1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This implements the TargetLoweringBase class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/BitVector.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/StringExtras.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/Triple.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/CodeGen/Analysis.h" 22 #include "llvm/CodeGen/ISDOpcodes.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineMemOperand.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/MachineValueType.h" 32 #include "llvm/CodeGen/RuntimeLibcalls.h" 33 #include "llvm/CodeGen/StackMaps.h" 34 #include "llvm/CodeGen/TargetLowering.h" 35 #include "llvm/CodeGen/TargetOpcodes.h" 36 #include "llvm/CodeGen/TargetRegisterInfo.h" 37 #include "llvm/CodeGen/ValueTypes.h" 38 #include "llvm/IR/Attributes.h" 39 #include "llvm/IR/CallingConv.h" 40 #include "llvm/IR/DataLayout.h" 41 #include "llvm/IR/DerivedTypes.h" 42 #include "llvm/IR/Function.h" 43 #include "llvm/IR/GlobalValue.h" 44 #include "llvm/IR/GlobalVariable.h" 45 #include "llvm/IR/IRBuilder.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/Type.h" 48 #include "llvm/Support/BranchProbability.h" 49 #include "llvm/Support/Casting.h" 50 #include "llvm/Support/CommandLine.h" 51 #include "llvm/Support/Compiler.h" 52 #include "llvm/Support/ErrorHandling.h" 53 #include "llvm/Support/MathExtras.h" 54 #include "llvm/Target/TargetMachine.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstddef> 58 #include <cstdint> 59 #include <cstring> 60 #include <iterator> 61 #include <string> 62 #include <tuple> 63 #include <utility> 64 65 using namespace llvm; 66 67 static cl::opt<bool> JumpIsExpensiveOverride( 68 "jump-is-expensive", cl::init(false), 69 cl::desc("Do not create extra branches to split comparison logic."), 70 cl::Hidden); 71 72 static cl::opt<unsigned> MinimumJumpTableEntries 73 ("min-jump-table-entries", cl::init(4), cl::Hidden, 74 cl::desc("Set minimum number of entries to use a jump table.")); 75 76 static cl::opt<unsigned> MaximumJumpTableSize 77 ("max-jump-table-size", cl::init(0), cl::Hidden, 78 cl::desc("Set maximum size of jump tables; zero for no limit.")); 79 80 /// Minimum jump table density for normal functions. 81 static cl::opt<unsigned> 82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 83 cl::desc("Minimum density for building a jump table in " 84 "a normal function")); 85 86 /// Minimum jump table density for -Os or -Oz functions. 87 static cl::opt<unsigned> OptsizeJumpTableDensity( 88 "optsize-jump-table-density", cl::init(40), cl::Hidden, 89 cl::desc("Minimum density for building a jump table in " 90 "an optsize function")); 91 92 static bool darwinHasSinCos(const Triple &TT) { 93 assert(TT.isOSDarwin() && "should be called with darwin triple"); 94 // Don't bother with 32 bit x86. 95 if (TT.getArch() == Triple::x86) 96 return false; 97 // Macos < 10.9 has no sincos_stret. 98 if (TT.isMacOSX()) 99 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 100 // iOS < 7.0 has no sincos_stret. 101 if (TT.isiOS()) 102 return !TT.isOSVersionLT(7, 0); 103 // Any other darwin such as WatchOS/TvOS is new enough. 104 return true; 105 } 106 107 // Although this default value is arbitrary, it is not random. It is assumed 108 // that a condition that evaluates the same way by a higher percentage than this 109 // is best represented as control flow. Therefore, the default value N should be 110 // set such that the win from N% correct executions is greater than the loss 111 // from (100 - N)% mispredicted executions for the majority of intended targets. 112 static cl::opt<int> MinPercentageForPredictableBranch( 113 "min-predictable-branch", cl::init(99), 114 cl::desc("Minimum percentage (0-100) that a condition must be either true " 115 "or false to assume that the condition is predictable"), 116 cl::Hidden); 117 118 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 119 #define HANDLE_LIBCALL(code, name) \ 120 setLibcallName(RTLIB::code, name); 121 #include "llvm/CodeGen/RuntimeLibcalls.def" 122 #undef HANDLE_LIBCALL 123 // Initialize calling conventions to their default. 124 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 125 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 126 127 // A few names are different on particular architectures or environments. 128 if (TT.isOSDarwin()) { 129 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 130 // of the gnueabi-style __gnu_*_ieee. 131 // FIXME: What about other targets? 132 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 133 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 134 135 // Darwin 10 and higher has an optimized __bzero. 136 if (!TT.isMacOSX() || !TT.isMacOSXVersionLT(10, 6) || TT.isArch64Bit()) { 137 setLibcallName(RTLIB::BZERO, TT.isAArch64() ? "bzero" : "__bzero"); 138 } 139 140 if (darwinHasSinCos(TT)) { 141 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 142 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 143 if (TT.isWatchABI()) { 144 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 145 CallingConv::ARM_AAPCS_VFP); 146 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 147 CallingConv::ARM_AAPCS_VFP); 148 } 149 } 150 } else { 151 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 152 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 153 } 154 155 if (TT.isGNUEnvironment() || TT.isOSFuchsia()) { 156 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 157 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 158 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 159 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 160 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 161 } 162 163 if (TT.isOSOpenBSD()) { 164 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 165 } 166 } 167 168 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 169 /// UNKNOWN_LIBCALL if there is none. 170 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 171 if (OpVT == MVT::f16) { 172 if (RetVT == MVT::f32) 173 return FPEXT_F16_F32; 174 } else if (OpVT == MVT::f32) { 175 if (RetVT == MVT::f64) 176 return FPEXT_F32_F64; 177 if (RetVT == MVT::f128) 178 return FPEXT_F32_F128; 179 if (RetVT == MVT::ppcf128) 180 return FPEXT_F32_PPCF128; 181 } else if (OpVT == MVT::f64) { 182 if (RetVT == MVT::f128) 183 return FPEXT_F64_F128; 184 else if (RetVT == MVT::ppcf128) 185 return FPEXT_F64_PPCF128; 186 } 187 188 return UNKNOWN_LIBCALL; 189 } 190 191 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 192 /// UNKNOWN_LIBCALL if there is none. 193 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 194 if (RetVT == MVT::f16) { 195 if (OpVT == MVT::f32) 196 return FPROUND_F32_F16; 197 if (OpVT == MVT::f64) 198 return FPROUND_F64_F16; 199 if (OpVT == MVT::f80) 200 return FPROUND_F80_F16; 201 if (OpVT == MVT::f128) 202 return FPROUND_F128_F16; 203 if (OpVT == MVT::ppcf128) 204 return FPROUND_PPCF128_F16; 205 } else if (RetVT == MVT::f32) { 206 if (OpVT == MVT::f64) 207 return FPROUND_F64_F32; 208 if (OpVT == MVT::f80) 209 return FPROUND_F80_F32; 210 if (OpVT == MVT::f128) 211 return FPROUND_F128_F32; 212 if (OpVT == MVT::ppcf128) 213 return FPROUND_PPCF128_F32; 214 } else if (RetVT == MVT::f64) { 215 if (OpVT == MVT::f80) 216 return FPROUND_F80_F64; 217 if (OpVT == MVT::f128) 218 return FPROUND_F128_F64; 219 if (OpVT == MVT::ppcf128) 220 return FPROUND_PPCF128_F64; 221 } 222 223 return UNKNOWN_LIBCALL; 224 } 225 226 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 227 /// UNKNOWN_LIBCALL if there is none. 228 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 229 if (OpVT == MVT::f32) { 230 if (RetVT == MVT::i32) 231 return FPTOSINT_F32_I32; 232 if (RetVT == MVT::i64) 233 return FPTOSINT_F32_I64; 234 if (RetVT == MVT::i128) 235 return FPTOSINT_F32_I128; 236 } else if (OpVT == MVT::f64) { 237 if (RetVT == MVT::i32) 238 return FPTOSINT_F64_I32; 239 if (RetVT == MVT::i64) 240 return FPTOSINT_F64_I64; 241 if (RetVT == MVT::i128) 242 return FPTOSINT_F64_I128; 243 } else if (OpVT == MVT::f80) { 244 if (RetVT == MVT::i32) 245 return FPTOSINT_F80_I32; 246 if (RetVT == MVT::i64) 247 return FPTOSINT_F80_I64; 248 if (RetVT == MVT::i128) 249 return FPTOSINT_F80_I128; 250 } else if (OpVT == MVT::f128) { 251 if (RetVT == MVT::i32) 252 return FPTOSINT_F128_I32; 253 if (RetVT == MVT::i64) 254 return FPTOSINT_F128_I64; 255 if (RetVT == MVT::i128) 256 return FPTOSINT_F128_I128; 257 } else if (OpVT == MVT::ppcf128) { 258 if (RetVT == MVT::i32) 259 return FPTOSINT_PPCF128_I32; 260 if (RetVT == MVT::i64) 261 return FPTOSINT_PPCF128_I64; 262 if (RetVT == MVT::i128) 263 return FPTOSINT_PPCF128_I128; 264 } 265 return UNKNOWN_LIBCALL; 266 } 267 268 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 269 /// UNKNOWN_LIBCALL if there is none. 270 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 271 if (OpVT == MVT::f32) { 272 if (RetVT == MVT::i32) 273 return FPTOUINT_F32_I32; 274 if (RetVT == MVT::i64) 275 return FPTOUINT_F32_I64; 276 if (RetVT == MVT::i128) 277 return FPTOUINT_F32_I128; 278 } else if (OpVT == MVT::f64) { 279 if (RetVT == MVT::i32) 280 return FPTOUINT_F64_I32; 281 if (RetVT == MVT::i64) 282 return FPTOUINT_F64_I64; 283 if (RetVT == MVT::i128) 284 return FPTOUINT_F64_I128; 285 } else if (OpVT == MVT::f80) { 286 if (RetVT == MVT::i32) 287 return FPTOUINT_F80_I32; 288 if (RetVT == MVT::i64) 289 return FPTOUINT_F80_I64; 290 if (RetVT == MVT::i128) 291 return FPTOUINT_F80_I128; 292 } else if (OpVT == MVT::f128) { 293 if (RetVT == MVT::i32) 294 return FPTOUINT_F128_I32; 295 if (RetVT == MVT::i64) 296 return FPTOUINT_F128_I64; 297 if (RetVT == MVT::i128) 298 return FPTOUINT_F128_I128; 299 } else if (OpVT == MVT::ppcf128) { 300 if (RetVT == MVT::i32) 301 return FPTOUINT_PPCF128_I32; 302 if (RetVT == MVT::i64) 303 return FPTOUINT_PPCF128_I64; 304 if (RetVT == MVT::i128) 305 return FPTOUINT_PPCF128_I128; 306 } 307 return UNKNOWN_LIBCALL; 308 } 309 310 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 311 /// UNKNOWN_LIBCALL if there is none. 312 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 313 if (OpVT == MVT::i32) { 314 if (RetVT == MVT::f32) 315 return SINTTOFP_I32_F32; 316 if (RetVT == MVT::f64) 317 return SINTTOFP_I32_F64; 318 if (RetVT == MVT::f80) 319 return SINTTOFP_I32_F80; 320 if (RetVT == MVT::f128) 321 return SINTTOFP_I32_F128; 322 if (RetVT == MVT::ppcf128) 323 return SINTTOFP_I32_PPCF128; 324 } else if (OpVT == MVT::i64) { 325 if (RetVT == MVT::f32) 326 return SINTTOFP_I64_F32; 327 if (RetVT == MVT::f64) 328 return SINTTOFP_I64_F64; 329 if (RetVT == MVT::f80) 330 return SINTTOFP_I64_F80; 331 if (RetVT == MVT::f128) 332 return SINTTOFP_I64_F128; 333 if (RetVT == MVT::ppcf128) 334 return SINTTOFP_I64_PPCF128; 335 } else if (OpVT == MVT::i128) { 336 if (RetVT == MVT::f32) 337 return SINTTOFP_I128_F32; 338 if (RetVT == MVT::f64) 339 return SINTTOFP_I128_F64; 340 if (RetVT == MVT::f80) 341 return SINTTOFP_I128_F80; 342 if (RetVT == MVT::f128) 343 return SINTTOFP_I128_F128; 344 if (RetVT == MVT::ppcf128) 345 return SINTTOFP_I128_PPCF128; 346 } 347 return UNKNOWN_LIBCALL; 348 } 349 350 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 351 /// UNKNOWN_LIBCALL if there is none. 352 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 353 if (OpVT == MVT::i32) { 354 if (RetVT == MVT::f32) 355 return UINTTOFP_I32_F32; 356 if (RetVT == MVT::f64) 357 return UINTTOFP_I32_F64; 358 if (RetVT == MVT::f80) 359 return UINTTOFP_I32_F80; 360 if (RetVT == MVT::f128) 361 return UINTTOFP_I32_F128; 362 if (RetVT == MVT::ppcf128) 363 return UINTTOFP_I32_PPCF128; 364 } else if (OpVT == MVT::i64) { 365 if (RetVT == MVT::f32) 366 return UINTTOFP_I64_F32; 367 if (RetVT == MVT::f64) 368 return UINTTOFP_I64_F64; 369 if (RetVT == MVT::f80) 370 return UINTTOFP_I64_F80; 371 if (RetVT == MVT::f128) 372 return UINTTOFP_I64_F128; 373 if (RetVT == MVT::ppcf128) 374 return UINTTOFP_I64_PPCF128; 375 } else if (OpVT == MVT::i128) { 376 if (RetVT == MVT::f32) 377 return UINTTOFP_I128_F32; 378 if (RetVT == MVT::f64) 379 return UINTTOFP_I128_F64; 380 if (RetVT == MVT::f80) 381 return UINTTOFP_I128_F80; 382 if (RetVT == MVT::f128) 383 return UINTTOFP_I128_F128; 384 if (RetVT == MVT::ppcf128) 385 return UINTTOFP_I128_PPCF128; 386 } 387 return UNKNOWN_LIBCALL; 388 } 389 390 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 391 #define OP_TO_LIBCALL(Name, Enum) \ 392 case Name: \ 393 switch (VT.SimpleTy) { \ 394 default: \ 395 return UNKNOWN_LIBCALL; \ 396 case MVT::i8: \ 397 return Enum##_1; \ 398 case MVT::i16: \ 399 return Enum##_2; \ 400 case MVT::i32: \ 401 return Enum##_4; \ 402 case MVT::i64: \ 403 return Enum##_8; \ 404 case MVT::i128: \ 405 return Enum##_16; \ 406 } 407 408 switch (Opc) { 409 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 410 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 411 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 412 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 413 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 414 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 415 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 416 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 417 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 418 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 419 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 420 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 421 } 422 423 #undef OP_TO_LIBCALL 424 425 return UNKNOWN_LIBCALL; 426 } 427 428 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 429 switch (ElementSize) { 430 case 1: 431 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 432 case 2: 433 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 434 case 4: 435 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 436 case 8: 437 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 438 case 16: 439 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 440 default: 441 return UNKNOWN_LIBCALL; 442 } 443 } 444 445 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 446 switch (ElementSize) { 447 case 1: 448 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 449 case 2: 450 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 451 case 4: 452 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 453 case 8: 454 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 455 case 16: 456 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 457 default: 458 return UNKNOWN_LIBCALL; 459 } 460 } 461 462 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 463 switch (ElementSize) { 464 case 1: 465 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 466 case 2: 467 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 468 case 4: 469 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 470 case 8: 471 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 472 case 16: 473 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 474 default: 475 return UNKNOWN_LIBCALL; 476 } 477 } 478 479 /// InitCmpLibcallCCs - Set default comparison libcall CC. 480 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 481 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 482 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 483 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 484 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 485 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 486 CCs[RTLIB::UNE_F32] = ISD::SETNE; 487 CCs[RTLIB::UNE_F64] = ISD::SETNE; 488 CCs[RTLIB::UNE_F128] = ISD::SETNE; 489 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 490 CCs[RTLIB::OGE_F32] = ISD::SETGE; 491 CCs[RTLIB::OGE_F64] = ISD::SETGE; 492 CCs[RTLIB::OGE_F128] = ISD::SETGE; 493 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 494 CCs[RTLIB::OLT_F32] = ISD::SETLT; 495 CCs[RTLIB::OLT_F64] = ISD::SETLT; 496 CCs[RTLIB::OLT_F128] = ISD::SETLT; 497 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 498 CCs[RTLIB::OLE_F32] = ISD::SETLE; 499 CCs[RTLIB::OLE_F64] = ISD::SETLE; 500 CCs[RTLIB::OLE_F128] = ISD::SETLE; 501 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 502 CCs[RTLIB::OGT_F32] = ISD::SETGT; 503 CCs[RTLIB::OGT_F64] = ISD::SETGT; 504 CCs[RTLIB::OGT_F128] = ISD::SETGT; 505 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 506 CCs[RTLIB::UO_F32] = ISD::SETNE; 507 CCs[RTLIB::UO_F64] = ISD::SETNE; 508 CCs[RTLIB::UO_F128] = ISD::SETNE; 509 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 510 CCs[RTLIB::O_F32] = ISD::SETEQ; 511 CCs[RTLIB::O_F64] = ISD::SETEQ; 512 CCs[RTLIB::O_F128] = ISD::SETEQ; 513 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 514 } 515 516 /// NOTE: The TargetMachine owns TLOF. 517 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 518 initActions(); 519 520 // Perform these initializations only once. 521 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 522 MaxLoadsPerMemcmp = 8; 523 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 524 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 525 UseUnderscoreSetJmp = false; 526 UseUnderscoreLongJmp = false; 527 HasMultipleConditionRegisters = false; 528 HasExtractBitsInsn = false; 529 JumpIsExpensive = JumpIsExpensiveOverride; 530 PredictableSelectIsExpensive = false; 531 EnableExtLdPromotion = false; 532 HasFloatingPointExceptions = true; 533 StackPointerRegisterToSaveRestore = 0; 534 BooleanContents = UndefinedBooleanContent; 535 BooleanFloatContents = UndefinedBooleanContent; 536 BooleanVectorContents = UndefinedBooleanContent; 537 SchedPreferenceInfo = Sched::ILP; 538 JumpBufSize = 0; 539 JumpBufAlignment = 0; 540 MinFunctionAlignment = 0; 541 PrefFunctionAlignment = 0; 542 PrefLoopAlignment = 0; 543 GatherAllAliasesMaxDepth = 18; 544 MinStackArgumentAlignment = 1; 545 // TODO: the default will be switched to 0 in the next commit, along 546 // with the Target-specific changes necessary. 547 MaxAtomicSizeInBitsSupported = 1024; 548 549 MinCmpXchgSizeInBits = 0; 550 SupportsUnalignedAtomics = false; 551 552 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 553 554 InitLibcalls(TM.getTargetTriple()); 555 InitCmpLibcallCCs(CmpLibcallCCs); 556 } 557 558 void TargetLoweringBase::initActions() { 559 // All operations default to being supported. 560 memset(OpActions, 0, sizeof(OpActions)); 561 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 562 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 563 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 564 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 565 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 566 std::fill(std::begin(TargetDAGCombineArray), 567 std::end(TargetDAGCombineArray), 0); 568 569 // Set default actions for various operations. 570 for (MVT VT : MVT::all_valuetypes()) { 571 // Default all indexed load / store to expand. 572 for (unsigned IM = (unsigned)ISD::PRE_INC; 573 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 574 setIndexedLoadAction(IM, VT, Expand); 575 setIndexedStoreAction(IM, VT, Expand); 576 } 577 578 // Most backends expect to see the node which just returns the value loaded. 579 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 580 581 // These operations default to expand. 582 setOperationAction(ISD::FGETSIGN, VT, Expand); 583 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 584 setOperationAction(ISD::FMINNUM, VT, Expand); 585 setOperationAction(ISD::FMAXNUM, VT, Expand); 586 setOperationAction(ISD::FMINNAN, VT, Expand); 587 setOperationAction(ISD::FMAXNAN, VT, Expand); 588 setOperationAction(ISD::FMAD, VT, Expand); 589 setOperationAction(ISD::SMIN, VT, Expand); 590 setOperationAction(ISD::SMAX, VT, Expand); 591 setOperationAction(ISD::UMIN, VT, Expand); 592 setOperationAction(ISD::UMAX, VT, Expand); 593 setOperationAction(ISD::ABS, VT, Expand); 594 595 // Overflow operations default to expand 596 setOperationAction(ISD::SADDO, VT, Expand); 597 setOperationAction(ISD::SSUBO, VT, Expand); 598 setOperationAction(ISD::UADDO, VT, Expand); 599 setOperationAction(ISD::USUBO, VT, Expand); 600 setOperationAction(ISD::SMULO, VT, Expand); 601 setOperationAction(ISD::UMULO, VT, Expand); 602 603 // ADDCARRY operations default to expand 604 setOperationAction(ISD::ADDCARRY, VT, Expand); 605 setOperationAction(ISD::SUBCARRY, VT, Expand); 606 setOperationAction(ISD::SETCCCARRY, VT, Expand); 607 608 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 609 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 610 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 611 612 setOperationAction(ISD::BITREVERSE, VT, Expand); 613 614 // These library functions default to expand. 615 setOperationAction(ISD::FROUND, VT, Expand); 616 setOperationAction(ISD::FPOWI, VT, Expand); 617 618 // These operations default to expand for vector types. 619 if (VT.isVector()) { 620 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 621 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 622 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 623 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 624 } 625 626 // For most targets @llvm.get.dynamic.area.offset just returns 0. 627 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 628 } 629 630 // Most targets ignore the @llvm.prefetch intrinsic. 631 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 632 633 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 634 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 635 636 // ConstantFP nodes default to expand. Targets can either change this to 637 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 638 // to optimize expansions for certain constants. 639 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 640 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 641 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 642 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 643 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 644 645 // These library functions default to expand. 646 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 647 setOperationAction(ISD::FLOG , VT, Expand); 648 setOperationAction(ISD::FLOG2, VT, Expand); 649 setOperationAction(ISD::FLOG10, VT, Expand); 650 setOperationAction(ISD::FEXP , VT, Expand); 651 setOperationAction(ISD::FEXP2, VT, Expand); 652 setOperationAction(ISD::FFLOOR, VT, Expand); 653 setOperationAction(ISD::FNEARBYINT, VT, Expand); 654 setOperationAction(ISD::FCEIL, VT, Expand); 655 setOperationAction(ISD::FRINT, VT, Expand); 656 setOperationAction(ISD::FTRUNC, VT, Expand); 657 setOperationAction(ISD::FROUND, VT, Expand); 658 } 659 660 // Default ISD::TRAP to expand (which turns it into abort). 661 setOperationAction(ISD::TRAP, MVT::Other, Expand); 662 663 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 664 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 665 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 666 } 667 668 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 669 EVT) const { 670 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 671 } 672 673 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, 674 const DataLayout &DL) const { 675 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 676 if (LHSTy.isVector()) 677 return LHSTy; 678 return getScalarShiftAmountTy(DL, LHSTy); 679 } 680 681 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 682 assert(isTypeLegal(VT)); 683 switch (Op) { 684 default: 685 return false; 686 case ISD::SDIV: 687 case ISD::UDIV: 688 case ISD::SREM: 689 case ISD::UREM: 690 return true; 691 } 692 } 693 694 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 695 // If the command-line option was specified, ignore this request. 696 if (!JumpIsExpensiveOverride.getNumOccurrences()) 697 JumpIsExpensive = isExpensive; 698 } 699 700 TargetLoweringBase::LegalizeKind 701 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 702 // If this is a simple type, use the ComputeRegisterProp mechanism. 703 if (VT.isSimple()) { 704 MVT SVT = VT.getSimpleVT(); 705 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 706 MVT NVT = TransformToType[SVT.SimpleTy]; 707 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 708 709 assert((LA == TypeLegal || LA == TypeSoftenFloat || 710 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 711 "Promote may not follow Expand or Promote"); 712 713 if (LA == TypeSplitVector) 714 return LegalizeKind(LA, 715 EVT::getVectorVT(Context, SVT.getVectorElementType(), 716 SVT.getVectorNumElements() / 2)); 717 if (LA == TypeScalarizeVector) 718 return LegalizeKind(LA, SVT.getVectorElementType()); 719 return LegalizeKind(LA, NVT); 720 } 721 722 // Handle Extended Scalar Types. 723 if (!VT.isVector()) { 724 assert(VT.isInteger() && "Float types must be simple"); 725 unsigned BitSize = VT.getSizeInBits(); 726 // First promote to a power-of-two size, then expand if necessary. 727 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 728 EVT NVT = VT.getRoundIntegerType(Context); 729 assert(NVT != VT && "Unable to round integer VT"); 730 LegalizeKind NextStep = getTypeConversion(Context, NVT); 731 // Avoid multi-step promotion. 732 if (NextStep.first == TypePromoteInteger) 733 return NextStep; 734 // Return rounded integer type. 735 return LegalizeKind(TypePromoteInteger, NVT); 736 } 737 738 return LegalizeKind(TypeExpandInteger, 739 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 740 } 741 742 // Handle vector types. 743 unsigned NumElts = VT.getVectorNumElements(); 744 EVT EltVT = VT.getVectorElementType(); 745 746 // Vectors with only one element are always scalarized. 747 if (NumElts == 1) 748 return LegalizeKind(TypeScalarizeVector, EltVT); 749 750 // Try to widen vector elements until the element type is a power of two and 751 // promote it to a legal type later on, for example: 752 // <3 x i8> -> <4 x i8> -> <4 x i32> 753 if (EltVT.isInteger()) { 754 // Vectors with a number of elements that is not a power of two are always 755 // widened, for example <3 x i8> -> <4 x i8>. 756 if (!VT.isPow2VectorType()) { 757 NumElts = (unsigned)NextPowerOf2(NumElts); 758 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 759 return LegalizeKind(TypeWidenVector, NVT); 760 } 761 762 // Examine the element type. 763 LegalizeKind LK = getTypeConversion(Context, EltVT); 764 765 // If type is to be expanded, split the vector. 766 // <4 x i140> -> <2 x i140> 767 if (LK.first == TypeExpandInteger) 768 return LegalizeKind(TypeSplitVector, 769 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 770 771 // Promote the integer element types until a legal vector type is found 772 // or until the element integer type is too big. If a legal type was not 773 // found, fallback to the usual mechanism of widening/splitting the 774 // vector. 775 EVT OldEltVT = EltVT; 776 while (true) { 777 // Increase the bitwidth of the element to the next pow-of-two 778 // (which is greater than 8 bits). 779 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 780 .getRoundIntegerType(Context); 781 782 // Stop trying when getting a non-simple element type. 783 // Note that vector elements may be greater than legal vector element 784 // types. Example: X86 XMM registers hold 64bit element on 32bit 785 // systems. 786 if (!EltVT.isSimple()) 787 break; 788 789 // Build a new vector type and check if it is legal. 790 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 791 // Found a legal promoted vector type. 792 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 793 return LegalizeKind(TypePromoteInteger, 794 EVT::getVectorVT(Context, EltVT, NumElts)); 795 } 796 797 // Reset the type to the unexpanded type if we did not find a legal vector 798 // type with a promoted vector element type. 799 EltVT = OldEltVT; 800 } 801 802 // Try to widen the vector until a legal type is found. 803 // If there is no wider legal type, split the vector. 804 while (true) { 805 // Round up to the next power of 2. 806 NumElts = (unsigned)NextPowerOf2(NumElts); 807 808 // If there is no simple vector type with this many elements then there 809 // cannot be a larger legal vector type. Note that this assumes that 810 // there are no skipped intermediate vector types in the simple types. 811 if (!EltVT.isSimple()) 812 break; 813 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 814 if (LargerVector == MVT()) 815 break; 816 817 // If this type is legal then widen the vector. 818 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 819 return LegalizeKind(TypeWidenVector, LargerVector); 820 } 821 822 // Widen odd vectors to next power of two. 823 if (!VT.isPow2VectorType()) { 824 EVT NVT = VT.getPow2VectorType(Context); 825 return LegalizeKind(TypeWidenVector, NVT); 826 } 827 828 // Vectors with illegal element types are expanded. 829 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 830 return LegalizeKind(TypeSplitVector, NVT); 831 } 832 833 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 834 unsigned &NumIntermediates, 835 MVT &RegisterVT, 836 TargetLoweringBase *TLI) { 837 // Figure out the right, legal destination reg to copy into. 838 unsigned NumElts = VT.getVectorNumElements(); 839 MVT EltTy = VT.getVectorElementType(); 840 841 unsigned NumVectorRegs = 1; 842 843 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 844 // could break down into LHS/RHS like LegalizeDAG does. 845 if (!isPowerOf2_32(NumElts)) { 846 NumVectorRegs = NumElts; 847 NumElts = 1; 848 } 849 850 // Divide the input until we get to a supported size. This will always 851 // end with a scalar if the target doesn't support vectors. 852 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 853 NumElts >>= 1; 854 NumVectorRegs <<= 1; 855 } 856 857 NumIntermediates = NumVectorRegs; 858 859 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 860 if (!TLI->isTypeLegal(NewVT)) 861 NewVT = EltTy; 862 IntermediateVT = NewVT; 863 864 unsigned NewVTSize = NewVT.getSizeInBits(); 865 866 // Convert sizes such as i33 to i64. 867 if (!isPowerOf2_32(NewVTSize)) 868 NewVTSize = NextPowerOf2(NewVTSize); 869 870 MVT DestVT = TLI->getRegisterType(NewVT); 871 RegisterVT = DestVT; 872 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 873 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 874 875 // Otherwise, promotion or legal types use the same number of registers as 876 // the vector decimated to the appropriate level. 877 return NumVectorRegs; 878 } 879 880 /// isLegalRC - Return true if the value types that can be represented by the 881 /// specified register class are all legal. 882 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 883 const TargetRegisterClass &RC) const { 884 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 885 if (isTypeLegal(*I)) 886 return true; 887 return false; 888 } 889 890 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 891 /// sequence of memory operands that is recognized by PrologEpilogInserter. 892 MachineBasicBlock * 893 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 894 MachineBasicBlock *MBB) const { 895 MachineInstr *MI = &InitialMI; 896 MachineFunction &MF = *MI->getMF(); 897 MachineFrameInfo &MFI = MF.getFrameInfo(); 898 899 // We're handling multiple types of operands here: 900 // PATCHPOINT MetaArgs - live-in, read only, direct 901 // STATEPOINT Deopt Spill - live-through, read only, indirect 902 // STATEPOINT Deopt Alloca - live-through, read only, direct 903 // (We're currently conservative and mark the deopt slots read/write in 904 // practice.) 905 // STATEPOINT GC Spill - live-through, read/write, indirect 906 // STATEPOINT GC Alloca - live-through, read/write, direct 907 // The live-in vs live-through is handled already (the live through ones are 908 // all stack slots), but we need to handle the different type of stackmap 909 // operands and memory effects here. 910 911 // MI changes inside this loop as we grow operands. 912 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 913 MachineOperand &MO = MI->getOperand(OperIdx); 914 if (!MO.isFI()) 915 continue; 916 917 // foldMemoryOperand builds a new MI after replacing a single FI operand 918 // with the canonical set of five x86 addressing-mode operands. 919 int FI = MO.getIndex(); 920 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 921 922 // Copy operands before the frame-index. 923 for (unsigned i = 0; i < OperIdx; ++i) 924 MIB.add(MI->getOperand(i)); 925 // Add frame index operands recognized by stackmaps.cpp 926 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 927 // indirect-mem-ref tag, size, #FI, offset. 928 // Used for spills inserted by StatepointLowering. This codepath is not 929 // used for patchpoints/stackmaps at all, for these spilling is done via 930 // foldMemoryOperand callback only. 931 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 932 MIB.addImm(StackMaps::IndirectMemRefOp); 933 MIB.addImm(MFI.getObjectSize(FI)); 934 MIB.add(MI->getOperand(OperIdx)); 935 MIB.addImm(0); 936 } else { 937 // direct-mem-ref tag, #FI, offset. 938 // Used by patchpoint, and direct alloca arguments to statepoints 939 MIB.addImm(StackMaps::DirectMemRefOp); 940 MIB.add(MI->getOperand(OperIdx)); 941 MIB.addImm(0); 942 } 943 // Copy the operands after the frame index. 944 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 945 MIB.add(MI->getOperand(i)); 946 947 // Inherit previous memory operands. 948 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 949 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 950 951 // Add a new memory operand for this FI. 952 assert(MFI.getObjectOffset(FI) != -1); 953 954 auto Flags = MachineMemOperand::MOLoad; 955 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 956 Flags |= MachineMemOperand::MOStore; 957 Flags |= MachineMemOperand::MOVolatile; 958 } 959 MachineMemOperand *MMO = MF.getMachineMemOperand( 960 MachinePointerInfo::getFixedStack(MF, FI), Flags, 961 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 962 MIB->addMemOperand(MF, MMO); 963 964 // Replace the instruction and update the operand index. 965 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 966 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 967 MI->eraseFromParent(); 968 MI = MIB; 969 } 970 return MBB; 971 } 972 973 /// findRepresentativeClass - Return the largest legal super-reg register class 974 /// of the register class for the specified type and its associated "cost". 975 // This function is in TargetLowering because it uses RegClassForVT which would 976 // need to be moved to TargetRegisterInfo and would necessitate moving 977 // isTypeLegal over as well - a massive change that would just require 978 // TargetLowering having a TargetRegisterInfo class member that it would use. 979 std::pair<const TargetRegisterClass *, uint8_t> 980 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 981 MVT VT) const { 982 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 983 if (!RC) 984 return std::make_pair(RC, 0); 985 986 // Compute the set of all super-register classes. 987 BitVector SuperRegRC(TRI->getNumRegClasses()); 988 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 989 SuperRegRC.setBitsInMask(RCI.getMask()); 990 991 // Find the first legal register class with the largest spill size. 992 const TargetRegisterClass *BestRC = RC; 993 for (unsigned i : SuperRegRC.set_bits()) { 994 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 995 // We want the largest possible spill size. 996 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 997 continue; 998 if (!isLegalRC(*TRI, *SuperRC)) 999 continue; 1000 BestRC = SuperRC; 1001 } 1002 return std::make_pair(BestRC, 1); 1003 } 1004 1005 /// computeRegisterProperties - Once all of the register classes are added, 1006 /// this allows us to compute derived properties we expose. 1007 void TargetLoweringBase::computeRegisterProperties( 1008 const TargetRegisterInfo *TRI) { 1009 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1010 "Too many value types for ValueTypeActions to hold!"); 1011 1012 // Everything defaults to needing one register. 1013 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1014 NumRegistersForVT[i] = 1; 1015 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1016 } 1017 // ...except isVoid, which doesn't need any registers. 1018 NumRegistersForVT[MVT::isVoid] = 0; 1019 1020 // Find the largest integer register class. 1021 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1022 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1023 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1024 1025 // Every integer value type larger than this largest register takes twice as 1026 // many registers to represent as the previous ValueType. 1027 for (unsigned ExpandedReg = LargestIntReg + 1; 1028 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1029 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1030 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1031 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1032 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1033 TypeExpandInteger); 1034 } 1035 1036 // Inspect all of the ValueType's smaller than the largest integer 1037 // register to see which ones need promotion. 1038 unsigned LegalIntReg = LargestIntReg; 1039 for (unsigned IntReg = LargestIntReg - 1; 1040 IntReg >= (unsigned)MVT::i1; --IntReg) { 1041 MVT IVT = (MVT::SimpleValueType)IntReg; 1042 if (isTypeLegal(IVT)) { 1043 LegalIntReg = IntReg; 1044 } else { 1045 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1046 (const MVT::SimpleValueType)LegalIntReg; 1047 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1048 } 1049 } 1050 1051 // ppcf128 type is really two f64's. 1052 if (!isTypeLegal(MVT::ppcf128)) { 1053 if (isTypeLegal(MVT::f64)) { 1054 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1055 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1056 TransformToType[MVT::ppcf128] = MVT::f64; 1057 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1058 } else { 1059 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1060 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1061 TransformToType[MVT::ppcf128] = MVT::i128; 1062 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1063 } 1064 } 1065 1066 // Decide how to handle f128. If the target does not have native f128 support, 1067 // expand it to i128 and we will be generating soft float library calls. 1068 if (!isTypeLegal(MVT::f128)) { 1069 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1070 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1071 TransformToType[MVT::f128] = MVT::i128; 1072 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1073 } 1074 1075 // Decide how to handle f64. If the target does not have native f64 support, 1076 // expand it to i64 and we will be generating soft float library calls. 1077 if (!isTypeLegal(MVT::f64)) { 1078 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1079 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1080 TransformToType[MVT::f64] = MVT::i64; 1081 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1082 } 1083 1084 // Decide how to handle f32. If the target does not have native f32 support, 1085 // expand it to i32 and we will be generating soft float library calls. 1086 if (!isTypeLegal(MVT::f32)) { 1087 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1088 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1089 TransformToType[MVT::f32] = MVT::i32; 1090 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1091 } 1092 1093 // Decide how to handle f16. If the target does not have native f16 support, 1094 // promote it to f32, because there are no f16 library calls (except for 1095 // conversions). 1096 if (!isTypeLegal(MVT::f16)) { 1097 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1098 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1099 TransformToType[MVT::f16] = MVT::f32; 1100 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1101 } 1102 1103 // Loop over all of the vector value types to see which need transformations. 1104 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1105 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1106 MVT VT = (MVT::SimpleValueType) i; 1107 if (isTypeLegal(VT)) 1108 continue; 1109 1110 MVT EltVT = VT.getVectorElementType(); 1111 unsigned NElts = VT.getVectorNumElements(); 1112 bool IsLegalWiderType = false; 1113 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1114 switch (PreferredAction) { 1115 case TypePromoteInteger: 1116 // Try to promote the elements of integer vectors. If no legal 1117 // promotion was found, fall through to the widen-vector method. 1118 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1119 MVT SVT = (MVT::SimpleValueType) nVT; 1120 // Promote vectors of integers to vectors with the same number 1121 // of elements, with a wider element type. 1122 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1123 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1124 TransformToType[i] = SVT; 1125 RegisterTypeForVT[i] = SVT; 1126 NumRegistersForVT[i] = 1; 1127 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1128 IsLegalWiderType = true; 1129 break; 1130 } 1131 } 1132 if (IsLegalWiderType) 1133 break; 1134 LLVM_FALLTHROUGH; 1135 1136 case TypeWidenVector: 1137 // Try to widen the vector. 1138 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1139 MVT SVT = (MVT::SimpleValueType) nVT; 1140 if (SVT.getVectorElementType() == EltVT 1141 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1142 TransformToType[i] = SVT; 1143 RegisterTypeForVT[i] = SVT; 1144 NumRegistersForVT[i] = 1; 1145 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1146 IsLegalWiderType = true; 1147 break; 1148 } 1149 } 1150 if (IsLegalWiderType) 1151 break; 1152 LLVM_FALLTHROUGH; 1153 1154 case TypeSplitVector: 1155 case TypeScalarizeVector: { 1156 MVT IntermediateVT; 1157 MVT RegisterVT; 1158 unsigned NumIntermediates; 1159 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1160 NumIntermediates, RegisterVT, this); 1161 RegisterTypeForVT[i] = RegisterVT; 1162 1163 MVT NVT = VT.getPow2VectorType(); 1164 if (NVT == VT) { 1165 // Type is already a power of 2. The default action is to split. 1166 TransformToType[i] = MVT::Other; 1167 if (PreferredAction == TypeScalarizeVector) 1168 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1169 else if (PreferredAction == TypeSplitVector) 1170 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1171 else 1172 // Set type action according to the number of elements. 1173 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1174 : TypeSplitVector); 1175 } else { 1176 TransformToType[i] = NVT; 1177 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1178 } 1179 break; 1180 } 1181 default: 1182 llvm_unreachable("Unknown vector legalization action!"); 1183 } 1184 } 1185 1186 // Determine the 'representative' register class for each value type. 1187 // An representative register class is the largest (meaning one which is 1188 // not a sub-register class / subreg register class) legal register class for 1189 // a group of value types. For example, on i386, i8, i16, and i32 1190 // representative would be GR32; while on x86_64 it's GR64. 1191 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1192 const TargetRegisterClass* RRC; 1193 uint8_t Cost; 1194 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1195 RepRegClassForVT[i] = RRC; 1196 RepRegClassCostForVT[i] = Cost; 1197 } 1198 } 1199 1200 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1201 EVT VT) const { 1202 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1203 return getPointerTy(DL).SimpleTy; 1204 } 1205 1206 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1207 return MVT::i32; // return the default value 1208 } 1209 1210 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1211 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1212 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1213 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1214 /// 1215 /// This method returns the number of registers needed, and the VT for each 1216 /// register. It also returns the VT and quantity of the intermediate values 1217 /// before they are promoted/expanded. 1218 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1219 EVT &IntermediateVT, 1220 unsigned &NumIntermediates, 1221 MVT &RegisterVT) const { 1222 unsigned NumElts = VT.getVectorNumElements(); 1223 1224 // If there is a wider vector type with the same element type as this one, 1225 // or a promoted vector type that has the same number of elements which 1226 // are wider, then we should convert to that legal vector type. 1227 // This handles things like <2 x float> -> <4 x float> and 1228 // <4 x i1> -> <4 x i32>. 1229 LegalizeTypeAction TA = getTypeAction(Context, VT); 1230 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1231 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1232 if (isTypeLegal(RegisterEVT)) { 1233 IntermediateVT = RegisterEVT; 1234 RegisterVT = RegisterEVT.getSimpleVT(); 1235 NumIntermediates = 1; 1236 return 1; 1237 } 1238 } 1239 1240 // Figure out the right, legal destination reg to copy into. 1241 EVT EltTy = VT.getVectorElementType(); 1242 1243 unsigned NumVectorRegs = 1; 1244 1245 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1246 // could break down into LHS/RHS like LegalizeDAG does. 1247 if (!isPowerOf2_32(NumElts)) { 1248 NumVectorRegs = NumElts; 1249 NumElts = 1; 1250 } 1251 1252 // Divide the input until we get to a supported size. This will always 1253 // end with a scalar if the target doesn't support vectors. 1254 while (NumElts > 1 && !isTypeLegal( 1255 EVT::getVectorVT(Context, EltTy, NumElts))) { 1256 NumElts >>= 1; 1257 NumVectorRegs <<= 1; 1258 } 1259 1260 NumIntermediates = NumVectorRegs; 1261 1262 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1263 if (!isTypeLegal(NewVT)) 1264 NewVT = EltTy; 1265 IntermediateVT = NewVT; 1266 1267 MVT DestVT = getRegisterType(Context, NewVT); 1268 RegisterVT = DestVT; 1269 unsigned NewVTSize = NewVT.getSizeInBits(); 1270 1271 // Convert sizes such as i33 to i64. 1272 if (!isPowerOf2_32(NewVTSize)) 1273 NewVTSize = NextPowerOf2(NewVTSize); 1274 1275 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1276 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1277 1278 // Otherwise, promotion or legal types use the same number of registers as 1279 // the vector decimated to the appropriate level. 1280 return NumVectorRegs; 1281 } 1282 1283 /// Get the EVTs and ArgFlags collections that represent the legalized return 1284 /// type of the given function. This does not require a DAG or a return value, 1285 /// and is suitable for use before any DAGs for the function are constructed. 1286 /// TODO: Move this out of TargetLowering.cpp. 1287 void llvm::GetReturnInfo(Type *ReturnType, AttributeList attr, 1288 SmallVectorImpl<ISD::OutputArg> &Outs, 1289 const TargetLowering &TLI, const DataLayout &DL) { 1290 SmallVector<EVT, 4> ValueVTs; 1291 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1292 unsigned NumValues = ValueVTs.size(); 1293 if (NumValues == 0) return; 1294 1295 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1296 EVT VT = ValueVTs[j]; 1297 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1298 1299 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1300 ExtendKind = ISD::SIGN_EXTEND; 1301 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1302 ExtendKind = ISD::ZERO_EXTEND; 1303 1304 // FIXME: C calling convention requires the return type to be promoted to 1305 // at least 32-bit. But this is not necessary for non-C calling 1306 // conventions. The frontend should mark functions whose return values 1307 // require promoting with signext or zeroext attributes. 1308 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1309 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1310 if (VT.bitsLT(MinVT)) 1311 VT = MinVT; 1312 } 1313 1314 unsigned NumParts = 1315 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT); 1316 MVT PartVT = 1317 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT); 1318 1319 // 'inreg' on function refers to return value 1320 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1321 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1322 Flags.setInReg(); 1323 1324 // Propagate extension type if any 1325 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1326 Flags.setSExt(); 1327 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1328 Flags.setZExt(); 1329 1330 for (unsigned i = 0; i < NumParts; ++i) 1331 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1332 } 1333 } 1334 1335 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1336 /// function arguments in the caller parameter area. This is the actual 1337 /// alignment, not its logarithm. 1338 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1339 const DataLayout &DL) const { 1340 return DL.getABITypeAlignment(Ty); 1341 } 1342 1343 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1344 const DataLayout &DL, EVT VT, 1345 unsigned AddrSpace, 1346 unsigned Alignment, 1347 bool *Fast) const { 1348 // Check if the specified alignment is sufficient based on the data layout. 1349 // TODO: While using the data layout works in practice, a better solution 1350 // would be to implement this check directly (make this a virtual function). 1351 // For example, the ABI alignment may change based on software platform while 1352 // this function should only be affected by hardware implementation. 1353 Type *Ty = VT.getTypeForEVT(Context); 1354 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1355 // Assume that an access that meets the ABI-specified alignment is fast. 1356 if (Fast != nullptr) 1357 *Fast = true; 1358 return true; 1359 } 1360 1361 // This is a misaligned access. 1362 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1363 } 1364 1365 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1366 return BranchProbability(MinPercentageForPredictableBranch, 100); 1367 } 1368 1369 //===----------------------------------------------------------------------===// 1370 // TargetTransformInfo Helpers 1371 //===----------------------------------------------------------------------===// 1372 1373 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1374 enum InstructionOpcodes { 1375 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1376 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1377 #include "llvm/IR/Instruction.def" 1378 }; 1379 switch (static_cast<InstructionOpcodes>(Opcode)) { 1380 case Ret: return 0; 1381 case Br: return 0; 1382 case Switch: return 0; 1383 case IndirectBr: return 0; 1384 case Invoke: return 0; 1385 case Resume: return 0; 1386 case Unreachable: return 0; 1387 case CleanupRet: return 0; 1388 case CatchRet: return 0; 1389 case CatchPad: return 0; 1390 case CatchSwitch: return 0; 1391 case CleanupPad: return 0; 1392 case Add: return ISD::ADD; 1393 case FAdd: return ISD::FADD; 1394 case Sub: return ISD::SUB; 1395 case FSub: return ISD::FSUB; 1396 case Mul: return ISD::MUL; 1397 case FMul: return ISD::FMUL; 1398 case UDiv: return ISD::UDIV; 1399 case SDiv: return ISD::SDIV; 1400 case FDiv: return ISD::FDIV; 1401 case URem: return ISD::UREM; 1402 case SRem: return ISD::SREM; 1403 case FRem: return ISD::FREM; 1404 case Shl: return ISD::SHL; 1405 case LShr: return ISD::SRL; 1406 case AShr: return ISD::SRA; 1407 case And: return ISD::AND; 1408 case Or: return ISD::OR; 1409 case Xor: return ISD::XOR; 1410 case Alloca: return 0; 1411 case Load: return ISD::LOAD; 1412 case Store: return ISD::STORE; 1413 case GetElementPtr: return 0; 1414 case Fence: return 0; 1415 case AtomicCmpXchg: return 0; 1416 case AtomicRMW: return 0; 1417 case Trunc: return ISD::TRUNCATE; 1418 case ZExt: return ISD::ZERO_EXTEND; 1419 case SExt: return ISD::SIGN_EXTEND; 1420 case FPToUI: return ISD::FP_TO_UINT; 1421 case FPToSI: return ISD::FP_TO_SINT; 1422 case UIToFP: return ISD::UINT_TO_FP; 1423 case SIToFP: return ISD::SINT_TO_FP; 1424 case FPTrunc: return ISD::FP_ROUND; 1425 case FPExt: return ISD::FP_EXTEND; 1426 case PtrToInt: return ISD::BITCAST; 1427 case IntToPtr: return ISD::BITCAST; 1428 case BitCast: return ISD::BITCAST; 1429 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1430 case ICmp: return ISD::SETCC; 1431 case FCmp: return ISD::SETCC; 1432 case PHI: return 0; 1433 case Call: return 0; 1434 case Select: return ISD::SELECT; 1435 case UserOp1: return 0; 1436 case UserOp2: return 0; 1437 case VAArg: return 0; 1438 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1439 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1440 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1441 case ExtractValue: return ISD::MERGE_VALUES; 1442 case InsertValue: return ISD::MERGE_VALUES; 1443 case LandingPad: return 0; 1444 } 1445 1446 llvm_unreachable("Unknown instruction type encountered!"); 1447 } 1448 1449 std::pair<int, MVT> 1450 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1451 Type *Ty) const { 1452 LLVMContext &C = Ty->getContext(); 1453 EVT MTy = getValueType(DL, Ty); 1454 1455 int Cost = 1; 1456 // We keep legalizing the type until we find a legal kind. We assume that 1457 // the only operation that costs anything is the split. After splitting 1458 // we need to handle two types. 1459 while (true) { 1460 LegalizeKind LK = getTypeConversion(C, MTy); 1461 1462 if (LK.first == TypeLegal) 1463 return std::make_pair(Cost, MTy.getSimpleVT()); 1464 1465 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1466 Cost *= 2; 1467 1468 // Do not loop with f128 type. 1469 if (MTy == LK.second) 1470 return std::make_pair(Cost, MTy.getSimpleVT()); 1471 1472 // Keep legalizing the type. 1473 MTy = LK.second; 1474 } 1475 } 1476 1477 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1478 bool UseTLS) const { 1479 // compiler-rt provides a variable with a magic name. Targets that do not 1480 // link with compiler-rt may also provide such a variable. 1481 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1482 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1483 auto UnsafeStackPtr = 1484 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1485 1486 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1487 1488 if (!UnsafeStackPtr) { 1489 auto TLSModel = UseTLS ? 1490 GlobalValue::InitialExecTLSModel : 1491 GlobalValue::NotThreadLocal; 1492 // The global variable is not defined yet, define it ourselves. 1493 // We use the initial-exec TLS model because we do not support the 1494 // variable living anywhere other than in the main executable. 1495 UnsafeStackPtr = new GlobalVariable( 1496 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1497 UnsafeStackPtrVar, nullptr, TLSModel); 1498 } else { 1499 // The variable exists, check its type and attributes. 1500 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1501 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1502 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1503 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1504 (UseTLS ? "" : "not ") + "be thread-local"); 1505 } 1506 return UnsafeStackPtr; 1507 } 1508 1509 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1510 if (!TM.getTargetTriple().isAndroid()) 1511 return getDefaultSafeStackPointerLocation(IRB, true); 1512 1513 // Android provides a libc function to retrieve the address of the current 1514 // thread's unsafe stack pointer. 1515 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1516 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1517 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1518 StackPtrTy->getPointerTo(0)); 1519 return IRB.CreateCall(Fn); 1520 } 1521 1522 //===----------------------------------------------------------------------===// 1523 // Loop Strength Reduction hooks 1524 //===----------------------------------------------------------------------===// 1525 1526 /// isLegalAddressingMode - Return true if the addressing mode represented 1527 /// by AM is legal for this target, for a load/store of the specified type. 1528 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1529 const AddrMode &AM, Type *Ty, 1530 unsigned AS, Instruction *I) const { 1531 // The default implementation of this implements a conservative RISCy, r+r and 1532 // r+i addr mode. 1533 1534 // Allows a sign-extended 16-bit immediate field. 1535 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1536 return false; 1537 1538 // No global is ever allowed as a base. 1539 if (AM.BaseGV) 1540 return false; 1541 1542 // Only support r+r, 1543 switch (AM.Scale) { 1544 case 0: // "r+i" or just "i", depending on HasBaseReg. 1545 break; 1546 case 1: 1547 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1548 return false; 1549 // Otherwise we have r+r or r+i. 1550 break; 1551 case 2: 1552 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1553 return false; 1554 // Allow 2*r as r+r. 1555 break; 1556 default: // Don't allow n * r 1557 return false; 1558 } 1559 1560 return true; 1561 } 1562 1563 //===----------------------------------------------------------------------===// 1564 // Stack Protector 1565 //===----------------------------------------------------------------------===// 1566 1567 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1568 // so that SelectionDAG handle SSP. 1569 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1570 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1571 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1572 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1573 return M.getOrInsertGlobal("__guard_local", PtrTy); 1574 } 1575 return nullptr; 1576 } 1577 1578 // Currently only support "standard" __stack_chk_guard. 1579 // TODO: add LOAD_STACK_GUARD support. 1580 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1581 M.getOrInsertGlobal("__stack_chk_guard", Type::getInt8PtrTy(M.getContext())); 1582 } 1583 1584 // Currently only support "standard" __stack_chk_guard. 1585 // TODO: add LOAD_STACK_GUARD support. 1586 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1587 return M.getGlobalVariable("__stack_chk_guard", true); 1588 } 1589 1590 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1591 return nullptr; 1592 } 1593 1594 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1595 return MinimumJumpTableEntries; 1596 } 1597 1598 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1599 MinimumJumpTableEntries = Val; 1600 } 1601 1602 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1603 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1604 } 1605 1606 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1607 return MaximumJumpTableSize; 1608 } 1609 1610 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1611 MaximumJumpTableSize = Val; 1612 } 1613 1614 //===----------------------------------------------------------------------===// 1615 // Reciprocal Estimates 1616 //===----------------------------------------------------------------------===// 1617 1618 /// Get the reciprocal estimate attribute string for a function that will 1619 /// override the target defaults. 1620 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1621 const Function &F = MF.getFunction(); 1622 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1623 } 1624 1625 /// Construct a string for the given reciprocal operation of the given type. 1626 /// This string should match the corresponding option to the front-end's 1627 /// "-mrecip" flag assuming those strings have been passed through in an 1628 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1629 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1630 std::string Name = VT.isVector() ? "vec-" : ""; 1631 1632 Name += IsSqrt ? "sqrt" : "div"; 1633 1634 // TODO: Handle "half" or other float types? 1635 if (VT.getScalarType() == MVT::f64) { 1636 Name += "d"; 1637 } else { 1638 assert(VT.getScalarType() == MVT::f32 && 1639 "Unexpected FP type for reciprocal estimate"); 1640 Name += "f"; 1641 } 1642 1643 return Name; 1644 } 1645 1646 /// Return the character position and value (a single numeric character) of a 1647 /// customized refinement operation in the input string if it exists. Return 1648 /// false if there is no customized refinement step count. 1649 static bool parseRefinementStep(StringRef In, size_t &Position, 1650 uint8_t &Value) { 1651 const char RefStepToken = ':'; 1652 Position = In.find(RefStepToken); 1653 if (Position == StringRef::npos) 1654 return false; 1655 1656 StringRef RefStepString = In.substr(Position + 1); 1657 // Allow exactly one numeric character for the additional refinement 1658 // step parameter. 1659 if (RefStepString.size() == 1) { 1660 char RefStepChar = RefStepString[0]; 1661 if (RefStepChar >= '0' && RefStepChar <= '9') { 1662 Value = RefStepChar - '0'; 1663 return true; 1664 } 1665 } 1666 report_fatal_error("Invalid refinement step for -recip."); 1667 } 1668 1669 /// For the input attribute string, return one of the ReciprocalEstimate enum 1670 /// status values (enabled, disabled, or not specified) for this operation on 1671 /// the specified data type. 1672 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1673 if (Override.empty()) 1674 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1675 1676 SmallVector<StringRef, 4> OverrideVector; 1677 SplitString(Override, OverrideVector, ","); 1678 unsigned NumArgs = OverrideVector.size(); 1679 1680 // Check if "all", "none", or "default" was specified. 1681 if (NumArgs == 1) { 1682 // Look for an optional setting of the number of refinement steps needed 1683 // for this type of reciprocal operation. 1684 size_t RefPos; 1685 uint8_t RefSteps; 1686 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1687 // Split the string for further processing. 1688 Override = Override.substr(0, RefPos); 1689 } 1690 1691 // All reciprocal types are enabled. 1692 if (Override == "all") 1693 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1694 1695 // All reciprocal types are disabled. 1696 if (Override == "none") 1697 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1698 1699 // Target defaults for enablement are used. 1700 if (Override == "default") 1701 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1702 } 1703 1704 // The attribute string may omit the size suffix ('f'/'d'). 1705 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1706 std::string VTNameNoSize = VTName; 1707 VTNameNoSize.pop_back(); 1708 static const char DisabledPrefix = '!'; 1709 1710 for (StringRef RecipType : OverrideVector) { 1711 size_t RefPos; 1712 uint8_t RefSteps; 1713 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1714 RecipType = RecipType.substr(0, RefPos); 1715 1716 // Ignore the disablement token for string matching. 1717 bool IsDisabled = RecipType[0] == DisabledPrefix; 1718 if (IsDisabled) 1719 RecipType = RecipType.substr(1); 1720 1721 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1722 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1723 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1724 } 1725 1726 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1727 } 1728 1729 /// For the input attribute string, return the customized refinement step count 1730 /// for this operation on the specified data type. If the step count does not 1731 /// exist, return the ReciprocalEstimate enum value for unspecified. 1732 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1733 if (Override.empty()) 1734 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1735 1736 SmallVector<StringRef, 4> OverrideVector; 1737 SplitString(Override, OverrideVector, ","); 1738 unsigned NumArgs = OverrideVector.size(); 1739 1740 // Check if "all", "default", or "none" was specified. 1741 if (NumArgs == 1) { 1742 // Look for an optional setting of the number of refinement steps needed 1743 // for this type of reciprocal operation. 1744 size_t RefPos; 1745 uint8_t RefSteps; 1746 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1747 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1748 1749 // Split the string for further processing. 1750 Override = Override.substr(0, RefPos); 1751 assert(Override != "none" && 1752 "Disabled reciprocals, but specifed refinement steps?"); 1753 1754 // If this is a general override, return the specified number of steps. 1755 if (Override == "all" || Override == "default") 1756 return RefSteps; 1757 } 1758 1759 // The attribute string may omit the size suffix ('f'/'d'). 1760 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1761 std::string VTNameNoSize = VTName; 1762 VTNameNoSize.pop_back(); 1763 1764 for (StringRef RecipType : OverrideVector) { 1765 size_t RefPos; 1766 uint8_t RefSteps; 1767 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1768 continue; 1769 1770 RecipType = RecipType.substr(0, RefPos); 1771 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1772 return RefSteps; 1773 } 1774 1775 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1776 } 1777 1778 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1779 MachineFunction &MF) const { 1780 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1781 } 1782 1783 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1784 MachineFunction &MF) const { 1785 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1786 } 1787 1788 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1789 MachineFunction &MF) const { 1790 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1791 } 1792 1793 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1794 MachineFunction &MF) const { 1795 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1796 } 1797 1798 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1799 MF.getRegInfo().freezeReservedRegs(MF); 1800 } 1801