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 setOperationAction(ISD::SPLAT_VECTOR, VT, Expand); 694 } 695 696 // Constrained floating-point operations default to expand. 697 setOperationAction(ISD::STRICT_FADD, VT, Expand); 698 setOperationAction(ISD::STRICT_FSUB, VT, Expand); 699 setOperationAction(ISD::STRICT_FMUL, VT, Expand); 700 setOperationAction(ISD::STRICT_FDIV, VT, Expand); 701 setOperationAction(ISD::STRICT_FREM, VT, Expand); 702 setOperationAction(ISD::STRICT_FMA, VT, Expand); 703 setOperationAction(ISD::STRICT_FSQRT, VT, Expand); 704 setOperationAction(ISD::STRICT_FPOW, VT, Expand); 705 setOperationAction(ISD::STRICT_FPOWI, VT, Expand); 706 setOperationAction(ISD::STRICT_FSIN, VT, Expand); 707 setOperationAction(ISD::STRICT_FCOS, VT, Expand); 708 setOperationAction(ISD::STRICT_FEXP, VT, Expand); 709 setOperationAction(ISD::STRICT_FEXP2, VT, Expand); 710 setOperationAction(ISD::STRICT_FLOG, VT, Expand); 711 setOperationAction(ISD::STRICT_FLOG10, VT, Expand); 712 setOperationAction(ISD::STRICT_FLOG2, VT, Expand); 713 setOperationAction(ISD::STRICT_LRINT, VT, Expand); 714 setOperationAction(ISD::STRICT_LLRINT, VT, Expand); 715 setOperationAction(ISD::STRICT_FRINT, VT, Expand); 716 setOperationAction(ISD::STRICT_FNEARBYINT, VT, Expand); 717 setOperationAction(ISD::STRICT_FCEIL, VT, Expand); 718 setOperationAction(ISD::STRICT_FFLOOR, VT, Expand); 719 setOperationAction(ISD::STRICT_LROUND, VT, Expand); 720 setOperationAction(ISD::STRICT_LLROUND, VT, Expand); 721 setOperationAction(ISD::STRICT_FROUND, VT, Expand); 722 setOperationAction(ISD::STRICT_FTRUNC, VT, Expand); 723 setOperationAction(ISD::STRICT_FMAXNUM, VT, Expand); 724 setOperationAction(ISD::STRICT_FMINNUM, VT, Expand); 725 setOperationAction(ISD::STRICT_FP_ROUND, VT, Expand); 726 setOperationAction(ISD::STRICT_FP_EXTEND, VT, Expand); 727 setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Expand); 728 setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Expand); 729 730 // For most targets @llvm.get.dynamic.area.offset just returns 0. 731 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 732 733 // Vector reduction default to expand. 734 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 735 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 736 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 737 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 738 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 739 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 740 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 741 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 742 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 743 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 744 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 745 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 746 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 747 } 748 749 // Most targets ignore the @llvm.prefetch intrinsic. 750 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 751 752 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 753 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 754 755 // ConstantFP nodes default to expand. Targets can either change this to 756 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 757 // to optimize expansions for certain constants. 758 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 759 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 760 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 761 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 762 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 763 764 // These library functions default to expand. 765 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 766 setOperationAction(ISD::FCBRT, VT, Expand); 767 setOperationAction(ISD::FLOG , VT, Expand); 768 setOperationAction(ISD::FLOG2, VT, Expand); 769 setOperationAction(ISD::FLOG10, VT, Expand); 770 setOperationAction(ISD::FEXP , VT, Expand); 771 setOperationAction(ISD::FEXP2, VT, Expand); 772 setOperationAction(ISD::FFLOOR, VT, Expand); 773 setOperationAction(ISD::FNEARBYINT, VT, Expand); 774 setOperationAction(ISD::FCEIL, VT, Expand); 775 setOperationAction(ISD::FRINT, VT, Expand); 776 setOperationAction(ISD::FTRUNC, VT, Expand); 777 setOperationAction(ISD::FROUND, VT, Expand); 778 setOperationAction(ISD::LROUND, VT, Expand); 779 setOperationAction(ISD::LLROUND, VT, Expand); 780 setOperationAction(ISD::LRINT, VT, Expand); 781 setOperationAction(ISD::LLRINT, VT, Expand); 782 } 783 784 // Default ISD::TRAP to expand (which turns it into abort). 785 setOperationAction(ISD::TRAP, MVT::Other, Expand); 786 787 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 788 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 789 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 790 } 791 792 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 793 EVT) const { 794 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 795 } 796 797 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 798 bool LegalTypes) const { 799 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 800 if (LHSTy.isVector()) 801 return LHSTy; 802 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 803 : getPointerTy(DL); 804 } 805 806 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 807 assert(isTypeLegal(VT)); 808 switch (Op) { 809 default: 810 return false; 811 case ISD::SDIV: 812 case ISD::UDIV: 813 case ISD::SREM: 814 case ISD::UREM: 815 return true; 816 } 817 } 818 819 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 820 // If the command-line option was specified, ignore this request. 821 if (!JumpIsExpensiveOverride.getNumOccurrences()) 822 JumpIsExpensive = isExpensive; 823 } 824 825 TargetLoweringBase::LegalizeKind 826 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 827 // If this is a simple type, use the ComputeRegisterProp mechanism. 828 if (VT.isSimple()) { 829 MVT SVT = VT.getSimpleVT(); 830 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 831 MVT NVT = TransformToType[SVT.SimpleTy]; 832 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 833 834 assert((LA == TypeLegal || LA == TypeSoftenFloat || 835 (NVT.isVector() || 836 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 837 "Promote may not follow Expand or Promote"); 838 839 if (LA == TypeSplitVector) 840 return LegalizeKind(LA, 841 EVT::getVectorVT(Context, SVT.getVectorElementType(), 842 SVT.getVectorNumElements() / 2)); 843 if (LA == TypeScalarizeVector) 844 return LegalizeKind(LA, SVT.getVectorElementType()); 845 return LegalizeKind(LA, NVT); 846 } 847 848 // Handle Extended Scalar Types. 849 if (!VT.isVector()) { 850 assert(VT.isInteger() && "Float types must be simple"); 851 unsigned BitSize = VT.getSizeInBits(); 852 // First promote to a power-of-two size, then expand if necessary. 853 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 854 EVT NVT = VT.getRoundIntegerType(Context); 855 assert(NVT != VT && "Unable to round integer VT"); 856 LegalizeKind NextStep = getTypeConversion(Context, NVT); 857 // Avoid multi-step promotion. 858 if (NextStep.first == TypePromoteInteger) 859 return NextStep; 860 // Return rounded integer type. 861 return LegalizeKind(TypePromoteInteger, NVT); 862 } 863 864 return LegalizeKind(TypeExpandInteger, 865 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 866 } 867 868 // Handle vector types. 869 unsigned NumElts = VT.getVectorNumElements(); 870 EVT EltVT = VT.getVectorElementType(); 871 872 // Vectors with only one element are always scalarized. 873 if (NumElts == 1) 874 return LegalizeKind(TypeScalarizeVector, EltVT); 875 876 // Try to widen vector elements until the element type is a power of two and 877 // promote it to a legal type later on, for example: 878 // <3 x i8> -> <4 x i8> -> <4 x i32> 879 if (EltVT.isInteger()) { 880 // Vectors with a number of elements that is not a power of two are always 881 // widened, for example <3 x i8> -> <4 x i8>. 882 if (!VT.isPow2VectorType()) { 883 NumElts = (unsigned)NextPowerOf2(NumElts); 884 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 885 return LegalizeKind(TypeWidenVector, NVT); 886 } 887 888 // Examine the element type. 889 LegalizeKind LK = getTypeConversion(Context, EltVT); 890 891 // If type is to be expanded, split the vector. 892 // <4 x i140> -> <2 x i140> 893 if (LK.first == TypeExpandInteger) 894 return LegalizeKind(TypeSplitVector, 895 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 896 897 // Promote the integer element types until a legal vector type is found 898 // or until the element integer type is too big. If a legal type was not 899 // found, fallback to the usual mechanism of widening/splitting the 900 // vector. 901 EVT OldEltVT = EltVT; 902 while (true) { 903 // Increase the bitwidth of the element to the next pow-of-two 904 // (which is greater than 8 bits). 905 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 906 .getRoundIntegerType(Context); 907 908 // Stop trying when getting a non-simple element type. 909 // Note that vector elements may be greater than legal vector element 910 // types. Example: X86 XMM registers hold 64bit element on 32bit 911 // systems. 912 if (!EltVT.isSimple()) 913 break; 914 915 // Build a new vector type and check if it is legal. 916 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 917 // Found a legal promoted vector type. 918 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 919 return LegalizeKind(TypePromoteInteger, 920 EVT::getVectorVT(Context, EltVT, NumElts)); 921 } 922 923 // Reset the type to the unexpanded type if we did not find a legal vector 924 // type with a promoted vector element type. 925 EltVT = OldEltVT; 926 } 927 928 // Try to widen the vector until a legal type is found. 929 // If there is no wider legal type, split the vector. 930 while (true) { 931 // Round up to the next power of 2. 932 NumElts = (unsigned)NextPowerOf2(NumElts); 933 934 // If there is no simple vector type with this many elements then there 935 // cannot be a larger legal vector type. Note that this assumes that 936 // there are no skipped intermediate vector types in the simple types. 937 if (!EltVT.isSimple()) 938 break; 939 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 940 if (LargerVector == MVT()) 941 break; 942 943 // If this type is legal then widen the vector. 944 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 945 return LegalizeKind(TypeWidenVector, LargerVector); 946 } 947 948 // Widen odd vectors to next power of two. 949 if (!VT.isPow2VectorType()) { 950 EVT NVT = VT.getPow2VectorType(Context); 951 return LegalizeKind(TypeWidenVector, NVT); 952 } 953 954 // Vectors with illegal element types are expanded. 955 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 956 return LegalizeKind(TypeSplitVector, NVT); 957 } 958 959 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 960 unsigned &NumIntermediates, 961 MVT &RegisterVT, 962 TargetLoweringBase *TLI) { 963 // Figure out the right, legal destination reg to copy into. 964 unsigned NumElts = VT.getVectorNumElements(); 965 MVT EltTy = VT.getVectorElementType(); 966 967 unsigned NumVectorRegs = 1; 968 969 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 970 // could break down into LHS/RHS like LegalizeDAG does. 971 if (!isPowerOf2_32(NumElts)) { 972 NumVectorRegs = NumElts; 973 NumElts = 1; 974 } 975 976 // Divide the input until we get to a supported size. This will always 977 // end with a scalar if the target doesn't support vectors. 978 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 979 NumElts >>= 1; 980 NumVectorRegs <<= 1; 981 } 982 983 NumIntermediates = NumVectorRegs; 984 985 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 986 if (!TLI->isTypeLegal(NewVT)) 987 NewVT = EltTy; 988 IntermediateVT = NewVT; 989 990 unsigned NewVTSize = NewVT.getSizeInBits(); 991 992 // Convert sizes such as i33 to i64. 993 if (!isPowerOf2_32(NewVTSize)) 994 NewVTSize = NextPowerOf2(NewVTSize); 995 996 MVT DestVT = TLI->getRegisterType(NewVT); 997 RegisterVT = DestVT; 998 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 999 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1000 1001 // Otherwise, promotion or legal types use the same number of registers as 1002 // the vector decimated to the appropriate level. 1003 return NumVectorRegs; 1004 } 1005 1006 /// isLegalRC - Return true if the value types that can be represented by the 1007 /// specified register class are all legal. 1008 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 1009 const TargetRegisterClass &RC) const { 1010 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 1011 if (isTypeLegal(*I)) 1012 return true; 1013 return false; 1014 } 1015 1016 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1017 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1018 MachineBasicBlock * 1019 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 1020 MachineBasicBlock *MBB) const { 1021 MachineInstr *MI = &InitialMI; 1022 MachineFunction &MF = *MI->getMF(); 1023 MachineFrameInfo &MFI = MF.getFrameInfo(); 1024 1025 // We're handling multiple types of operands here: 1026 // PATCHPOINT MetaArgs - live-in, read only, direct 1027 // STATEPOINT Deopt Spill - live-through, read only, indirect 1028 // STATEPOINT Deopt Alloca - live-through, read only, direct 1029 // (We're currently conservative and mark the deopt slots read/write in 1030 // practice.) 1031 // STATEPOINT GC Spill - live-through, read/write, indirect 1032 // STATEPOINT GC Alloca - live-through, read/write, direct 1033 // The live-in vs live-through is handled already (the live through ones are 1034 // all stack slots), but we need to handle the different type of stackmap 1035 // operands and memory effects here. 1036 1037 // MI changes inside this loop as we grow operands. 1038 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1039 MachineOperand &MO = MI->getOperand(OperIdx); 1040 if (!MO.isFI()) 1041 continue; 1042 1043 // foldMemoryOperand builds a new MI after replacing a single FI operand 1044 // with the canonical set of five x86 addressing-mode operands. 1045 int FI = MO.getIndex(); 1046 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1047 1048 // Copy operands before the frame-index. 1049 for (unsigned i = 0; i < OperIdx; ++i) 1050 MIB.add(MI->getOperand(i)); 1051 // Add frame index operands recognized by stackmaps.cpp 1052 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1053 // indirect-mem-ref tag, size, #FI, offset. 1054 // Used for spills inserted by StatepointLowering. This codepath is not 1055 // used for patchpoints/stackmaps at all, for these spilling is done via 1056 // foldMemoryOperand callback only. 1057 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1058 MIB.addImm(StackMaps::IndirectMemRefOp); 1059 MIB.addImm(MFI.getObjectSize(FI)); 1060 MIB.add(MI->getOperand(OperIdx)); 1061 MIB.addImm(0); 1062 } else { 1063 // direct-mem-ref tag, #FI, offset. 1064 // Used by patchpoint, and direct alloca arguments to statepoints 1065 MIB.addImm(StackMaps::DirectMemRefOp); 1066 MIB.add(MI->getOperand(OperIdx)); 1067 MIB.addImm(0); 1068 } 1069 // Copy the operands after the frame index. 1070 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1071 MIB.add(MI->getOperand(i)); 1072 1073 // Inherit previous memory operands. 1074 MIB.cloneMemRefs(*MI); 1075 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1076 1077 // Add a new memory operand for this FI. 1078 assert(MFI.getObjectOffset(FI) != -1); 1079 1080 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1081 // PATCHPOINT should be updated to do the same. (TODO) 1082 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1083 auto Flags = MachineMemOperand::MOLoad; 1084 MachineMemOperand *MMO = MF.getMachineMemOperand( 1085 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1086 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1087 MIB->addMemOperand(MF, MMO); 1088 } 1089 1090 // Replace the instruction and update the operand index. 1091 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1092 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1093 MI->eraseFromParent(); 1094 MI = MIB; 1095 } 1096 return MBB; 1097 } 1098 1099 MachineBasicBlock * 1100 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1101 MachineBasicBlock *MBB) const { 1102 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1103 "Called emitXRayCustomEvent on the wrong MI!"); 1104 auto &MF = *MI.getMF(); 1105 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1106 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1107 MIB.add(MI.getOperand(OpIdx)); 1108 1109 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1110 MI.eraseFromParent(); 1111 return MBB; 1112 } 1113 1114 MachineBasicBlock * 1115 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1116 MachineBasicBlock *MBB) const { 1117 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1118 "Called emitXRayTypedEvent on the wrong MI!"); 1119 auto &MF = *MI.getMF(); 1120 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1121 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1122 MIB.add(MI.getOperand(OpIdx)); 1123 1124 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1125 MI.eraseFromParent(); 1126 return MBB; 1127 } 1128 1129 /// findRepresentativeClass - Return the largest legal super-reg register class 1130 /// of the register class for the specified type and its associated "cost". 1131 // This function is in TargetLowering because it uses RegClassForVT which would 1132 // need to be moved to TargetRegisterInfo and would necessitate moving 1133 // isTypeLegal over as well - a massive change that would just require 1134 // TargetLowering having a TargetRegisterInfo class member that it would use. 1135 std::pair<const TargetRegisterClass *, uint8_t> 1136 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1137 MVT VT) const { 1138 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1139 if (!RC) 1140 return std::make_pair(RC, 0); 1141 1142 // Compute the set of all super-register classes. 1143 BitVector SuperRegRC(TRI->getNumRegClasses()); 1144 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1145 SuperRegRC.setBitsInMask(RCI.getMask()); 1146 1147 // Find the first legal register class with the largest spill size. 1148 const TargetRegisterClass *BestRC = RC; 1149 for (unsigned i : SuperRegRC.set_bits()) { 1150 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1151 // We want the largest possible spill size. 1152 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1153 continue; 1154 if (!isLegalRC(*TRI, *SuperRC)) 1155 continue; 1156 BestRC = SuperRC; 1157 } 1158 return std::make_pair(BestRC, 1); 1159 } 1160 1161 /// computeRegisterProperties - Once all of the register classes are added, 1162 /// this allows us to compute derived properties we expose. 1163 void TargetLoweringBase::computeRegisterProperties( 1164 const TargetRegisterInfo *TRI) { 1165 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1166 "Too many value types for ValueTypeActions to hold!"); 1167 1168 // Everything defaults to needing one register. 1169 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1170 NumRegistersForVT[i] = 1; 1171 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1172 } 1173 // ...except isVoid, which doesn't need any registers. 1174 NumRegistersForVT[MVT::isVoid] = 0; 1175 1176 // Find the largest integer register class. 1177 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1178 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1179 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1180 1181 // Every integer value type larger than this largest register takes twice as 1182 // many registers to represent as the previous ValueType. 1183 for (unsigned ExpandedReg = LargestIntReg + 1; 1184 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1185 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1186 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1187 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1188 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1189 TypeExpandInteger); 1190 } 1191 1192 // Inspect all of the ValueType's smaller than the largest integer 1193 // register to see which ones need promotion. 1194 unsigned LegalIntReg = LargestIntReg; 1195 for (unsigned IntReg = LargestIntReg - 1; 1196 IntReg >= (unsigned)MVT::i1; --IntReg) { 1197 MVT IVT = (MVT::SimpleValueType)IntReg; 1198 if (isTypeLegal(IVT)) { 1199 LegalIntReg = IntReg; 1200 } else { 1201 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1202 (MVT::SimpleValueType)LegalIntReg; 1203 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1204 } 1205 } 1206 1207 // ppcf128 type is really two f64's. 1208 if (!isTypeLegal(MVT::ppcf128)) { 1209 if (isTypeLegal(MVT::f64)) { 1210 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1211 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1212 TransformToType[MVT::ppcf128] = MVT::f64; 1213 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1214 } else { 1215 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1216 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1217 TransformToType[MVT::ppcf128] = MVT::i128; 1218 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1219 } 1220 } 1221 1222 // Decide how to handle f128. If the target does not have native f128 support, 1223 // expand it to i128 and we will be generating soft float library calls. 1224 if (!isTypeLegal(MVT::f128)) { 1225 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1226 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1227 TransformToType[MVT::f128] = MVT::i128; 1228 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1229 } 1230 1231 // Decide how to handle f64. If the target does not have native f64 support, 1232 // expand it to i64 and we will be generating soft float library calls. 1233 if (!isTypeLegal(MVT::f64)) { 1234 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1235 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1236 TransformToType[MVT::f64] = MVT::i64; 1237 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1238 } 1239 1240 // Decide how to handle f32. If the target does not have native f32 support, 1241 // expand it to i32 and we will be generating soft float library calls. 1242 if (!isTypeLegal(MVT::f32)) { 1243 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1244 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1245 TransformToType[MVT::f32] = MVT::i32; 1246 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1247 } 1248 1249 // Decide how to handle f16. If the target does not have native f16 support, 1250 // promote it to f32, because there are no f16 library calls (except for 1251 // conversions). 1252 if (!isTypeLegal(MVT::f16)) { 1253 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1254 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1255 TransformToType[MVT::f16] = MVT::f32; 1256 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1257 } 1258 1259 // Loop over all of the vector value types to see which need transformations. 1260 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1261 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1262 MVT VT = (MVT::SimpleValueType) i; 1263 if (isTypeLegal(VT)) 1264 continue; 1265 1266 MVT EltVT = VT.getVectorElementType(); 1267 unsigned NElts = VT.getVectorNumElements(); 1268 bool IsLegalWiderType = false; 1269 bool IsScalable = VT.isScalableVector(); 1270 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1271 switch (PreferredAction) { 1272 case TypePromoteInteger: { 1273 MVT::SimpleValueType EndVT = IsScalable ? 1274 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE : 1275 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE; 1276 // Try to promote the elements of integer vectors. If no legal 1277 // promotion was found, fall through to the widen-vector method. 1278 for (unsigned nVT = i + 1; 1279 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) { 1280 MVT SVT = (MVT::SimpleValueType) nVT; 1281 // Promote vectors of integers to vectors with the same number 1282 // of elements, with a wider element type. 1283 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1284 SVT.getVectorNumElements() == NElts && 1285 SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1286 TransformToType[i] = SVT; 1287 RegisterTypeForVT[i] = SVT; 1288 NumRegistersForVT[i] = 1; 1289 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1290 IsLegalWiderType = true; 1291 break; 1292 } 1293 } 1294 if (IsLegalWiderType) 1295 break; 1296 LLVM_FALLTHROUGH; 1297 } 1298 1299 case TypeWidenVector: 1300 if (isPowerOf2_32(NElts)) { 1301 // Try to widen the vector. 1302 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1303 MVT SVT = (MVT::SimpleValueType) nVT; 1304 if (SVT.getVectorElementType() == EltVT 1305 && SVT.getVectorNumElements() > NElts 1306 && SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1307 TransformToType[i] = SVT; 1308 RegisterTypeForVT[i] = SVT; 1309 NumRegistersForVT[i] = 1; 1310 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1311 IsLegalWiderType = true; 1312 break; 1313 } 1314 } 1315 if (IsLegalWiderType) 1316 break; 1317 } else { 1318 // Only widen to the next power of 2 to keep consistency with EVT. 1319 MVT NVT = VT.getPow2VectorType(); 1320 if (isTypeLegal(NVT)) { 1321 TransformToType[i] = NVT; 1322 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1323 RegisterTypeForVT[i] = NVT; 1324 NumRegistersForVT[i] = 1; 1325 break; 1326 } 1327 } 1328 LLVM_FALLTHROUGH; 1329 1330 case TypeSplitVector: 1331 case TypeScalarizeVector: { 1332 MVT IntermediateVT; 1333 MVT RegisterVT; 1334 unsigned NumIntermediates; 1335 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1336 NumIntermediates, RegisterVT, this); 1337 RegisterTypeForVT[i] = RegisterVT; 1338 1339 MVT NVT = VT.getPow2VectorType(); 1340 if (NVT == VT) { 1341 // Type is already a power of 2. The default action is to split. 1342 TransformToType[i] = MVT::Other; 1343 if (PreferredAction == TypeScalarizeVector) 1344 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1345 else if (PreferredAction == TypeSplitVector) 1346 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1347 else 1348 // Set type action according to the number of elements. 1349 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1350 : TypeSplitVector); 1351 } else { 1352 TransformToType[i] = NVT; 1353 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1354 } 1355 break; 1356 } 1357 default: 1358 llvm_unreachable("Unknown vector legalization action!"); 1359 } 1360 } 1361 1362 // Determine the 'representative' register class for each value type. 1363 // An representative register class is the largest (meaning one which is 1364 // not a sub-register class / subreg register class) legal register class for 1365 // a group of value types. For example, on i386, i8, i16, and i32 1366 // representative would be GR32; while on x86_64 it's GR64. 1367 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1368 const TargetRegisterClass* RRC; 1369 uint8_t Cost; 1370 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1371 RepRegClassForVT[i] = RRC; 1372 RepRegClassCostForVT[i] = Cost; 1373 } 1374 } 1375 1376 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1377 EVT VT) const { 1378 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1379 return getPointerTy(DL).SimpleTy; 1380 } 1381 1382 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1383 return MVT::i32; // return the default value 1384 } 1385 1386 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1387 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1388 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1389 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1390 /// 1391 /// This method returns the number of registers needed, and the VT for each 1392 /// register. It also returns the VT and quantity of the intermediate values 1393 /// before they are promoted/expanded. 1394 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1395 EVT &IntermediateVT, 1396 unsigned &NumIntermediates, 1397 MVT &RegisterVT) const { 1398 unsigned NumElts = VT.getVectorNumElements(); 1399 1400 // If there is a wider vector type with the same element type as this one, 1401 // or a promoted vector type that has the same number of elements which 1402 // are wider, then we should convert to that legal vector type. 1403 // This handles things like <2 x float> -> <4 x float> and 1404 // <4 x i1> -> <4 x i32>. 1405 LegalizeTypeAction TA = getTypeAction(Context, VT); 1406 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1407 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1408 if (isTypeLegal(RegisterEVT)) { 1409 IntermediateVT = RegisterEVT; 1410 RegisterVT = RegisterEVT.getSimpleVT(); 1411 NumIntermediates = 1; 1412 return 1; 1413 } 1414 } 1415 1416 // Figure out the right, legal destination reg to copy into. 1417 EVT EltTy = VT.getVectorElementType(); 1418 1419 unsigned NumVectorRegs = 1; 1420 1421 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1422 // could break down into LHS/RHS like LegalizeDAG does. 1423 if (!isPowerOf2_32(NumElts)) { 1424 NumVectorRegs = NumElts; 1425 NumElts = 1; 1426 } 1427 1428 // Divide the input until we get to a supported size. This will always 1429 // end with a scalar if the target doesn't support vectors. 1430 while (NumElts > 1 && !isTypeLegal( 1431 EVT::getVectorVT(Context, EltTy, NumElts))) { 1432 NumElts >>= 1; 1433 NumVectorRegs <<= 1; 1434 } 1435 1436 NumIntermediates = NumVectorRegs; 1437 1438 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1439 if (!isTypeLegal(NewVT)) 1440 NewVT = EltTy; 1441 IntermediateVT = NewVT; 1442 1443 MVT DestVT = getRegisterType(Context, NewVT); 1444 RegisterVT = DestVT; 1445 unsigned NewVTSize = NewVT.getSizeInBits(); 1446 1447 // Convert sizes such as i33 to i64. 1448 if (!isPowerOf2_32(NewVTSize)) 1449 NewVTSize = NextPowerOf2(NewVTSize); 1450 1451 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1452 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1453 1454 // Otherwise, promotion or legal types use the same number of registers as 1455 // the vector decimated to the appropriate level. 1456 return NumVectorRegs; 1457 } 1458 1459 /// Get the EVTs and ArgFlags collections that represent the legalized return 1460 /// type of the given function. This does not require a DAG or a return value, 1461 /// and is suitable for use before any DAGs for the function are constructed. 1462 /// TODO: Move this out of TargetLowering.cpp. 1463 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1464 AttributeList attr, 1465 SmallVectorImpl<ISD::OutputArg> &Outs, 1466 const TargetLowering &TLI, const DataLayout &DL) { 1467 SmallVector<EVT, 4> ValueVTs; 1468 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1469 unsigned NumValues = ValueVTs.size(); 1470 if (NumValues == 0) return; 1471 1472 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1473 EVT VT = ValueVTs[j]; 1474 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1475 1476 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1477 ExtendKind = ISD::SIGN_EXTEND; 1478 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1479 ExtendKind = ISD::ZERO_EXTEND; 1480 1481 // FIXME: C calling convention requires the return type to be promoted to 1482 // at least 32-bit. But this is not necessary for non-C calling 1483 // conventions. The frontend should mark functions whose return values 1484 // require promoting with signext or zeroext attributes. 1485 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1486 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1487 if (VT.bitsLT(MinVT)) 1488 VT = MinVT; 1489 } 1490 1491 unsigned NumParts = 1492 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1493 MVT PartVT = 1494 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1495 1496 // 'inreg' on function refers to return value 1497 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1498 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1499 Flags.setInReg(); 1500 1501 // Propagate extension type if any 1502 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1503 Flags.setSExt(); 1504 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1505 Flags.setZExt(); 1506 1507 for (unsigned i = 0; i < NumParts; ++i) 1508 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1509 } 1510 } 1511 1512 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1513 /// function arguments in the caller parameter area. This is the actual 1514 /// alignment, not its logarithm. 1515 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1516 const DataLayout &DL) const { 1517 return DL.getABITypeAlignment(Ty); 1518 } 1519 1520 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1521 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1522 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1523 // Check if the specified alignment is sufficient based on the data layout. 1524 // TODO: While using the data layout works in practice, a better solution 1525 // would be to implement this check directly (make this a virtual function). 1526 // For example, the ABI alignment may change based on software platform while 1527 // this function should only be affected by hardware implementation. 1528 Type *Ty = VT.getTypeForEVT(Context); 1529 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1530 // Assume that an access that meets the ABI-specified alignment is fast. 1531 if (Fast != nullptr) 1532 *Fast = true; 1533 return true; 1534 } 1535 1536 // This is a misaligned access. 1537 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1538 } 1539 1540 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1541 LLVMContext &Context, const DataLayout &DL, EVT VT, 1542 const MachineMemOperand &MMO, bool *Fast) const { 1543 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(), 1544 MMO.getAlignment(), MMO.getFlags(), 1545 Fast); 1546 } 1547 1548 bool TargetLoweringBase::allowsMemoryAccess( 1549 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1550 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1551 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment, 1552 Flags, Fast); 1553 } 1554 1555 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1556 const DataLayout &DL, EVT VT, 1557 const MachineMemOperand &MMO, 1558 bool *Fast) const { 1559 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1560 MMO.getAlignment(), MMO.getFlags(), Fast); 1561 } 1562 1563 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1564 return BranchProbability(MinPercentageForPredictableBranch, 100); 1565 } 1566 1567 //===----------------------------------------------------------------------===// 1568 // TargetTransformInfo Helpers 1569 //===----------------------------------------------------------------------===// 1570 1571 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1572 enum InstructionOpcodes { 1573 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1574 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1575 #include "llvm/IR/Instruction.def" 1576 }; 1577 switch (static_cast<InstructionOpcodes>(Opcode)) { 1578 case Ret: return 0; 1579 case Br: return 0; 1580 case Switch: return 0; 1581 case IndirectBr: return 0; 1582 case Invoke: return 0; 1583 case CallBr: return 0; 1584 case Resume: return 0; 1585 case Unreachable: return 0; 1586 case CleanupRet: return 0; 1587 case CatchRet: return 0; 1588 case CatchPad: return 0; 1589 case CatchSwitch: return 0; 1590 case CleanupPad: return 0; 1591 case FNeg: return ISD::FNEG; 1592 case Add: return ISD::ADD; 1593 case FAdd: return ISD::FADD; 1594 case Sub: return ISD::SUB; 1595 case FSub: return ISD::FSUB; 1596 case Mul: return ISD::MUL; 1597 case FMul: return ISD::FMUL; 1598 case UDiv: return ISD::UDIV; 1599 case SDiv: return ISD::SDIV; 1600 case FDiv: return ISD::FDIV; 1601 case URem: return ISD::UREM; 1602 case SRem: return ISD::SREM; 1603 case FRem: return ISD::FREM; 1604 case Shl: return ISD::SHL; 1605 case LShr: return ISD::SRL; 1606 case AShr: return ISD::SRA; 1607 case And: return ISD::AND; 1608 case Or: return ISD::OR; 1609 case Xor: return ISD::XOR; 1610 case Alloca: return 0; 1611 case Load: return ISD::LOAD; 1612 case Store: return ISD::STORE; 1613 case GetElementPtr: return 0; 1614 case Fence: return 0; 1615 case AtomicCmpXchg: return 0; 1616 case AtomicRMW: return 0; 1617 case Trunc: return ISD::TRUNCATE; 1618 case ZExt: return ISD::ZERO_EXTEND; 1619 case SExt: return ISD::SIGN_EXTEND; 1620 case FPToUI: return ISD::FP_TO_UINT; 1621 case FPToSI: return ISD::FP_TO_SINT; 1622 case UIToFP: return ISD::UINT_TO_FP; 1623 case SIToFP: return ISD::SINT_TO_FP; 1624 case FPTrunc: return ISD::FP_ROUND; 1625 case FPExt: return ISD::FP_EXTEND; 1626 case PtrToInt: return ISD::BITCAST; 1627 case IntToPtr: return ISD::BITCAST; 1628 case BitCast: return ISD::BITCAST; 1629 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1630 case ICmp: return ISD::SETCC; 1631 case FCmp: return ISD::SETCC; 1632 case PHI: return 0; 1633 case Call: return 0; 1634 case Select: return ISD::SELECT; 1635 case UserOp1: return 0; 1636 case UserOp2: return 0; 1637 case VAArg: return 0; 1638 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1639 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1640 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1641 case ExtractValue: return ISD::MERGE_VALUES; 1642 case InsertValue: return ISD::MERGE_VALUES; 1643 case LandingPad: return 0; 1644 } 1645 1646 llvm_unreachable("Unknown instruction type encountered!"); 1647 } 1648 1649 std::pair<int, MVT> 1650 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1651 Type *Ty) const { 1652 LLVMContext &C = Ty->getContext(); 1653 EVT MTy = getValueType(DL, Ty); 1654 1655 int Cost = 1; 1656 // We keep legalizing the type until we find a legal kind. We assume that 1657 // the only operation that costs anything is the split. After splitting 1658 // we need to handle two types. 1659 while (true) { 1660 LegalizeKind LK = getTypeConversion(C, MTy); 1661 1662 if (LK.first == TypeLegal) 1663 return std::make_pair(Cost, MTy.getSimpleVT()); 1664 1665 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1666 Cost *= 2; 1667 1668 // Do not loop with f128 type. 1669 if (MTy == LK.second) 1670 return std::make_pair(Cost, MTy.getSimpleVT()); 1671 1672 // Keep legalizing the type. 1673 MTy = LK.second; 1674 } 1675 } 1676 1677 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1678 bool UseTLS) const { 1679 // compiler-rt provides a variable with a magic name. Targets that do not 1680 // link with compiler-rt may also provide such a variable. 1681 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1682 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1683 auto UnsafeStackPtr = 1684 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1685 1686 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1687 1688 if (!UnsafeStackPtr) { 1689 auto TLSModel = UseTLS ? 1690 GlobalValue::InitialExecTLSModel : 1691 GlobalValue::NotThreadLocal; 1692 // The global variable is not defined yet, define it ourselves. 1693 // We use the initial-exec TLS model because we do not support the 1694 // variable living anywhere other than in the main executable. 1695 UnsafeStackPtr = new GlobalVariable( 1696 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1697 UnsafeStackPtrVar, nullptr, TLSModel); 1698 } else { 1699 // The variable exists, check its type and attributes. 1700 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1701 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1702 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1703 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1704 (UseTLS ? "" : "not ") + "be thread-local"); 1705 } 1706 return UnsafeStackPtr; 1707 } 1708 1709 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1710 if (!TM.getTargetTriple().isAndroid()) 1711 return getDefaultSafeStackPointerLocation(IRB, true); 1712 1713 // Android provides a libc function to retrieve the address of the current 1714 // thread's unsafe stack pointer. 1715 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1716 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1717 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1718 StackPtrTy->getPointerTo(0)); 1719 return IRB.CreateCall(Fn); 1720 } 1721 1722 //===----------------------------------------------------------------------===// 1723 // Loop Strength Reduction hooks 1724 //===----------------------------------------------------------------------===// 1725 1726 /// isLegalAddressingMode - Return true if the addressing mode represented 1727 /// by AM is legal for this target, for a load/store of the specified type. 1728 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1729 const AddrMode &AM, Type *Ty, 1730 unsigned AS, Instruction *I) const { 1731 // The default implementation of this implements a conservative RISCy, r+r and 1732 // r+i addr mode. 1733 1734 // Allows a sign-extended 16-bit immediate field. 1735 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1736 return false; 1737 1738 // No global is ever allowed as a base. 1739 if (AM.BaseGV) 1740 return false; 1741 1742 // Only support r+r, 1743 switch (AM.Scale) { 1744 case 0: // "r+i" or just "i", depending on HasBaseReg. 1745 break; 1746 case 1: 1747 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1748 return false; 1749 // Otherwise we have r+r or r+i. 1750 break; 1751 case 2: 1752 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1753 return false; 1754 // Allow 2*r as r+r. 1755 break; 1756 default: // Don't allow n * r 1757 return false; 1758 } 1759 1760 return true; 1761 } 1762 1763 //===----------------------------------------------------------------------===// 1764 // Stack Protector 1765 //===----------------------------------------------------------------------===// 1766 1767 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1768 // so that SelectionDAG handle SSP. 1769 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1770 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1771 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1772 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1773 return M.getOrInsertGlobal("__guard_local", PtrTy); 1774 } 1775 return nullptr; 1776 } 1777 1778 // Currently only support "standard" __stack_chk_guard. 1779 // TODO: add LOAD_STACK_GUARD support. 1780 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1781 if (!M.getNamedValue("__stack_chk_guard")) 1782 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1783 GlobalVariable::ExternalLinkage, 1784 nullptr, "__stack_chk_guard"); 1785 } 1786 1787 // Currently only support "standard" __stack_chk_guard. 1788 // TODO: add LOAD_STACK_GUARD support. 1789 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1790 return M.getNamedValue("__stack_chk_guard"); 1791 } 1792 1793 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1794 return nullptr; 1795 } 1796 1797 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1798 return MinimumJumpTableEntries; 1799 } 1800 1801 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1802 MinimumJumpTableEntries = Val; 1803 } 1804 1805 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1806 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1807 } 1808 1809 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1810 return MaximumJumpTableSize; 1811 } 1812 1813 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1814 MaximumJumpTableSize = Val; 1815 } 1816 1817 //===----------------------------------------------------------------------===// 1818 // Reciprocal Estimates 1819 //===----------------------------------------------------------------------===// 1820 1821 /// Get the reciprocal estimate attribute string for a function that will 1822 /// override the target defaults. 1823 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1824 const Function &F = MF.getFunction(); 1825 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1826 } 1827 1828 /// Construct a string for the given reciprocal operation of the given type. 1829 /// This string should match the corresponding option to the front-end's 1830 /// "-mrecip" flag assuming those strings have been passed through in an 1831 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1832 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1833 std::string Name = VT.isVector() ? "vec-" : ""; 1834 1835 Name += IsSqrt ? "sqrt" : "div"; 1836 1837 // TODO: Handle "half" or other float types? 1838 if (VT.getScalarType() == MVT::f64) { 1839 Name += "d"; 1840 } else { 1841 assert(VT.getScalarType() == MVT::f32 && 1842 "Unexpected FP type for reciprocal estimate"); 1843 Name += "f"; 1844 } 1845 1846 return Name; 1847 } 1848 1849 /// Return the character position and value (a single numeric character) of a 1850 /// customized refinement operation in the input string if it exists. Return 1851 /// false if there is no customized refinement step count. 1852 static bool parseRefinementStep(StringRef In, size_t &Position, 1853 uint8_t &Value) { 1854 const char RefStepToken = ':'; 1855 Position = In.find(RefStepToken); 1856 if (Position == StringRef::npos) 1857 return false; 1858 1859 StringRef RefStepString = In.substr(Position + 1); 1860 // Allow exactly one numeric character for the additional refinement 1861 // step parameter. 1862 if (RefStepString.size() == 1) { 1863 char RefStepChar = RefStepString[0]; 1864 if (RefStepChar >= '0' && RefStepChar <= '9') { 1865 Value = RefStepChar - '0'; 1866 return true; 1867 } 1868 } 1869 report_fatal_error("Invalid refinement step for -recip."); 1870 } 1871 1872 /// For the input attribute string, return one of the ReciprocalEstimate enum 1873 /// status values (enabled, disabled, or not specified) for this operation on 1874 /// the specified data type. 1875 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1876 if (Override.empty()) 1877 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1878 1879 SmallVector<StringRef, 4> OverrideVector; 1880 Override.split(OverrideVector, ','); 1881 unsigned NumArgs = OverrideVector.size(); 1882 1883 // Check if "all", "none", or "default" was specified. 1884 if (NumArgs == 1) { 1885 // Look for an optional setting of the number of refinement steps needed 1886 // for this type of reciprocal operation. 1887 size_t RefPos; 1888 uint8_t RefSteps; 1889 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1890 // Split the string for further processing. 1891 Override = Override.substr(0, RefPos); 1892 } 1893 1894 // All reciprocal types are enabled. 1895 if (Override == "all") 1896 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1897 1898 // All reciprocal types are disabled. 1899 if (Override == "none") 1900 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1901 1902 // Target defaults for enablement are used. 1903 if (Override == "default") 1904 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1905 } 1906 1907 // The attribute string may omit the size suffix ('f'/'d'). 1908 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1909 std::string VTNameNoSize = VTName; 1910 VTNameNoSize.pop_back(); 1911 static const char DisabledPrefix = '!'; 1912 1913 for (StringRef RecipType : OverrideVector) { 1914 size_t RefPos; 1915 uint8_t RefSteps; 1916 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1917 RecipType = RecipType.substr(0, RefPos); 1918 1919 // Ignore the disablement token for string matching. 1920 bool IsDisabled = RecipType[0] == DisabledPrefix; 1921 if (IsDisabled) 1922 RecipType = RecipType.substr(1); 1923 1924 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1925 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1926 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1927 } 1928 1929 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1930 } 1931 1932 /// For the input attribute string, return the customized refinement step count 1933 /// for this operation on the specified data type. If the step count does not 1934 /// exist, return the ReciprocalEstimate enum value for unspecified. 1935 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1936 if (Override.empty()) 1937 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1938 1939 SmallVector<StringRef, 4> OverrideVector; 1940 Override.split(OverrideVector, ','); 1941 unsigned NumArgs = OverrideVector.size(); 1942 1943 // Check if "all", "default", or "none" was specified. 1944 if (NumArgs == 1) { 1945 // Look for an optional setting of the number of refinement steps needed 1946 // for this type of reciprocal operation. 1947 size_t RefPos; 1948 uint8_t RefSteps; 1949 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1950 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1951 1952 // Split the string for further processing. 1953 Override = Override.substr(0, RefPos); 1954 assert(Override != "none" && 1955 "Disabled reciprocals, but specifed refinement steps?"); 1956 1957 // If this is a general override, return the specified number of steps. 1958 if (Override == "all" || Override == "default") 1959 return RefSteps; 1960 } 1961 1962 // The attribute string may omit the size suffix ('f'/'d'). 1963 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1964 std::string VTNameNoSize = VTName; 1965 VTNameNoSize.pop_back(); 1966 1967 for (StringRef RecipType : OverrideVector) { 1968 size_t RefPos; 1969 uint8_t RefSteps; 1970 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1971 continue; 1972 1973 RecipType = RecipType.substr(0, RefPos); 1974 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1975 return RefSteps; 1976 } 1977 1978 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1979 } 1980 1981 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1982 MachineFunction &MF) const { 1983 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1984 } 1985 1986 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1987 MachineFunction &MF) const { 1988 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1989 } 1990 1991 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1992 MachineFunction &MF) const { 1993 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1994 } 1995 1996 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1997 MachineFunction &MF) const { 1998 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1999 } 2000 2001 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 2002 MF.getRegInfo().freezeReservedRegs(MF); 2003 } 2004