISD::UMULO, VT, Expand); 665 666 // ADDCARRY operations default to expand 667 setOperationAction(ISD::ADDCARRY, VT, Expand); 668 setOperationAction(ISD::SUBCARRY, VT, Expand); 669 setOperationAction(ISD::SETCCCARRY, VT, Expand); 670 671 // ADDC/ADDE/SUBC/SUBE default to expand. 672 setOperationAction(ISD::ADDC, VT, Expand); 673 setOperationAction(ISD::ADDE, VT, Expand); 674 setOperationAction(ISD::SUBC, VT, Expand); 675 setOperationAction(ISD::SUBE, VT, Expand); 676 677 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 678 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 679 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 680 681 setOperationAction(ISD::BITREVERSE, VT, Expand); 682 683 // These library functions default to expand. 684 setOperationAction(ISD::FROUND, VT, Expand); 685 setOperationAction(ISD::FPOWI, VT, Expand); 686 687 // These operations default to expand for vector types. 688 if (VT.isVector()) { 689 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 690 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 691 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 692 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 693 } 694 695 // Constrained floating-point operations default to expand. 696 setOperationAction(ISD::STRICT_FADD, VT, Expand); 697 setOperationAction(ISD::STRICT_FSUB, VT, Expand); 698 setOperationAction(ISD::STRICT_FMUL, VT, Expand); 699 setOperationAction(ISD::STRICT_FDIV, VT, Expand); 700 setOperationAction(ISD::STRICT_FREM, VT, Expand); 701 setOperationAction(ISD::STRICT_FMA, VT, Expand); 702 setOperationAction(ISD::STRICT_FSQRT, VT, Expand); 703 setOperationAction(ISD::STRICT_FPOW, VT, Expand); 704 setOperationAction(ISD::STRICT_FPOWI, VT, Expand); 705 setOperationAction(ISD::STRICT_FSIN, VT, Expand); 706 setOperationAction(ISD::STRICT_FCOS, VT, Expand); 707 setOperationAction(ISD::STRICT_FEXP, VT, Expand); 708 setOperationAction(ISD::STRICT_FEXP2, VT, Expand); 709 setOperationAction(ISD::STRICT_FLOG, VT, Expand); 710 setOperationAction(ISD::STRICT_FLOG10, VT, Expand); 711 setOperationAction(ISD::STRICT_FLOG2, VT, Expand); 712 setOperationAction(ISD::STRICT_LRINT, VT, Expand); 713 setOperationAction(ISD::STRICT_LLRINT, VT, Expand); 714 setOperationAction(ISD::STRICT_FRINT, VT, Expand); 715 setOperationAction(ISD::STRICT_FNEARBYINT, VT, Expand); 716 setOperationAction(ISD::STRICT_FCEIL, VT, Expand); 717 setOperationAction(ISD::STRICT_FFLOOR, VT, Expand); 718 setOperationAction(ISD::STRICT_LROUND, VT, Expand); 719 setOperationAction(ISD::STRICT_LLROUND, VT, Expand); 720 setOperationAction(ISD::STRICT_FROUND, VT, Expand); 721 setOperationAction(ISD::STRICT_FTRUNC, VT, Expand); 722 setOperationAction(ISD::STRICT_FMAXNUM, VT, Expand); 723 setOperationAction(ISD::STRICT_FMINNUM, VT, Expand); 724 setOperationAction(ISD::STRICT_FP_ROUND, VT, Expand); 725 setOperationAction(ISD::STRICT_FP_EXTEND, VT, Expand); 726 setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Expand); 727 setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Expand); 728 729 // For most targets @llvm.get.dynamic.area.offset just returns 0. 730 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 731 732 // Vector reduction default to expand. 733 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 734 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 735 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 736 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 737 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 738 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 739 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 740 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 741 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 742 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 743 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 744 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 745 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 746 } 747 748 // Most targets ignore the @llvm.prefetch intrinsic. 749 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 750 751 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 752 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 753 754 // ConstantFP nodes default to expand. Targets can either change this to 755 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 756 // to optimize expansions for certain constants. 757 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 758 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 759 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 760 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 761 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 762 763 // These library functions default to expand. 764 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 765 setOperationAction(ISD::FCBRT, VT, Expand); 766 setOperationAction(ISD::FLOG , VT, Expand); 767 setOperationAction(ISD::FLOG2, VT, Expand); 768 setOperationAction(ISD::FLOG10, VT, Expand); 769 setOperationAction(ISD::FEXP , VT, Expand); 770 setOperationAction(ISD::FEXP2, VT, Expand); 771 setOperationAction(ISD::FFLOOR, VT, Expand); 772 setOperationAction(ISD::FNEARBYINT, VT, Expand); 773 setOperationAction(ISD::FCEIL, VT, Expand); 774 setOperationAction(ISD::FRINT, VT, Expand); 775 setOperationAction(ISD::FTRUNC, VT, Expand); 776 setOperationAction(ISD::FROUND, VT, Expand); 777 setOperationAction(ISD::LROUND, VT, Expand); 778 setOperationAction(ISD::LLROUND, VT, Expand); 779 setOperationAction(ISD::LRINT, VT, Expand); 780 setOperationAction(ISD::LLRINT, VT, Expand); 781 } 782 783 // Default ISD::TRAP to expand (which turns it into abort). 784 setOperationAction(ISD::TRAP, MVT::Other, Expand); 785 786 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 787 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 788 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 789 } 790 791 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 792 EVT) const { 793 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 794 } 795 796 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 797 bool LegalTypes) const { 798 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 799 if (LHSTy.isVector()) 800 return LHSTy; 801 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 802 : getPointerTy(DL); 803 } 804 805 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 806 assert(isTypeLegal(VT)); 807 switch (Op) { 808 default: 809 return false; 810 case ISD::SDIV: 811 case ISD::UDIV: 812 case ISD::SREM: 813 case ISD::UREM: 814 return true; 815 } 816 } 817 818 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 819 // If the command-line option was specified, ignore this request. 820 if (!JumpIsExpensiveOverride.getNumOccurrences()) 821 JumpIsExpensive = isExpensive; 822 } 823 824 TargetLoweringBase::LegalizeKind 825 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 826 // If this is a simple type, use the ComputeRegisterProp mechanism. 827 if (VT.isSimple()) { 828 MVT SVT = VT.getSimpleVT(); 829 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 830 MVT NVT = TransformToType[SVT.SimpleTy]; 831 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 832 833 assert((LA == TypeLegal || LA == TypeSoftenFloat || 834 (NVT.isVector() || 835 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 836 "Promote may not follow Expand or Promote"); 837 838 if (LA == TypeSplitVector) 839 return LegalizeKind(LA, 840 EVT::getVectorVT(Context, SVT.getVectorElementType(), 841 SVT.getVectorNumElements() / 2)); 842 if (LA == TypeScalarizeVector) 843 return LegalizeKind(LA, SVT.getVectorElementType()); 844 return LegalizeKind(LA, NVT); 845 } 846 847 // Handle Extended Scalar Types. 848 if (!VT.isVector()) { 849 assert(VT.isInteger() && "Float types must be simple"); 850 unsigned BitSize = VT.getSizeInBits(); 851 // First promote to a power-of-two size, then expand if necessary. 852 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 853 EVT NVT = VT.getRoundIntegerType(Context); 854 assert(NVT != VT && "Unable to round integer VT"); 855 LegalizeKind NextStep = getTypeConversion(Context, NVT); 856 // Avoid multi-step promotion. 857 if (NextStep.first == TypePromoteInteger) 858 return NextStep; 859 // Return rounded integer type. 860 return LegalizeKind(TypePromoteInteger, NVT); 861 } 862 863 return LegalizeKind(TypeExpandInteger, 864 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 865 } 866 867 // Handle vector types. 868 unsigned NumElts = VT.getVectorNumElements(); 869 EVT EltVT = VT.getVectorElementType(); 870 871 // Vectors with only one element are always scalarized. 872 if (NumElts == 1) 873 return LegalizeKind(TypeScalarizeVector, EltVT); 874 875 // Try to widen vector elements until the element type is a power of two and 876 // promote it to a legal type later on, for example: 877 // <3 x i8> -> <4 x i8> -> <4 x i32> 878 if (EltVT.isInteger()) { 879 // Vectors with a number of elements that is not a power of two are always 880 // widened, for example <3 x i8> -> <4 x i8>. 881 if (!VT.isPow2VectorType()) { 882 NumElts = (unsigned)NextPowerOf2(NumElts); 883 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 884 return LegalizeKind(TypeWidenVector, NVT); 885 } 886 887 // Examine the element type. 888 LegalizeKind LK = getTypeConversion(Context, EltVT); 889 890 // If type is to be expanded, split the vector. 891 // <4 x i140> -> <2 x i140> 892 if (LK.first == TypeExpandInteger) 893 return LegalizeKind(TypeSplitVector, 894 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 895 896 // Promote the integer element types until a legal vector type is found 897 // or until the element integer type is too big. If a legal type was not 898 // found, fallback to the usual mechanism of widening/splitting the 899 // vector. 900 EVT OldEltVT = EltVT; 901 while (true) { 902 // Increase the bitwidth of the element to the next pow-of-two 903 // (which is greater than 8 bits). 904 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 905 .getRoundIntegerType(Context); 906 907 // Stop trying when getting a non-simple element type. 908 // Note that vector elements may be greater than legal vector element 909 // types. Example: X86 XMM registers hold 64bit element on 32bit 910 // systems. 911 if (!EltVT.isSimple()) 912 break; 913 914 // Build a new vector type and check if it is legal. 915 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 916 // Found a legal promoted vector type. 917 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 918 return LegalizeKind(TypePromoteInteger, 919 EVT::getVectorVT(Context, EltVT, NumElts)); 920 } 921 922 // Reset the type to the unexpanded type if we did not find a legal vector 923 // type with a promoted vector element type. 924 EltVT = OldEltVT; 925 } 926 927 // Try to widen the vector until a legal type is found. 928 // If there is no wider legal type, split the vector. 929 while (true) { 930 // Round up to the next power of 2. 931 NumElts = (unsigned)NextPowerOf2(NumElts); 932 933 // If there is no simple vector type with this many elements then there 934 // cannot be a larger legal vector type. Note that this assumes that 935 // there are no skipped intermediate vector types in the simple types. 936 if (!EltVT.isSimple()) 937 break; 938 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 939 if (LargerVector == MVT()) 940 break; 941 942 // If this type is legal then widen the vector. 943 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 944 return LegalizeKind(TypeWidenVector, LargerVector); 945 } 946 947 // Widen odd vectors to next power of two. 948 if (!VT.isPow2VectorType()) { 949 EVT NVT = VT.getPow2VectorType(Context); 950 return LegalizeKind(TypeWidenVector, NVT); 951 } 952 953 // Vectors with illegal element types are expanded. 954 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 955 return LegalizeKind(TypeSplitVector, NVT); 956 } 957 958 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 959 unsigned &NumIntermediates, 960 MVT &RegisterVT, 961 TargetLoweringBase *TLI) { 962 // Figure out the right, legal destination reg to copy into. 963 unsigned NumElts = VT.getVectorNumElements(); 964 MVT EltTy = VT.getVectorElementType(); 965 966 unsigned NumVectorRegs = 1; 967 968 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 969 // could break down into LHS/RHS like LegalizeDAG does. 970 if (!isPowerOf2_32(NumElts)) { 971 NumVectorRegs = NumElts; 972 NumElts = 1; 973 } 974 975 // Divide the input until we get to a supported size. This will always 976 // end with a scalar if the target doesn't support vectors. 977 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 978 NumElts >>= 1; 979 NumVectorRegs <<= 1; 980 } 981 982 NumIntermediates = NumVectorRegs; 983 984 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 985 if (!TLI->isTypeLegal(NewVT)) 986 NewVT = EltTy; 987 IntermediateVT = NewVT; 988 989 unsigned NewVTSize = NewVT.getSizeInBits(); 990 991 // Convert sizes such as i33 to i64. 992 if (!isPowerOf2_32(NewVTSize)) 993 NewVTSize = NextPowerOf2(NewVTSize); 994 995 MVT DestVT = TLI->getRegisterType(NewVT); 996 RegisterVT = DestVT; 997 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 998 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 999 1000 // Otherwise, promotion or legal types use the same number of registers as 1001 // the vector decimated to the appropriate level. 1002 return NumVectorRegs; 1003 } 1004 1005 /// isLegalRC - Return true if the value types that can be represented by the 1006 /// specified register class are all legal. 1007 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 1008 const TargetRegisterClass &RC) const { 1009 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 1010 if (isTypeLegal(*I)) 1011 return true; 1012 return false; 1013 } 1014 1015 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1016 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1017 MachineBasicBlock * 1018 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 1019 MachineBasicBlock *MBB) const { 1020 MachineInstr *MI = &InitialMI; 1021 MachineFunction &MF = *MI->getMF(); 1022 MachineFrameInfo &MFI = MF.getFrameInfo(); 1023 1024 // We're handling multiple types of operands here: 1025 // PATCHPOINT MetaArgs - live-in, read only, direct 1026 // STATEPOINT Deopt Spill - live-through, read only, indirect 1027 // STATEPOINT Deopt Alloca - live-through, read only, direct 1028 // (We're currently conservative and mark the deopt slots read/write in 1029 // practice.) 1030 // STATEPOINT GC Spill - live-through, read/write, indirect 1031 // STATEPOINT GC Alloca - live-through, read/write, direct 1032 // The live-in vs live-through is handled already (the live through ones are 1033 // all stack slots), but we need to handle the different type of stackmap 1034 // operands and memory effects here. 1035 1036 // MI changes inside this loop as we grow operands. 1037 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1038 MachineOperand &MO = MI->getOperand(OperIdx); 1039 if (!MO.isFI()) 1040 continue; 1041 1042 // foldMemoryOperand builds a new MI after replacing a single FI operand 1043 // with the canonical set of five x86 addressing-mode operands. 1044 int FI = MO.getIndex(); 1045 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1046 1047 // Copy operands before the frame-index. 1048 for (unsigned i = 0; i < OperIdx; ++i) 1049 MIB.add(MI->getOperand(i)); 1050 // Add frame index operands recognized by stackmaps.cpp 1051 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1052 // indirect-mem-ref tag, size, #FI, offset. 1053 // Used for spills inserted by StatepointLowering. This codepath is not 1054 // used for patchpoints/stackmaps at all, for these spilling is done via 1055 // foldMemoryOperand callback only. 1056 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1057 MIB.addImm(StackMaps::IndirectMemRefOp); 1058 MIB.addImm(MFI.getObjectSize(FI)); 1059 MIB.add(MI->getOperand(OperIdx)); 1060 MIB.addImm(0); 1061 } else { 1062 // direct-mem-ref tag, #FI, offset. 1063 // Used by patchpoint, and direct alloca arguments to statepoints 1064 MIB.addImm(StackMaps::DirectMemRefOp); 1065 MIB.add(MI->getOperand(OperIdx)); 1066 MIB.addImm(0); 1067 } 1068 // Copy the operands after the frame index. 1069 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1070 MIB.add(MI->getOperand(i)); 1071 1072 // Inherit previous memory operands. 1073 MIB.cloneMemRefs(*MI); 1074 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1075 1076 // Add a new memory operand for this FI. 1077 assert(MFI.getObjectOffset(FI) != -1); 1078 1079 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1080 // PATCHPOINT should be updated to do the same. (TODO) 1081 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1082 auto Flags = MachineMemOperand::MOLoad; 1083 MachineMemOperand *MMO = MF.getMachineMemOperand( 1084 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1085 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1086 MIB->addMemOperand(MF, MMO); 1087 } 1088 1089 // Replace the instruction and update the operand index. 1090 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1091 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1092 MI->eraseFromParent(); 1093 MI = MIB; 1094 } 1095 return MBB; 1096 } 1097 1098 MachineBasicBlock * 1099 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1100 MachineBasicBlock *MBB) const { 1101 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1102 "Called emitXRayCustomEvent on the wrong MI!"); 1103 auto &MF = *MI.getMF(); 1104 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1105 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1106 MIB.add(MI.getOperand(OpIdx)); 1107 1108 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1109 MI.eraseFromParent(); 1110 return MBB; 1111 } 1112 1113 MachineBasicBlock * 1114 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1115 MachineBasicBlock *MBB) const { 1116 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1117 "Called emitXRayTypedEvent on the wrong MI!"); 1118 auto &MF = *MI.getMF(); 1119 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1120 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1121 MIB.add(MI.getOperand(OpIdx)); 1122 1123 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1124 MI.eraseFromParent(); 1125 return MBB; 1126 } 1127 1128 /// findRepresentativeClass - Return the largest legal super-reg register class 1129 /// of the register class for the specified type and its associated "cost". 1130 // This function is in TargetLowering because it uses RegClassForVT which would 1131 // need to be moved to TargetRegisterInfo and would necessitate moving 1132 // isTypeLegal over as well - a massive change that would just require 1133 // TargetLowering having a TargetRegisterInfo class member that it would use. 1134 std::pair<const TargetRegisterClass *, uint8_t> 1135 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1136 MVT VT) const { 1137 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1138 if (!RC) 1139 return std::make_pair(RC, 0); 1140 1141 // Compute the set of all super-register classes. 1142 BitVector SuperRegRC(TRI->getNumRegClasses()); 1143 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1144 SuperRegRC.setBitsInMask(RCI.getMask()); 1145 1146 // Find the first legal register class with the largest spill size. 1147 const TargetRegisterClass *BestRC = RC; 1148 for (unsigned i : SuperRegRC.set_bits()) { 1149 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1150 // We want the largest possible spill size. 1151 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1152 continue; 1153 if (!isLegalRC(*TRI, *SuperRC)) 1154 continue; 1155 BestRC = SuperRC; 1156 } 1157 return std::make_pair(BestRC, 1); 1158 } 1159 1160 /// computeRegisterProperties - Once all of the register classes are added, 1161 /// this allows us to compute derived properties we expose. 1162 void TargetLoweringBase::computeRegisterProperties( 1163 const TargetRegisterInfo *TRI) { 1164 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1165 "Too many value types for ValueTypeActions to hold!"); 1166 1167 // Everything defaults to needing one register. 1168 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1169 NumRegistersForVT[i] = 1; 1170 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1171 } 1172 // ...except isVoid, which doesn't need any registers. 1173 NumRegistersForVT[MVT::isVoid] = 0; 1174 1175 // Find the largest integer register class. 1176 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1177 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1178 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1179 1180 // Every integer value type larger than this largest register takes twice as 1181 // many registers to represent as the previous ValueType. 1182 for (unsigned ExpandedReg = LargestIntReg + 1; 1183 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1184 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1185 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1186 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1187 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1188 TypeExpandInteger); 1189 } 1190 1191 // Inspect all of the ValueType's smaller than the largest integer 1192 // register to see which ones need promotion. 1193 unsigned LegalIntReg = LargestIntReg; 1194 for (unsigned IntReg = LargestIntReg - 1; 1195 IntReg >= (unsigned)MVT::i1; --IntReg) { 1196 MVT IVT = (MVT::SimpleValueType)IntReg; 1197 if (isTypeLegal(IVT)) { 1198 LegalIntReg = IntReg; 1199 } else { 1200 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1201 (MVT::SimpleValueType)LegalIntReg; 1202 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1203 } 1204 } 1205 1206 // ppcf128 type is really two f64's. 1207 if (!isTypeLegal(MVT::ppcf128)) { 1208 if (isTypeLegal(MVT::f64)) { 1209 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1210 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1211 TransformToType[MVT::ppcf128] = MVT::f64; 1212 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1213 } else { 1214 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1215 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1216 TransformToType[MVT::ppcf128] = MVT::i128; 1217 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1218 } 1219 } 1220 1221 // Decide how to handle f128. If the target does not have native f128 support, 1222 // expand it to i128 and we will be generating soft float library calls. 1223 if (!isTypeLegal(MVT::f128)) { 1224 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1225 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1226 TransformToType[MVT::f128] = MVT::i128; 1227 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1228 } 1229 1230 // Decide how to handle f64. If the target does not have native f64 support, 1231 // expand it to i64 and we will be generating soft float library calls. 1232 if (!isTypeLegal(MVT::f64)) { 1233 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1234 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1235 TransformToType[MVT::f64] = MVT::i64; 1236 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1237 } 1238 1239 // Decide how to handle f32. If the target does not have native f32 support, 1240 // expand it to i32 and we will be generating soft float library calls. 1241 if (!isTypeLegal(MVT::f32)) { 1242 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1243 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1244 TransformToType[MVT::f32] = MVT::i32; 1245 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1246 } 1247 1248 // Decide how to handle f16. If the target does not have native f16 support, 1249 // promote it to f32, because there are no f16 library calls (except for 1250 // conversions). 1251 if (!isTypeLegal(MVT::f16)) { 1252 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1253 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1254 TransformToType[MVT::f16] = MVT::f32; 1255 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1256 } 1257 1258 // Loop over all of the vector value types to see which need transformations. 1259 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1260 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1261 MVT VT = (MVT::SimpleValueType) i; 1262 if (isTypeLegal(VT)) 1263 continue; 1264 1265 MVT EltVT = VT.getVectorElementType(); 1266 unsigned NElts = VT.getVectorNumElements(); 1267 bool IsLegalWiderType = false; 1268 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1269 switch (PreferredAction) { 1270 case TypePromoteInteger: 1271 // Try to promote the elements of integer vectors. If no legal 1272 // promotion was found, fall through to the widen-vector method. 1273 for (unsigned nVT = i + 1; 1274 nVT <= MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE; ++nVT) { 1275 MVT SVT = (MVT::SimpleValueType) nVT; 1276 // Promote vectors of integers to vectors with the same number 1277 // of elements, with a wider element type. 1278 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1279 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1280 TransformToType[i] = SVT; 1281 RegisterTypeForVT[i] = SVT; 1282 NumRegistersForVT[i] = 1; 1283 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1284 IsLegalWiderType = true; 1285 break; 1286 } 1287 } 1288 if (IsLegalWiderType) 1289 break; 1290 LLVM_FALLTHROUGH; 1291 1292 case TypeWidenVector: 1293 if (isPowerOf2_32(NElts)) { 1294 // Try to widen the vector. 1295 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1296 MVT SVT = (MVT::SimpleValueType) nVT; 1297 if (SVT.getVectorElementType() == EltVT 1298 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1299 TransformToType[i] = SVT; 1300 RegisterTypeForVT[i] = SVT; 1301 NumRegistersForVT[i] = 1; 1302 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1303 IsLegalWiderType = true; 1304 break; 1305 } 1306 } 1307 if (IsLegalWiderType) 1308 break; 1309 } else { 1310 // Only widen to the next power of 2 to keep consistency with EVT. 1311 MVT NVT = VT.getPow2VectorType(); 1312 if (isTypeLegal(NVT)) { 1313 TransformToType[i] = NVT; 1314 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1315 RegisterTypeForVT[i] = NVT; 1316 NumRegistersForVT[i] = 1; 1317 break; 1318 } 1319 } 1320 LLVM_FALLTHROUGH; 1321 1322 case TypeSplitVector: 1323 case TypeScalarizeVector: { 1324 MVT IntermediateVT; 1325 MVT RegisterVT; 1326 unsigned NumIntermediates; 1327 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1328 NumIntermediates, RegisterVT, this); 1329 RegisterTypeForVT[i] = RegisterVT; 1330 1331 MVT NVT = VT.getPow2VectorType(); 1332 if (NVT == VT) { 1333 // Type is already a power of 2. The default action is to split. 1334 TransformToType[i] = MVT::Other; 1335 if (PreferredAction == TypeScalarizeVector) 1336 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1337 else if (PreferredAction == TypeSplitVector) 1338 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1339 else 1340 // Set type action according to the number of elements. 1341 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1342 : TypeSplitVector); 1343 } else { 1344 TransformToType[i] = NVT; 1345 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1346 } 1347 break; 1348 } 1349 default: 1350 llvm_unreachable("Unknown vector legalization action!"); 1351 } 1352 } 1353 1354 // Determine the 'representative' register class for each value type. 1355 // An representative register class is the largest (meaning one which is 1356 // not a sub-register class / subreg register class) legal register class for 1357 // a group of value types. For example, on i386, i8, i16, and i32 1358 // representative would be GR32; while on x86_64 it's GR64. 1359 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1360 const TargetRegisterClass* RRC; 1361 uint8_t Cost; 1362 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1363 RepRegClassForVT[i] = RRC; 1364 RepRegClassCostForVT[i] = Cost; 1365 } 1366 } 1367 1368 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1369 EVT VT) const { 1370 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1371 return getPointerTy(DL).SimpleTy; 1372 } 1373 1374 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1375 return MVT::i32; // return the default value 1376 } 1377 1378 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1379 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1380 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1381 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1382 /// 1383 /// This method returns the number of registers needed, and the VT for each 1384 /// register. It also returns the VT and quantity of the intermediate values 1385 /// before they are promoted/expanded. 1386 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1387 EVT &IntermediateVT, 1388 unsigned &NumIntermediates, 1389 MVT &RegisterVT) const { 1390 unsigned NumElts = VT.getVectorNumElements(); 1391 1392 // If there is a wider vector type with the same element type as this one, 1393 // or a promoted vector type that has the same number of elements which 1394 // are wider, then we should convert to that legal vector type. 1395 // This handles things like <2 x float> -> <4 x float> and 1396 // <4 x i1> -> <4 x i32>. 1397 LegalizeTypeAction TA = getTypeAction(Context, VT); 1398 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1399 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1400 if (isTypeLegal(RegisterEVT)) { 1401 IntermediateVT = RegisterEVT; 1402 RegisterVT = RegisterEVT.getSimpleVT(); 1403 NumIntermediates = 1; 1404 return 1; 1405 } 1406 } 1407 1408 // Figure out the right, legal destination reg to copy into. 1409 EVT EltTy = VT.getVectorElementType(); 1410 1411 unsigned NumVectorRegs = 1; 1412 1413 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1414 // could break down into LHS/RHS like LegalizeDAG does. 1415 if (!isPowerOf2_32(NumElts)) { 1416 NumVectorRegs = NumElts; 1417 NumElts = 1; 1418 } 1419 1420 // Divide the input until we get to a supported size. This will always 1421 // end with a scalar if the target doesn't support vectors. 1422 while (NumElts > 1 && !isTypeLegal( 1423 EVT::getVectorVT(Context, EltTy, NumElts))) { 1424 NumElts >>= 1; 1425 NumVectorRegs <<= 1; 1426 } 1427 1428 NumIntermediates = NumVectorRegs; 1429 1430 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1431 if (!isTypeLegal(NewVT)) 1432 NewVT = EltTy; 1433 IntermediateVT = NewVT; 1434 1435 MVT DestVT = getRegisterType(Context, NewVT); 1436 RegisterVT = DestVT; 1437 unsigned NewVTSize = NewVT.getSizeInBits(); 1438 1439 // Convert sizes such as i33 to i64. 1440 if (!isPowerOf2_32(NewVTSize)) 1441 NewVTSize = NextPowerOf2(NewVTSize); 1442 1443 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1444 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1445 1446 // Otherwise, promotion or legal types use the same number of registers as 1447 // the vector decimated to the appropriate level. 1448 return NumVectorRegs; 1449 } 1450 1451 /// Get the EVTs and ArgFlags collections that represent the legalized return 1452 /// type of the given function. This does not require a DAG or a return value, 1453 /// and is suitable for use before any DAGs for the function are constructed. 1454 /// TODO: Move this out of TargetLowering.cpp. 1455 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1456 AttributeList attr, 1457 SmallVectorImpl<ISD::OutputArg> &Outs, 1458 const TargetLowering &TLI, const DataLayout &DL) { 1459 SmallVector<EVT, 4> ValueVTs; 1460 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1461 unsigned NumValues = ValueVTs.size(); 1462 if (NumValues == 0) return; 1463 1464 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1465 EVT VT = ValueVTs[j]; 1466 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1467 1468 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1469 ExtendKind = ISD::SIGN_EXTEND; 1470 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1471 ExtendKind = ISD::ZERO_EXTEND; 1472 1473 // FIXME: C calling convention requires the return type to be promoted to 1474 // at least 32-bit. But this is not necessary for non-C calling 1475 // conventions. The frontend should mark functions whose return values 1476 // require promoting with signext or zeroext attributes. 1477 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1478 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1479 if (VT.bitsLT(MinVT)) 1480 VT = MinVT; 1481 } 1482 1483 unsigned NumParts = 1484 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1485 MVT PartVT = 1486 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1487 1488 // 'inreg' on function refers to return value 1489 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1490 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1491 Flags.setInReg(); 1492 1493 // Propagate extension type if any 1494 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1495 Flags.setSExt(); 1496 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1497 Flags.setZExt(); 1498 1499 for (unsigned i = 0; i < NumParts; ++i) 1500 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1501 } 1502 } 1503 1504 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1505 /// function arguments in the caller parameter area. This is the actual 1506 /// alignment, not its logarithm. 1507 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1508 const DataLayout &DL) const { 1509 return DL.getABITypeAlignment(Ty); 1510 } 1511 1512 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1513 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1514 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1515 // Check if the specified alignment is sufficient based on the data layout. 1516 // TODO: While using the data layout works in practice, a better solution 1517 // would be to implement this check directly (make this a virtual function). 1518 // For example, the ABI alignment may change based on software platform while 1519 // this function should only be affected by hardware implementation. 1520 Type *Ty = VT.getTypeForEVT(Context); 1521 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1522 // Assume that an access that meets the ABI-specified alignment is fast. 1523 if (Fast != nullptr) 1524 *Fast = true; 1525 return true; 1526 } 1527 1528 // This is a misaligned access. 1529 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1530 } 1531 1532 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1533 LLVMContext &Context, const DataLayout &DL, EVT VT, 1534 const MachineMemOperand &MMO, bool *Fast) const { 1535 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(), 1536 MMO.getAlignment(), MMO.getFlags(), 1537 Fast); 1538 } 1539 1540 bool TargetLoweringBase::allowsMemoryAccess( 1541 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1542 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1543 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment, 1544 Flags, Fast); 1545 } 1546 1547 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1548 const DataLayout &DL, EVT VT, 1549 const MachineMemOperand &MMO, 1550 bool *Fast) const { 1551 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1552 MMO.getAlignment(), MMO.getFlags(), Fast); 1553 } 1554 1555 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1556 return BranchProbability(MinPercentageForPredictableBranch, 100); 1557 } 1558 1559 //===----------------------------------------------------------------------===// 1560 // TargetTransformInfo Helpers 1561 //===----------------------------------------------------------------------===// 1562 1563 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1564 enum InstructionOpcodes { 1565 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1566 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1567 #include "llvm/IR/Instruction.def" 1568 }; 1569 switch (static_cast<InstructionOpcodes>(Opcode)) { 1570 case Ret: return 0; 1571 case Br: return 0; 1572 case Switch: return 0; 1573 case IndirectBr: return 0; 1574 case Invoke: return 0; 1575 case CallBr: return 0; 1576 case Resume: return 0; 1577 case Unreachable: return 0; 1578 case CleanupRet: return 0; 1579 case CatchRet: return 0; 1580 case CatchPad: return 0; 1581 case CatchSwitch: return 0; 1582 case CleanupPad: return 0; 1583 case FNeg: return ISD::FNEG; 1584 case Add: return ISD::ADD; 1585 case FAdd: return ISD::FADD; 1586 case Sub: return ISD::SUB; 1587 case FSub: return ISD::FSUB; 1588 case Mul: return ISD::MUL; 1589 case FMul: return ISD::FMUL; 1590 case UDiv: return ISD::UDIV; 1591 case SDiv: return ISD::SDIV; 1592 case FDiv: return ISD::FDIV; 1593 case URem: return ISD::UREM; 1594 case SRem: return ISD::SREM; 1595 case FRem: return ISD::FREM; 1596 case Shl: return ISD::SHL; 1597 case LShr: return ISD::SRL; 1598 case AShr: return ISD::SRA; 1599 case And: return ISD::AND; 1600 case Or: return ISD::OR; 1601 case Xor: return ISD::XOR; 1602 case Alloca: return 0; 1603 case Load: return ISD::LOAD; 1604 case Store: return ISD::STORE; 1605 case GetElementPtr: return 0; 1606 case Fence: return 0; 1607 case AtomicCmpXchg: return 0; 1608 case AtomicRMW: return 0; 1609 case Trunc: return ISD::TRUNCATE; 1610 case ZExt: return ISD::ZERO_EXTEND; 1611 case SExt: return ISD::SIGN_EXTEND; 1612 case FPToUI: return ISD::FP_TO_UINT; 1613 case FPToSI: return ISD::FP_TO_SINT; 1614 case UIToFP: return ISD::UINT_TO_FP; 1615 case SIToFP: return ISD::SINT_TO_FP; 1616 case FPTrunc: return ISD::FP_ROUND; 1617 case FPExt: return ISD::FP_EXTEND; 1618 case PtrToInt: return ISD::BITCAST; 1619 case IntToPtr: return ISD::BITCAST; 1620 case BitCast: return ISD::BITCAST; 1621 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1622 case ICmp: return ISD::SETCC; 1623 case FCmp: return ISD::SETCC; 1624 case PHI: return 0; 1625 case Call: return 0; 1626 case Select: return ISD::SELECT; 1627 case UserOp1: return 0; 1628 case UserOp2: return 0; 1629 case VAArg: return 0; 1630 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1631 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1632 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1633 case ExtractValue: return ISD::MERGE_VALUES; 1634 case InsertValue: return ISD::MERGE_VALUES; 1635 case LandingPad: return 0; 1636 } 1637 1638 llvm_unreachable("Unknown instruction type encountered!"); 1639 } 1640 1641 std::pair<int, MVT> 1642 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1643 Type *Ty) const { 1644 LLVMContext &C = Ty->getContext(); 1645 EVT MTy = getValueType(DL, Ty); 1646 1647 int Cost = 1; 1648 // We keep legalizing the type until we find a legal kind. We assume that 1649 // the only operation that costs anything is the split. After splitting 1650 // we need to handle two types. 1651 while (true) { 1652 LegalizeKind LK = getTypeConversion(C, MTy); 1653 1654 if (LK.first == TypeLegal) 1655 return std::make_pair(Cost, MTy.getSimpleVT()); 1656 1657 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1658 Cost *= 2; 1659 1660 // Do not loop with f128 type. 1661 if (MTy == LK.second) 1662 return std::make_pair(Cost, MTy.getSimpleVT()); 1663 1664 // Keep legalizing the type. 1665 MTy = LK.second; 1666 } 1667 } 1668 1669 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1670 bool UseTLS) const { 1671 // compiler-rt provides a variable with a magic name. Targets that do not 1672 // link with compiler-rt may also provide such a variable. 1673 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1674 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1675 auto UnsafeStackPtr = 1676 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1677 1678 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1679 1680 if (!UnsafeStackPtr) { 1681 auto TLSModel = UseTLS ? 1682 GlobalValue::InitialExecTLSModel : 1683 GlobalValue::NotThreadLocal; 1684 // The global variable is not defined yet, define it ourselves. 1685 // We use the initial-exec TLS model because we do not support the 1686 // variable living anywhere other than in the main executable. 1687 UnsafeStackPtr = new GlobalVariable( 1688 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1689 UnsafeStackPtrVar, nullptr, TLSModel); 1690 } else { 1691 // The variable exists, check its type and attributes. 1692 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1693 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1694 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1695 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1696 (UseTLS ? "" : "not ") + "be thread-local"); 1697 } 1698 return UnsafeStackPtr; 1699 } 1700 1701 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1702 if (!TM.getTargetTriple().isAndroid()) 1703 return getDefaultSafeStackPointerLocation(IRB, true); 1704 1705 // Android provides a libc function to retrieve the address of the current 1706 // thread's unsafe stack pointer. 1707 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1708 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1709 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1710 StackPtrTy->getPointerTo(0)); 1711 return IRB.CreateCall(Fn); 1712 } 1713 1714 //===----------------------------------------------------------------------===// 1715 // Loop Strength Reduction hooks 1716 //===----------------------------------------------------------------------===// 1717 1718 /// isLegalAddressingMode - Return true if the addressing mode represented 1719 /// by AM is legal for this target, for a load/store of the specified type. 1720 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1721 const AddrMode &AM, Type *Ty, 1722 unsigned AS, Instruction *I) const { 1723 // The default implementation of this implements a conservative RISCy, r+r and 1724 // r+i addr mode. 1725 1726 // Allows a sign-extended 16-bit immediate field. 1727 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1728 return false; 1729 1730 // No global is ever allowed as a base. 1731 if (AM.BaseGV) 1732 return false; 1733 1734 // Only support r+r, 1735 switch (AM.Scale) { 1736 case 0: // "r+i" or just "i", depending on HasBaseReg. 1737 break; 1738 case 1: 1739 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1740 return false; 1741 // Otherwise we have r+r or r+i. 1742 break; 1743 case 2: 1744 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1745 return false; 1746 // Allow 2*r as r+r. 1747 break; 1748 default: // Don't allow n * r 1749 return false; 1750 } 1751 1752 return true; 1753 } 1754 1755 //===----------------------------------------------------------------------===// 1756 // Stack Protector 1757 //===----------------------------------------------------------------------===// 1758 1759 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1760 // so that SelectionDAG handle SSP. 1761 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1762 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1763 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1764 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1765 return M.getOrInsertGlobal("__guard_local", PtrTy); 1766 } 1767 return nullptr; 1768 } 1769 1770 // Currently only support "standard" __stack_chk_guard. 1771 // TODO: add LOAD_STACK_GUARD support. 1772 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1773 if (!M.getNamedValue("__stack_chk_guard")) 1774 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1775 GlobalVariable::ExternalLinkage, 1776 nullptr, "__stack_chk_guard"); 1777 } 1778 1779 // Currently only support "standard" __stack_chk_guard. 1780 // TODO: add LOAD_STACK_GUARD support. 1781 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1782 return M.getNamedValue("__stack_chk_guard"); 1783 } 1784 1785 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1786 return nullptr; 1787 } 1788 1789 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1790 return MinimumJumpTableEntries; 1791 } 1792 1793 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1794 MinimumJumpTableEntries = Val; 1795 } 1796 1797 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1798 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1799 } 1800 1801 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1802 return MaximumJumpTableSize; 1803 } 1804 1805 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1806 MaximumJumpTableSize = Val; 1807 } 1808 1809 //===----------------------------------------------------------------------===// 1810 // Reciprocal Estimates 1811 //===----------------------------------------------------------------------===// 1812 1813 /// Get the reciprocal estimate attribute string for a function that will 1814 /// override the target defaults. 1815 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1816 const Function &F = MF.getFunction(); 1817 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1818 } 1819 1820 /// Construct a string for the given reciprocal operation of the given type. 1821 /// This string should match the corresponding option to the front-end's 1822 /// "-mrecip" flag assuming those strings have been passed through in an 1823 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1824 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1825 std::string Name = VT.isVector() ? "vec-" : ""; 1826 1827 Name += IsSqrt ? "sqrt" : "div"; 1828 1829 // TODO: Handle "half" or other float types? 1830 if (VT.getScalarType() == MVT::f64) { 1831 Name += "d"; 1832 } else { 1833 assert(VT.getScalarType() == MVT::f32 && 1834 "Unexpected FP type for reciprocal estimate"); 1835 Name += "f"; 1836 } 1837 1838 return Name; 1839 } 1840 1841 /// Return the character position and value (a single numeric character) of a 1842 /// customized refinement operation in the input string if it exists. Return 1843 /// false if there is no customized refinement step count. 1844 static bool parseRefinementStep(StringRef In, size_t &Position, 1845 uint8_t &Value) { 1846 const char RefStepToken = ':'; 1847 Position = In.find(RefStepToken); 1848 if (Position == StringRef::npos) 1849 return false; 1850 1851 StringRef RefStepString = In.substr(Position + 1); 1852 // Allow exactly one numeric character for the additional refinement 1853 // step parameter. 1854 if (RefStepString.size() == 1) { 1855 char RefStepChar = RefStepString[0]; 1856 if (RefStepChar >= '0' && RefStepChar <= '9') { 1857 Value = RefStepChar - '0'; 1858 return true; 1859 } 1860 } 1861 report_fatal_error("Invalid refinement step for -recip."); 1862 } 1863 1864 /// For the input attribute string, return one of the ReciprocalEstimate enum 1865 /// status values (enabled, disabled, or not specified) for this operation on 1866 /// the specified data type. 1867 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1868 if (Override.empty()) 1869 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1870 1871 SmallVector<StringRef, 4> OverrideVector; 1872 Override.split(OverrideVector, ','); 1873 unsigned NumArgs = OverrideVector.size(); 1874 1875 // Check if "all", "none", or "default" was specified. 1876 if (NumArgs == 1) { 1877 // Look for an optional setting of the number of refinement steps needed 1878 // for this type of reciprocal operation. 1879 size_t RefPos; 1880 uint8_t RefSteps; 1881 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1882 // Split the string for further processing. 1883 Override = Override.substr(0, RefPos); 1884 } 1885 1886 // All reciprocal types are enabled. 1887 if (Override == "all") 1888 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1889 1890 // All reciprocal types are disabled. 1891 if (Override == "none") 1892 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1893 1894 // Target defaults for enablement are used. 1895 if (Override == "default") 1896 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1897 } 1898 1899 // The attribute string may omit the size suffix ('f'/'d'). 1900 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1901 std::string VTNameNoSize = VTName; 1902 VTNameNoSize.pop_back(); 1903 static const char DisabledPrefix = '!'; 1904 1905 for (StringRef RecipType : OverrideVector) { 1906 size_t RefPos; 1907 uint8_t RefSteps; 1908 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1909 RecipType = RecipType.substr(0, RefPos); 1910 1911 // Ignore the disablement token for string matching. 1912 bool IsDisabled = RecipType[0] == DisabledPrefix; 1913 if (IsDisabled) 1914 RecipType = RecipType.substr(1); 1915 1916 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1917 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1918 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1919 } 1920 1921 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1922 } 1923 1924 /// For the input attribute string, return the customized refinement step count 1925 /// for this operation on the specified data type. If the step count does not 1926 /// exist, return the ReciprocalEstimate enum value for unspecified. 1927 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1928 if (Override.empty()) 1929 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1930 1931 SmallVector<StringRef, 4> OverrideVector; 1932 Override.split(OverrideVector, ','); 1933 unsigned NumArgs = OverrideVector.size(); 1934 1935 // Check if "all", "default", or "none" was specified. 1936 if (NumArgs == 1) { 1937 // Look for an optional setting of the number of refinement steps needed 1938 // for this type of reciprocal operation. 1939 size_t RefPos; 1940 uint8_t RefSteps; 1941 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1942 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1943 1944 // Split the string for further processing. 1945 Override = Override.substr(0, RefPos); 1946 assert(Override != "none" && 1947 "Disabled reciprocals, but specifed refinement steps?"); 1948 1949 // If this is a general override, return the specified number of steps. 1950 if (Override == "all" || Override == "default") 1951 return RefSteps; 1952 } 1953 1954 // The attribute string may omit the size suffix ('f'/'d'). 1955 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1956 std::string VTNameNoSize = VTName; 1957 VTNameNoSize.pop_back(); 1958 1959 for (StringRef RecipType : OverrideVector) { 1960 size_t RefPos; 1961 uint8_t RefSteps; 1962 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1963 continue; 1964 1965 RecipType = RecipType.substr(0, RefPos); 1966 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1967 return RefSteps; 1968 } 1969 1970 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1971 } 1972 1973 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1974 MachineFunction &MF) const { 1975 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1976 } 1977 1978 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1979 MachineFunction &MF) const { 1980 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1981 } 1982 1983 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1984 MachineFunction &MF) const { 1985 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1986 } 1987 1988 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1989 MachineFunction &MF) const { 1990 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1991 } 1992 1993 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1994 MF.getRegInfo().freezeReservedRegs(MF); 1995 } 1996