1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This implements the TargetLoweringBase class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/BitVector.h" 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/StringExtras.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/ADT/Triple.h" 19 #include "llvm/ADT/Twine.h" 20 #include "llvm/CodeGen/Analysis.h" 21 #include "llvm/CodeGen/ISDOpcodes.h" 22 #include "llvm/CodeGen/MachineBasicBlock.h" 23 #include "llvm/CodeGen/MachineFrameInfo.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/RuntimeLibcalls.h" 31 #include "llvm/CodeGen/StackMaps.h" 32 #include "llvm/CodeGen/TargetLowering.h" 33 #include "llvm/CodeGen/TargetOpcodes.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/ValueTypes.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/CallingConv.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DerivedTypes.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/IR/GlobalValue.h" 42 #include "llvm/IR/GlobalVariable.h" 43 #include "llvm/IR/IRBuilder.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/IR/Type.h" 46 #include "llvm/Support/BranchProbability.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Compiler.h" 50 #include "llvm/Support/ErrorHandling.h" 51 #include "llvm/Support/MachineValueType.h" 52 #include "llvm/Support/MathExtras.h" 53 #include "llvm/Target/TargetMachine.h" 54 #include <algorithm> 55 #include <cassert> 56 #include <cstddef> 57 #include <cstdint> 58 #include <cstring> 59 #include <iterator> 60 #include <string> 61 #include <tuple> 62 #include <utility> 63 64 using namespace llvm; 65 66 static cl::opt<bool> JumpIsExpensiveOverride( 67 "jump-is-expensive", cl::init(false), 68 cl::desc("Do not create extra branches to split comparison logic."), 69 cl::Hidden); 70 71 static cl::opt<unsigned> MinimumJumpTableEntries 72 ("min-jump-table-entries", cl::init(4), cl::Hidden, 73 cl::desc("Set minimum number of entries to use a jump table.")); 74 75 static cl::opt<unsigned> MaximumJumpTableSize 76 ("max-jump-table-size", cl::init(UINT_MAX), cl::Hidden, 77 cl::desc("Set maximum size of jump tables.")); 78 79 /// Minimum jump table density for normal functions. 80 static cl::opt<unsigned> 81 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 82 cl::desc("Minimum density for building a jump table in " 83 "a normal function")); 84 85 /// Minimum jump table density for -Os or -Oz functions. 86 static cl::opt<unsigned> OptsizeJumpTableDensity( 87 "optsize-jump-table-density", cl::init(40), cl::Hidden, 88 cl::desc("Minimum density for building a jump table in " 89 "an optsize function")); 90 91 static bool darwinHasSinCos(const Triple &TT) { 92 assert(TT.isOSDarwin() && "should be called with darwin triple"); 93 // Don't bother with 32 bit x86. 94 if (TT.getArch() == Triple::x86) 95 return false; 96 // Macos < 10.9 has no sincos_stret. 97 if (TT.isMacOSX()) 98 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 99 // iOS < 7.0 has no sincos_stret. 100 if (TT.isiOS()) 101 return !TT.isOSVersionLT(7, 0); 102 // Any other darwin such as WatchOS/TvOS is new enough. 103 return true; 104 } 105 106 // Although this default value is arbitrary, it is not random. It is assumed 107 // that a condition that evaluates the same way by a higher percentage than this 108 // is best represented as control flow. Therefore, the default value N should be 109 // set such that the win from N% correct executions is greater than the loss 110 // from (100 - N)% mispredicted executions for the majority of intended targets. 111 static cl::opt<int> MinPercentageForPredictableBranch( 112 "min-predictable-branch", cl::init(99), 113 cl::desc("Minimum percentage (0-100) that a condition must be either true " 114 "or false to assume that the condition is predictable"), 115 cl::Hidden); 116 117 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 118 #define HANDLE_LIBCALL(code, name) \ 119 setLibcallName(RTLIB::code, name); 120 #include "llvm/IR/RuntimeLibcalls.def" 121 #undef HANDLE_LIBCALL 122 // Initialize calling conventions to their default. 123 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 124 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 125 126 // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf". 127 if (TT.getArch() == Triple::ppc || TT.isPPC64()) { 128 setLibcallName(RTLIB::ADD_F128, "__addkf3"); 129 setLibcallName(RTLIB::SUB_F128, "__subkf3"); 130 setLibcallName(RTLIB::MUL_F128, "__mulkf3"); 131 setLibcallName(RTLIB::DIV_F128, "__divkf3"); 132 setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2"); 133 setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2"); 134 setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2"); 135 setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2"); 136 setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi"); 137 setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi"); 138 setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi"); 139 setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi"); 140 setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf"); 141 setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf"); 142 setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf"); 143 setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf"); 144 setLibcallName(RTLIB::OEQ_F128, "__eqkf2"); 145 setLibcallName(RTLIB::UNE_F128, "__nekf2"); 146 setLibcallName(RTLIB::OGE_F128, "__gekf2"); 147 setLibcallName(RTLIB::OLT_F128, "__ltkf2"); 148 setLibcallName(RTLIB::OLE_F128, "__lekf2"); 149 setLibcallName(RTLIB::OGT_F128, "__gtkf2"); 150 setLibcallName(RTLIB::UO_F128, "__unordkf2"); 151 setLibcallName(RTLIB::O_F128, "__unordkf2"); 152 } 153 154 // A few names are different on particular architectures or environments. 155 if (TT.isOSDarwin()) { 156 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 157 // of the gnueabi-style __gnu_*_ieee. 158 // FIXME: What about other targets? 159 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 160 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 161 162 // Some darwins have an optimized __bzero/bzero function. 163 switch (TT.getArch()) { 164 case Triple::x86: 165 case Triple::x86_64: 166 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 167 setLibcallName(RTLIB::BZERO, "__bzero"); 168 break; 169 case Triple::aarch64: 170 case Triple::aarch64_32: 171 setLibcallName(RTLIB::BZERO, "bzero"); 172 break; 173 default: 174 break; 175 } 176 177 if (darwinHasSinCos(TT)) { 178 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 179 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 180 if (TT.isWatchABI()) { 181 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 182 CallingConv::ARM_AAPCS_VFP); 183 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 184 CallingConv::ARM_AAPCS_VFP); 185 } 186 } 187 } else { 188 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 189 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 190 } 191 192 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 193 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 194 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 195 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 196 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 197 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 198 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 199 } 200 201 if (TT.isPS4CPU()) { 202 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 203 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 204 } 205 206 if (TT.isOSOpenBSD()) { 207 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 208 } 209 } 210 211 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 212 /// UNKNOWN_LIBCALL if there is none. 213 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 214 if (OpVT == MVT::f16) { 215 if (RetVT == MVT::f32) 216 return FPEXT_F16_F32; 217 } else if (OpVT == MVT::f32) { 218 if (RetVT == MVT::f64) 219 return FPEXT_F32_F64; 220 if (RetVT == MVT::f128) 221 return FPEXT_F32_F128; 222 if (RetVT == MVT::ppcf128) 223 return FPEXT_F32_PPCF128; 224 } else if (OpVT == MVT::f64) { 225 if (RetVT == MVT::f128) 226 return FPEXT_F64_F128; 227 else if (RetVT == MVT::ppcf128) 228 return FPEXT_F64_PPCF128; 229 } else if (OpVT == MVT::f80) { 230 if (RetVT == MVT::f128) 231 return FPEXT_F80_F128; 232 } 233 234 return UNKNOWN_LIBCALL; 235 } 236 237 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 238 /// UNKNOWN_LIBCALL if there is none. 239 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 240 if (RetVT == MVT::f16) { 241 if (OpVT == MVT::f32) 242 return FPROUND_F32_F16; 243 if (OpVT == MVT::f64) 244 return FPROUND_F64_F16; 245 if (OpVT == MVT::f80) 246 return FPROUND_F80_F16; 247 if (OpVT == MVT::f128) 248 return FPROUND_F128_F16; 249 if (OpVT == MVT::ppcf128) 250 return FPROUND_PPCF128_F16; 251 } else if (RetVT == MVT::f32) { 252 if (OpVT == MVT::f64) 253 return FPROUND_F64_F32; 254 if (OpVT == MVT::f80) 255 return FPROUND_F80_F32; 256 if (OpVT == MVT::f128) 257 return FPROUND_F128_F32; 258 if (OpVT == MVT::ppcf128) 259 return FPROUND_PPCF128_F32; 260 } else if (RetVT == MVT::f64) { 261 if (OpVT == MVT::f80) 262 return FPROUND_F80_F64; 263 if (OpVT == MVT::f128) 264 return FPROUND_F128_F64; 265 if (OpVT == MVT::ppcf128) 266 return FPROUND_PPCF128_F64; 267 } else if (RetVT == MVT::f80) { 268 if (OpVT == MVT::f128) 269 return FPROUND_F128_F80; 270 } 271 272 return UNKNOWN_LIBCALL; 273 } 274 275 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 276 /// UNKNOWN_LIBCALL if there is none. 277 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 278 if (OpVT == MVT::f32) { 279 if (RetVT == MVT::i32) 280 return FPTOSINT_F32_I32; 281 if (RetVT == MVT::i64) 282 return FPTOSINT_F32_I64; 283 if (RetVT == MVT::i128) 284 return FPTOSINT_F32_I128; 285 } else if (OpVT == MVT::f64) { 286 if (RetVT == MVT::i32) 287 return FPTOSINT_F64_I32; 288 if (RetVT == MVT::i64) 289 return FPTOSINT_F64_I64; 290 if (RetVT == MVT::i128) 291 return FPTOSINT_F64_I128; 292 } else if (OpVT == MVT::f80) { 293 if (RetVT == MVT::i32) 294 return FPTOSINT_F80_I32; 295 if (RetVT == MVT::i64) 296 return FPTOSINT_F80_I64; 297 if (RetVT == MVT::i128) 298 return FPTOSINT_F80_I128; 299 } else if (OpVT == MVT::f128) { 300 if (RetVT == MVT::i32) 301 return FPTOSINT_F128_I32; 302 if (RetVT == MVT::i64) 303 return FPTOSINT_F128_I64; 304 if (RetVT == MVT::i128) 305 return FPTOSINT_F128_I128; 306 } else if (OpVT == MVT::ppcf128) { 307 if (RetVT == MVT::i32) 308 return FPTOSINT_PPCF128_I32; 309 if (RetVT == MVT::i64) 310 return FPTOSINT_PPCF128_I64; 311 if (RetVT == MVT::i128) 312 return FPTOSINT_PPCF128_I128; 313 } 314 return UNKNOWN_LIBCALL; 315 } 316 317 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 318 /// UNKNOWN_LIBCALL if there is none. 319 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 320 if (OpVT == MVT::f32) { 321 if (RetVT == MVT::i32) 322 return FPTOUINT_F32_I32; 323 if (RetVT == MVT::i64) 324 return FPTOUINT_F32_I64; 325 if (RetVT == MVT::i128) 326 return FPTOUINT_F32_I128; 327 } else if (OpVT == MVT::f64) { 328 if (RetVT == MVT::i32) 329 return FPTOUINT_F64_I32; 330 if (RetVT == MVT::i64) 331 return FPTOUINT_F64_I64; 332 if (RetVT == MVT::i128) 333 return FPTOUINT_F64_I128; 334 } else if (OpVT == MVT::f80) { 335 if (RetVT == MVT::i32) 336 return FPTOUINT_F80_I32; 337 if (RetVT == MVT::i64) 338 return FPTOUINT_F80_I64; 339 if (RetVT == MVT::i128) 340 return FPTOUINT_F80_I128; 341 } else if (OpVT == MVT::f128) { 342 if (RetVT == MVT::i32) 343 return FPTOUINT_F128_I32; 344 if (RetVT == MVT::i64) 345 return FPTOUINT_F128_I64; 346 if (RetVT == MVT::i128) 347 return FPTOUINT_F128_I128; 348 } else if (OpVT == MVT::ppcf128) { 349 if (RetVT == MVT::i32) 350 return FPTOUINT_PPCF128_I32; 351 if (RetVT == MVT::i64) 352 return FPTOUINT_PPCF128_I64; 353 if (RetVT == MVT::i128) 354 return FPTOUINT_PPCF128_I128; 355 } 356 return UNKNOWN_LIBCALL; 357 } 358 359 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 360 /// UNKNOWN_LIBCALL if there is none. 361 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 362 if (OpVT == MVT::i32) { 363 if (RetVT == MVT::f32) 364 return SINTTOFP_I32_F32; 365 if (RetVT == MVT::f64) 366 return SINTTOFP_I32_F64; 367 if (RetVT == MVT::f80) 368 return SINTTOFP_I32_F80; 369 if (RetVT == MVT::f128) 370 return SINTTOFP_I32_F128; 371 if (RetVT == MVT::ppcf128) 372 return SINTTOFP_I32_PPCF128; 373 } else if (OpVT == MVT::i64) { 374 if (RetVT == MVT::f32) 375 return SINTTOFP_I64_F32; 376 if (RetVT == MVT::f64) 377 return SINTTOFP_I64_F64; 378 if (RetVT == MVT::f80) 379 return SINTTOFP_I64_F80; 380 if (RetVT == MVT::f128) 381 return SINTTOFP_I64_F128; 382 if (RetVT == MVT::ppcf128) 383 return SINTTOFP_I64_PPCF128; 384 } else if (OpVT == MVT::i128) { 385 if (RetVT == MVT::f32) 386 return SINTTOFP_I128_F32; 387 if (RetVT == MVT::f64) 388 return SINTTOFP_I128_F64; 389 if (RetVT == MVT::f80) 390 return SINTTOFP_I128_F80; 391 if (RetVT == MVT::f128) 392 return SINTTOFP_I128_F128; 393 if (RetVT == MVT::ppcf128) 394 return SINTTOFP_I128_PPCF128; 395 } 396 return UNKNOWN_LIBCALL; 397 } 398 399 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 400 /// UNKNOWN_LIBCALL if there is none. 401 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 402 if (OpVT == MVT::i32) { 403 if (RetVT == MVT::f32) 404 return UINTTOFP_I32_F32; 405 if (RetVT == MVT::f64) 406 return UINTTOFP_I32_F64; 407 if (RetVT == MVT::f80) 408 return UINTTOFP_I32_F80; 409 if (RetVT == MVT::f128) 410 return UINTTOFP_I32_F128; 411 if (RetVT == MVT::ppcf128) 412 return UINTTOFP_I32_PPCF128; 413 } else if (OpVT == MVT::i64) { 414 if (RetVT == MVT::f32) 415 return UINTTOFP_I64_F32; 416 if (RetVT == MVT::f64) 417 return UINTTOFP_I64_F64; 418 if (RetVT == MVT::f80) 419 return UINTTOFP_I64_F80; 420 if (RetVT == MVT::f128) 421 return UINTTOFP_I64_F128; 422 if (RetVT == MVT::ppcf128) 423 return UINTTOFP_I64_PPCF128; 424 } else if (OpVT == MVT::i128) { 425 if (RetVT == MVT::f32) 426 return UINTTOFP_I128_F32; 427 if (RetVT == MVT::f64) 428 return UINTTOFP_I128_F64; 429 if (RetVT == MVT::f80) 430 return UINTTOFP_I128_F80; 431 if (RetVT == MVT::f128) 432 return UINTTOFP_I128_F128; 433 if (RetVT == MVT::ppcf128) 434 return UINTTOFP_I128_PPCF128; 435 } 436 return UNKNOWN_LIBCALL; 437 } 438 439 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 440 #define OP_TO_LIBCALL(Name, Enum) \ 441 case Name: \ 442 switch (VT.SimpleTy) { \ 443 default: \ 444 return UNKNOWN_LIBCALL; \ 445 case MVT::i8: \ 446 return Enum##_1; \ 447 case MVT::i16: \ 448 return Enum##_2; \ 449 case MVT::i32: \ 450 return Enum##_4; \ 451 case MVT::i64: \ 452 return Enum##_8; \ 453 case MVT::i128: \ 454 return Enum##_16; \ 455 } 456 457 switch (Opc) { 458 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 459 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 460 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 461 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 462 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 463 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 464 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 465 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 466 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 467 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 469 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 470 } 471 472 #undef OP_TO_LIBCALL 473 474 return UNKNOWN_LIBCALL; 475 } 476 477 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 495 switch (ElementSize) { 496 case 1: 497 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 498 case 2: 499 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 500 case 4: 501 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 502 case 8: 503 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 504 case 16: 505 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 506 default: 507 return UNKNOWN_LIBCALL; 508 } 509 } 510 511 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 512 switch (ElementSize) { 513 case 1: 514 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 515 case 2: 516 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 517 case 4: 518 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 519 case 8: 520 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 521 case 16: 522 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 523 default: 524 return UNKNOWN_LIBCALL; 525 } 526 } 527 528 /// InitCmpLibcallCCs - Set default comparison libcall CC. 529 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 530 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 531 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 532 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 533 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 534 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 535 CCs[RTLIB::UNE_F32] = ISD::SETNE; 536 CCs[RTLIB::UNE_F64] = ISD::SETNE; 537 CCs[RTLIB::UNE_F128] = ISD::SETNE; 538 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 539 CCs[RTLIB::OGE_F32] = ISD::SETGE; 540 CCs[RTLIB::OGE_F64] = ISD::SETGE; 541 CCs[RTLIB::OGE_F128] = ISD::SETGE; 542 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 543 CCs[RTLIB::OLT_F32] = ISD::SETLT; 544 CCs[RTLIB::OLT_F64] = ISD::SETLT; 545 CCs[RTLIB::OLT_F128] = ISD::SETLT; 546 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 547 CCs[RTLIB::OLE_F32] = ISD::SETLE; 548 CCs[RTLIB::OLE_F64] = ISD::SETLE; 549 CCs[RTLIB::OLE_F128] = ISD::SETLE; 550 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 551 CCs[RTLIB::OGT_F32] = ISD::SETGT; 552 CCs[RTLIB::OGT_F64] = ISD::SETGT; 553 CCs[RTLIB::OGT_F128] = ISD::SETGT; 554 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 555 CCs[RTLIB::UO_F32] = ISD::SETNE; 556 CCs[RTLIB::UO_F64] = ISD::SETNE; 557 CCs[RTLIB::UO_F128] = ISD::SETNE; 558 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 559 CCs[RTLIB::O_F32] = ISD::SETEQ; 560 CCs[RTLIB::O_F64] = ISD::SETEQ; 561 CCs[RTLIB::O_F128] = ISD::SETEQ; 562 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 563 } 564 565 /// NOTE: The TargetMachine owns TLOF. 566 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 567 initActions(); 568 569 // Perform these initializations only once. 570 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 571 MaxLoadsPerMemcmp = 8; 572 MaxGluedStoresPerMemcpy = 0; 573 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 574 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 575 UseUnderscoreSetJmp = false; 576 UseUnderscoreLongJmp = false; 577 HasMultipleConditionRegisters = false; 578 HasExtractBitsInsn = false; 579 JumpIsExpensive = JumpIsExpensiveOverride; 580 PredictableSelectIsExpensive = false; 581 EnableExtLdPromotion = false; 582 StackPointerRegisterToSaveRestore = 0; 583 BooleanContents = UndefinedBooleanContent; 584 BooleanFloatContents = UndefinedBooleanContent; 585 BooleanVectorContents = UndefinedBooleanContent; 586 SchedPreferenceInfo = Sched::ILP; 587 GatherAllAliasesMaxDepth = 18; 588 // TODO: the default will be switched to 0 in the next commit, along 589 // with the Target-specific changes necessary. 590 MaxAtomicSizeInBitsSupported = 1024; 591 592 MinCmpXchgSizeInBits = 0; 593 SupportsUnalignedAtomics = false; 594 595 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 596 597 InitLibcalls(TM.getTargetTriple()); 598 InitCmpLibcallCCs(CmpLibcallCCs); 599 } 600 601 void TargetLoweringBase::initActions() { 602 // All operations default to being supported. 603 memset(OpActions, 0, sizeof(OpActions)); 604 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 605 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 606 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 607 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 608 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 609 std::fill(std::begin(TargetDAGCombineArray), 610 std::end(TargetDAGCombineArray), 0); 611 612 for (MVT VT : MVT::fp_valuetypes()) { 613 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 614 if (IntVT.isValid()) { 615 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 616 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 617 } 618 } 619 620 // Set default actions for various operations. 621 for (MVT VT : MVT::all_valuetypes()) { 622 // Default all indexed load / store to expand. 623 for (unsigned IM = (unsigned)ISD::PRE_INC; 624 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 625 setIndexedLoadAction(IM, VT, Expand); 626 setIndexedStoreAction(IM, VT, Expand); 627 } 628 629 // Most backends expect to see the node which just returns the value loaded. 630 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 631 632 // These operations default to expand. 633 setOperationAction(ISD::FGETSIGN, VT, Expand); 634 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 635 setOperationAction(ISD::FMINNUM, VT, Expand); 636 setOperationAction(ISD::FMAXNUM, VT, Expand); 637 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 638 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 639 setOperationAction(ISD::FMINIMUM, VT, Expand); 640 setOperationAction(ISD::FMAXIMUM, VT, Expand); 641 setOperationAction(ISD::FMAD, VT, Expand); 642 setOperationAction(ISD::SMIN, VT, Expand); 643 setOperationAction(ISD::SMAX, VT, Expand); 644 setOperationAction(ISD::UMIN, VT, Expand); 645 setOperationAction(ISD::UMAX, VT, Expand); 646 setOperationAction(ISD::ABS, VT, Expand); 647 setOperationAction(ISD::FSHL, VT, Expand); 648 setOperationAction(ISD::FSHR, VT, Expand); 649 setOperationAction(ISD::SADDSAT, VT, Expand); 650 setOperationAction(ISD::UADDSAT, VT, Expand); 651 setOperationAction(ISD::SSUBSAT, VT, Expand); 652 setOperationAction(ISD::USUBSAT, VT, Expand); 653 setOperationAction(ISD::SMULFIX, VT, Expand); 654 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 655 setOperationAction(ISD::UMULFIX, VT, Expand); 656 setOperationAction(ISD::UMULFIXSAT, VT, Expand); 657 658 // Overflow operations default to expand 659 setOperationAction(ISD::SADDO, VT, Expand); 660 setOperationAction(ISD::SSUBO, VT, Expand); 661 setOperationAction(ISD::UADDO, VT, Expand); 662 setOperationAction(ISD::USUBO, VT, Expand); 663 setOperationAction(ISD::SMULO, VT, Expand); 664 setOperationAction(ISD::UMULO, VT, Expand); 665 666 // ADDCARRY operations default to expand 667 setOperationAction(ISD::ADDCARRY, VT, Expand); 668 setOperationAction(ISD::SUBCARRY, VT, Expand); 669 setOperationAction(ISD::SETCCCARRY, VT, Expand); 670 671 // ADDC/ADDE/SUBC/SUBE default to expand. 672 setOperationAction(ISD::ADDC, VT, Expand); 673 setOperationAction(ISD::ADDE, VT, Expand); 674 setOperationAction(ISD::SUBC, VT, Expand); 675 setOperationAction(ISD::SUBE, VT, Expand); 676 677 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 678 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 679 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 680 681 setOperationAction(ISD::BITREVERSE, VT, Expand); 682 683 // These library functions default to expand. 684 setOperationAction(ISD::FROUND, VT, Expand); 685 setOperationAction(ISD::FPOWI, VT, Expand); 686 687 // These operations default to expand for vector types. 688 if (VT.isVector()) { 689 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 690 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 691 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 692 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 693 } 694 695 // Constrained floating-point operations default to expand. 696 setOperationAction(ISD::STRICT_FADD, VT, Expand); 697 setOperationAction(ISD::STRICT_FSUB, VT, Expand); 698 setOperationAction(ISD::STRICT_FMUL, VT, Expand); 699 setOperationAction(ISD::STRICT_FDIV, VT, Expand); 700 setOperationAction(ISD::STRICT_FREM, VT, Expand); 701 setOperationAction(ISD::STRICT_FMA, VT, Expand); 702 setOperationAction(ISD::STRICT_FSQRT, VT, Expand); 703 setOperationAction(ISD::STRICT_FPOW, VT, Expand); 704 setOperationAction(ISD::STRICT_FPOWI, VT, Expand); 705 setOperationAction(ISD::STRICT_FSIN, VT, Expand); 706 setOperationAction(ISD::STRICT_FCOS, VT, Expand); 707 setOperationAction(ISD::STRICT_FEXP, VT, Expand); 708 setOperationAction(ISD::STRICT_FEXP2, VT, Expand); 709 setOperationAction(ISD::STRICT_FLOG, VT, Expand); 710 setOperationAction(ISD::STRICT_FLOG10, VT, Expand); 711 setOperationAction(ISD::STRICT_FLOG2, VT, Expand); 712 setOperationAction(ISD::STRICT_FRINT, VT, Expand); 713 setOperationAction(ISD::STRICT_FNEARBYINT, VT, Expand); 714 setOperationAction(ISD::STRICT_FCEIL, VT, Expand); 715 setOperationAction(ISD::STRICT_FFLOOR, VT, Expand); 716 setOperationAction(ISD::STRICT_FROUND, VT, Expand); 717 setOperationAction(ISD::STRICT_FTRUNC, VT, Expand); 718 setOperationAction(ISD::STRICT_FMAXNUM, VT, Expand); 719 setOperationAction(ISD::STRICT_FMINNUM, VT, Expand); 720 setOperationAction(ISD::STRICT_FP_ROUND, VT, Expand); 721 setOperationAction(ISD::STRICT_FP_EXTEND, VT, Expand); 722 setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Expand); 723 setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Expand); 724 725 // For most targets @llvm.get.dynamic.area.offset just returns 0. 726 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 727 728 // Vector reduction default to expand. 729 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 730 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 731 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 732 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 733 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 734 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 735 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 736 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 737 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 738 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 739 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 740 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 741 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 742 } 743 744 // Most targets ignore the @llvm.prefetch intrinsic. 745 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 746 747 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 748 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 749 750 // ConstantFP nodes default to expand. Targets can either change this to 751 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 752 // to optimize expansions for certain constants. 753 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 754 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 755 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 756 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 757 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 758 759 // These library functions default to expand. 760 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 761 setOperationAction(ISD::FCBRT, VT, Expand); 762 setOperationAction(ISD::FLOG , VT, Expand); 763 setOperationAction(ISD::FLOG2, VT, Expand); 764 setOperationAction(ISD::FLOG10, VT, Expand); 765 setOperationAction(ISD::FEXP , VT, Expand); 766 setOperationAction(ISD::FEXP2, VT, Expand); 767 setOperationAction(ISD::FFLOOR, VT, Expand); 768 setOperationAction(ISD::FNEARBYINT, VT, Expand); 769 setOperationAction(ISD::FCEIL, VT, Expand); 770 setOperationAction(ISD::FRINT, VT, Expand); 771 setOperationAction(ISD::FTRUNC, VT, Expand); 772 setOperationAction(ISD::FROUND, VT, Expand); 773 setOperationAction(ISD::LROUND, VT, Expand); 774 setOperationAction(ISD::LLROUND, VT, Expand); 775 setOperationAction(ISD::LRINT, VT, Expand); 776 setOperationAction(ISD::LLRINT, VT, Expand); 777 } 778 779 // Default ISD::TRAP to expand (which turns it into abort). 780 setOperationAction(ISD::TRAP, MVT::Other, Expand); 781 782 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 783 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 784 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 785 } 786 787 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 788 EVT) const { 789 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 790 } 791 792 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 793 bool LegalTypes) const { 794 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 795 if (LHSTy.isVector()) 796 return LHSTy; 797 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 798 : getPointerTy(DL); 799 } 800 801 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 802 assert(isTypeLegal(VT)); 803 switch (Op) { 804 default: 805 return false; 806 case ISD::SDIV: 807 case ISD::UDIV: 808 case ISD::SREM: 809 case ISD::UREM: 810 return true; 811 } 812 } 813 814 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 815 // If the command-line option was specified, ignore this request. 816 if (!JumpIsExpensiveOverride.getNumOccurrences()) 817 JumpIsExpensive = isExpensive; 818 } 819 820 TargetLoweringBase::LegalizeKind 821 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 822 // If this is a simple type, use the ComputeRegisterProp mechanism. 823 if (VT.isSimple()) { 824 MVT SVT = VT.getSimpleVT(); 825 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 826 MVT NVT = TransformToType[SVT.SimpleTy]; 827 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 828 829 assert((LA == TypeLegal || LA == TypeSoftenFloat || 830 (NVT.isVector() || 831 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 832 "Promote may not follow Expand or Promote"); 833 834 if (LA == TypeSplitVector) 835 return LegalizeKind(LA, 836 EVT::getVectorVT(Context, SVT.getVectorElementType(), 837 SVT.getVectorNumElements() / 2)); 838 if (LA == TypeScalarizeVector) 839 return LegalizeKind(LA, SVT.getVectorElementType()); 840 return LegalizeKind(LA, NVT); 841 } 842 843 // Handle Extended Scalar Types. 844 if (!VT.isVector()) { 845 assert(VT.isInteger() && "Float types must be simple"); 846 unsigned BitSize = VT.getSizeInBits(); 847 // First promote to a power-of-two size, then expand if necessary. 848 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 849 EVT NVT = VT.getRoundIntegerType(Context); 850 assert(NVT != VT && "Unable to round integer VT"); 851 LegalizeKind NextStep = getTypeConversion(Context, NVT); 852 // Avoid multi-step promotion. 853 if (NextStep.first == TypePromoteInteger) 854 return NextStep; 855 // Return rounded integer type. 856 return LegalizeKind(TypePromoteInteger, NVT); 857 } 858 859 return LegalizeKind(TypeExpandInteger, 860 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 861 } 862 863 // Handle vector types. 864 unsigned NumElts = VT.getVectorNumElements(); 865 EVT EltVT = VT.getVectorElementType(); 866 867 // Vectors with only one element are always scalarized. 868 if (NumElts == 1) 869 return LegalizeKind(TypeScalarizeVector, EltVT); 870 871 // Try to widen vector elements until the element type is a power of two and 872 // promote it to a legal type later on, for example: 873 // <3 x i8> -> <4 x i8> -> <4 x i32> 874 if (EltVT.isInteger()) { 875 // Vectors with a number of elements that is not a power of two are always 876 // widened, for example <3 x i8> -> <4 x i8>. 877 if (!VT.isPow2VectorType()) { 878 NumElts = (unsigned)NextPowerOf2(NumElts); 879 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 880 return LegalizeKind(TypeWidenVector, NVT); 881 } 882 883 // Examine the element type. 884 LegalizeKind LK = getTypeConversion(Context, EltVT); 885 886 // If type is to be expanded, split the vector. 887 // <4 x i140> -> <2 x i140> 888 if (LK.first == TypeExpandInteger) 889 return LegalizeKind(TypeSplitVector, 890 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 891 892 // Promote the integer element types until a legal vector type is found 893 // or until the element integer type is too big. If a legal type was not 894 // found, fallback to the usual mechanism of widening/splitting the 895 // vector. 896 EVT OldEltVT = EltVT; 897 while (true) { 898 // Increase the bitwidth of the element to the next pow-of-two 899 // (which is greater than 8 bits). 900 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 901 .getRoundIntegerType(Context); 902 903 // Stop trying when getting a non-simple element type. 904 // Note that vector elements may be greater than legal vector element 905 // types. Example: X86 XMM registers hold 64bit element on 32bit 906 // systems. 907 if (!EltVT.isSimple()) 908 break; 909 910 // Build a new vector type and check if it is legal. 911 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 912 // Found a legal promoted vector type. 913 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 914 return LegalizeKind(TypePromoteInteger, 915 EVT::getVectorVT(Context, EltVT, NumElts)); 916 } 917 918 // Reset the type to the unexpanded type if we did not find a legal vector 919 // type with a promoted vector element type. 920 EltVT = OldEltVT; 921 } 922 923 // Try to widen the vector until a legal type is found. 924 // If there is no wider legal type, split the vector. 925 while (true) { 926 // Round up to the next power of 2. 927 NumElts = (unsigned)NextPowerOf2(NumElts); 928 929 // If there is no simple vector type with this many elements then there 930 // cannot be a larger legal vector type. Note that this assumes that 931 // there are no skipped intermediate vector types in the simple types. 932 if (!EltVT.isSimple()) 933 break; 934 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 935 if (LargerVector == MVT()) 936 break; 937 938 // If this type is legal then widen the vector. 939 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 940 return LegalizeKind(TypeWidenVector, LargerVector); 941 } 942 943 // Widen odd vectors to next power of two. 944 if (!VT.isPow2VectorType()) { 945 EVT NVT = VT.getPow2VectorType(Context); 946 return LegalizeKind(TypeWidenVector, NVT); 947 } 948 949 // Vectors with illegal element types are expanded. 950 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 951 return LegalizeKind(TypeSplitVector, NVT); 952 } 953 954 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 955 unsigned &NumIntermediates, 956 MVT &RegisterVT, 957 TargetLoweringBase *TLI) { 958 // Figure out the right, legal destination reg to copy into. 959 unsigned NumElts = VT.getVectorNumElements(); 960 MVT EltTy = VT.getVectorElementType(); 961 962 unsigned NumVectorRegs = 1; 963 964 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 965 // could break down into LHS/RHS like LegalizeDAG does. 966 if (!isPowerOf2_32(NumElts)) { 967 NumVectorRegs = NumElts; 968 NumElts = 1; 969 } 970 971 // Divide the input until we get to a supported size. This will always 972 // end with a scalar if the target doesn't support vectors. 973 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 974 NumElts >>= 1; 975 NumVectorRegs <<= 1; 976 } 977 978 NumIntermediates = NumVectorRegs; 979 980 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 981 if (!TLI->isTypeLegal(NewVT)) 982 NewVT = EltTy; 983 IntermediateVT = NewVT; 984 985 unsigned NewVTSize = NewVT.getSizeInBits(); 986 987 // Convert sizes such as i33 to i64. 988 if (!isPowerOf2_32(NewVTSize)) 989 NewVTSize = NextPowerOf2(NewVTSize); 990 991 MVT DestVT = TLI->getRegisterType(NewVT); 992 RegisterVT = DestVT; 993 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 994 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 995 996 // Otherwise, promotion or legal types use the same number of registers as 997 // the vector decimated to the appropriate level. 998 return NumVectorRegs; 999 } 1000 1001 /// isLegalRC - Return true if the value types that can be represented by the 1002 /// specified register class are all legal. 1003 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 1004 const TargetRegisterClass &RC) const { 1005 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 1006 if (isTypeLegal(*I)) 1007 return true; 1008 return false; 1009 } 1010 1011 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1012 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1013 MachineBasicBlock * 1014 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 1015 MachineBasicBlock *MBB) const { 1016 MachineInstr *MI = &InitialMI; 1017 MachineFunction &MF = *MI->getMF(); 1018 MachineFrameInfo &MFI = MF.getFrameInfo(); 1019 1020 // We're handling multiple types of operands here: 1021 // PATCHPOINT MetaArgs - live-in, read only, direct 1022 // STATEPOINT Deopt Spill - live-through, read only, indirect 1023 // STATEPOINT Deopt Alloca - live-through, read only, direct 1024 // (We're currently conservative and mark the deopt slots read/write in 1025 // practice.) 1026 // STATEPOINT GC Spill - live-through, read/write, indirect 1027 // STATEPOINT GC Alloca - live-through, read/write, direct 1028 // The live-in vs live-through is handled already (the live through ones are 1029 // all stack slots), but we need to handle the different type of stackmap 1030 // operands and memory effects here. 1031 1032 // MI changes inside this loop as we grow operands. 1033 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1034 MachineOperand &MO = MI->getOperand(OperIdx); 1035 if (!MO.isFI()) 1036 continue; 1037 1038 // foldMemoryOperand builds a new MI after replacing a single FI operand 1039 // with the canonical set of five x86 addressing-mode operands. 1040 int FI = MO.getIndex(); 1041 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1042 1043 // Copy operands before the frame-index. 1044 for (unsigned i = 0; i < OperIdx; ++i) 1045 MIB.add(MI->getOperand(i)); 1046 // Add frame index operands recognized by stackmaps.cpp 1047 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1048 // indirect-mem-ref tag, size, #FI, offset. 1049 // Used for spills inserted by StatepointLowering. This codepath is not 1050 // used for patchpoints/stackmaps at all, for these spilling is done via 1051 // foldMemoryOperand callback only. 1052 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1053 MIB.addImm(StackMaps::IndirectMemRefOp); 1054 MIB.addImm(MFI.getObjectSize(FI)); 1055 MIB.add(MI->getOperand(OperIdx)); 1056 MIB.addImm(0); 1057 } else { 1058 // direct-mem-ref tag, #FI, offset. 1059 // Used by patchpoint, and direct alloca arguments to statepoints 1060 MIB.addImm(StackMaps::DirectMemRefOp); 1061 MIB.add(MI->getOperand(OperIdx)); 1062 MIB.addImm(0); 1063 } 1064 // Copy the operands after the frame index. 1065 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1066 MIB.add(MI->getOperand(i)); 1067 1068 // Inherit previous memory operands. 1069 MIB.cloneMemRefs(*MI); 1070 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1071 1072 // Add a new memory operand for this FI. 1073 assert(MFI.getObjectOffset(FI) != -1); 1074 1075 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1076 // PATCHPOINT should be updated to do the same. (TODO) 1077 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1078 auto Flags = MachineMemOperand::MOLoad; 1079 MachineMemOperand *MMO = MF.getMachineMemOperand( 1080 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1081 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1082 MIB->addMemOperand(MF, MMO); 1083 } 1084 1085 // Replace the instruction and update the operand index. 1086 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1087 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1088 MI->eraseFromParent(); 1089 MI = MIB; 1090 } 1091 return MBB; 1092 } 1093 1094 MachineBasicBlock * 1095 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1096 MachineBasicBlock *MBB) const { 1097 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1098 "Called emitXRayCustomEvent on the wrong MI!"); 1099 auto &MF = *MI.getMF(); 1100 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1101 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1102 MIB.add(MI.getOperand(OpIdx)); 1103 1104 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1105 MI.eraseFromParent(); 1106 return MBB; 1107 } 1108 1109 MachineBasicBlock * 1110 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1111 MachineBasicBlock *MBB) const { 1112 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1113 "Called emitXRayTypedEvent on the wrong MI!"); 1114 auto &MF = *MI.getMF(); 1115 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1116 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1117 MIB.add(MI.getOperand(OpIdx)); 1118 1119 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1120 MI.eraseFromParent(); 1121 return MBB; 1122 } 1123 1124 /// findRepresentativeClass - Return the largest legal super-reg register class 1125 /// of the register class for the specified type and its associated "cost". 1126 // This function is in TargetLowering because it uses RegClassForVT which would 1127 // need to be moved to TargetRegisterInfo and would necessitate moving 1128 // isTypeLegal over as well - a massive change that would just require 1129 // TargetLowering having a TargetRegisterInfo class member that it would use. 1130 std::pair<const TargetRegisterClass *, uint8_t> 1131 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1132 MVT VT) const { 1133 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1134 if (!RC) 1135 return std::make_pair(RC, 0); 1136 1137 // Compute the set of all super-register classes. 1138 BitVector SuperRegRC(TRI->getNumRegClasses()); 1139 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1140 SuperRegRC.setBitsInMask(RCI.getMask()); 1141 1142 // Find the first legal register class with the largest spill size. 1143 const TargetRegisterClass *BestRC = RC; 1144 for (unsigned i : SuperRegRC.set_bits()) { 1145 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1146 // We want the largest possible spill size. 1147 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1148 continue; 1149 if (!isLegalRC(*TRI, *SuperRC)) 1150 continue; 1151 BestRC = SuperRC; 1152 } 1153 return std::make_pair(BestRC, 1); 1154 } 1155 1156 /// computeRegisterProperties - Once all of the register classes are added, 1157 /// this allows us to compute derived properties we expose. 1158 void TargetLoweringBase::computeRegisterProperties( 1159 const TargetRegisterInfo *TRI) { 1160 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1161 "Too many value types for ValueTypeActions to hold!"); 1162 1163 // Everything defaults to needing one register. 1164 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1165 NumRegistersForVT[i] = 1; 1166 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1167 } 1168 // ...except isVoid, which doesn't need any registers. 1169 NumRegistersForVT[MVT::isVoid] = 0; 1170 1171 // Find the largest integer register class. 1172 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1173 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1174 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1175 1176 // Every integer value type larger than this largest register takes twice as 1177 // many registers to represent as the previous ValueType. 1178 for (unsigned ExpandedReg = LargestIntReg + 1; 1179 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1180 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1181 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1182 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1183 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1184 TypeExpandInteger); 1185 } 1186 1187 // Inspect all of the ValueType's smaller than the largest integer 1188 // register to see which ones need promotion. 1189 unsigned LegalIntReg = LargestIntReg; 1190 for (unsigned IntReg = LargestIntReg - 1; 1191 IntReg >= (unsigned)MVT::i1; --IntReg) { 1192 MVT IVT = (MVT::SimpleValueType)IntReg; 1193 if (isTypeLegal(IVT)) { 1194 LegalIntReg = IntReg; 1195 } else { 1196 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1197 (MVT::SimpleValueType)LegalIntReg; 1198 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1199 } 1200 } 1201 1202 // ppcf128 type is really two f64's. 1203 if (!isTypeLegal(MVT::ppcf128)) { 1204 if (isTypeLegal(MVT::f64)) { 1205 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1206 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1207 TransformToType[MVT::ppcf128] = MVT::f64; 1208 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1209 } else { 1210 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1211 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1212 TransformToType[MVT::ppcf128] = MVT::i128; 1213 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1214 } 1215 } 1216 1217 // Decide how to handle f128. If the target does not have native f128 support, 1218 // expand it to i128 and we will be generating soft float library calls. 1219 if (!isTypeLegal(MVT::f128)) { 1220 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1221 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1222 TransformToType[MVT::f128] = MVT::i128; 1223 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1224 } 1225 1226 // Decide how to handle f64. If the target does not have native f64 support, 1227 // expand it to i64 and we will be generating soft float library calls. 1228 if (!isTypeLegal(MVT::f64)) { 1229 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1230 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1231 TransformToType[MVT::f64] = MVT::i64; 1232 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1233 } 1234 1235 // Decide how to handle f32. If the target does not have native f32 support, 1236 // expand it to i32 and we will be generating soft float library calls. 1237 if (!isTypeLegal(MVT::f32)) { 1238 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1239 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1240 TransformToType[MVT::f32] = MVT::i32; 1241 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1242 } 1243 1244 // Decide how to handle f16. If the target does not have native f16 support, 1245 // promote it to f32, because there are no f16 library calls (except for 1246 // conversions). 1247 if (!isTypeLegal(MVT::f16)) { 1248 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1249 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1250 TransformToType[MVT::f16] = MVT::f32; 1251 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1252 } 1253 1254 // Loop over all of the vector value types to see which need transformations. 1255 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1256 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1257 MVT VT = (MVT::SimpleValueType) i; 1258 if (isTypeLegal(VT)) 1259 continue; 1260 1261 MVT EltVT = VT.getVectorElementType(); 1262 unsigned NElts = VT.getVectorNumElements(); 1263 bool IsLegalWiderType = false; 1264 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1265 switch (PreferredAction) { 1266 case TypePromoteInteger: 1267 // Try to promote the elements of integer vectors. If no legal 1268 // promotion was found, fall through to the widen-vector method. 1269 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1270 MVT SVT = (MVT::SimpleValueType) nVT; 1271 // Promote vectors of integers to vectors with the same number 1272 // of elements, with a wider element type. 1273 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1274 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1275 TransformToType[i] = SVT; 1276 RegisterTypeForVT[i] = SVT; 1277 NumRegistersForVT[i] = 1; 1278 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1279 IsLegalWiderType = true; 1280 break; 1281 } 1282 } 1283 if (IsLegalWiderType) 1284 break; 1285 LLVM_FALLTHROUGH; 1286 1287 case TypeWidenVector: 1288 if (isPowerOf2_32(NElts)) { 1289 // Try to widen the vector. 1290 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1291 MVT SVT = (MVT::SimpleValueType) nVT; 1292 if (SVT.getVectorElementType() == EltVT 1293 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1294 TransformToType[i] = SVT; 1295 RegisterTypeForVT[i] = SVT; 1296 NumRegistersForVT[i] = 1; 1297 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1298 IsLegalWiderType = true; 1299 break; 1300 } 1301 } 1302 if (IsLegalWiderType) 1303 break; 1304 } else { 1305 // Only widen to the next power of 2 to keep consistency with EVT. 1306 MVT NVT = VT.getPow2VectorType(); 1307 if (isTypeLegal(NVT)) { 1308 TransformToType[i] = NVT; 1309 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1310 RegisterTypeForVT[i] = NVT; 1311 NumRegistersForVT[i] = 1; 1312 break; 1313 } 1314 } 1315 LLVM_FALLTHROUGH; 1316 1317 case TypeSplitVector: 1318 case TypeScalarizeVector: { 1319 MVT IntermediateVT; 1320 MVT RegisterVT; 1321 unsigned NumIntermediates; 1322 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1323 NumIntermediates, RegisterVT, this); 1324 RegisterTypeForVT[i] = RegisterVT; 1325 1326 MVT NVT = VT.getPow2VectorType(); 1327 if (NVT == VT) { 1328 // Type is already a power of 2. The default action is to split. 1329 TransformToType[i] = MVT::Other; 1330 if (PreferredAction == TypeScalarizeVector) 1331 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1332 else if (PreferredAction == TypeSplitVector) 1333 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1334 else 1335 // Set type action according to the number of elements. 1336 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1337 : TypeSplitVector); 1338 } else { 1339 TransformToType[i] = NVT; 1340 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1341 } 1342 break; 1343 } 1344 default: 1345 llvm_unreachable("Unknown vector legalization action!"); 1346 } 1347 } 1348 1349 // Determine the 'representative' register class for each value type. 1350 // An representative register class is the largest (meaning one which is 1351 // not a sub-register class / subreg register class) legal register class for 1352 // a group of value types. For example, on i386, i8, i16, and i32 1353 // representative would be GR32; while on x86_64 it's GR64. 1354 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1355 const TargetRegisterClass* RRC; 1356 uint8_t Cost; 1357 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1358 RepRegClassForVT[i] = RRC; 1359 RepRegClassCostForVT[i] = Cost; 1360 } 1361 } 1362 1363 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1364 EVT VT) const { 1365 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1366 return getPointerTy(DL).SimpleTy; 1367 } 1368 1369 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1370 return MVT::i32; // return the default value 1371 } 1372 1373 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1374 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1375 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1376 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1377 /// 1378 /// This method returns the number of registers needed, and the VT for each 1379 /// register. It also returns the VT and quantity of the intermediate values 1380 /// before they are promoted/expanded. 1381 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1382 EVT &IntermediateVT, 1383 unsigned &NumIntermediates, 1384 MVT &RegisterVT) const { 1385 unsigned NumElts = VT.getVectorNumElements(); 1386 1387 // If there is a wider vector type with the same element type as this one, 1388 // or a promoted vector type that has the same number of elements which 1389 // are wider, then we should convert to that legal vector type. 1390 // This handles things like <2 x float> -> <4 x float> and 1391 // <4 x i1> -> <4 x i32>. 1392 LegalizeTypeAction TA = getTypeAction(Context, VT); 1393 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1394 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1395 if (isTypeLegal(RegisterEVT)) { 1396 IntermediateVT = RegisterEVT; 1397 RegisterVT = RegisterEVT.getSimpleVT(); 1398 NumIntermediates = 1; 1399 return 1; 1400 } 1401 } 1402 1403 // Figure out the right, legal destination reg to copy into. 1404 EVT EltTy = VT.getVectorElementType(); 1405 1406 unsigned NumVectorRegs = 1; 1407 1408 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1409 // could break down into LHS/RHS like LegalizeDAG does. 1410 if (!isPowerOf2_32(NumElts)) { 1411 NumVectorRegs = NumElts; 1412 NumElts = 1; 1413 } 1414 1415 // Divide the input until we get to a supported size. This will always 1416 // end with a scalar if the target doesn't support vectors. 1417 while (NumElts > 1 && !isTypeLegal( 1418 EVT::getVectorVT(Context, EltTy, NumElts))) { 1419 NumElts >>= 1; 1420 NumVectorRegs <<= 1; 1421 } 1422 1423 NumIntermediates = NumVectorRegs; 1424 1425 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1426 if (!isTypeLegal(NewVT)) 1427 NewVT = EltTy; 1428 IntermediateVT = NewVT; 1429 1430 MVT DestVT = getRegisterType(Context, NewVT); 1431 RegisterVT = DestVT; 1432 unsigned NewVTSize = NewVT.getSizeInBits(); 1433 1434 // Convert sizes such as i33 to i64. 1435 if (!isPowerOf2_32(NewVTSize)) 1436 NewVTSize = NextPowerOf2(NewVTSize); 1437 1438 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1439 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1440 1441 // Otherwise, promotion or legal types use the same number of registers as 1442 // the vector decimated to the appropriate level. 1443 return NumVectorRegs; 1444 } 1445 1446 /// Get the EVTs and ArgFlags collections that represent the legalized return 1447 /// type of the given function. This does not require a DAG or a return value, 1448 /// and is suitable for use before any DAGs for the function are constructed. 1449 /// TODO: Move this out of TargetLowering.cpp. 1450 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1451 AttributeList attr, 1452 SmallVectorImpl<ISD::OutputArg> &Outs, 1453 const TargetLowering &TLI, const DataLayout &DL) { 1454 SmallVector<EVT, 4> ValueVTs; 1455 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1456 unsigned NumValues = ValueVTs.size(); 1457 if (NumValues == 0) return; 1458 1459 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1460 EVT VT = ValueVTs[j]; 1461 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1462 1463 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1464 ExtendKind = ISD::SIGN_EXTEND; 1465 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1466 ExtendKind = ISD::ZERO_EXTEND; 1467 1468 // FIXME: C calling convention requires the return type to be promoted to 1469 // at least 32-bit. But this is not necessary for non-C calling 1470 // conventions. The frontend should mark functions whose return values 1471 // require promoting with signext or zeroext attributes. 1472 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1473 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1474 if (VT.bitsLT(MinVT)) 1475 VT = MinVT; 1476 } 1477 1478 unsigned NumParts = 1479 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1480 MVT PartVT = 1481 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1482 1483 // 'inreg' on function refers to return value 1484 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1485 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1486 Flags.setInReg(); 1487 1488 // Propagate extension type if any 1489 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1490 Flags.setSExt(); 1491 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1492 Flags.setZExt(); 1493 1494 for (unsigned i = 0; i < NumParts; ++i) 1495 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1496 } 1497 } 1498 1499 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1500 /// function arguments in the caller parameter area. This is the actual 1501 /// alignment, not its logarithm. 1502 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1503 const DataLayout &DL) const { 1504 return DL.getABITypeAlignment(Ty); 1505 } 1506 1507 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1508 const DataLayout &DL, EVT VT, 1509 unsigned AddrSpace, 1510 unsigned Alignment, 1511 MachineMemOperand::Flags Flags, 1512 bool *Fast) const { 1513 // Check if the specified alignment is sufficient based on the data layout. 1514 // TODO: While using the data layout works in practice, a better solution 1515 // would be to implement this check directly (make this a virtual function). 1516 // For example, the ABI alignment may change based on software platform while 1517 // this function should only be affected by hardware implementation. 1518 Type *Ty = VT.getTypeForEVT(Context); 1519 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1520 // Assume that an access that meets the ABI-specified alignment is fast. 1521 if (Fast != nullptr) 1522 *Fast = true; 1523 return true; 1524 } 1525 1526 // This is a misaligned access. 1527 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1528 } 1529 1530 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1531 const DataLayout &DL, EVT VT, 1532 const MachineMemOperand &MMO, 1533 bool *Fast) const { 1534 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1535 MMO.getAlignment(), MMO.getFlags(), Fast); 1536 } 1537 1538 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1539 return BranchProbability(MinPercentageForPredictableBranch, 100); 1540 } 1541 1542 //===----------------------------------------------------------------------===// 1543 // TargetTransformInfo Helpers 1544 //===----------------------------------------------------------------------===// 1545 1546 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1547 enum InstructionOpcodes { 1548 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1549 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1550 #include "llvm/IR/Instruction.def" 1551 }; 1552 switch (static_cast<InstructionOpcodes>(Opcode)) { 1553 case Ret: return 0; 1554 case Br: return 0; 1555 case Switch: return 0; 1556 case IndirectBr: return 0; 1557 case Invoke: return 0; 1558 case CallBr: return 0; 1559 case Resume: return 0; 1560 case Unreachable: return 0; 1561 case CleanupRet: return 0; 1562 case CatchRet: return 0; 1563 case CatchPad: return 0; 1564 case CatchSwitch: return 0; 1565 case CleanupPad: return 0; 1566 case FNeg: return ISD::FNEG; 1567 case Add: return ISD::ADD; 1568 case FAdd: return ISD::FADD; 1569 case Sub: return ISD::SUB; 1570 case FSub: return ISD::FSUB; 1571 case Mul: return ISD::MUL; 1572 case FMul: return ISD::FMUL; 1573 case UDiv: return ISD::UDIV; 1574 case SDiv: return ISD::SDIV; 1575 case FDiv: return ISD::FDIV; 1576 case URem: return ISD::UREM; 1577 case SRem: return ISD::SREM; 1578 case FRem: return ISD::FREM; 1579 case Shl: return ISD::SHL; 1580 case LShr: return ISD::SRL; 1581 case AShr: return ISD::SRA; 1582 case And: return ISD::AND; 1583 case Or: return ISD::OR; 1584 case Xor: return ISD::XOR; 1585 case Alloca: return 0; 1586 case Load: return ISD::LOAD; 1587 case Store: return ISD::STORE; 1588 case GetElementPtr: return 0; 1589 case Fence: return 0; 1590 case AtomicCmpXchg: return 0; 1591 case AtomicRMW: return 0; 1592 case Trunc: return ISD::TRUNCATE; 1593 case ZExt: return ISD::ZERO_EXTEND; 1594 case SExt: return ISD::SIGN_EXTEND; 1595 case FPToUI: return ISD::FP_TO_UINT; 1596 case FPToSI: return ISD::FP_TO_SINT; 1597 case UIToFP: return ISD::UINT_TO_FP; 1598 case SIToFP: return ISD::SINT_TO_FP; 1599 case FPTrunc: return ISD::FP_ROUND; 1600 case FPExt: return ISD::FP_EXTEND; 1601 case PtrToInt: return ISD::BITCAST; 1602 case IntToPtr: return ISD::BITCAST; 1603 case BitCast: return ISD::BITCAST; 1604 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1605 case ICmp: return ISD::SETCC; 1606 case FCmp: return ISD::SETCC; 1607 case PHI: return 0; 1608 case Call: return 0; 1609 case Select: return ISD::SELECT; 1610 case UserOp1: return 0; 1611 case UserOp2: return 0; 1612 case VAArg: return 0; 1613 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1614 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1615 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1616 case ExtractValue: return ISD::MERGE_VALUES; 1617 case InsertValue: return ISD::MERGE_VALUES; 1618 case LandingPad: return 0; 1619 } 1620 1621 llvm_unreachable("Unknown instruction type encountered!"); 1622 } 1623 1624 std::pair<int, MVT> 1625 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1626 Type *Ty) const { 1627 LLVMContext &C = Ty->getContext(); 1628 EVT MTy = getValueType(DL, Ty); 1629 1630 int Cost = 1; 1631 // We keep legalizing the type until we find a legal kind. We assume that 1632 // the only operation that costs anything is the split. After splitting 1633 // we need to handle two types. 1634 while (true) { 1635 LegalizeKind LK = getTypeConversion(C, MTy); 1636 1637 if (LK.first == TypeLegal) 1638 return std::make_pair(Cost, MTy.getSimpleVT()); 1639 1640 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1641 Cost *= 2; 1642 1643 // Do not loop with f128 type. 1644 if (MTy == LK.second) 1645 return std::make_pair(Cost, MTy.getSimpleVT()); 1646 1647 // Keep legalizing the type. 1648 MTy = LK.second; 1649 } 1650 } 1651 1652 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1653 bool UseTLS) const { 1654 // compiler-rt provides a variable with a magic name. Targets that do not 1655 // link with compiler-rt may also provide such a variable. 1656 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1657 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1658 auto UnsafeStackPtr = 1659 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1660 1661 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1662 1663 if (!UnsafeStackPtr) { 1664 auto TLSModel = UseTLS ? 1665 GlobalValue::InitialExecTLSModel : 1666 GlobalValue::NotThreadLocal; 1667 // The global variable is not defined yet, define it ourselves. 1668 // We use the initial-exec TLS model because we do not support the 1669 // variable living anywhere other than in the main executable. 1670 UnsafeStackPtr = new GlobalVariable( 1671 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1672 UnsafeStackPtrVar, nullptr, TLSModel); 1673 } else { 1674 // The variable exists, check its type and attributes. 1675 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1676 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1677 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1678 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1679 (UseTLS ? "" : "not ") + "be thread-local"); 1680 } 1681 return UnsafeStackPtr; 1682 } 1683 1684 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1685 if (!TM.getTargetTriple().isAndroid()) 1686 return getDefaultSafeStackPointerLocation(IRB, true); 1687 1688 // Android provides a libc function to retrieve the address of the current 1689 // thread's unsafe stack pointer. 1690 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1691 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1692 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1693 StackPtrTy->getPointerTo(0)); 1694 return IRB.CreateCall(Fn); 1695 } 1696 1697 //===----------------------------------------------------------------------===// 1698 // Loop Strength Reduction hooks 1699 //===----------------------------------------------------------------------===// 1700 1701 /// isLegalAddressingMode - Return true if the addressing mode represented 1702 /// by AM is legal for this target, for a load/store of the specified type. 1703 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1704 const AddrMode &AM, Type *Ty, 1705 unsigned AS, Instruction *I) const { 1706 // The default implementation of this implements a conservative RISCy, r+r and 1707 // r+i addr mode. 1708 1709 // Allows a sign-extended 16-bit immediate field. 1710 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1711 return false; 1712 1713 // No global is ever allowed as a base. 1714 if (AM.BaseGV) 1715 return false; 1716 1717 // Only support r+r, 1718 switch (AM.Scale) { 1719 case 0: // "r+i" or just "i", depending on HasBaseReg. 1720 break; 1721 case 1: 1722 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1723 return false; 1724 // Otherwise we have r+r or r+i. 1725 break; 1726 case 2: 1727 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1728 return false; 1729 // Allow 2*r as r+r. 1730 break; 1731 default: // Don't allow n * r 1732 return false; 1733 } 1734 1735 return true; 1736 } 1737 1738 //===----------------------------------------------------------------------===// 1739 // Stack Protector 1740 //===----------------------------------------------------------------------===// 1741 1742 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1743 // so that SelectionDAG handle SSP. 1744 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1745 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1746 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1747 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1748 return M.getOrInsertGlobal("__guard_local", PtrTy); 1749 } 1750 return nullptr; 1751 } 1752 1753 // Currently only support "standard" __stack_chk_guard. 1754 // TODO: add LOAD_STACK_GUARD support. 1755 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1756 if (!M.getNamedValue("__stack_chk_guard")) 1757 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1758 GlobalVariable::ExternalLinkage, 1759 nullptr, "__stack_chk_guard"); 1760 } 1761 1762 // Currently only support "standard" __stack_chk_guard. 1763 // TODO: add LOAD_STACK_GUARD support. 1764 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1765 return M.getNamedValue("__stack_chk_guard"); 1766 } 1767 1768 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1769 return nullptr; 1770 } 1771 1772 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1773 return MinimumJumpTableEntries; 1774 } 1775 1776 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1777 MinimumJumpTableEntries = Val; 1778 } 1779 1780 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1781 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1782 } 1783 1784 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1785 return MaximumJumpTableSize; 1786 } 1787 1788 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1789 MaximumJumpTableSize = Val; 1790 } 1791 1792 //===----------------------------------------------------------------------===// 1793 // Reciprocal Estimates 1794 //===----------------------------------------------------------------------===// 1795 1796 /// Get the reciprocal estimate attribute string for a function that will 1797 /// override the target defaults. 1798 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1799 const Function &F = MF.getFunction(); 1800 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1801 } 1802 1803 /// Construct a string for the given reciprocal operation of the given type. 1804 /// This string should match the corresponding option to the front-end's 1805 /// "-mrecip" flag assuming those strings have been passed through in an 1806 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1807 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1808 std::string Name = VT.isVector() ? "vec-" : ""; 1809 1810 Name += IsSqrt ? "sqrt" : "div"; 1811 1812 // TODO: Handle "half" or other float types? 1813 if (VT.getScalarType() == MVT::f64) { 1814 Name += "d"; 1815 } else { 1816 assert(VT.getScalarType() == MVT::f32 && 1817 "Unexpected FP type for reciprocal estimate"); 1818 Name += "f"; 1819 } 1820 1821 return Name; 1822 } 1823 1824 /// Return the character position and value (a single numeric character) of a 1825 /// customized refinement operation in the input string if it exists. Return 1826 /// false if there is no customized refinement step count. 1827 static bool parseRefinementStep(StringRef In, size_t &Position, 1828 uint8_t &Value) { 1829 const char RefStepToken = ':'; 1830 Position = In.find(RefStepToken); 1831 if (Position == StringRef::npos) 1832 return false; 1833 1834 StringRef RefStepString = In.substr(Position + 1); 1835 // Allow exactly one numeric character for the additional refinement 1836 // step parameter. 1837 if (RefStepString.size() == 1) { 1838 char RefStepChar = RefStepString[0]; 1839 if (RefStepChar >= '0' && RefStepChar <= '9') { 1840 Value = RefStepChar - '0'; 1841 return true; 1842 } 1843 } 1844 report_fatal_error("Invalid refinement step for -recip."); 1845 } 1846 1847 /// For the input attribute string, return one of the ReciprocalEstimate enum 1848 /// status values (enabled, disabled, or not specified) for this operation on 1849 /// the specified data type. 1850 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1851 if (Override.empty()) 1852 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1853 1854 SmallVector<StringRef, 4> OverrideVector; 1855 Override.split(OverrideVector, ','); 1856 unsigned NumArgs = OverrideVector.size(); 1857 1858 // Check if "all", "none", or "default" was specified. 1859 if (NumArgs == 1) { 1860 // Look for an optional setting of the number of refinement steps needed 1861 // for this type of reciprocal operation. 1862 size_t RefPos; 1863 uint8_t RefSteps; 1864 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1865 // Split the string for further processing. 1866 Override = Override.substr(0, RefPos); 1867 } 1868 1869 // All reciprocal types are enabled. 1870 if (Override == "all") 1871 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1872 1873 // All reciprocal types are disabled. 1874 if (Override == "none") 1875 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1876 1877 // Target defaults for enablement are used. 1878 if (Override == "default") 1879 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1880 } 1881 1882 // The attribute string may omit the size suffix ('f'/'d'). 1883 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1884 std::string VTNameNoSize = VTName; 1885 VTNameNoSize.pop_back(); 1886 static const char DisabledPrefix = '!'; 1887 1888 for (StringRef RecipType : OverrideVector) { 1889 size_t RefPos; 1890 uint8_t RefSteps; 1891 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1892 RecipType = RecipType.substr(0, RefPos); 1893 1894 // Ignore the disablement token for string matching. 1895 bool IsDisabled = RecipType[0] == DisabledPrefix; 1896 if (IsDisabled) 1897 RecipType = RecipType.substr(1); 1898 1899 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1900 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1901 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1902 } 1903 1904 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1905 } 1906 1907 /// For the input attribute string, return the customized refinement step count 1908 /// for this operation on the specified data type. If the step count does not 1909 /// exist, return the ReciprocalEstimate enum value for unspecified. 1910 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1911 if (Override.empty()) 1912 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1913 1914 SmallVector<StringRef, 4> OverrideVector; 1915 Override.split(OverrideVector, ','); 1916 unsigned NumArgs = OverrideVector.size(); 1917 1918 // Check if "all", "default", or "none" was specified. 1919 if (NumArgs == 1) { 1920 // Look for an optional setting of the number of refinement steps needed 1921 // for this type of reciprocal operation. 1922 size_t RefPos; 1923 uint8_t RefSteps; 1924 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1925 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1926 1927 // Split the string for further processing. 1928 Override = Override.substr(0, RefPos); 1929 assert(Override != "none" && 1930 "Disabled reciprocals, but specifed refinement steps?"); 1931 1932 // If this is a general override, return the specified number of steps. 1933 if (Override == "all" || Override == "default") 1934 return RefSteps; 1935 } 1936 1937 // The attribute string may omit the size suffix ('f'/'d'). 1938 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1939 std::string VTNameNoSize = VTName; 1940 VTNameNoSize.pop_back(); 1941 1942 for (StringRef RecipType : OverrideVector) { 1943 size_t RefPos; 1944 uint8_t RefSteps; 1945 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1946 continue; 1947 1948 RecipType = RecipType.substr(0, RefPos); 1949 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1950 return RefSteps; 1951 } 1952 1953 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1954 } 1955 1956 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1957 MachineFunction &MF) const { 1958 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1959 } 1960 1961 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1962 MachineFunction &MF) const { 1963 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1964 } 1965 1966 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1967 MachineFunction &MF) const { 1968 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1969 } 1970 1971 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1972 MachineFunction &MF) const { 1973 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1974 } 1975 1976 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1977 MF.getRegInfo().freezeReservedRegs(MF); 1978 } 1979