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 // FIXME: This option is only to test if the strict fp operation processed 92 // correctly by preventing mutating strict fp operation to normal fp operation 93 // during development. When the backend supports strict float operation, this 94 // option will be meaningless. 95 static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation", 96 cl::desc("Don't mutate strict-float node to a legalize node"), 97 cl::init(false), cl::Hidden); 98 99 static bool darwinHasSinCos(const Triple &TT) { 100 assert(TT.isOSDarwin() && "should be called with darwin triple"); 101 // Don't bother with 32 bit x86. 102 if (TT.getArch() == Triple::x86) 103 return false; 104 // Macos < 10.9 has no sincos_stret. 105 if (TT.isMacOSX()) 106 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 107 // iOS < 7.0 has no sincos_stret. 108 if (TT.isiOS()) 109 return !TT.isOSVersionLT(7, 0); 110 // Any other darwin such as WatchOS/TvOS is new enough. 111 return true; 112 } 113 114 // Although this default value is arbitrary, it is not random. It is assumed 115 // that a condition that evaluates the same way by a higher percentage than this 116 // is best represented as control flow. Therefore, the default value N should be 117 // set such that the win from N% correct executions is greater than the loss 118 // from (100 - N)% mispredicted executions for the majority of intended targets. 119 static cl::opt<int> MinPercentageForPredictableBranch( 120 "min-predictable-branch", cl::init(99), 121 cl::desc("Minimum percentage (0-100) that a condition must be either true " 122 "or false to assume that the condition is predictable"), 123 cl::Hidden); 124 125 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 126 #define HANDLE_LIBCALL(code, name) \ 127 setLibcallName(RTLIB::code, name); 128 #include "llvm/IR/RuntimeLibcalls.def" 129 #undef HANDLE_LIBCALL 130 // Initialize calling conventions to their default. 131 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 132 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 133 134 // For IEEE quad-precision libcall names, PPC uses "kf" instead of "tf". 135 if (TT.getArch() == Triple::ppc || TT.isPPC64()) { 136 setLibcallName(RTLIB::ADD_F128, "__addkf3"); 137 setLibcallName(RTLIB::SUB_F128, "__subkf3"); 138 setLibcallName(RTLIB::MUL_F128, "__mulkf3"); 139 setLibcallName(RTLIB::DIV_F128, "__divkf3"); 140 setLibcallName(RTLIB::FPEXT_F32_F128, "__extendsfkf2"); 141 setLibcallName(RTLIB::FPEXT_F64_F128, "__extenddfkf2"); 142 setLibcallName(RTLIB::FPROUND_F128_F32, "__trunckfsf2"); 143 setLibcallName(RTLIB::FPROUND_F128_F64, "__trunckfdf2"); 144 setLibcallName(RTLIB::FPTOSINT_F128_I32, "__fixkfsi"); 145 setLibcallName(RTLIB::FPTOSINT_F128_I64, "__fixkfdi"); 146 setLibcallName(RTLIB::FPTOUINT_F128_I32, "__fixunskfsi"); 147 setLibcallName(RTLIB::FPTOUINT_F128_I64, "__fixunskfdi"); 148 setLibcallName(RTLIB::SINTTOFP_I32_F128, "__floatsikf"); 149 setLibcallName(RTLIB::SINTTOFP_I64_F128, "__floatdikf"); 150 setLibcallName(RTLIB::UINTTOFP_I32_F128, "__floatunsikf"); 151 setLibcallName(RTLIB::UINTTOFP_I64_F128, "__floatundikf"); 152 setLibcallName(RTLIB::OEQ_F128, "__eqkf2"); 153 setLibcallName(RTLIB::UNE_F128, "__nekf2"); 154 setLibcallName(RTLIB::OGE_F128, "__gekf2"); 155 setLibcallName(RTLIB::OLT_F128, "__ltkf2"); 156 setLibcallName(RTLIB::OLE_F128, "__lekf2"); 157 setLibcallName(RTLIB::OGT_F128, "__gtkf2"); 158 setLibcallName(RTLIB::UO_F128, "__unordkf2"); 159 setLibcallName(RTLIB::O_F128, "__unordkf2"); 160 } 161 162 // A few names are different on particular architectures or environments. 163 if (TT.isOSDarwin()) { 164 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 165 // of the gnueabi-style __gnu_*_ieee. 166 // FIXME: What about other targets? 167 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 168 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 169 170 // Some darwins have an optimized __bzero/bzero function. 171 switch (TT.getArch()) { 172 case Triple::x86: 173 case Triple::x86_64: 174 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 175 setLibcallName(RTLIB::BZERO, "__bzero"); 176 break; 177 case Triple::aarch64: 178 case Triple::aarch64_32: 179 setLibcallName(RTLIB::BZERO, "bzero"); 180 break; 181 default: 182 break; 183 } 184 185 if (darwinHasSinCos(TT)) { 186 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 187 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 188 if (TT.isWatchABI()) { 189 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 190 CallingConv::ARM_AAPCS_VFP); 191 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 192 CallingConv::ARM_AAPCS_VFP); 193 } 194 } 195 } else { 196 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 197 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 198 } 199 200 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 201 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 202 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 203 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 204 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 205 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 206 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 207 } 208 209 if (TT.isPS4CPU()) { 210 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 211 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 212 } 213 214 if (TT.isOSOpenBSD()) { 215 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 216 } 217 } 218 219 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 220 /// UNKNOWN_LIBCALL if there is none. 221 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 222 if (OpVT == MVT::f16) { 223 if (RetVT == MVT::f32) 224 return FPEXT_F16_F32; 225 } else if (OpVT == MVT::f32) { 226 if (RetVT == MVT::f64) 227 return FPEXT_F32_F64; 228 if (RetVT == MVT::f128) 229 return FPEXT_F32_F128; 230 if (RetVT == MVT::ppcf128) 231 return FPEXT_F32_PPCF128; 232 } else if (OpVT == MVT::f64) { 233 if (RetVT == MVT::f128) 234 return FPEXT_F64_F128; 235 else if (RetVT == MVT::ppcf128) 236 return FPEXT_F64_PPCF128; 237 } else if (OpVT == MVT::f80) { 238 if (RetVT == MVT::f128) 239 return FPEXT_F80_F128; 240 } 241 242 return UNKNOWN_LIBCALL; 243 } 244 245 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 246 /// UNKNOWN_LIBCALL if there is none. 247 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 248 if (RetVT == MVT::f16) { 249 if (OpVT == MVT::f32) 250 return FPROUND_F32_F16; 251 if (OpVT == MVT::f64) 252 return FPROUND_F64_F16; 253 if (OpVT == MVT::f80) 254 return FPROUND_F80_F16; 255 if (OpVT == MVT::f128) 256 return FPROUND_F128_F16; 257 if (OpVT == MVT::ppcf128) 258 return FPROUND_PPCF128_F16; 259 } else if (RetVT == MVT::f32) { 260 if (OpVT == MVT::f64) 261 return FPROUND_F64_F32; 262 if (OpVT == MVT::f80) 263 return FPROUND_F80_F32; 264 if (OpVT == MVT::f128) 265 return FPROUND_F128_F32; 266 if (OpVT == MVT::ppcf128) 267 return FPROUND_PPCF128_F32; 268 } else if (RetVT == MVT::f64) { 269 if (OpVT == MVT::f80) 270 return FPROUND_F80_F64; 271 if (OpVT == MVT::f128) 272 return FPROUND_F128_F64; 273 if (OpVT == MVT::ppcf128) 274 return FPROUND_PPCF128_F64; 275 } else if (RetVT == MVT::f80) { 276 if (OpVT == MVT::f128) 277 return FPROUND_F128_F80; 278 } 279 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOSINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOSINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOSINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOSINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOSINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOSINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOSINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOSINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOSINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOSINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOSINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOSINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOSINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOSINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOSINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::f32) { 329 if (RetVT == MVT::i32) 330 return FPTOUINT_F32_I32; 331 if (RetVT == MVT::i64) 332 return FPTOUINT_F32_I64; 333 if (RetVT == MVT::i128) 334 return FPTOUINT_F32_I128; 335 } else if (OpVT == MVT::f64) { 336 if (RetVT == MVT::i32) 337 return FPTOUINT_F64_I32; 338 if (RetVT == MVT::i64) 339 return FPTOUINT_F64_I64; 340 if (RetVT == MVT::i128) 341 return FPTOUINT_F64_I128; 342 } else if (OpVT == MVT::f80) { 343 if (RetVT == MVT::i32) 344 return FPTOUINT_F80_I32; 345 if (RetVT == MVT::i64) 346 return FPTOUINT_F80_I64; 347 if (RetVT == MVT::i128) 348 return FPTOUINT_F80_I128; 349 } else if (OpVT == MVT::f128) { 350 if (RetVT == MVT::i32) 351 return FPTOUINT_F128_I32; 352 if (RetVT == MVT::i64) 353 return FPTOUINT_F128_I64; 354 if (RetVT == MVT::i128) 355 return FPTOUINT_F128_I128; 356 } else if (OpVT == MVT::ppcf128) { 357 if (RetVT == MVT::i32) 358 return FPTOUINT_PPCF128_I32; 359 if (RetVT == MVT::i64) 360 return FPTOUINT_PPCF128_I64; 361 if (RetVT == MVT::i128) 362 return FPTOUINT_PPCF128_I128; 363 } 364 return UNKNOWN_LIBCALL; 365 } 366 367 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 368 /// UNKNOWN_LIBCALL if there is none. 369 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 370 if (OpVT == MVT::i32) { 371 if (RetVT == MVT::f32) 372 return SINTTOFP_I32_F32; 373 if (RetVT == MVT::f64) 374 return SINTTOFP_I32_F64; 375 if (RetVT == MVT::f80) 376 return SINTTOFP_I32_F80; 377 if (RetVT == MVT::f128) 378 return SINTTOFP_I32_F128; 379 if (RetVT == MVT::ppcf128) 380 return SINTTOFP_I32_PPCF128; 381 } else if (OpVT == MVT::i64) { 382 if (RetVT == MVT::f32) 383 return SINTTOFP_I64_F32; 384 if (RetVT == MVT::f64) 385 return SINTTOFP_I64_F64; 386 if (RetVT == MVT::f80) 387 return SINTTOFP_I64_F80; 388 if (RetVT == MVT::f128) 389 return SINTTOFP_I64_F128; 390 if (RetVT == MVT::ppcf128) 391 return SINTTOFP_I64_PPCF128; 392 } else if (OpVT == MVT::i128) { 393 if (RetVT == MVT::f32) 394 return SINTTOFP_I128_F32; 395 if (RetVT == MVT::f64) 396 return SINTTOFP_I128_F64; 397 if (RetVT == MVT::f80) 398 return SINTTOFP_I128_F80; 399 if (RetVT == MVT::f128) 400 return SINTTOFP_I128_F128; 401 if (RetVT == MVT::ppcf128) 402 return SINTTOFP_I128_PPCF128; 403 } 404 return UNKNOWN_LIBCALL; 405 } 406 407 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 408 /// UNKNOWN_LIBCALL if there is none. 409 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 410 if (OpVT == MVT::i32) { 411 if (RetVT == MVT::f32) 412 return UINTTOFP_I32_F32; 413 if (RetVT == MVT::f64) 414 return UINTTOFP_I32_F64; 415 if (RetVT == MVT::f80) 416 return UINTTOFP_I32_F80; 417 if (RetVT == MVT::f128) 418 return UINTTOFP_I32_F128; 419 if (RetVT == MVT::ppcf128) 420 return UINTTOFP_I32_PPCF128; 421 } else if (OpVT == MVT::i64) { 422 if (RetVT == MVT::f32) 423 return UINTTOFP_I64_F32; 424 if (RetVT == MVT::f64) 425 return UINTTOFP_I64_F64; 426 if (RetVT == MVT::f80) 427 return UINTTOFP_I64_F80; 428 if (RetVT == MVT::f128) 429 return UINTTOFP_I64_F128; 430 if (RetVT == MVT::ppcf128) 431 return UINTTOFP_I64_PPCF128; 432 } else if (OpVT == MVT::i128) { 433 if (RetVT == MVT::f32) 434 return UINTTOFP_I128_F32; 435 if (RetVT == MVT::f64) 436 return UINTTOFP_I128_F64; 437 if (RetVT == MVT::f80) 438 return UINTTOFP_I128_F80; 439 if (RetVT == MVT::f128) 440 return UINTTOFP_I128_F128; 441 if (RetVT == MVT::ppcf128) 442 return UINTTOFP_I128_PPCF128; 443 } 444 return UNKNOWN_LIBCALL; 445 } 446 447 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 448 #define OP_TO_LIBCALL(Name, Enum) \ 449 case Name: \ 450 switch (VT.SimpleTy) { \ 451 default: \ 452 return UNKNOWN_LIBCALL; \ 453 case MVT::i8: \ 454 return Enum##_1; \ 455 case MVT::i16: \ 456 return Enum##_2; \ 457 case MVT::i32: \ 458 return Enum##_4; \ 459 case MVT::i64: \ 460 return Enum##_8; \ 461 case MVT::i128: \ 462 return Enum##_16; \ 463 } 464 465 switch (Opc) { 466 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 467 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 468 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 469 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 470 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 471 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 472 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 473 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 474 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 475 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 476 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 477 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 478 } 479 480 #undef OP_TO_LIBCALL 481 482 return UNKNOWN_LIBCALL; 483 } 484 485 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 486 switch (ElementSize) { 487 case 1: 488 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 489 case 2: 490 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 491 case 4: 492 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 493 case 8: 494 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 495 case 16: 496 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 497 default: 498 return UNKNOWN_LIBCALL; 499 } 500 } 501 502 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 503 switch (ElementSize) { 504 case 1: 505 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 506 case 2: 507 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 508 case 4: 509 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 510 case 8: 511 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 512 case 16: 513 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 514 default: 515 return UNKNOWN_LIBCALL; 516 } 517 } 518 519 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 520 switch (ElementSize) { 521 case 1: 522 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 523 case 2: 524 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 525 case 4: 526 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 527 case 8: 528 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 529 case 16: 530 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 531 default: 532 return UNKNOWN_LIBCALL; 533 } 534 } 535 536 /// InitCmpLibcallCCs - Set default comparison libcall CC. 537 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 538 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 539 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 540 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 541 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 542 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 543 CCs[RTLIB::UNE_F32] = ISD::SETNE; 544 CCs[RTLIB::UNE_F64] = ISD::SETNE; 545 CCs[RTLIB::UNE_F128] = ISD::SETNE; 546 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 547 CCs[RTLIB::OGE_F32] = ISD::SETGE; 548 CCs[RTLIB::OGE_F64] = ISD::SETGE; 549 CCs[RTLIB::OGE_F128] = ISD::SETGE; 550 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 551 CCs[RTLIB::OLT_F32] = ISD::SETLT; 552 CCs[RTLIB::OLT_F64] = ISD::SETLT; 553 CCs[RTLIB::OLT_F128] = ISD::SETLT; 554 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 555 CCs[RTLIB::OLE_F32] = ISD::SETLE; 556 CCs[RTLIB::OLE_F64] = ISD::SETLE; 557 CCs[RTLIB::OLE_F128] = ISD::SETLE; 558 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 559 CCs[RTLIB::OGT_F32] = ISD::SETGT; 560 CCs[RTLIB::OGT_F64] = ISD::SETGT; 561 CCs[RTLIB::OGT_F128] = ISD::SETGT; 562 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 563 CCs[RTLIB::UO_F32] = ISD::SETNE; 564 CCs[RTLIB::UO_F64] = ISD::SETNE; 565 CCs[RTLIB::UO_F128] = ISD::SETNE; 566 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 567 CCs[RTLIB::O_F32] = ISD::SETEQ; 568 CCs[RTLIB::O_F64] = ISD::SETEQ; 569 CCs[RTLIB::O_F128] = ISD::SETEQ; 570 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 571 } 572 573 /// NOTE: The TargetMachine owns TLOF. 574 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 575 initActions(); 576 577 // Perform these initializations only once. 578 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 579 MaxLoadsPerMemcmp = 8; 580 MaxGluedStoresPerMemcpy = 0; 581 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 582 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 583 HasMultipleConditionRegisters = false; 584 HasExtractBitsInsn = false; 585 JumpIsExpensive = JumpIsExpensiveOverride; 586 PredictableSelectIsExpensive = false; 587 EnableExtLdPromotion = false; 588 StackPointerRegisterToSaveRestore = 0; 589 BooleanContents = UndefinedBooleanContent; 590 BooleanFloatContents = UndefinedBooleanContent; 591 BooleanVectorContents = UndefinedBooleanContent; 592 SchedPreferenceInfo = Sched::ILP; 593 GatherAllAliasesMaxDepth = 18; 594 IsStrictFPEnabled = DisableStrictNodeMutation; 595 // TODO: the default will be switched to 0 in the next commit, along 596 // with the Target-specific changes necessary. 597 MaxAtomicSizeInBitsSupported = 1024; 598 599 MinCmpXchgSizeInBits = 0; 600 SupportsUnalignedAtomics = false; 601 602 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 603 604 InitLibcalls(TM.getTargetTriple()); 605 InitCmpLibcallCCs(CmpLibcallCCs); 606 } 607 608 void TargetLoweringBase::initActions() { 609 // All operations default to being supported. 610 memset(OpActions, 0, sizeof(OpActions)); 611 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 612 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 613 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 614 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 615 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 616 std::fill(std::begin(TargetDAGCombineArray), 617 std::end(TargetDAGCombineArray), 0); 618 619 for (MVT VT : MVT::fp_valuetypes()) { 620 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 621 if (IntVT.isValid()) { 622 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 623 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 624 } 625 } 626 627 // Set default actions for various operations. 628 for (MVT VT : MVT::all_valuetypes()) { 629 // Default all indexed load / store to expand. 630 for (unsigned IM = (unsigned)ISD::PRE_INC; 631 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 632 setIndexedLoadAction(IM, VT, Expand); 633 setIndexedStoreAction(IM, VT, Expand); 634 setIndexedMaskedLoadAction(IM, VT, Expand); 635 setIndexedMaskedStoreAction(IM, VT, Expand); 636 } 637 638 // Most backends expect to see the node which just returns the value loaded. 639 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 640 641 // These operations default to expand. 642 setOperationAction(ISD::FGETSIGN, VT, Expand); 643 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 644 setOperationAction(ISD::FMINNUM, VT, Expand); 645 setOperationAction(ISD::FMAXNUM, VT, Expand); 646 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 647 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 648 setOperationAction(ISD::FMINIMUM, VT, Expand); 649 setOperationAction(ISD::FMAXIMUM, VT, Expand); 650 setOperationAction(ISD::FMAD, VT, Expand); 651 setOperationAction(ISD::SMIN, VT, Expand); 652 setOperationAction(ISD::SMAX, VT, Expand); 653 setOperationAction(ISD::UMIN, VT, Expand); 654 setOperationAction(ISD::UMAX, VT, Expand); 655 setOperationAction(ISD::ABS, VT, Expand); 656 setOperationAction(ISD::FSHL, VT, Expand); 657 setOperationAction(ISD::FSHR, VT, Expand); 658 setOperationAction(ISD::SADDSAT, VT, Expand); 659 setOperationAction(ISD::UADDSAT, VT, Expand); 660 setOperationAction(ISD::SSUBSAT, VT, Expand); 661 setOperationAction(ISD::USUBSAT, VT, Expand); 662 setOperationAction(ISD::SMULFIX, VT, Expand); 663 setOperationAction(ISD::SMULFIXSAT, VT, Expand); 664 setOperationAction(ISD::UMULFIX, VT, Expand); 665 setOperationAction(ISD::UMULFIXSAT, VT, Expand); 666 667 // Overflow operations default to expand 668 setOperationAction(ISD::SADDO, VT, Expand); 669 setOperationAction(ISD::SSUBO, VT, Expand); 670 setOperationAction(ISD::UADDO, VT, Expand); 671 setOperationAction(ISD::USUBO, VT, Expand); 672 setOperationAction(ISD::SMULO, VT, Expand); 673 setOperationAction(ISD::UMULO, VT, Expand); 674 675 // ADDCARRY operations default to expand 676 setOperationAction(ISD::ADDCARRY, VT, Expand); 677 setOperationAction(ISD::SUBCARRY, VT, Expand); 678 setOperationAction(ISD::SETCCCARRY, VT, Expand); 679 680 // ADDC/ADDE/SUBC/SUBE default to expand. 681 setOperationAction(ISD::ADDC, VT, Expand); 682 setOperationAction(ISD::ADDE, VT, Expand); 683 setOperationAction(ISD::SUBC, VT, Expand); 684 setOperationAction(ISD::SUBE, VT, Expand); 685 686 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 687 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 688 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 689 690 setOperationAction(ISD::BITREVERSE, VT, Expand); 691 692 // These library functions default to expand. 693 setOperationAction(ISD::FROUND, VT, Expand); 694 setOperationAction(ISD::FPOWI, VT, Expand); 695 696 // These operations default to expand for vector types. 697 if (VT.isVector()) { 698 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 699 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); 700 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 701 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 702 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 703 setOperationAction(ISD::SPLAT_VECTOR, VT, Expand); 704 } 705 706 // Constrained floating-point operations default to expand. 707 #define INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ 708 setOperationAction(ISD::STRICT_##DAGN, VT, Expand); 709 #include "llvm/IR/ConstrainedOps.def" 710 711 // For most targets @llvm.get.dynamic.area.offset just returns 0. 712 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 713 714 // Vector reduction default to expand. 715 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 716 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 717 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 718 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 719 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 720 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 721 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 722 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 723 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 724 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 725 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 726 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 727 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 728 } 729 730 // Most targets ignore the @llvm.prefetch intrinsic. 731 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 732 733 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 734 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 735 736 // ConstantFP nodes default to expand. Targets can either change this to 737 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 738 // to optimize expansions for certain constants. 739 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 740 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 741 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 742 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 743 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 744 745 // These library functions default to expand. 746 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 747 setOperationAction(ISD::FCBRT, VT, Expand); 748 setOperationAction(ISD::FLOG , VT, Expand); 749 setOperationAction(ISD::FLOG2, VT, Expand); 750 setOperationAction(ISD::FLOG10, VT, Expand); 751 setOperationAction(ISD::FEXP , VT, Expand); 752 setOperationAction(ISD::FEXP2, VT, Expand); 753 setOperationAction(ISD::FFLOOR, VT, Expand); 754 setOperationAction(ISD::FNEARBYINT, VT, Expand); 755 setOperationAction(ISD::FCEIL, VT, Expand); 756 setOperationAction(ISD::FRINT, VT, Expand); 757 setOperationAction(ISD::FTRUNC, VT, Expand); 758 setOperationAction(ISD::FROUND, VT, Expand); 759 setOperationAction(ISD::LROUND, VT, Expand); 760 setOperationAction(ISD::LLROUND, VT, Expand); 761 setOperationAction(ISD::LRINT, VT, Expand); 762 setOperationAction(ISD::LLRINT, VT, Expand); 763 } 764 765 // Default ISD::TRAP to expand (which turns it into abort). 766 setOperationAction(ISD::TRAP, MVT::Other, Expand); 767 768 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 769 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 770 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 771 } 772 773 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 774 EVT) const { 775 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 776 } 777 778 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 779 bool LegalTypes) const { 780 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 781 if (LHSTy.isVector()) 782 return LHSTy; 783 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 784 : getPointerTy(DL); 785 } 786 787 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 788 assert(isTypeLegal(VT)); 789 switch (Op) { 790 default: 791 return false; 792 case ISD::SDIV: 793 case ISD::UDIV: 794 case ISD::SREM: 795 case ISD::UREM: 796 return true; 797 } 798 } 799 800 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 801 // If the command-line option was specified, ignore this request. 802 if (!JumpIsExpensiveOverride.getNumOccurrences()) 803 JumpIsExpensive = isExpensive; 804 } 805 806 TargetLoweringBase::LegalizeKind 807 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 808 // If this is a simple type, use the ComputeRegisterProp mechanism. 809 if (VT.isSimple()) { 810 MVT SVT = VT.getSimpleVT(); 811 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 812 MVT NVT = TransformToType[SVT.SimpleTy]; 813 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 814 815 assert((LA == TypeLegal || LA == TypeSoftenFloat || 816 (NVT.isVector() || 817 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && 818 "Promote may not follow Expand or Promote"); 819 820 if (LA == TypeSplitVector) 821 return LegalizeKind(LA, 822 EVT::getVectorVT(Context, SVT.getVectorElementType(), 823 SVT.getVectorNumElements() / 2)); 824 if (LA == TypeScalarizeVector) 825 return LegalizeKind(LA, SVT.getVectorElementType()); 826 return LegalizeKind(LA, NVT); 827 } 828 829 // Handle Extended Scalar Types. 830 if (!VT.isVector()) { 831 assert(VT.isInteger() && "Float types must be simple"); 832 unsigned BitSize = VT.getSizeInBits(); 833 // First promote to a power-of-two size, then expand if necessary. 834 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 835 EVT NVT = VT.getRoundIntegerType(Context); 836 assert(NVT != VT && "Unable to round integer VT"); 837 LegalizeKind NextStep = getTypeConversion(Context, NVT); 838 // Avoid multi-step promotion. 839 if (NextStep.first == TypePromoteInteger) 840 return NextStep; 841 // Return rounded integer type. 842 return LegalizeKind(TypePromoteInteger, NVT); 843 } 844 845 return LegalizeKind(TypeExpandInteger, 846 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 847 } 848 849 // Handle vector types. 850 unsigned NumElts = VT.getVectorNumElements(); 851 EVT EltVT = VT.getVectorElementType(); 852 853 // Vectors with only one element are always scalarized. 854 if (NumElts == 1) 855 return LegalizeKind(TypeScalarizeVector, EltVT); 856 857 // Try to widen vector elements until the element type is a power of two and 858 // promote it to a legal type later on, for example: 859 // <3 x i8> -> <4 x i8> -> <4 x i32> 860 if (EltVT.isInteger()) { 861 // Vectors with a number of elements that is not a power of two are always 862 // widened, for example <3 x i8> -> <4 x i8>. 863 if (!VT.isPow2VectorType()) { 864 NumElts = (unsigned)NextPowerOf2(NumElts); 865 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 866 return LegalizeKind(TypeWidenVector, NVT); 867 } 868 869 // Examine the element type. 870 LegalizeKind LK = getTypeConversion(Context, EltVT); 871 872 // If type is to be expanded, split the vector. 873 // <4 x i140> -> <2 x i140> 874 if (LK.first == TypeExpandInteger) 875 return LegalizeKind(TypeSplitVector, 876 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 877 878 // Promote the integer element types until a legal vector type is found 879 // or until the element integer type is too big. If a legal type was not 880 // found, fallback to the usual mechanism of widening/splitting the 881 // vector. 882 EVT OldEltVT = EltVT; 883 while (true) { 884 // Increase the bitwidth of the element to the next pow-of-two 885 // (which is greater than 8 bits). 886 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 887 .getRoundIntegerType(Context); 888 889 // Stop trying when getting a non-simple element type. 890 // Note that vector elements may be greater than legal vector element 891 // types. Example: X86 XMM registers hold 64bit element on 32bit 892 // systems. 893 if (!EltVT.isSimple()) 894 break; 895 896 // Build a new vector type and check if it is legal. 897 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 898 // Found a legal promoted vector type. 899 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 900 return LegalizeKind(TypePromoteInteger, 901 EVT::getVectorVT(Context, EltVT, NumElts)); 902 } 903 904 // Reset the type to the unexpanded type if we did not find a legal vector 905 // type with a promoted vector element type. 906 EltVT = OldEltVT; 907 } 908 909 // Try to widen the vector until a legal type is found. 910 // If there is no wider legal type, split the vector. 911 while (true) { 912 // Round up to the next power of 2. 913 NumElts = (unsigned)NextPowerOf2(NumElts); 914 915 // If there is no simple vector type with this many elements then there 916 // cannot be a larger legal vector type. Note that this assumes that 917 // there are no skipped intermediate vector types in the simple types. 918 if (!EltVT.isSimple()) 919 break; 920 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 921 if (LargerVector == MVT()) 922 break; 923 924 // If this type is legal then widen the vector. 925 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 926 return LegalizeKind(TypeWidenVector, LargerVector); 927 } 928 929 // Widen odd vectors to next power of two. 930 if (!VT.isPow2VectorType()) { 931 EVT NVT = VT.getPow2VectorType(Context); 932 return LegalizeKind(TypeWidenVector, NVT); 933 } 934 935 // Vectors with illegal element types are expanded. 936 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 937 return LegalizeKind(TypeSplitVector, NVT); 938 } 939 940 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 941 unsigned &NumIntermediates, 942 MVT &RegisterVT, 943 TargetLoweringBase *TLI) { 944 // Figure out the right, legal destination reg to copy into. 945 unsigned NumElts = VT.getVectorNumElements(); 946 MVT EltTy = VT.getVectorElementType(); 947 948 unsigned NumVectorRegs = 1; 949 950 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 951 // could break down into LHS/RHS like LegalizeDAG does. 952 if (!isPowerOf2_32(NumElts)) { 953 NumVectorRegs = NumElts; 954 NumElts = 1; 955 } 956 957 // Divide the input until we get to a supported size. This will always 958 // end with a scalar if the target doesn't support vectors. 959 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 960 NumElts >>= 1; 961 NumVectorRegs <<= 1; 962 } 963 964 NumIntermediates = NumVectorRegs; 965 966 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 967 if (!TLI->isTypeLegal(NewVT)) 968 NewVT = EltTy; 969 IntermediateVT = NewVT; 970 971 unsigned NewVTSize = NewVT.getSizeInBits(); 972 973 // Convert sizes such as i33 to i64. 974 if (!isPowerOf2_32(NewVTSize)) 975 NewVTSize = NextPowerOf2(NewVTSize); 976 977 MVT DestVT = TLI->getRegisterType(NewVT); 978 RegisterVT = DestVT; 979 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 980 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 981 982 // Otherwise, promotion or legal types use the same number of registers as 983 // the vector decimated to the appropriate level. 984 return NumVectorRegs; 985 } 986 987 /// isLegalRC - Return true if the value types that can be represented by the 988 /// specified register class are all legal. 989 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 990 const TargetRegisterClass &RC) const { 991 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 992 if (isTypeLegal(*I)) 993 return true; 994 return false; 995 } 996 997 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 998 /// sequence of memory operands that is recognized by PrologEpilogInserter. 999 MachineBasicBlock * 1000 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 1001 MachineBasicBlock *MBB) const { 1002 MachineInstr *MI = &InitialMI; 1003 MachineFunction &MF = *MI->getMF(); 1004 MachineFrameInfo &MFI = MF.getFrameInfo(); 1005 1006 // We're handling multiple types of operands here: 1007 // PATCHPOINT MetaArgs - live-in, read only, direct 1008 // STATEPOINT Deopt Spill - live-through, read only, indirect 1009 // STATEPOINT Deopt Alloca - live-through, read only, direct 1010 // (We're currently conservative and mark the deopt slots read/write in 1011 // practice.) 1012 // STATEPOINT GC Spill - live-through, read/write, indirect 1013 // STATEPOINT GC Alloca - live-through, read/write, direct 1014 // The live-in vs live-through is handled already (the live through ones are 1015 // all stack slots), but we need to handle the different type of stackmap 1016 // operands and memory effects here. 1017 1018 // MI changes inside this loop as we grow operands. 1019 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1020 MachineOperand &MO = MI->getOperand(OperIdx); 1021 if (!MO.isFI()) 1022 continue; 1023 1024 // foldMemoryOperand builds a new MI after replacing a single FI operand 1025 // with the canonical set of five x86 addressing-mode operands. 1026 int FI = MO.getIndex(); 1027 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1028 1029 // Copy operands before the frame-index. 1030 for (unsigned i = 0; i < OperIdx; ++i) 1031 MIB.add(MI->getOperand(i)); 1032 // Add frame index operands recognized by stackmaps.cpp 1033 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1034 // indirect-mem-ref tag, size, #FI, offset. 1035 // Used for spills inserted by StatepointLowering. This codepath is not 1036 // used for patchpoints/stackmaps at all, for these spilling is done via 1037 // foldMemoryOperand callback only. 1038 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1039 MIB.addImm(StackMaps::IndirectMemRefOp); 1040 MIB.addImm(MFI.getObjectSize(FI)); 1041 MIB.add(MI->getOperand(OperIdx)); 1042 MIB.addImm(0); 1043 } else { 1044 // direct-mem-ref tag, #FI, offset. 1045 // Used by patchpoint, and direct alloca arguments to statepoints 1046 MIB.addImm(StackMaps::DirectMemRefOp); 1047 MIB.add(MI->getOperand(OperIdx)); 1048 MIB.addImm(0); 1049 } 1050 // Copy the operands after the frame index. 1051 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1052 MIB.add(MI->getOperand(i)); 1053 1054 // Inherit previous memory operands. 1055 MIB.cloneMemRefs(*MI); 1056 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1057 1058 // Add a new memory operand for this FI. 1059 assert(MFI.getObjectOffset(FI) != -1); 1060 1061 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1062 // PATCHPOINT should be updated to do the same. (TODO) 1063 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1064 auto Flags = MachineMemOperand::MOLoad; 1065 MachineMemOperand *MMO = MF.getMachineMemOperand( 1066 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1067 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1068 MIB->addMemOperand(MF, MMO); 1069 } 1070 1071 // Replace the instruction and update the operand index. 1072 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1073 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1074 MI->eraseFromParent(); 1075 MI = MIB; 1076 } 1077 return MBB; 1078 } 1079 1080 MachineBasicBlock * 1081 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1082 MachineBasicBlock *MBB) const { 1083 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1084 "Called emitXRayCustomEvent on the wrong MI!"); 1085 auto &MF = *MI.getMF(); 1086 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1087 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1088 MIB.add(MI.getOperand(OpIdx)); 1089 1090 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1091 MI.eraseFromParent(); 1092 return MBB; 1093 } 1094 1095 MachineBasicBlock * 1096 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1097 MachineBasicBlock *MBB) const { 1098 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1099 "Called emitXRayTypedEvent on the wrong MI!"); 1100 auto &MF = *MI.getMF(); 1101 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1102 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1103 MIB.add(MI.getOperand(OpIdx)); 1104 1105 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1106 MI.eraseFromParent(); 1107 return MBB; 1108 } 1109 1110 /// findRepresentativeClass - Return the largest legal super-reg register class 1111 /// of the register class for the specified type and its associated "cost". 1112 // This function is in TargetLowering because it uses RegClassForVT which would 1113 // need to be moved to TargetRegisterInfo and would necessitate moving 1114 // isTypeLegal over as well - a massive change that would just require 1115 // TargetLowering having a TargetRegisterInfo class member that it would use. 1116 std::pair<const TargetRegisterClass *, uint8_t> 1117 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1118 MVT VT) const { 1119 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1120 if (!RC) 1121 return std::make_pair(RC, 0); 1122 1123 // Compute the set of all super-register classes. 1124 BitVector SuperRegRC(TRI->getNumRegClasses()); 1125 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1126 SuperRegRC.setBitsInMask(RCI.getMask()); 1127 1128 // Find the first legal register class with the largest spill size. 1129 const TargetRegisterClass *BestRC = RC; 1130 for (unsigned i : SuperRegRC.set_bits()) { 1131 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1132 // We want the largest possible spill size. 1133 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1134 continue; 1135 if (!isLegalRC(*TRI, *SuperRC)) 1136 continue; 1137 BestRC = SuperRC; 1138 } 1139 return std::make_pair(BestRC, 1); 1140 } 1141 1142 /// computeRegisterProperties - Once all of the register classes are added, 1143 /// this allows us to compute derived properties we expose. 1144 void TargetLoweringBase::computeRegisterProperties( 1145 const TargetRegisterInfo *TRI) { 1146 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1147 "Too many value types for ValueTypeActions to hold!"); 1148 1149 // Everything defaults to needing one register. 1150 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1151 NumRegistersForVT[i] = 1; 1152 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1153 } 1154 // ...except isVoid, which doesn't need any registers. 1155 NumRegistersForVT[MVT::isVoid] = 0; 1156 1157 // Find the largest integer register class. 1158 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1159 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1160 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1161 1162 // Every integer value type larger than this largest register takes twice as 1163 // many registers to represent as the previous ValueType. 1164 for (unsigned ExpandedReg = LargestIntReg + 1; 1165 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1166 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1167 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1168 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1169 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1170 TypeExpandInteger); 1171 } 1172 1173 // Inspect all of the ValueType's smaller than the largest integer 1174 // register to see which ones need promotion. 1175 unsigned LegalIntReg = LargestIntReg; 1176 for (unsigned IntReg = LargestIntReg - 1; 1177 IntReg >= (unsigned)MVT::i1; --IntReg) { 1178 MVT IVT = (MVT::SimpleValueType)IntReg; 1179 if (isTypeLegal(IVT)) { 1180 LegalIntReg = IntReg; 1181 } else { 1182 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1183 (MVT::SimpleValueType)LegalIntReg; 1184 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1185 } 1186 } 1187 1188 // ppcf128 type is really two f64's. 1189 if (!isTypeLegal(MVT::ppcf128)) { 1190 if (isTypeLegal(MVT::f64)) { 1191 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1192 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1193 TransformToType[MVT::ppcf128] = MVT::f64; 1194 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1195 } else { 1196 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1197 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1198 TransformToType[MVT::ppcf128] = MVT::i128; 1199 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1200 } 1201 } 1202 1203 // Decide how to handle f128. If the target does not have native f128 support, 1204 // expand it to i128 and we will be generating soft float library calls. 1205 if (!isTypeLegal(MVT::f128)) { 1206 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1207 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1208 TransformToType[MVT::f128] = MVT::i128; 1209 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1210 } 1211 1212 // Decide how to handle f64. If the target does not have native f64 support, 1213 // expand it to i64 and we will be generating soft float library calls. 1214 if (!isTypeLegal(MVT::f64)) { 1215 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1216 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1217 TransformToType[MVT::f64] = MVT::i64; 1218 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1219 } 1220 1221 // Decide how to handle f32. If the target does not have native f32 support, 1222 // expand it to i32 and we will be generating soft float library calls. 1223 if (!isTypeLegal(MVT::f32)) { 1224 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1225 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1226 TransformToType[MVT::f32] = MVT::i32; 1227 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1228 } 1229 1230 // Decide how to handle f16. If the target does not have native f16 support, 1231 // promote it to f32, because there are no f16 library calls (except for 1232 // conversions). 1233 if (!isTypeLegal(MVT::f16)) { 1234 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1235 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1236 TransformToType[MVT::f16] = MVT::f32; 1237 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1238 } 1239 1240 // Loop over all of the vector value types to see which need transformations. 1241 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1242 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1243 MVT VT = (MVT::SimpleValueType) i; 1244 if (isTypeLegal(VT)) 1245 continue; 1246 1247 MVT EltVT = VT.getVectorElementType(); 1248 unsigned NElts = VT.getVectorNumElements(); 1249 bool IsLegalWiderType = false; 1250 bool IsScalable = VT.isScalableVector(); 1251 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1252 switch (PreferredAction) { 1253 case TypePromoteInteger: { 1254 MVT::SimpleValueType EndVT = IsScalable ? 1255 MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE : 1256 MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE; 1257 // Try to promote the elements of integer vectors. If no legal 1258 // promotion was found, fall through to the widen-vector method. 1259 for (unsigned nVT = i + 1; 1260 (MVT::SimpleValueType)nVT <= EndVT; ++nVT) { 1261 MVT SVT = (MVT::SimpleValueType) nVT; 1262 // Promote vectors of integers to vectors with the same number 1263 // of elements, with a wider element type. 1264 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1265 SVT.getVectorNumElements() == NElts && 1266 SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1267 TransformToType[i] = SVT; 1268 RegisterTypeForVT[i] = SVT; 1269 NumRegistersForVT[i] = 1; 1270 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1271 IsLegalWiderType = true; 1272 break; 1273 } 1274 } 1275 if (IsLegalWiderType) 1276 break; 1277 LLVM_FALLTHROUGH; 1278 } 1279 1280 case TypeWidenVector: 1281 if (isPowerOf2_32(NElts)) { 1282 // Try to widen the vector. 1283 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1284 MVT SVT = (MVT::SimpleValueType) nVT; 1285 if (SVT.getVectorElementType() == EltVT 1286 && SVT.getVectorNumElements() > NElts 1287 && SVT.isScalableVector() == IsScalable && isTypeLegal(SVT)) { 1288 TransformToType[i] = SVT; 1289 RegisterTypeForVT[i] = SVT; 1290 NumRegistersForVT[i] = 1; 1291 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1292 IsLegalWiderType = true; 1293 break; 1294 } 1295 } 1296 if (IsLegalWiderType) 1297 break; 1298 } else { 1299 // Only widen to the next power of 2 to keep consistency with EVT. 1300 MVT NVT = VT.getPow2VectorType(); 1301 if (isTypeLegal(NVT)) { 1302 TransformToType[i] = NVT; 1303 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1304 RegisterTypeForVT[i] = NVT; 1305 NumRegistersForVT[i] = 1; 1306 break; 1307 } 1308 } 1309 LLVM_FALLTHROUGH; 1310 1311 case TypeSplitVector: 1312 case TypeScalarizeVector: { 1313 MVT IntermediateVT; 1314 MVT RegisterVT; 1315 unsigned NumIntermediates; 1316 unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1317 NumIntermediates, RegisterVT, this); 1318 NumRegistersForVT[i] = NumRegisters; 1319 assert(NumRegistersForVT[i] == NumRegisters && 1320 "NumRegistersForVT size cannot represent NumRegisters!"); 1321 RegisterTypeForVT[i] = RegisterVT; 1322 1323 MVT NVT = VT.getPow2VectorType(); 1324 if (NVT == VT) { 1325 // Type is already a power of 2. The default action is to split. 1326 TransformToType[i] = MVT::Other; 1327 if (PreferredAction == TypeScalarizeVector) 1328 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1329 else if (PreferredAction == TypeSplitVector) 1330 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1331 else 1332 // Set type action according to the number of elements. 1333 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1334 : TypeSplitVector); 1335 } else { 1336 TransformToType[i] = NVT; 1337 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1338 } 1339 break; 1340 } 1341 default: 1342 llvm_unreachable("Unknown vector legalization action!"); 1343 } 1344 } 1345 1346 // Determine the 'representative' register class for each value type. 1347 // An representative register class is the largest (meaning one which is 1348 // not a sub-register class / subreg register class) legal register class for 1349 // a group of value types. For example, on i386, i8, i16, and i32 1350 // representative would be GR32; while on x86_64 it's GR64. 1351 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1352 const TargetRegisterClass* RRC; 1353 uint8_t Cost; 1354 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1355 RepRegClassForVT[i] = RRC; 1356 RepRegClassCostForVT[i] = Cost; 1357 } 1358 } 1359 1360 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1361 EVT VT) const { 1362 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1363 return getPointerTy(DL).SimpleTy; 1364 } 1365 1366 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1367 return MVT::i32; // return the default value 1368 } 1369 1370 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1371 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1372 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1373 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1374 /// 1375 /// This method returns the number of registers needed, and the VT for each 1376 /// register. It also returns the VT and quantity of the intermediate values 1377 /// before they are promoted/expanded. 1378 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1379 EVT &IntermediateVT, 1380 unsigned &NumIntermediates, 1381 MVT &RegisterVT) const { 1382 unsigned NumElts = VT.getVectorNumElements(); 1383 1384 // If there is a wider vector type with the same element type as this one, 1385 // or a promoted vector type that has the same number of elements which 1386 // are wider, then we should convert to that legal vector type. 1387 // This handles things like <2 x float> -> <4 x float> and 1388 // <4 x i1> -> <4 x i32>. 1389 LegalizeTypeAction TA = getTypeAction(Context, VT); 1390 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1391 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1392 if (isTypeLegal(RegisterEVT)) { 1393 IntermediateVT = RegisterEVT; 1394 RegisterVT = RegisterEVT.getSimpleVT(); 1395 NumIntermediates = 1; 1396 return 1; 1397 } 1398 } 1399 1400 // Figure out the right, legal destination reg to copy into. 1401 EVT EltTy = VT.getVectorElementType(); 1402 1403 unsigned NumVectorRegs = 1; 1404 1405 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1406 // could break down into LHS/RHS like LegalizeDAG does. 1407 if (!isPowerOf2_32(NumElts)) { 1408 NumVectorRegs = NumElts; 1409 NumElts = 1; 1410 } 1411 1412 // Divide the input until we get to a supported size. This will always 1413 // end with a scalar if the target doesn't support vectors. 1414 while (NumElts > 1 && !isTypeLegal( 1415 EVT::getVectorVT(Context, EltTy, NumElts))) { 1416 NumElts >>= 1; 1417 NumVectorRegs <<= 1; 1418 } 1419 1420 NumIntermediates = NumVectorRegs; 1421 1422 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1423 if (!isTypeLegal(NewVT)) 1424 NewVT = EltTy; 1425 IntermediateVT = NewVT; 1426 1427 MVT DestVT = getRegisterType(Context, NewVT); 1428 RegisterVT = DestVT; 1429 unsigned NewVTSize = NewVT.getSizeInBits(); 1430 1431 // Convert sizes such as i33 to i64. 1432 if (!isPowerOf2_32(NewVTSize)) 1433 NewVTSize = NextPowerOf2(NewVTSize); 1434 1435 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1436 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1437 1438 // Otherwise, promotion or legal types use the same number of registers as 1439 // the vector decimated to the appropriate level. 1440 return NumVectorRegs; 1441 } 1442 1443 bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI, 1444 uint64_t NumCases, 1445 uint64_t Range, 1446 ProfileSummaryInfo *PSI, 1447 BlockFrequencyInfo *BFI) const { 1448 // FIXME: This function check the maximum table size and density, but the 1449 // minimum size is not checked. It would be nice if the minimum size is 1450 // also combined within this function. Currently, the minimum size check is 1451 // performed in findJumpTable() in SelectionDAGBuiler and 1452 // getEstimatedNumberOfCaseClusters() in BasicTTIImpl. 1453 const bool OptForSize = 1454 SI->getParent()->getParent()->hasOptSize() || 1455 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI); 1456 const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize); 1457 const unsigned MaxJumpTableSize = getMaximumJumpTableSize(); 1458 1459 // Check whether the number of cases is small enough and 1460 // the range is dense enough for a jump table. 1461 return (OptForSize || Range <= MaxJumpTableSize) && 1462 (NumCases * 100 >= Range * MinDensity); 1463 } 1464 1465 /// Get the EVTs and ArgFlags collections that represent the legalized return 1466 /// type of the given function. This does not require a DAG or a return value, 1467 /// and is suitable for use before any DAGs for the function are constructed. 1468 /// TODO: Move this out of TargetLowering.cpp. 1469 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1470 AttributeList attr, 1471 SmallVectorImpl<ISD::OutputArg> &Outs, 1472 const TargetLowering &TLI, const DataLayout &DL) { 1473 SmallVector<EVT, 4> ValueVTs; 1474 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1475 unsigned NumValues = ValueVTs.size(); 1476 if (NumValues == 0) return; 1477 1478 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1479 EVT VT = ValueVTs[j]; 1480 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1481 1482 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1483 ExtendKind = ISD::SIGN_EXTEND; 1484 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1485 ExtendKind = ISD::ZERO_EXTEND; 1486 1487 // FIXME: C calling convention requires the return type to be promoted to 1488 // at least 32-bit. But this is not necessary for non-C calling 1489 // conventions. The frontend should mark functions whose return values 1490 // require promoting with signext or zeroext attributes. 1491 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1492 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1493 if (VT.bitsLT(MinVT)) 1494 VT = MinVT; 1495 } 1496 1497 unsigned NumParts = 1498 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1499 MVT PartVT = 1500 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1501 1502 // 'inreg' on function refers to return value 1503 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1504 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1505 Flags.setInReg(); 1506 1507 // Propagate extension type if any 1508 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1509 Flags.setSExt(); 1510 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1511 Flags.setZExt(); 1512 1513 for (unsigned i = 0; i < NumParts; ++i) 1514 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isfixed=*/true, 0, 0)); 1515 } 1516 } 1517 1518 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1519 /// function arguments in the caller parameter area. This is the actual 1520 /// alignment, not its logarithm. 1521 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1522 const DataLayout &DL) const { 1523 return DL.getABITypeAlignment(Ty); 1524 } 1525 1526 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1527 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1528 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1529 // Check if the specified alignment is sufficient based on the data layout. 1530 // TODO: While using the data layout works in practice, a better solution 1531 // would be to implement this check directly (make this a virtual function). 1532 // For example, the ABI alignment may change based on software platform while 1533 // this function should only be affected by hardware implementation. 1534 Type *Ty = VT.getTypeForEVT(Context); 1535 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1536 // Assume that an access that meets the ABI-specified alignment is fast. 1537 if (Fast != nullptr) 1538 *Fast = true; 1539 return true; 1540 } 1541 1542 // This is a misaligned access. 1543 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); 1544 } 1545 1546 bool TargetLoweringBase::allowsMemoryAccessForAlignment( 1547 LLVMContext &Context, const DataLayout &DL, EVT VT, 1548 const MachineMemOperand &MMO, bool *Fast) const { 1549 return allowsMemoryAccessForAlignment(Context, DL, VT, MMO.getAddrSpace(), 1550 MMO.getAlignment(), MMO.getFlags(), 1551 Fast); 1552 } 1553 1554 bool TargetLoweringBase::allowsMemoryAccess( 1555 LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, 1556 unsigned Alignment, MachineMemOperand::Flags Flags, bool *Fast) const { 1557 return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment, 1558 Flags, Fast); 1559 } 1560 1561 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1562 const DataLayout &DL, EVT VT, 1563 const MachineMemOperand &MMO, 1564 bool *Fast) const { 1565 return allowsMemoryAccess(Context, DL, VT, MMO.getAddrSpace(), 1566 MMO.getAlignment(), MMO.getFlags(), Fast); 1567 } 1568 1569 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1570 return BranchProbability(MinPercentageForPredictableBranch, 100); 1571 } 1572 1573 //===----------------------------------------------------------------------===// 1574 // TargetTransformInfo Helpers 1575 //===----------------------------------------------------------------------===// 1576 1577 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1578 enum InstructionOpcodes { 1579 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1580 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1581 #include "llvm/IR/Instruction.def" 1582 }; 1583 switch (static_cast<InstructionOpcodes>(Opcode)) { 1584 case Ret: return 0; 1585 case Br: return 0; 1586 case Switch: return 0; 1587 case IndirectBr: return 0; 1588 case Invoke: return 0; 1589 case CallBr: return 0; 1590 case Resume: return 0; 1591 case Unreachable: return 0; 1592 case CleanupRet: return 0; 1593 case CatchRet: return 0; 1594 case CatchPad: return 0; 1595 case CatchSwitch: return 0; 1596 case CleanupPad: return 0; 1597 case FNeg: return ISD::FNEG; 1598 case Add: return ISD::ADD; 1599 case FAdd: return ISD::FADD; 1600 case Sub: return ISD::SUB; 1601 case FSub: return ISD::FSUB; 1602 case Mul: return ISD::MUL; 1603 case FMul: return ISD::FMUL; 1604 case UDiv: return ISD::UDIV; 1605 case SDiv: return ISD::SDIV; 1606 case FDiv: return ISD::FDIV; 1607 case URem: return ISD::UREM; 1608 case SRem: return ISD::SREM; 1609 case FRem: return ISD::FREM; 1610 case Shl: return ISD::SHL; 1611 case LShr: return ISD::SRL; 1612 case AShr: return ISD::SRA; 1613 case And: return ISD::AND; 1614 case Or: return ISD::OR; 1615 case Xor: return ISD::XOR; 1616 case Alloca: return 0; 1617 case Load: return ISD::LOAD; 1618 case Store: return ISD::STORE; 1619 case GetElementPtr: return 0; 1620 case Fence: return 0; 1621 case AtomicCmpXchg: return 0; 1622 case AtomicRMW: return 0; 1623 case Trunc: return ISD::TRUNCATE; 1624 case ZExt: return ISD::ZERO_EXTEND; 1625 case SExt: return ISD::SIGN_EXTEND; 1626 case FPToUI: return ISD::FP_TO_UINT; 1627 case FPToSI: return ISD::FP_TO_SINT; 1628 case UIToFP: return ISD::UINT_TO_FP; 1629 case SIToFP: return ISD::SINT_TO_FP; 1630 case FPTrunc: return ISD::FP_ROUND; 1631 case FPExt: return ISD::FP_EXTEND; 1632 case PtrToInt: return ISD::BITCAST; 1633 case IntToPtr: return ISD::BITCAST; 1634 case BitCast: return ISD::BITCAST; 1635 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1636 case ICmp: return ISD::SETCC; 1637 case FCmp: return ISD::SETCC; 1638 case PHI: return 0; 1639 case Call: return 0; 1640 case Select: return ISD::SELECT; 1641 case UserOp1: return 0; 1642 case UserOp2: return 0; 1643 case VAArg: return 0; 1644 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1645 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1646 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1647 case ExtractValue: return ISD::MERGE_VALUES; 1648 case InsertValue: return ISD::MERGE_VALUES; 1649 case LandingPad: return 0; 1650 case Freeze: return 0; 1651 } 1652 1653 llvm_unreachable("Unknown instruction type encountered!"); 1654 } 1655 1656 std::pair<int, MVT> 1657 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1658 Type *Ty) const { 1659 LLVMContext &C = Ty->getContext(); 1660 EVT MTy = getValueType(DL, Ty); 1661 1662 int Cost = 1; 1663 // We keep legalizing the type until we find a legal kind. We assume that 1664 // the only operation that costs anything is the split. After splitting 1665 // we need to handle two types. 1666 while (true) { 1667 LegalizeKind LK = getTypeConversion(C, MTy); 1668 1669 if (LK.first == TypeLegal) 1670 return std::make_pair(Cost, MTy.getSimpleVT()); 1671 1672 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1673 Cost *= 2; 1674 1675 // Do not loop with f128 type. 1676 if (MTy == LK.second) 1677 return std::make_pair(Cost, MTy.getSimpleVT()); 1678 1679 // Keep legalizing the type. 1680 MTy = LK.second; 1681 } 1682 } 1683 1684 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1685 bool UseTLS) const { 1686 // compiler-rt provides a variable with a magic name. Targets that do not 1687 // link with compiler-rt may also provide such a variable. 1688 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1689 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1690 auto UnsafeStackPtr = 1691 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1692 1693 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1694 1695 if (!UnsafeStackPtr) { 1696 auto TLSModel = UseTLS ? 1697 GlobalValue::InitialExecTLSModel : 1698 GlobalValue::NotThreadLocal; 1699 // The global variable is not defined yet, define it ourselves. 1700 // We use the initial-exec TLS model because we do not support the 1701 // variable living anywhere other than in the main executable. 1702 UnsafeStackPtr = new GlobalVariable( 1703 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1704 UnsafeStackPtrVar, nullptr, TLSModel); 1705 } else { 1706 // The variable exists, check its type and attributes. 1707 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1708 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1709 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1710 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1711 (UseTLS ? "" : "not ") + "be thread-local"); 1712 } 1713 return UnsafeStackPtr; 1714 } 1715 1716 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1717 if (!TM.getTargetTriple().isAndroid()) 1718 return getDefaultSafeStackPointerLocation(IRB, true); 1719 1720 // Android provides a libc function to retrieve the address of the current 1721 // thread's unsafe stack pointer. 1722 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1723 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1724 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1725 StackPtrTy->getPointerTo(0)); 1726 return IRB.CreateCall(Fn); 1727 } 1728 1729 //===----------------------------------------------------------------------===// 1730 // Loop Strength Reduction hooks 1731 //===----------------------------------------------------------------------===// 1732 1733 /// isLegalAddressingMode - Return true if the addressing mode represented 1734 /// by AM is legal for this target, for a load/store of the specified type. 1735 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1736 const AddrMode &AM, Type *Ty, 1737 unsigned AS, Instruction *I) const { 1738 // The default implementation of this implements a conservative RISCy, r+r and 1739 // r+i addr mode. 1740 1741 // Allows a sign-extended 16-bit immediate field. 1742 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1743 return false; 1744 1745 // No global is ever allowed as a base. 1746 if (AM.BaseGV) 1747 return false; 1748 1749 // Only support r+r, 1750 switch (AM.Scale) { 1751 case 0: // "r+i" or just "i", depending on HasBaseReg. 1752 break; 1753 case 1: 1754 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1755 return false; 1756 // Otherwise we have r+r or r+i. 1757 break; 1758 case 2: 1759 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1760 return false; 1761 // Allow 2*r as r+r. 1762 break; 1763 default: // Don't allow n * r 1764 return false; 1765 } 1766 1767 return true; 1768 } 1769 1770 //===----------------------------------------------------------------------===// 1771 // Stack Protector 1772 //===----------------------------------------------------------------------===// 1773 1774 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1775 // so that SelectionDAG handle SSP. 1776 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1777 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1778 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1779 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1780 return M.getOrInsertGlobal("__guard_local", PtrTy); 1781 } 1782 return nullptr; 1783 } 1784 1785 // Currently only support "standard" __stack_chk_guard. 1786 // TODO: add LOAD_STACK_GUARD support. 1787 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1788 if (!M.getNamedValue("__stack_chk_guard")) 1789 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1790 GlobalVariable::ExternalLinkage, 1791 nullptr, "__stack_chk_guard"); 1792 } 1793 1794 // Currently only support "standard" __stack_chk_guard. 1795 // TODO: add LOAD_STACK_GUARD support. 1796 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1797 return M.getNamedValue("__stack_chk_guard"); 1798 } 1799 1800 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1801 return nullptr; 1802 } 1803 1804 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1805 return MinimumJumpTableEntries; 1806 } 1807 1808 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1809 MinimumJumpTableEntries = Val; 1810 } 1811 1812 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1813 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1814 } 1815 1816 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1817 return MaximumJumpTableSize; 1818 } 1819 1820 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1821 MaximumJumpTableSize = Val; 1822 } 1823 1824 //===----------------------------------------------------------------------===// 1825 // Reciprocal Estimates 1826 //===----------------------------------------------------------------------===// 1827 1828 /// Get the reciprocal estimate attribute string for a function that will 1829 /// override the target defaults. 1830 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1831 const Function &F = MF.getFunction(); 1832 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1833 } 1834 1835 /// Construct a string for the given reciprocal operation of the given type. 1836 /// This string should match the corresponding option to the front-end's 1837 /// "-mrecip" flag assuming those strings have been passed through in an 1838 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1839 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1840 std::string Name = VT.isVector() ? "vec-" : ""; 1841 1842 Name += IsSqrt ? "sqrt" : "div"; 1843 1844 // TODO: Handle "half" or other float types? 1845 if (VT.getScalarType() == MVT::f64) { 1846 Name += "d"; 1847 } else { 1848 assert(VT.getScalarType() == MVT::f32 && 1849 "Unexpected FP type for reciprocal estimate"); 1850 Name += "f"; 1851 } 1852 1853 return Name; 1854 } 1855 1856 /// Return the character position and value (a single numeric character) of a 1857 /// customized refinement operation in the input string if it exists. Return 1858 /// false if there is no customized refinement step count. 1859 static bool parseRefinementStep(StringRef In, size_t &Position, 1860 uint8_t &Value) { 1861 const char RefStepToken = ':'; 1862 Position = In.find(RefStepToken); 1863 if (Position == StringRef::npos) 1864 return false; 1865 1866 StringRef RefStepString = In.substr(Position + 1); 1867 // Allow exactly one numeric character for the additional refinement 1868 // step parameter. 1869 if (RefStepString.size() == 1) { 1870 char RefStepChar = RefStepString[0]; 1871 if (RefStepChar >= '0' && RefStepChar <= '9') { 1872 Value = RefStepChar - '0'; 1873 return true; 1874 } 1875 } 1876 report_fatal_error("Invalid refinement step for -recip."); 1877 } 1878 1879 /// For the input attribute string, return one of the ReciprocalEstimate enum 1880 /// status values (enabled, disabled, or not specified) for this operation on 1881 /// the specified data type. 1882 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1883 if (Override.empty()) 1884 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1885 1886 SmallVector<StringRef, 4> OverrideVector; 1887 Override.split(OverrideVector, ','); 1888 unsigned NumArgs = OverrideVector.size(); 1889 1890 // Check if "all", "none", or "default" was specified. 1891 if (NumArgs == 1) { 1892 // Look for an optional setting of the number of refinement steps needed 1893 // for this type of reciprocal operation. 1894 size_t RefPos; 1895 uint8_t RefSteps; 1896 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1897 // Split the string for further processing. 1898 Override = Override.substr(0, RefPos); 1899 } 1900 1901 // All reciprocal types are enabled. 1902 if (Override == "all") 1903 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1904 1905 // All reciprocal types are disabled. 1906 if (Override == "none") 1907 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1908 1909 // Target defaults for enablement are used. 1910 if (Override == "default") 1911 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1912 } 1913 1914 // The attribute string may omit the size suffix ('f'/'d'). 1915 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1916 std::string VTNameNoSize = VTName; 1917 VTNameNoSize.pop_back(); 1918 static const char DisabledPrefix = '!'; 1919 1920 for (StringRef RecipType : OverrideVector) { 1921 size_t RefPos; 1922 uint8_t RefSteps; 1923 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1924 RecipType = RecipType.substr(0, RefPos); 1925 1926 // Ignore the disablement token for string matching. 1927 bool IsDisabled = RecipType[0] == DisabledPrefix; 1928 if (IsDisabled) 1929 RecipType = RecipType.substr(1); 1930 1931 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1932 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1933 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1934 } 1935 1936 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1937 } 1938 1939 /// For the input attribute string, return the customized refinement step count 1940 /// for this operation on the specified data type. If the step count does not 1941 /// exist, return the ReciprocalEstimate enum value for unspecified. 1942 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1943 if (Override.empty()) 1944 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1945 1946 SmallVector<StringRef, 4> OverrideVector; 1947 Override.split(OverrideVector, ','); 1948 unsigned NumArgs = OverrideVector.size(); 1949 1950 // Check if "all", "default", or "none" was specified. 1951 if (NumArgs == 1) { 1952 // Look for an optional setting of the number of refinement steps needed 1953 // for this type of reciprocal operation. 1954 size_t RefPos; 1955 uint8_t RefSteps; 1956 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1957 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1958 1959 // Split the string for further processing. 1960 Override = Override.substr(0, RefPos); 1961 assert(Override != "none" && 1962 "Disabled reciprocals, but specifed refinement steps?"); 1963 1964 // If this is a general override, return the specified number of steps. 1965 if (Override == "all" || Override == "default") 1966 return RefSteps; 1967 } 1968 1969 // The attribute string may omit the size suffix ('f'/'d'). 1970 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1971 std::string VTNameNoSize = VTName; 1972 VTNameNoSize.pop_back(); 1973 1974 for (StringRef RecipType : OverrideVector) { 1975 size_t RefPos; 1976 uint8_t RefSteps; 1977 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1978 continue; 1979 1980 RecipType = RecipType.substr(0, RefPos); 1981 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1982 return RefSteps; 1983 } 1984 1985 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1986 } 1987 1988 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1989 MachineFunction &MF) const { 1990 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1991 } 1992 1993 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1994 MachineFunction &MF) const { 1995 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1996 } 1997 1998 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1999 MachineFunction &MF) const { 2000 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 2001 } 2002 2003 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 2004 MachineFunction &MF) const { 2005 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 2006 } 2007 2008 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 2009 MF.getRegInfo().freezeReservedRegs(MF); 2010 } 2011