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