| 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/DenseMap.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/Twine.h" |
| 20 | #include "llvm/Analysis/Loads.h" |
| 21 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 22 | #include "llvm/CodeGen/Analysis.h" |
| 23 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 24 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 26 | #include "llvm/CodeGen/MachineFunction.h" |
| 27 | #include "llvm/CodeGen/MachineInstr.h" |
| 28 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 29 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 30 | #include "llvm/CodeGen/MachineOperand.h" |
| 31 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 32 | #include "llvm/CodeGen/RuntimeLibcallUtil.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/CodeGenTypes/MachineValueType.h" |
| 39 | #include "llvm/IR/Attributes.h" |
| 40 | #include "llvm/IR/CallingConv.h" |
| 41 | #include "llvm/IR/DataLayout.h" |
| 42 | #include "llvm/IR/DerivedTypes.h" |
| 43 | #include "llvm/IR/Function.h" |
| 44 | #include "llvm/IR/GlobalValue.h" |
| 45 | #include "llvm/IR/GlobalVariable.h" |
| 46 | #include "llvm/IR/IRBuilder.h" |
| 47 | #include "llvm/IR/Module.h" |
| 48 | #include "llvm/IR/Type.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 "llvm/Target/TargetOptions.h" |
| 56 | #include "llvm/TargetParser/Triple.h" |
| 57 | #include "llvm/Transforms/Utils/SizeOpts.h" |
| 58 | #include <algorithm> |
| 59 | #include <cassert> |
| 60 | #include <cstdint> |
| 61 | #include <cstring> |
| 62 | #include <string> |
| 63 | #include <tuple> |
| 64 | #include <utility> |
| 65 | |
| 66 | using namespace llvm; |
| 67 | |
| 68 | static cl::opt<bool> JumpIsExpensiveOverride( |
| 69 | "jump-is-expensive" , cl::init(Val: false), |
| 70 | cl::desc("Do not create extra branches to split comparison logic." ), |
| 71 | cl::Hidden); |
| 72 | |
| 73 | static cl::opt<unsigned> MinimumJumpTableEntries |
| 74 | ("min-jump-table-entries" , cl::init(Val: 4), cl::Hidden, |
| 75 | cl::desc("Set minimum number of entries to use a jump table." )); |
| 76 | |
| 77 | static cl::opt<unsigned> MaximumJumpTableSize |
| 78 | ("max-jump-table-size" , cl::init(UINT_MAX), cl::Hidden, |
| 79 | cl::desc("Set maximum size of jump tables." )); |
| 80 | |
| 81 | /// Minimum jump table density for normal functions. |
| 82 | static cl::opt<unsigned> |
| 83 | JumpTableDensity("jump-table-density" , cl::init(Val: 10), cl::Hidden, |
| 84 | cl::desc("Minimum density for building a jump table in " |
| 85 | "a normal function" )); |
| 86 | |
| 87 | /// Minimum jump table density for -Os or -Oz functions. |
| 88 | static cl::opt<unsigned> OptsizeJumpTableDensity( |
| 89 | "optsize-jump-table-density" , cl::init(Val: 40), cl::Hidden, |
| 90 | cl::desc("Minimum density for building a jump table in " |
| 91 | "an optsize function" )); |
| 92 | |
| 93 | static cl::opt<unsigned> MinimumBitTestCmpsOverride( |
| 94 | "min-bit-test-cmps" , cl::init(Val: 2), cl::Hidden, |
| 95 | cl::desc("Set minimum of largest number of comparisons " |
| 96 | "to use bit test for switch." )); |
| 97 | |
| 98 | static cl::opt<unsigned> MaxStoresPerMemsetOverride( |
| 99 | "max-store-memset" , cl::init(Val: 0), cl::Hidden, |
| 100 | cl::desc("Override target's MaxStoresPerMemset and " |
| 101 | "MaxStoresPerMemsetOptSize. " |
| 102 | "Set to 0 to use the target default." )); |
| 103 | |
| 104 | static cl::opt<unsigned> MaxStoresPerMemcpyOverride( |
| 105 | "max-store-memcpy" , cl::init(Val: 0), cl::Hidden, |
| 106 | cl::desc("Override target's MaxStoresPerMemcpy and " |
| 107 | "MaxStoresPerMemcpyOptSize. " |
| 108 | "Set to 0 to use the target default." )); |
| 109 | |
| 110 | static cl::opt<unsigned> MaxStoresPerMemmoveOverride( |
| 111 | "max-store-memmove" , cl::init(Val: 0), cl::Hidden, |
| 112 | cl::desc("Override target's MaxStoresPerMemmove and " |
| 113 | "MaxStoresPerMemmoveOptSize. " |
| 114 | "Set to 0 to use the target default." )); |
| 115 | |
| 116 | // FIXME: This option is only to test if the strict fp operation processed |
| 117 | // correctly by preventing mutating strict fp operation to normal fp operation |
| 118 | // during development. When the backend supports strict float operation, this |
| 119 | // option will be meaningless. |
| 120 | static cl::opt<bool> DisableStrictNodeMutation("disable-strictnode-mutation" , |
| 121 | cl::desc("Don't mutate strict-float node to a legalize node" ), |
| 122 | cl::init(Val: false), cl::Hidden); |
| 123 | |
| 124 | LLVM_ABI RTLIB::Libcall RTLIB::getSHL(EVT VT) { |
| 125 | if (VT == MVT::i16) |
| 126 | return RTLIB::SHL_I16; |
| 127 | if (VT == MVT::i32) |
| 128 | return RTLIB::SHL_I32; |
| 129 | if (VT == MVT::i64) |
| 130 | return RTLIB::SHL_I64; |
| 131 | if (VT == MVT::i128) |
| 132 | return RTLIB::SHL_I128; |
| 133 | |
| 134 | return RTLIB::UNKNOWN_LIBCALL; |
| 135 | } |
| 136 | |
| 137 | LLVM_ABI RTLIB::Libcall RTLIB::getSRL(EVT VT) { |
| 138 | if (VT == MVT::i16) |
| 139 | return RTLIB::SRL_I16; |
| 140 | if (VT == MVT::i32) |
| 141 | return RTLIB::SRL_I32; |
| 142 | if (VT == MVT::i64) |
| 143 | return RTLIB::SRL_I64; |
| 144 | if (VT == MVT::i128) |
| 145 | return RTLIB::SRL_I128; |
| 146 | |
| 147 | return RTLIB::UNKNOWN_LIBCALL; |
| 148 | } |
| 149 | |
| 150 | LLVM_ABI RTLIB::Libcall RTLIB::getSRA(EVT VT) { |
| 151 | if (VT == MVT::i16) |
| 152 | return RTLIB::SRA_I16; |
| 153 | if (VT == MVT::i32) |
| 154 | return RTLIB::SRA_I32; |
| 155 | if (VT == MVT::i64) |
| 156 | return RTLIB::SRA_I64; |
| 157 | if (VT == MVT::i128) |
| 158 | return RTLIB::SRA_I128; |
| 159 | |
| 160 | return RTLIB::UNKNOWN_LIBCALL; |
| 161 | } |
| 162 | |
| 163 | LLVM_ABI RTLIB::Libcall RTLIB::getMUL(EVT VT) { |
| 164 | if (VT == MVT::i16) |
| 165 | return RTLIB::MUL_I16; |
| 166 | if (VT == MVT::i32) |
| 167 | return RTLIB::MUL_I32; |
| 168 | if (VT == MVT::i64) |
| 169 | return RTLIB::MUL_I64; |
| 170 | if (VT == MVT::i128) |
| 171 | return RTLIB::MUL_I128; |
| 172 | return RTLIB::UNKNOWN_LIBCALL; |
| 173 | } |
| 174 | |
| 175 | LLVM_ABI RTLIB::Libcall RTLIB::getMULO(EVT VT) { |
| 176 | if (VT == MVT::i32) |
| 177 | return RTLIB::MULO_I32; |
| 178 | if (VT == MVT::i64) |
| 179 | return RTLIB::MULO_I64; |
| 180 | if (VT == MVT::i128) |
| 181 | return RTLIB::MULO_I128; |
| 182 | return RTLIB::UNKNOWN_LIBCALL; |
| 183 | } |
| 184 | |
| 185 | LLVM_ABI RTLIB::Libcall RTLIB::getSDIV(EVT VT) { |
| 186 | if (VT == MVT::i16) |
| 187 | return RTLIB::SDIV_I16; |
| 188 | if (VT == MVT::i32) |
| 189 | return RTLIB::SDIV_I32; |
| 190 | if (VT == MVT::i64) |
| 191 | return RTLIB::SDIV_I64; |
| 192 | if (VT == MVT::i128) |
| 193 | return RTLIB::SDIV_I128; |
| 194 | return RTLIB::UNKNOWN_LIBCALL; |
| 195 | } |
| 196 | |
| 197 | LLVM_ABI RTLIB::Libcall RTLIB::getUDIV(EVT VT) { |
| 198 | if (VT == MVT::i16) |
| 199 | return RTLIB::UDIV_I16; |
| 200 | if (VT == MVT::i32) |
| 201 | return RTLIB::UDIV_I32; |
| 202 | if (VT == MVT::i64) |
| 203 | return RTLIB::UDIV_I64; |
| 204 | if (VT == MVT::i128) |
| 205 | return RTLIB::UDIV_I128; |
| 206 | return RTLIB::UNKNOWN_LIBCALL; |
| 207 | } |
| 208 | |
| 209 | LLVM_ABI RTLIB::Libcall RTLIB::getSREM(EVT VT) { |
| 210 | if (VT == MVT::i16) |
| 211 | return RTLIB::SREM_I16; |
| 212 | if (VT == MVT::i32) |
| 213 | return RTLIB::SREM_I32; |
| 214 | if (VT == MVT::i64) |
| 215 | return RTLIB::SREM_I64; |
| 216 | if (VT == MVT::i128) |
| 217 | return RTLIB::SREM_I128; |
| 218 | return RTLIB::UNKNOWN_LIBCALL; |
| 219 | } |
| 220 | |
| 221 | LLVM_ABI RTLIB::Libcall RTLIB::getUREM(EVT VT) { |
| 222 | if (VT == MVT::i16) |
| 223 | return RTLIB::UREM_I16; |
| 224 | if (VT == MVT::i32) |
| 225 | return RTLIB::UREM_I32; |
| 226 | if (VT == MVT::i64) |
| 227 | return RTLIB::UREM_I64; |
| 228 | if (VT == MVT::i128) |
| 229 | return RTLIB::UREM_I128; |
| 230 | return RTLIB::UNKNOWN_LIBCALL; |
| 231 | } |
| 232 | |
| 233 | LLVM_ABI RTLIB::Libcall RTLIB::getCTPOP(EVT VT) { |
| 234 | if (VT == MVT::i32) |
| 235 | return RTLIB::CTPOP_I32; |
| 236 | if (VT == MVT::i64) |
| 237 | return RTLIB::CTPOP_I64; |
| 238 | if (VT == MVT::i128) |
| 239 | return RTLIB::CTPOP_I128; |
| 240 | return RTLIB::UNKNOWN_LIBCALL; |
| 241 | } |
| 242 | |
| 243 | /// GetFPLibCall - Helper to return the right libcall for the given floating |
| 244 | /// point type, or UNKNOWN_LIBCALL if there is none. |
| 245 | RTLIB::Libcall RTLIB::getFPLibCall(EVT VT, |
| 246 | RTLIB::Libcall Call_F32, |
| 247 | RTLIB::Libcall Call_F64, |
| 248 | RTLIB::Libcall Call_F80, |
| 249 | RTLIB::Libcall Call_F128, |
| 250 | RTLIB::Libcall Call_PPCF128) { |
| 251 | return |
| 252 | VT == MVT::f32 ? Call_F32 : |
| 253 | VT == MVT::f64 ? Call_F64 : |
| 254 | VT == MVT::f80 ? Call_F80 : |
| 255 | VT == MVT::f128 ? Call_F128 : |
| 256 | VT == MVT::ppcf128 ? Call_PPCF128 : |
| 257 | RTLIB::UNKNOWN_LIBCALL; |
| 258 | } |
| 259 | |
| 260 | /// getFPEXT - Return the FPEXT_*_* value for the given types, or |
| 261 | /// UNKNOWN_LIBCALL if there is none. |
| 262 | RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { |
| 263 | if (OpVT == MVT::f16) { |
| 264 | if (RetVT == MVT::f32) |
| 265 | return FPEXT_F16_F32; |
| 266 | if (RetVT == MVT::f64) |
| 267 | return FPEXT_F16_F64; |
| 268 | if (RetVT == MVT::f80) |
| 269 | return FPEXT_F16_F80; |
| 270 | if (RetVT == MVT::f128) |
| 271 | return FPEXT_F16_F128; |
| 272 | } else if (OpVT == MVT::f32) { |
| 273 | if (RetVT == MVT::f64) |
| 274 | return FPEXT_F32_F64; |
| 275 | if (RetVT == MVT::f128) |
| 276 | return FPEXT_F32_F128; |
| 277 | if (RetVT == MVT::ppcf128) |
| 278 | return FPEXT_F32_PPCF128; |
| 279 | } else if (OpVT == MVT::f64) { |
| 280 | if (RetVT == MVT::f128) |
| 281 | return FPEXT_F64_F128; |
| 282 | else if (RetVT == MVT::ppcf128) |
| 283 | return FPEXT_F64_PPCF128; |
| 284 | } else if (OpVT == MVT::f80) { |
| 285 | if (RetVT == MVT::f128) |
| 286 | return FPEXT_F80_F128; |
| 287 | } else if (OpVT == MVT::bf16) { |
| 288 | if (RetVT == MVT::f32) |
| 289 | return FPEXT_BF16_F32; |
| 290 | } |
| 291 | |
| 292 | return UNKNOWN_LIBCALL; |
| 293 | } |
| 294 | |
| 295 | /// getFPROUND - Return the FPROUND_*_* value for the given types, or |
| 296 | /// UNKNOWN_LIBCALL if there is none. |
| 297 | RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { |
| 298 | if (RetVT == MVT::f16) { |
| 299 | if (OpVT == MVT::f32) |
| 300 | return FPROUND_F32_F16; |
| 301 | if (OpVT == MVT::f64) |
| 302 | return FPROUND_F64_F16; |
| 303 | if (OpVT == MVT::f80) |
| 304 | return FPROUND_F80_F16; |
| 305 | if (OpVT == MVT::f128) |
| 306 | return FPROUND_F128_F16; |
| 307 | if (OpVT == MVT::ppcf128) |
| 308 | return FPROUND_PPCF128_F16; |
| 309 | } else if (RetVT == MVT::bf16) { |
| 310 | if (OpVT == MVT::f32) |
| 311 | return FPROUND_F32_BF16; |
| 312 | if (OpVT == MVT::f64) |
| 313 | return FPROUND_F64_BF16; |
| 314 | if (OpVT == MVT::f80) |
| 315 | return FPROUND_F80_BF16; |
| 316 | if (OpVT == MVT::f128) |
| 317 | return FPROUND_F128_BF16; |
| 318 | } else if (RetVT == MVT::f32) { |
| 319 | if (OpVT == MVT::f64) |
| 320 | return FPROUND_F64_F32; |
| 321 | if (OpVT == MVT::f80) |
| 322 | return FPROUND_F80_F32; |
| 323 | if (OpVT == MVT::f128) |
| 324 | return FPROUND_F128_F32; |
| 325 | if (OpVT == MVT::ppcf128) |
| 326 | return FPROUND_PPCF128_F32; |
| 327 | } else if (RetVT == MVT::f64) { |
| 328 | if (OpVT == MVT::f80) |
| 329 | return FPROUND_F80_F64; |
| 330 | if (OpVT == MVT::f128) |
| 331 | return FPROUND_F128_F64; |
| 332 | if (OpVT == MVT::ppcf128) |
| 333 | return FPROUND_PPCF128_F64; |
| 334 | } else if (RetVT == MVT::f80) { |
| 335 | if (OpVT == MVT::f128) |
| 336 | return FPROUND_F128_F80; |
| 337 | } |
| 338 | |
| 339 | return UNKNOWN_LIBCALL; |
| 340 | } |
| 341 | |
| 342 | /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or |
| 343 | /// UNKNOWN_LIBCALL if there is none. |
| 344 | RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { |
| 345 | if (OpVT == MVT::f16) { |
| 346 | if (RetVT == MVT::i32) |
| 347 | return FPTOSINT_F16_I32; |
| 348 | if (RetVT == MVT::i64) |
| 349 | return FPTOSINT_F16_I64; |
| 350 | if (RetVT == MVT::i128) |
| 351 | return FPTOSINT_F16_I128; |
| 352 | } else if (OpVT == MVT::f32) { |
| 353 | if (RetVT == MVT::i32) |
| 354 | return FPTOSINT_F32_I32; |
| 355 | if (RetVT == MVT::i64) |
| 356 | return FPTOSINT_F32_I64; |
| 357 | if (RetVT == MVT::i128) |
| 358 | return FPTOSINT_F32_I128; |
| 359 | } else if (OpVT == MVT::f64) { |
| 360 | if (RetVT == MVT::i32) |
| 361 | return FPTOSINT_F64_I32; |
| 362 | if (RetVT == MVT::i64) |
| 363 | return FPTOSINT_F64_I64; |
| 364 | if (RetVT == MVT::i128) |
| 365 | return FPTOSINT_F64_I128; |
| 366 | } else if (OpVT == MVT::f80) { |
| 367 | if (RetVT == MVT::i32) |
| 368 | return FPTOSINT_F80_I32; |
| 369 | if (RetVT == MVT::i64) |
| 370 | return FPTOSINT_F80_I64; |
| 371 | if (RetVT == MVT::i128) |
| 372 | return FPTOSINT_F80_I128; |
| 373 | } else if (OpVT == MVT::f128) { |
| 374 | if (RetVT == MVT::i32) |
| 375 | return FPTOSINT_F128_I32; |
| 376 | if (RetVT == MVT::i64) |
| 377 | return FPTOSINT_F128_I64; |
| 378 | if (RetVT == MVT::i128) |
| 379 | return FPTOSINT_F128_I128; |
| 380 | } else if (OpVT == MVT::ppcf128) { |
| 381 | if (RetVT == MVT::i32) |
| 382 | return FPTOSINT_PPCF128_I32; |
| 383 | if (RetVT == MVT::i64) |
| 384 | return FPTOSINT_PPCF128_I64; |
| 385 | if (RetVT == MVT::i128) |
| 386 | return FPTOSINT_PPCF128_I128; |
| 387 | } |
| 388 | return UNKNOWN_LIBCALL; |
| 389 | } |
| 390 | |
| 391 | /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or |
| 392 | /// UNKNOWN_LIBCALL if there is none. |
| 393 | RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { |
| 394 | if (OpVT == MVT::f16) { |
| 395 | if (RetVT == MVT::i32) |
| 396 | return FPTOUINT_F16_I32; |
| 397 | if (RetVT == MVT::i64) |
| 398 | return FPTOUINT_F16_I64; |
| 399 | if (RetVT == MVT::i128) |
| 400 | return FPTOUINT_F16_I128; |
| 401 | } else if (OpVT == MVT::f32) { |
| 402 | if (RetVT == MVT::i32) |
| 403 | return FPTOUINT_F32_I32; |
| 404 | if (RetVT == MVT::i64) |
| 405 | return FPTOUINT_F32_I64; |
| 406 | if (RetVT == MVT::i128) |
| 407 | return FPTOUINT_F32_I128; |
| 408 | } else if (OpVT == MVT::f64) { |
| 409 | if (RetVT == MVT::i32) |
| 410 | return FPTOUINT_F64_I32; |
| 411 | if (RetVT == MVT::i64) |
| 412 | return FPTOUINT_F64_I64; |
| 413 | if (RetVT == MVT::i128) |
| 414 | return FPTOUINT_F64_I128; |
| 415 | } else if (OpVT == MVT::f80) { |
| 416 | if (RetVT == MVT::i32) |
| 417 | return FPTOUINT_F80_I32; |
| 418 | if (RetVT == MVT::i64) |
| 419 | return FPTOUINT_F80_I64; |
| 420 | if (RetVT == MVT::i128) |
| 421 | return FPTOUINT_F80_I128; |
| 422 | } else if (OpVT == MVT::f128) { |
| 423 | if (RetVT == MVT::i32) |
| 424 | return FPTOUINT_F128_I32; |
| 425 | if (RetVT == MVT::i64) |
| 426 | return FPTOUINT_F128_I64; |
| 427 | if (RetVT == MVT::i128) |
| 428 | return FPTOUINT_F128_I128; |
| 429 | } else if (OpVT == MVT::ppcf128) { |
| 430 | if (RetVT == MVT::i32) |
| 431 | return FPTOUINT_PPCF128_I32; |
| 432 | if (RetVT == MVT::i64) |
| 433 | return FPTOUINT_PPCF128_I64; |
| 434 | if (RetVT == MVT::i128) |
| 435 | return FPTOUINT_PPCF128_I128; |
| 436 | } |
| 437 | return UNKNOWN_LIBCALL; |
| 438 | } |
| 439 | |
| 440 | /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or |
| 441 | /// UNKNOWN_LIBCALL if there is none. |
| 442 | RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { |
| 443 | if (OpVT == MVT::i32) { |
| 444 | if (RetVT == MVT::f16) |
| 445 | return SINTTOFP_I32_F16; |
| 446 | if (RetVT == MVT::f32) |
| 447 | return SINTTOFP_I32_F32; |
| 448 | if (RetVT == MVT::f64) |
| 449 | return SINTTOFP_I32_F64; |
| 450 | if (RetVT == MVT::f80) |
| 451 | return SINTTOFP_I32_F80; |
| 452 | if (RetVT == MVT::f128) |
| 453 | return SINTTOFP_I32_F128; |
| 454 | if (RetVT == MVT::ppcf128) |
| 455 | return SINTTOFP_I32_PPCF128; |
| 456 | } else if (OpVT == MVT::i64) { |
| 457 | if (RetVT == MVT::bf16) |
| 458 | return SINTTOFP_I64_BF16; |
| 459 | if (RetVT == MVT::f16) |
| 460 | return SINTTOFP_I64_F16; |
| 461 | if (RetVT == MVT::f32) |
| 462 | return SINTTOFP_I64_F32; |
| 463 | if (RetVT == MVT::f64) |
| 464 | return SINTTOFP_I64_F64; |
| 465 | if (RetVT == MVT::f80) |
| 466 | return SINTTOFP_I64_F80; |
| 467 | if (RetVT == MVT::f128) |
| 468 | return SINTTOFP_I64_F128; |
| 469 | if (RetVT == MVT::ppcf128) |
| 470 | return SINTTOFP_I64_PPCF128; |
| 471 | } else if (OpVT == MVT::i128) { |
| 472 | if (RetVT == MVT::f16) |
| 473 | return SINTTOFP_I128_F16; |
| 474 | if (RetVT == MVT::f32) |
| 475 | return SINTTOFP_I128_F32; |
| 476 | if (RetVT == MVT::f64) |
| 477 | return SINTTOFP_I128_F64; |
| 478 | if (RetVT == MVT::f80) |
| 479 | return SINTTOFP_I128_F80; |
| 480 | if (RetVT == MVT::f128) |
| 481 | return SINTTOFP_I128_F128; |
| 482 | if (RetVT == MVT::ppcf128) |
| 483 | return SINTTOFP_I128_PPCF128; |
| 484 | } |
| 485 | return UNKNOWN_LIBCALL; |
| 486 | } |
| 487 | |
| 488 | /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or |
| 489 | /// UNKNOWN_LIBCALL if there is none. |
| 490 | RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { |
| 491 | if (OpVT == MVT::i32) { |
| 492 | if (RetVT == MVT::f16) |
| 493 | return UINTTOFP_I32_F16; |
| 494 | if (RetVT == MVT::f32) |
| 495 | return UINTTOFP_I32_F32; |
| 496 | if (RetVT == MVT::f64) |
| 497 | return UINTTOFP_I32_F64; |
| 498 | if (RetVT == MVT::f80) |
| 499 | return UINTTOFP_I32_F80; |
| 500 | if (RetVT == MVT::f128) |
| 501 | return UINTTOFP_I32_F128; |
| 502 | if (RetVT == MVT::ppcf128) |
| 503 | return UINTTOFP_I32_PPCF128; |
| 504 | } else if (OpVT == MVT::i64) { |
| 505 | if (RetVT == MVT::bf16) |
| 506 | return UINTTOFP_I64_BF16; |
| 507 | if (RetVT == MVT::f16) |
| 508 | return UINTTOFP_I64_F16; |
| 509 | if (RetVT == MVT::f32) |
| 510 | return UINTTOFP_I64_F32; |
| 511 | if (RetVT == MVT::f64) |
| 512 | return UINTTOFP_I64_F64; |
| 513 | if (RetVT == MVT::f80) |
| 514 | return UINTTOFP_I64_F80; |
| 515 | if (RetVT == MVT::f128) |
| 516 | return UINTTOFP_I64_F128; |
| 517 | if (RetVT == MVT::ppcf128) |
| 518 | return UINTTOFP_I64_PPCF128; |
| 519 | } else if (OpVT == MVT::i128) { |
| 520 | if (RetVT == MVT::f16) |
| 521 | return UINTTOFP_I128_F16; |
| 522 | if (RetVT == MVT::f32) |
| 523 | return UINTTOFP_I128_F32; |
| 524 | if (RetVT == MVT::f64) |
| 525 | return UINTTOFP_I128_F64; |
| 526 | if (RetVT == MVT::f80) |
| 527 | return UINTTOFP_I128_F80; |
| 528 | if (RetVT == MVT::f128) |
| 529 | return UINTTOFP_I128_F128; |
| 530 | if (RetVT == MVT::ppcf128) |
| 531 | return UINTTOFP_I128_PPCF128; |
| 532 | } |
| 533 | return UNKNOWN_LIBCALL; |
| 534 | } |
| 535 | |
| 536 | RTLIB::Libcall RTLIB::getPOWI(EVT RetVT) { |
| 537 | return getFPLibCall(VT: RetVT, Call_F32: POWI_F32, Call_F64: POWI_F64, Call_F80: POWI_F80, Call_F128: POWI_F128, |
| 538 | Call_PPCF128: POWI_PPCF128); |
| 539 | } |
| 540 | |
| 541 | RTLIB::Libcall RTLIB::getPOW(EVT RetVT) { |
| 542 | return getFPLibCall(VT: RetVT, Call_F32: POW_F32, Call_F64: POW_F64, Call_F80: POW_F80, Call_F128: POW_F128, Call_PPCF128: POW_PPCF128); |
| 543 | } |
| 544 | |
| 545 | RTLIB::Libcall RTLIB::getLDEXP(EVT RetVT) { |
| 546 | return getFPLibCall(VT: RetVT, Call_F32: LDEXP_F32, Call_F64: LDEXP_F64, Call_F80: LDEXP_F80, Call_F128: LDEXP_F128, |
| 547 | Call_PPCF128: LDEXP_PPCF128); |
| 548 | } |
| 549 | |
| 550 | RTLIB::Libcall RTLIB::getFREXP(EVT RetVT) { |
| 551 | return getFPLibCall(VT: RetVT, Call_F32: FREXP_F32, Call_F64: FREXP_F64, Call_F80: FREXP_F80, Call_F128: FREXP_F128, |
| 552 | Call_PPCF128: FREXP_PPCF128); |
| 553 | } |
| 554 | |
| 555 | RTLIB::Libcall RTLIB::getSIN(EVT RetVT) { |
| 556 | return getFPLibCall(VT: RetVT, Call_F32: SIN_F32, Call_F64: SIN_F64, Call_F80: SIN_F80, Call_F128: SIN_F128, Call_PPCF128: SIN_PPCF128); |
| 557 | } |
| 558 | |
| 559 | RTLIB::Libcall RTLIB::getCOS(EVT RetVT) { |
| 560 | return getFPLibCall(VT: RetVT, Call_F32: COS_F32, Call_F64: COS_F64, Call_F80: COS_F80, Call_F128: COS_F128, Call_PPCF128: COS_PPCF128); |
| 561 | } |
| 562 | |
| 563 | RTLIB::Libcall RTLIB::getSINCOS(EVT RetVT) { |
| 564 | // TODO: Tablegen should generate this function |
| 565 | if (RetVT.isVector()) { |
| 566 | if (!RetVT.isSimple()) |
| 567 | return RTLIB::UNKNOWN_LIBCALL; |
| 568 | switch (RetVT.getSimpleVT().SimpleTy) { |
| 569 | case MVT::v4f32: |
| 570 | return RTLIB::SINCOS_V4F32; |
| 571 | case MVT::v2f64: |
| 572 | return RTLIB::SINCOS_V2F64; |
| 573 | case MVT::nxv4f32: |
| 574 | return RTLIB::SINCOS_NXV4F32; |
| 575 | case MVT::nxv2f64: |
| 576 | return RTLIB::SINCOS_NXV2F64; |
| 577 | default: |
| 578 | return RTLIB::UNKNOWN_LIBCALL; |
| 579 | } |
| 580 | } |
| 581 | |
| 582 | return getFPLibCall(VT: RetVT, Call_F32: SINCOS_F32, Call_F64: SINCOS_F64, Call_F80: SINCOS_F80, Call_F128: SINCOS_F128, |
| 583 | Call_PPCF128: SINCOS_PPCF128); |
| 584 | } |
| 585 | |
| 586 | RTLIB::Libcall RTLIB::getSINCOSPI(EVT RetVT) { |
| 587 | // TODO: Tablegen should generate this function |
| 588 | if (RetVT.isVector()) { |
| 589 | if (!RetVT.isSimple()) |
| 590 | return RTLIB::UNKNOWN_LIBCALL; |
| 591 | switch (RetVT.getSimpleVT().SimpleTy) { |
| 592 | case MVT::v4f32: |
| 593 | return RTLIB::SINCOSPI_V4F32; |
| 594 | case MVT::v2f64: |
| 595 | return RTLIB::SINCOSPI_V2F64; |
| 596 | case MVT::nxv4f32: |
| 597 | return RTLIB::SINCOSPI_NXV4F32; |
| 598 | case MVT::nxv2f64: |
| 599 | return RTLIB::SINCOSPI_NXV2F64; |
| 600 | default: |
| 601 | return RTLIB::UNKNOWN_LIBCALL; |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | return getFPLibCall(VT: RetVT, Call_F32: SINCOSPI_F32, Call_F64: SINCOSPI_F64, Call_F80: SINCOSPI_F80, |
| 606 | Call_F128: SINCOSPI_F128, Call_PPCF128: SINCOSPI_PPCF128); |
| 607 | } |
| 608 | |
| 609 | RTLIB::Libcall RTLIB::getSINCOS_STRET(EVT RetVT) { |
| 610 | return getFPLibCall(VT: RetVT, Call_F32: SINCOS_STRET_F32, Call_F64: SINCOS_STRET_F64, |
| 611 | Call_F80: UNKNOWN_LIBCALL, Call_F128: UNKNOWN_LIBCALL, Call_PPCF128: UNKNOWN_LIBCALL); |
| 612 | } |
| 613 | |
| 614 | RTLIB::Libcall RTLIB::getREM(EVT VT) { |
| 615 | // TODO: Tablegen should generate this function |
| 616 | if (VT.isVector()) { |
| 617 | if (!VT.isSimple()) |
| 618 | return RTLIB::UNKNOWN_LIBCALL; |
| 619 | switch (VT.getSimpleVT().SimpleTy) { |
| 620 | case MVT::v4f32: |
| 621 | return RTLIB::REM_V4F32; |
| 622 | case MVT::v2f64: |
| 623 | return RTLIB::REM_V2F64; |
| 624 | case MVT::nxv4f32: |
| 625 | return RTLIB::REM_NXV4F32; |
| 626 | case MVT::nxv2f64: |
| 627 | return RTLIB::REM_NXV2F64; |
| 628 | default: |
| 629 | return RTLIB::UNKNOWN_LIBCALL; |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | return getFPLibCall(VT, Call_F32: REM_F32, Call_F64: REM_F64, Call_F80: REM_F80, Call_F128: REM_F128, Call_PPCF128: REM_PPCF128); |
| 634 | } |
| 635 | |
| 636 | RTLIB::Libcall RTLIB::getMODF(EVT RetVT) { |
| 637 | // TODO: Tablegen should generate this function |
| 638 | if (RetVT.isVector()) { |
| 639 | if (!RetVT.isSimple()) |
| 640 | return RTLIB::UNKNOWN_LIBCALL; |
| 641 | switch (RetVT.getSimpleVT().SimpleTy) { |
| 642 | case MVT::v4f32: |
| 643 | return RTLIB::MODF_V4F32; |
| 644 | case MVT::v2f64: |
| 645 | return RTLIB::MODF_V2F64; |
| 646 | case MVT::nxv4f32: |
| 647 | return RTLIB::MODF_NXV4F32; |
| 648 | case MVT::nxv2f64: |
| 649 | return RTLIB::MODF_NXV2F64; |
| 650 | default: |
| 651 | return RTLIB::UNKNOWN_LIBCALL; |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | return getFPLibCall(VT: RetVT, Call_F32: MODF_F32, Call_F64: MODF_F64, Call_F80: MODF_F80, Call_F128: MODF_F128, |
| 656 | Call_PPCF128: MODF_PPCF128); |
| 657 | } |
| 658 | |
| 659 | RTLIB::Libcall RTLIB::getLROUND(EVT VT) { |
| 660 | if (VT == MVT::f32) |
| 661 | return RTLIB::LROUND_F32; |
| 662 | if (VT == MVT::f64) |
| 663 | return RTLIB::LROUND_F64; |
| 664 | if (VT == MVT::f80) |
| 665 | return RTLIB::LROUND_F80; |
| 666 | if (VT == MVT::f128) |
| 667 | return RTLIB::LROUND_F128; |
| 668 | if (VT == MVT::ppcf128) |
| 669 | return RTLIB::LROUND_PPCF128; |
| 670 | |
| 671 | return RTLIB::UNKNOWN_LIBCALL; |
| 672 | } |
| 673 | |
| 674 | RTLIB::Libcall RTLIB::getLLROUND(EVT VT) { |
| 675 | if (VT == MVT::f32) |
| 676 | return RTLIB::LLROUND_F32; |
| 677 | if (VT == MVT::f64) |
| 678 | return RTLIB::LLROUND_F64; |
| 679 | if (VT == MVT::f80) |
| 680 | return RTLIB::LLROUND_F80; |
| 681 | if (VT == MVT::f128) |
| 682 | return RTLIB::LLROUND_F128; |
| 683 | if (VT == MVT::ppcf128) |
| 684 | return RTLIB::LLROUND_PPCF128; |
| 685 | |
| 686 | return RTLIB::UNKNOWN_LIBCALL; |
| 687 | } |
| 688 | |
| 689 | RTLIB::Libcall RTLIB::getLRINT(EVT VT) { |
| 690 | if (VT == MVT::f32) |
| 691 | return RTLIB::LRINT_F32; |
| 692 | if (VT == MVT::f64) |
| 693 | return RTLIB::LRINT_F64; |
| 694 | if (VT == MVT::f80) |
| 695 | return RTLIB::LRINT_F80; |
| 696 | if (VT == MVT::f128) |
| 697 | return RTLIB::LRINT_F128; |
| 698 | if (VT == MVT::ppcf128) |
| 699 | return RTLIB::LRINT_PPCF128; |
| 700 | return RTLIB::UNKNOWN_LIBCALL; |
| 701 | } |
| 702 | |
| 703 | RTLIB::Libcall RTLIB::getLLRINT(EVT VT) { |
| 704 | if (VT == MVT::f32) |
| 705 | return RTLIB::LLRINT_F32; |
| 706 | if (VT == MVT::f64) |
| 707 | return RTLIB::LLRINT_F64; |
| 708 | if (VT == MVT::f80) |
| 709 | return RTLIB::LLRINT_F80; |
| 710 | if (VT == MVT::f128) |
| 711 | return RTLIB::LLRINT_F128; |
| 712 | if (VT == MVT::ppcf128) |
| 713 | return RTLIB::LLRINT_PPCF128; |
| 714 | return RTLIB::UNKNOWN_LIBCALL; |
| 715 | } |
| 716 | |
| 717 | RTLIB::Libcall RTLIB::getOutlineAtomicHelper(const Libcall (&LC)[5][4], |
| 718 | AtomicOrdering Order, |
| 719 | uint64_t MemSize) { |
| 720 | unsigned ModeN, ModelN; |
| 721 | switch (MemSize) { |
| 722 | case 1: |
| 723 | ModeN = 0; |
| 724 | break; |
| 725 | case 2: |
| 726 | ModeN = 1; |
| 727 | break; |
| 728 | case 4: |
| 729 | ModeN = 2; |
| 730 | break; |
| 731 | case 8: |
| 732 | ModeN = 3; |
| 733 | break; |
| 734 | case 16: |
| 735 | ModeN = 4; |
| 736 | break; |
| 737 | default: |
| 738 | return RTLIB::UNKNOWN_LIBCALL; |
| 739 | } |
| 740 | |
| 741 | switch (Order) { |
| 742 | case AtomicOrdering::Monotonic: |
| 743 | ModelN = 0; |
| 744 | break; |
| 745 | case AtomicOrdering::Acquire: |
| 746 | ModelN = 1; |
| 747 | break; |
| 748 | case AtomicOrdering::Release: |
| 749 | ModelN = 2; |
| 750 | break; |
| 751 | case AtomicOrdering::AcquireRelease: |
| 752 | case AtomicOrdering::SequentiallyConsistent: |
| 753 | ModelN = 3; |
| 754 | break; |
| 755 | default: |
| 756 | return UNKNOWN_LIBCALL; |
| 757 | } |
| 758 | |
| 759 | return LC[ModeN][ModelN]; |
| 760 | } |
| 761 | |
| 762 | RTLIB::Libcall RTLIB::getOUTLINE_ATOMIC(unsigned Opc, AtomicOrdering Order, |
| 763 | MVT VT) { |
| 764 | if (!VT.isScalarInteger()) |
| 765 | return UNKNOWN_LIBCALL; |
| 766 | uint64_t MemSize = VT.getScalarSizeInBits() / 8; |
| 767 | |
| 768 | #define LCALLS(A, B) \ |
| 769 | { A##B##_RELAX, A##B##_ACQ, A##B##_REL, A##B##_ACQ_REL } |
| 770 | #define LCALL5(A) \ |
| 771 | LCALLS(A, 1), LCALLS(A, 2), LCALLS(A, 4), LCALLS(A, 8), LCALLS(A, 16) |
| 772 | switch (Opc) { |
| 773 | case ISD::ATOMIC_CMP_SWAP: { |
| 774 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_CAS)}; |
| 775 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 776 | } |
| 777 | case ISD::ATOMIC_SWAP: { |
| 778 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_SWP)}; |
| 779 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 780 | } |
| 781 | case ISD::ATOMIC_LOAD_ADD: { |
| 782 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDADD)}; |
| 783 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 784 | } |
| 785 | case ISD::ATOMIC_LOAD_OR: { |
| 786 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDSET)}; |
| 787 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 788 | } |
| 789 | case ISD::ATOMIC_LOAD_CLR: { |
| 790 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDCLR)}; |
| 791 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 792 | } |
| 793 | case ISD::ATOMIC_LOAD_XOR: { |
| 794 | const Libcall LC[5][4] = {LCALL5(OUTLINE_ATOMIC_LDEOR)}; |
| 795 | return getOutlineAtomicHelper(LC, Order, MemSize); |
| 796 | } |
| 797 | default: |
| 798 | return UNKNOWN_LIBCALL; |
| 799 | } |
| 800 | #undef LCALLS |
| 801 | #undef LCALL5 |
| 802 | } |
| 803 | |
| 804 | RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { |
| 805 | #define OP_TO_LIBCALL(Name, Enum) \ |
| 806 | case Name: \ |
| 807 | switch (VT.SimpleTy) { \ |
| 808 | default: \ |
| 809 | return UNKNOWN_LIBCALL; \ |
| 810 | case MVT::i8: \ |
| 811 | return Enum##_1; \ |
| 812 | case MVT::i16: \ |
| 813 | return Enum##_2; \ |
| 814 | case MVT::i32: \ |
| 815 | return Enum##_4; \ |
| 816 | case MVT::i64: \ |
| 817 | return Enum##_8; \ |
| 818 | case MVT::i128: \ |
| 819 | return Enum##_16; \ |
| 820 | } |
| 821 | |
| 822 | switch (Opc) { |
| 823 | OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) |
| 824 | OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) |
| 825 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) |
| 826 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) |
| 827 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) |
| 828 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) |
| 829 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) |
| 830 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) |
| 831 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) |
| 832 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) |
| 833 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) |
| 834 | OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) |
| 835 | } |
| 836 | |
| 837 | #undef OP_TO_LIBCALL |
| 838 | |
| 839 | return UNKNOWN_LIBCALL; |
| 840 | } |
| 841 | |
| 842 | RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { |
| 843 | switch (ElementSize) { |
| 844 | case 1: |
| 845 | return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; |
| 846 | case 2: |
| 847 | return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; |
| 848 | case 4: |
| 849 | return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; |
| 850 | case 8: |
| 851 | return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; |
| 852 | case 16: |
| 853 | return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; |
| 854 | default: |
| 855 | return UNKNOWN_LIBCALL; |
| 856 | } |
| 857 | } |
| 858 | |
| 859 | RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { |
| 860 | switch (ElementSize) { |
| 861 | case 1: |
| 862 | return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; |
| 863 | case 2: |
| 864 | return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; |
| 865 | case 4: |
| 866 | return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; |
| 867 | case 8: |
| 868 | return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; |
| 869 | case 16: |
| 870 | return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; |
| 871 | default: |
| 872 | return UNKNOWN_LIBCALL; |
| 873 | } |
| 874 | } |
| 875 | |
| 876 | RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { |
| 877 | switch (ElementSize) { |
| 878 | case 1: |
| 879 | return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; |
| 880 | case 2: |
| 881 | return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; |
| 882 | case 4: |
| 883 | return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; |
| 884 | case 8: |
| 885 | return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; |
| 886 | case 16: |
| 887 | return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; |
| 888 | default: |
| 889 | return UNKNOWN_LIBCALL; |
| 890 | } |
| 891 | } |
| 892 | |
| 893 | ISD::CondCode TargetLoweringBase::getSoftFloatCmpLibcallPredicate( |
| 894 | RTLIB::LibcallImpl Impl) const { |
| 895 | switch (Impl) { |
| 896 | case RTLIB::impl___aeabi_dcmpeq__une: |
| 897 | case RTLIB::impl___aeabi_fcmpeq__une: |
| 898 | // Usage in the eq case, so we have to invert the comparison. |
| 899 | return ISD::SETEQ; |
| 900 | case RTLIB::impl___aeabi_dcmpeq__oeq: |
| 901 | case RTLIB::impl___aeabi_fcmpeq__oeq: |
| 902 | // Normal comparison to boolean value. |
| 903 | return ISD::SETNE; |
| 904 | case RTLIB::impl___aeabi_dcmplt: |
| 905 | case RTLIB::impl___aeabi_dcmple: |
| 906 | case RTLIB::impl___aeabi_dcmpge: |
| 907 | case RTLIB::impl___aeabi_dcmpgt: |
| 908 | case RTLIB::impl___aeabi_dcmpun: |
| 909 | case RTLIB::impl___aeabi_fcmplt: |
| 910 | case RTLIB::impl___aeabi_fcmple: |
| 911 | case RTLIB::impl___aeabi_fcmpge: |
| 912 | case RTLIB::impl___aeabi_fcmpgt: |
| 913 | /// The AEABI versions return a typical boolean value, so we can compare |
| 914 | /// against the integer result as simply != 0. |
| 915 | return ISD::SETNE; |
| 916 | default: |
| 917 | break; |
| 918 | } |
| 919 | |
| 920 | // Assume libgcc/compiler-rt behavior. Most of the cases are really aliases of |
| 921 | // each other, and return a 3-way comparison style result of -1, 0, or 1 |
| 922 | // depending on lt/eq/gt. |
| 923 | // |
| 924 | // FIXME: It would be cleaner to directly express this as a 3-way comparison |
| 925 | // soft FP libcall instead of individual compares. |
| 926 | RTLIB::Libcall LC = RTLIB::RuntimeLibcallsInfo::getLibcallFromImpl(Impl); |
| 927 | switch (LC) { |
| 928 | case RTLIB::OEQ_F32: |
| 929 | case RTLIB::OEQ_F64: |
| 930 | case RTLIB::OEQ_F128: |
| 931 | case RTLIB::OEQ_PPCF128: |
| 932 | return ISD::SETEQ; |
| 933 | case RTLIB::UNE_F32: |
| 934 | case RTLIB::UNE_F64: |
| 935 | case RTLIB::UNE_F128: |
| 936 | case RTLIB::UNE_PPCF128: |
| 937 | return ISD::SETNE; |
| 938 | case RTLIB::OGE_F32: |
| 939 | case RTLIB::OGE_F64: |
| 940 | case RTLIB::OGE_F128: |
| 941 | case RTLIB::OGE_PPCF128: |
| 942 | return ISD::SETGE; |
| 943 | case RTLIB::OLT_F32: |
| 944 | case RTLIB::OLT_F64: |
| 945 | case RTLIB::OLT_F128: |
| 946 | case RTLIB::OLT_PPCF128: |
| 947 | return ISD::SETLT; |
| 948 | case RTLIB::OLE_F32: |
| 949 | case RTLIB::OLE_F64: |
| 950 | case RTLIB::OLE_F128: |
| 951 | case RTLIB::OLE_PPCF128: |
| 952 | return ISD::SETLE; |
| 953 | case RTLIB::OGT_F32: |
| 954 | case RTLIB::OGT_F64: |
| 955 | case RTLIB::OGT_F128: |
| 956 | case RTLIB::OGT_PPCF128: |
| 957 | return ISD::SETGT; |
| 958 | case RTLIB::UO_F32: |
| 959 | case RTLIB::UO_F64: |
| 960 | case RTLIB::UO_F128: |
| 961 | case RTLIB::UO_PPCF128: |
| 962 | return ISD::SETNE; |
| 963 | default: |
| 964 | llvm_unreachable("not a compare libcall" ); |
| 965 | } |
| 966 | } |
| 967 | |
| 968 | /// NOTE: The TargetMachine owns TLOF. |
| 969 | TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm, |
| 970 | const TargetSubtargetInfo &STI) |
| 971 | : TM(tm), |
| 972 | RuntimeLibcallInfo(TM.getTargetTriple(), TM.Options.ExceptionModel, |
| 973 | TM.Options.FloatABIType, TM.Options.EABIVersion, |
| 974 | TM.Options.MCOptions.getABIName(), TM.Options.VecLib), |
| 975 | Libcalls(RuntimeLibcallInfo, STI) { |
| 976 | initActions(); |
| 977 | |
| 978 | // Perform these initializations only once. |
| 979 | MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = |
| 980 | MaxLoadsPerMemcmp = 8; |
| 981 | MaxGluedStoresPerMemcpy = 0; |
| 982 | MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = |
| 983 | MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; |
| 984 | HasExtractBitsInsn = false; |
| 985 | JumpIsExpensive = JumpIsExpensiveOverride; |
| 986 | PredictableSelectIsExpensive = false; |
| 987 | EnableExtLdPromotion = false; |
| 988 | StackPointerRegisterToSaveRestore = 0; |
| 989 | BooleanContents = UndefinedBooleanContent; |
| 990 | BooleanFloatContents = UndefinedBooleanContent; |
| 991 | BooleanVectorContents = UndefinedBooleanContent; |
| 992 | SchedPreferenceInfo = Sched::ILP; |
| 993 | GatherAllAliasesMaxDepth = 18; |
| 994 | IsStrictFPEnabled = DisableStrictNodeMutation; |
| 995 | MaxBytesForAlignment = 0; |
| 996 | MaxAtomicSizeInBitsSupported = 0; |
| 997 | |
| 998 | // Assume that even with libcalls, no target supports wider than 128 bit |
| 999 | // division. |
| 1000 | MaxDivRemBitWidthSupported = 128; |
| 1001 | |
| 1002 | MaxLargeFPConvertBitWidthSupported = 128; |
| 1003 | |
| 1004 | MinCmpXchgSizeInBits = 0; |
| 1005 | SupportsUnalignedAtomics = false; |
| 1006 | |
| 1007 | MinimumBitTestCmps = MinimumBitTestCmpsOverride; |
| 1008 | } |
| 1009 | |
| 1010 | // Define the virtual destructor out-of-line to act as a key method to anchor |
| 1011 | // debug info (see coding standards). |
| 1012 | TargetLoweringBase::~TargetLoweringBase() = default; |
| 1013 | |
| 1014 | void TargetLoweringBase::initActions() { |
| 1015 | // All operations default to being supported. |
| 1016 | memset(s: OpActions, c: 0, n: sizeof(OpActions)); |
| 1017 | memset(s: LoadExtActions, c: 0, n: sizeof(LoadExtActions)); |
| 1018 | memset(s: TruncStoreActions, c: 0, n: sizeof(TruncStoreActions)); |
| 1019 | memset(s: IndexedModeActions, c: 0, n: sizeof(IndexedModeActions)); |
| 1020 | memset(s: CondCodeActions, c: 0, n: sizeof(CondCodeActions)); |
| 1021 | llvm::fill(Range&: RegClassForVT, Value: nullptr); |
| 1022 | llvm::fill(Range&: TargetDAGCombineArray, Value: 0); |
| 1023 | |
| 1024 | // Let extending atomic loads be unsupported by default. |
| 1025 | for (MVT ValVT : MVT::all_valuetypes()) |
| 1026 | for (MVT MemVT : MVT::all_valuetypes()) |
| 1027 | setAtomicLoadExtAction(ExtTypes: {ISD::SEXTLOAD, ISD::ZEXTLOAD}, ValVT, MemVT, |
| 1028 | Action: Expand); |
| 1029 | |
| 1030 | // We're somewhat special casing MVT::i2 and MVT::i4. Ideally we want to |
| 1031 | // remove this and targets should individually set these types if not legal. |
| 1032 | for (ISD::NodeType NT : enum_seq(Begin: ISD::DELETED_NODE, End: ISD::BUILTIN_OP_END, |
| 1033 | force_iteration_on_noniterable_enum)) { |
| 1034 | for (MVT VT : {MVT::i2, MVT::i4}) |
| 1035 | OpActions[(unsigned)VT.SimpleTy][NT] = Expand; |
| 1036 | } |
| 1037 | for (MVT AVT : MVT::all_valuetypes()) { |
| 1038 | for (MVT VT : {MVT::i2, MVT::i4, MVT::v128i2, MVT::v64i4}) { |
| 1039 | setTruncStoreAction(ValVT: AVT, MemVT: VT, Action: Expand); |
| 1040 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: AVT, MemVT: VT, Action: Expand); |
| 1041 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: AVT, MemVT: VT, Action: Expand); |
| 1042 | } |
| 1043 | } |
| 1044 | for (unsigned IM = (unsigned)ISD::PRE_INC; |
| 1045 | IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { |
| 1046 | for (MVT VT : {MVT::i2, MVT::i4}) { |
| 1047 | setIndexedLoadAction(IdxModes: IM, VT, Action: Expand); |
| 1048 | setIndexedStoreAction(IdxModes: IM, VT, Action: Expand); |
| 1049 | setIndexedMaskedLoadAction(IdxMode: IM, VT, Action: Expand); |
| 1050 | setIndexedMaskedStoreAction(IdxMode: IM, VT, Action: Expand); |
| 1051 | } |
| 1052 | } |
| 1053 | |
| 1054 | for (MVT VT : MVT::fp_valuetypes()) { |
| 1055 | MVT IntVT = MVT::getIntegerVT(BitWidth: VT.getFixedSizeInBits()); |
| 1056 | if (IntVT.isValid()) { |
| 1057 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT, Action: Promote); |
| 1058 | AddPromotedToType(Opc: ISD::ATOMIC_SWAP, OrigVT: VT, DestVT: IntVT); |
| 1059 | } |
| 1060 | } |
| 1061 | |
| 1062 | // If f16 fma is not natively supported, the value must be promoted to an f64 |
| 1063 | // (and not to f32!) to prevent double rounding issues. |
| 1064 | AddPromotedToType(Opc: ISD::FMA, OrigVT: MVT::f16, DestVT: MVT::f64); |
| 1065 | AddPromotedToType(Opc: ISD::STRICT_FMA, OrigVT: MVT::f16, DestVT: MVT::f64); |
| 1066 | |
| 1067 | // Set default actions for various operations. |
| 1068 | for (MVT VT : MVT::all_valuetypes()) { |
| 1069 | // Default all indexed load / store to expand. |
| 1070 | for (unsigned IM = (unsigned)ISD::PRE_INC; |
| 1071 | IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { |
| 1072 | setIndexedLoadAction(IdxModes: IM, VT, Action: Expand); |
| 1073 | setIndexedStoreAction(IdxModes: IM, VT, Action: Expand); |
| 1074 | setIndexedMaskedLoadAction(IdxMode: IM, VT, Action: Expand); |
| 1075 | setIndexedMaskedStoreAction(IdxMode: IM, VT, Action: Expand); |
| 1076 | } |
| 1077 | |
| 1078 | // Most backends expect to see the node which just returns the value loaded. |
| 1079 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Action: Expand); |
| 1080 | |
| 1081 | // These operations default to expand. |
| 1082 | setOperationAction(Ops: {ISD::FGETSIGN, ISD::CONCAT_VECTORS, |
| 1083 | ISD::FMINNUM, ISD::FMAXNUM, |
| 1084 | ISD::FMINNUM_IEEE, ISD::FMAXNUM_IEEE, |
| 1085 | ISD::FMINIMUM, ISD::FMAXIMUM, |
| 1086 | ISD::FMINIMUMNUM, ISD::FMAXIMUMNUM, |
| 1087 | ISD::FMAD, ISD::SMIN, |
| 1088 | ISD::SMAX, ISD::UMIN, |
| 1089 | ISD::UMAX, ISD::ABS, |
| 1090 | ISD::FSHL, ISD::FSHR, |
| 1091 | ISD::SADDSAT, ISD::UADDSAT, |
| 1092 | ISD::SSUBSAT, ISD::USUBSAT, |
| 1093 | ISD::SSHLSAT, ISD::USHLSAT, |
| 1094 | ISD::SMULFIX, ISD::SMULFIXSAT, |
| 1095 | ISD::UMULFIX, ISD::UMULFIXSAT, |
| 1096 | ISD::SDIVFIX, ISD::SDIVFIXSAT, |
| 1097 | ISD::UDIVFIX, ISD::UDIVFIXSAT, |
| 1098 | ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT, |
| 1099 | ISD::IS_FPCLASS, ISD::FCBRT, |
| 1100 | ISD::FLOG, ISD::FLOG2, |
| 1101 | ISD::FLOG10, ISD::FEXP, |
| 1102 | ISD::FEXP2, ISD::FEXP10, |
| 1103 | ISD::FFLOOR, ISD::FNEARBYINT, |
| 1104 | ISD::FCEIL, ISD::FRINT, |
| 1105 | ISD::FTRUNC, ISD::FROUNDEVEN, |
| 1106 | ISD::FTAN, ISD::FACOS, |
| 1107 | ISD::FASIN, ISD::FATAN, |
| 1108 | ISD::FCOSH, ISD::FSINH, |
| 1109 | ISD::FTANH, ISD::FATAN2, |
| 1110 | ISD::FMULADD}, |
| 1111 | VT, Action: Expand); |
| 1112 | |
| 1113 | // Overflow operations default to expand |
| 1114 | setOperationAction(Ops: {ISD::SADDO, ISD::SSUBO, ISD::UADDO, ISD::USUBO, |
| 1115 | ISD::SMULO, ISD::UMULO}, |
| 1116 | VT, Action: Expand); |
| 1117 | |
| 1118 | // Carry-using overflow operations default to expand. |
| 1119 | setOperationAction(Ops: {ISD::UADDO_CARRY, ISD::USUBO_CARRY, ISD::SETCCCARRY, |
| 1120 | ISD::SADDO_CARRY, ISD::SSUBO_CARRY}, |
| 1121 | VT, Action: Expand); |
| 1122 | |
| 1123 | // ADDC/ADDE/SUBC/SUBE default to expand. |
| 1124 | setOperationAction(Ops: {ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}, VT, |
| 1125 | Action: Expand); |
| 1126 | |
| 1127 | // [US]CMP default to expand |
| 1128 | setOperationAction(Ops: {ISD::UCMP, ISD::SCMP}, VT, Action: Expand); |
| 1129 | |
| 1130 | // Halving adds |
| 1131 | setOperationAction( |
| 1132 | Ops: {ISD::AVGFLOORS, ISD::AVGFLOORU, ISD::AVGCEILS, ISD::AVGCEILU}, VT, |
| 1133 | Action: Expand); |
| 1134 | |
| 1135 | // Absolute difference |
| 1136 | setOperationAction(Ops: {ISD::ABDS, ISD::ABDU}, VT, Action: Expand); |
| 1137 | |
| 1138 | // Carry-less multiply |
| 1139 | setOperationAction(Ops: {ISD::CLMUL, ISD::CLMULR, ISD::CLMULH}, VT, Action: Expand); |
| 1140 | |
| 1141 | // Saturated trunc |
| 1142 | setOperationAction(Op: ISD::TRUNCATE_SSAT_S, VT, Action: Expand); |
| 1143 | setOperationAction(Op: ISD::TRUNCATE_SSAT_U, VT, Action: Expand); |
| 1144 | setOperationAction(Op: ISD::TRUNCATE_USAT_U, VT, Action: Expand); |
| 1145 | |
| 1146 | // These default to Expand so they will be expanded to CTLZ/CTTZ by default. |
| 1147 | setOperationAction(Ops: {ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT, |
| 1148 | Action: Expand); |
| 1149 | setOperationAction(Op: ISD::CTLS, VT, Action: Expand); |
| 1150 | |
| 1151 | setOperationAction(Ops: {ISD::BITREVERSE, ISD::PARITY}, VT, Action: Expand); |
| 1152 | |
| 1153 | // These library functions default to expand. |
| 1154 | setOperationAction(Ops: {ISD::FROUND, ISD::FPOWI, ISD::FLDEXP, ISD::FFREXP, |
| 1155 | ISD::FSINCOS, ISD::FSINCOSPI, ISD::FMODF}, |
| 1156 | VT, Action: Expand); |
| 1157 | |
| 1158 | // These operations default to expand for vector types. |
| 1159 | if (VT.isVector()) |
| 1160 | setOperationAction(Ops: {ISD::FCOPYSIGN, ISD::SIGN_EXTEND_INREG, |
| 1161 | ISD::ANY_EXTEND_VECTOR_INREG, |
| 1162 | ISD::SIGN_EXTEND_VECTOR_INREG, |
| 1163 | ISD::ZERO_EXTEND_VECTOR_INREG, ISD::SPLAT_VECTOR, |
| 1164 | ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND}, |
| 1165 | VT, Action: Expand); |
| 1166 | |
| 1167 | // Constrained floating-point operations default to expand. |
| 1168 | #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ |
| 1169 | setOperationAction(ISD::STRICT_##DAGN, VT, Expand); |
| 1170 | #include "llvm/IR/ConstrainedOps.def" |
| 1171 | |
| 1172 | // For most targets @llvm.get.dynamic.area.offset just returns 0. |
| 1173 | setOperationAction(Op: ISD::GET_DYNAMIC_AREA_OFFSET, VT, Action: Expand); |
| 1174 | |
| 1175 | // Vector reduction default to expand. |
| 1176 | setOperationAction( |
| 1177 | Ops: {ISD::VECREDUCE_FADD, ISD::VECREDUCE_FMUL, ISD::VECREDUCE_ADD, |
| 1178 | ISD::VECREDUCE_MUL, ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, |
| 1179 | ISD::VECREDUCE_XOR, ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN, |
| 1180 | ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN, ISD::VECREDUCE_FMAX, |
| 1181 | ISD::VECREDUCE_FMIN, ISD::VECREDUCE_FMAXIMUM, ISD::VECREDUCE_FMINIMUM, |
| 1182 | ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_SEQ_FMUL}, |
| 1183 | VT, Action: Expand); |
| 1184 | |
| 1185 | // Named vector shuffles default to expand. |
| 1186 | setOperationAction(Ops: {ISD::VECTOR_SPLICE_LEFT, ISD::VECTOR_SPLICE_RIGHT}, VT, |
| 1187 | Action: Expand); |
| 1188 | |
| 1189 | // Only some target support this vector operation. Most need to expand it. |
| 1190 | setOperationAction(Op: ISD::VECTOR_COMPRESS, VT, Action: Expand); |
| 1191 | |
| 1192 | // VP operations default to expand. |
| 1193 | #define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) \ |
| 1194 | setOperationAction(ISD::SDOPC, VT, Expand); |
| 1195 | #include "llvm/IR/VPIntrinsics.def" |
| 1196 | |
| 1197 | // Masked vector extracts default to expand. |
| 1198 | setOperationAction(Op: ISD::VECTOR_FIND_LAST_ACTIVE, VT, Action: Expand); |
| 1199 | |
| 1200 | setOperationAction(Op: ISD::LOOP_DEPENDENCE_RAW_MASK, VT, Action: Expand); |
| 1201 | setOperationAction(Op: ISD::LOOP_DEPENDENCE_WAR_MASK, VT, Action: Expand); |
| 1202 | |
| 1203 | // FP environment operations default to expand. |
| 1204 | setOperationAction(Op: ISD::GET_FPENV, VT, Action: Expand); |
| 1205 | setOperationAction(Op: ISD::SET_FPENV, VT, Action: Expand); |
| 1206 | setOperationAction(Op: ISD::RESET_FPENV, VT, Action: Expand); |
| 1207 | |
| 1208 | setOperationAction(Op: ISD::MSTORE, VT, Action: Expand); |
| 1209 | } |
| 1210 | |
| 1211 | // Most targets ignore the @llvm.prefetch intrinsic. |
| 1212 | setOperationAction(Op: ISD::PREFETCH, VT: MVT::Other, Action: Expand); |
| 1213 | |
| 1214 | // Most targets also ignore the @llvm.readcyclecounter intrinsic. |
| 1215 | setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Expand); |
| 1216 | |
| 1217 | // Most targets also ignore the @llvm.readsteadycounter intrinsic. |
| 1218 | setOperationAction(Op: ISD::READSTEADYCOUNTER, VT: MVT::i64, Action: Expand); |
| 1219 | |
| 1220 | // ConstantFP nodes default to expand. Targets can either change this to |
| 1221 | // Legal, in which case all fp constants are legal, or use isFPImmLegal() |
| 1222 | // to optimize expansions for certain constants. |
| 1223 | setOperationAction(Ops: ISD::ConstantFP, |
| 1224 | VTs: {MVT::bf16, MVT::f16, MVT::f32, MVT::f64, MVT::f80, MVT::f128}, |
| 1225 | Action: Expand); |
| 1226 | |
| 1227 | // Insert custom handling default for llvm.canonicalize.*. |
| 1228 | setOperationAction(Ops: ISD::FCANONICALIZE, |
| 1229 | VTs: {MVT::f16, MVT::f32, MVT::f64, MVT::f128}, Action: Expand); |
| 1230 | |
| 1231 | // FIXME: Query RuntimeLibCalls to make the decision. |
| 1232 | setOperationAction(Ops: {ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND}, |
| 1233 | VTs: {MVT::f32, MVT::f64, MVT::f128}, Action: LibCall); |
| 1234 | |
| 1235 | setOperationAction(Ops: {ISD::FTAN, ISD::FACOS, ISD::FASIN, ISD::FATAN, ISD::FCOSH, |
| 1236 | ISD::FSINH, ISD::FTANH, ISD::FATAN2}, |
| 1237 | VT: MVT::f16, Action: Promote); |
| 1238 | // Default ISD::TRAP to expand (which turns it into abort). |
| 1239 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Expand); |
| 1240 | |
| 1241 | // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" |
| 1242 | // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. |
| 1243 | setOperationAction(Op: ISD::DEBUGTRAP, VT: MVT::Other, Action: Expand); |
| 1244 | |
| 1245 | setOperationAction(Op: ISD::UBSANTRAP, VT: MVT::Other, Action: Expand); |
| 1246 | |
| 1247 | setOperationAction(Op: ISD::GET_FPENV_MEM, VT: MVT::Other, Action: Expand); |
| 1248 | setOperationAction(Op: ISD::SET_FPENV_MEM, VT: MVT::Other, Action: Expand); |
| 1249 | |
| 1250 | for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) { |
| 1251 | setOperationAction(Op: ISD::GET_FPMODE, VT, Action: Expand); |
| 1252 | setOperationAction(Op: ISD::SET_FPMODE, VT, Action: Expand); |
| 1253 | } |
| 1254 | setOperationAction(Op: ISD::RESET_FPMODE, VT: MVT::Other, Action: Expand); |
| 1255 | |
| 1256 | // This one by default will call __clear_cache unless the target |
| 1257 | // wants something different. |
| 1258 | setOperationAction(Op: ISD::CLEAR_CACHE, VT: MVT::Other, Action: LibCall); |
| 1259 | |
| 1260 | // By default, STACKADDRESS nodes are expanded like STACKSAVE nodes. |
| 1261 | // On SPARC targets, custom lowering is required. |
| 1262 | setOperationAction(Op: ISD::STACKADDRESS, VT: MVT::Other, Action: Expand); |
| 1263 | } |
| 1264 | |
| 1265 | MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, |
| 1266 | EVT) const { |
| 1267 | return MVT::getIntegerVT(BitWidth: DL.getPointerSizeInBits(AS: 0)); |
| 1268 | } |
| 1269 | |
| 1270 | EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, |
| 1271 | const DataLayout &DL) const { |
| 1272 | assert(LHSTy.isInteger() && "Shift amount is not an integer type!" ); |
| 1273 | if (LHSTy.isVector()) |
| 1274 | return LHSTy; |
| 1275 | MVT ShiftVT = getScalarShiftAmountTy(DL, LHSTy); |
| 1276 | // If any possible shift value won't fit in the prefered type, just use |
| 1277 | // something safe. Assume it will be legalized when the shift is expanded. |
| 1278 | if (ShiftVT.getSizeInBits() < Log2_32_Ceil(Value: LHSTy.getSizeInBits())) |
| 1279 | ShiftVT = MVT::i32; |
| 1280 | assert(ShiftVT.getSizeInBits() >= Log2_32_Ceil(LHSTy.getSizeInBits()) && |
| 1281 | "ShiftVT is still too small!" ); |
| 1282 | return ShiftVT; |
| 1283 | } |
| 1284 | |
| 1285 | bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { |
| 1286 | assert(isTypeLegal(VT)); |
| 1287 | switch (Op) { |
| 1288 | default: |
| 1289 | return false; |
| 1290 | case ISD::SDIV: |
| 1291 | case ISD::UDIV: |
| 1292 | case ISD::SREM: |
| 1293 | case ISD::UREM: |
| 1294 | return true; |
| 1295 | } |
| 1296 | } |
| 1297 | |
| 1298 | bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS, |
| 1299 | unsigned DestAS) const { |
| 1300 | return TM.isNoopAddrSpaceCast(SrcAS, DestAS); |
| 1301 | } |
| 1302 | |
| 1303 | unsigned TargetLoweringBase::getBitWidthForCttzElements( |
| 1304 | Type *RetTy, ElementCount EC, bool ZeroIsPoison, |
| 1305 | const ConstantRange *VScaleRange) const { |
| 1306 | // Find the smallest "sensible" element type to use for the expansion. |
| 1307 | ConstantRange CR(APInt(64, EC.getKnownMinValue())); |
| 1308 | if (EC.isScalable()) |
| 1309 | CR = CR.umul_sat(Other: *VScaleRange); |
| 1310 | |
| 1311 | if (ZeroIsPoison) |
| 1312 | CR = CR.subtract(CI: APInt(64, 1)); |
| 1313 | |
| 1314 | unsigned EltWidth = RetTy->getScalarSizeInBits(); |
| 1315 | EltWidth = std::min(a: EltWidth, b: CR.getActiveBits()); |
| 1316 | EltWidth = std::max(a: llvm::bit_ceil(Value: EltWidth), b: (unsigned)8); |
| 1317 | |
| 1318 | return EltWidth; |
| 1319 | } |
| 1320 | |
| 1321 | void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { |
| 1322 | // If the command-line option was specified, ignore this request. |
| 1323 | if (!JumpIsExpensiveOverride.getNumOccurrences()) |
| 1324 | JumpIsExpensive = isExpensive; |
| 1325 | } |
| 1326 | |
| 1327 | TargetLoweringBase::LegalizeKind |
| 1328 | TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { |
| 1329 | // If this is a simple type, use the ComputeRegisterProp mechanism. |
| 1330 | if (VT.isSimple()) { |
| 1331 | MVT SVT = VT.getSimpleVT(); |
| 1332 | assert((unsigned)SVT.SimpleTy < std::size(TransformToType)); |
| 1333 | MVT NVT = TransformToType[SVT.SimpleTy]; |
| 1334 | LegalizeTypeAction LA = ValueTypeActions.getTypeAction(VT: SVT); |
| 1335 | |
| 1336 | assert((LA == TypeLegal || LA == TypeSoftenFloat || |
| 1337 | LA == TypeSoftPromoteHalf || |
| 1338 | (NVT.isVector() || |
| 1339 | ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger)) && |
| 1340 | "Promote may not follow Expand or Promote" ); |
| 1341 | |
| 1342 | if (LA == TypeSplitVector) |
| 1343 | return LegalizeKind(LA, EVT(SVT).getHalfNumVectorElementsVT(Context)); |
| 1344 | if (LA == TypeScalarizeVector) |
| 1345 | return LegalizeKind(LA, SVT.getVectorElementType()); |
| 1346 | return LegalizeKind(LA, NVT); |
| 1347 | } |
| 1348 | |
| 1349 | // Handle Extended Scalar Types. |
| 1350 | if (!VT.isVector()) { |
| 1351 | assert(VT.isInteger() && "Float types must be simple" ); |
| 1352 | unsigned BitSize = VT.getSizeInBits(); |
| 1353 | // First promote to a power-of-two size, then expand if necessary. |
| 1354 | if (BitSize < 8 || !isPowerOf2_32(Value: BitSize)) { |
| 1355 | EVT NVT = VT.getRoundIntegerType(Context); |
| 1356 | assert(NVT != VT && "Unable to round integer VT" ); |
| 1357 | LegalizeKind NextStep = getTypeConversion(Context, VT: NVT); |
| 1358 | // Avoid multi-step promotion. |
| 1359 | if (NextStep.first == TypePromoteInteger) |
| 1360 | return NextStep; |
| 1361 | // Return rounded integer type. |
| 1362 | return LegalizeKind(TypePromoteInteger, NVT); |
| 1363 | } |
| 1364 | |
| 1365 | return LegalizeKind(TypeExpandInteger, |
| 1366 | EVT::getIntegerVT(Context, BitWidth: VT.getSizeInBits() / 2)); |
| 1367 | } |
| 1368 | |
| 1369 | // Handle vector types. |
| 1370 | ElementCount NumElts = VT.getVectorElementCount(); |
| 1371 | EVT EltVT = VT.getVectorElementType(); |
| 1372 | |
| 1373 | // Vectors with only one element are always scalarized. |
| 1374 | if (NumElts.isScalar()) |
| 1375 | return LegalizeKind(TypeScalarizeVector, EltVT); |
| 1376 | |
| 1377 | // Try to widen vector elements until the element type is a power of two and |
| 1378 | // promote it to a legal type later on, for example: |
| 1379 | // <3 x i8> -> <4 x i8> -> <4 x i32> |
| 1380 | if (EltVT.isInteger()) { |
| 1381 | // Vectors with a number of elements that is not a power of two are always |
| 1382 | // widened, for example <3 x i8> -> <4 x i8>. |
| 1383 | if (!VT.isPow2VectorType()) { |
| 1384 | NumElts = NumElts.coefficientNextPowerOf2(); |
| 1385 | EVT NVT = EVT::getVectorVT(Context, VT: EltVT, EC: NumElts); |
| 1386 | return LegalizeKind(TypeWidenVector, NVT); |
| 1387 | } |
| 1388 | |
| 1389 | // Examine the element type. |
| 1390 | LegalizeKind LK = getTypeConversion(Context, VT: EltVT); |
| 1391 | |
| 1392 | // If type is to be expanded, split the vector. |
| 1393 | // <4 x i140> -> <2 x i140> |
| 1394 | if (LK.first == TypeExpandInteger) { |
| 1395 | if (NumElts.isScalable() && NumElts.getKnownMinValue() == 1) |
| 1396 | return LegalizeKind(TypeScalarizeScalableVector, EltVT); |
| 1397 | return LegalizeKind(TypeSplitVector, |
| 1398 | VT.getHalfNumVectorElementsVT(Context)); |
| 1399 | } |
| 1400 | |
| 1401 | // Promote the integer element types until a legal vector type is found |
| 1402 | // or until the element integer type is too big. If a legal type was not |
| 1403 | // found, fallback to the usual mechanism of widening/splitting the |
| 1404 | // vector. |
| 1405 | EVT OldEltVT = EltVT; |
| 1406 | while (true) { |
| 1407 | // Increase the bitwidth of the element to the next pow-of-two |
| 1408 | // (which is greater than 8 bits). |
| 1409 | EltVT = EVT::getIntegerVT(Context, BitWidth: 1 + EltVT.getSizeInBits()) |
| 1410 | .getRoundIntegerType(Context); |
| 1411 | |
| 1412 | // Stop trying when getting a non-simple element type. |
| 1413 | // Note that vector elements may be greater than legal vector element |
| 1414 | // types. Example: X86 XMM registers hold 64bit element on 32bit |
| 1415 | // systems. |
| 1416 | if (!EltVT.isSimple()) |
| 1417 | break; |
| 1418 | |
| 1419 | // Build a new vector type and check if it is legal. |
| 1420 | MVT NVT = MVT::getVectorVT(VT: EltVT.getSimpleVT(), EC: NumElts); |
| 1421 | // Found a legal promoted vector type. |
| 1422 | if (NVT != MVT() && ValueTypeActions.getTypeAction(VT: NVT) == TypeLegal) |
| 1423 | return LegalizeKind(TypePromoteInteger, |
| 1424 | EVT::getVectorVT(Context, VT: EltVT, EC: NumElts)); |
| 1425 | } |
| 1426 | |
| 1427 | // Reset the type to the unexpanded type if we did not find a legal vector |
| 1428 | // type with a promoted vector element type. |
| 1429 | EltVT = OldEltVT; |
| 1430 | } |
| 1431 | |
| 1432 | // Try to widen the vector until a legal type is found. |
| 1433 | // If there is no wider legal type, split the vector. |
| 1434 | while (true) { |
| 1435 | // Round up to the next power of 2. |
| 1436 | NumElts = NumElts.coefficientNextPowerOf2(); |
| 1437 | |
| 1438 | // If there is no simple vector type with this many elements then there |
| 1439 | // cannot be a larger legal vector type. Note that this assumes that |
| 1440 | // there are no skipped intermediate vector types in the simple types. |
| 1441 | if (!EltVT.isSimple()) |
| 1442 | break; |
| 1443 | MVT LargerVector = MVT::getVectorVT(VT: EltVT.getSimpleVT(), EC: NumElts); |
| 1444 | if (LargerVector == MVT()) |
| 1445 | break; |
| 1446 | |
| 1447 | // If this type is legal then widen the vector. |
| 1448 | if (ValueTypeActions.getTypeAction(VT: LargerVector) == TypeLegal) |
| 1449 | return LegalizeKind(TypeWidenVector, LargerVector); |
| 1450 | } |
| 1451 | |
| 1452 | // Widen odd vectors to next power of two. |
| 1453 | if (!VT.isPow2VectorType()) { |
| 1454 | EVT NVT = VT.getPow2VectorType(Context); |
| 1455 | return LegalizeKind(TypeWidenVector, NVT); |
| 1456 | } |
| 1457 | |
| 1458 | if (VT.getVectorElementCount() == ElementCount::getScalable(MinVal: 1)) |
| 1459 | return LegalizeKind(TypeScalarizeScalableVector, EltVT); |
| 1460 | |
| 1461 | // Vectors with illegal element types are expanded. |
| 1462 | EVT NVT = EVT::getVectorVT(Context, VT: EltVT, |
| 1463 | EC: VT.getVectorElementCount().divideCoefficientBy(RHS: 2)); |
| 1464 | return LegalizeKind(TypeSplitVector, NVT); |
| 1465 | } |
| 1466 | |
| 1467 | static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, |
| 1468 | unsigned &NumIntermediates, |
| 1469 | MVT &RegisterVT, |
| 1470 | TargetLoweringBase *TLI) { |
| 1471 | // Figure out the right, legal destination reg to copy into. |
| 1472 | ElementCount EC = VT.getVectorElementCount(); |
| 1473 | MVT EltTy = VT.getVectorElementType(); |
| 1474 | |
| 1475 | unsigned NumVectorRegs = 1; |
| 1476 | |
| 1477 | // Scalable vectors cannot be scalarized, so splitting or widening is |
| 1478 | // required. |
| 1479 | if (VT.isScalableVector() && !isPowerOf2_32(Value: EC.getKnownMinValue())) |
| 1480 | llvm_unreachable( |
| 1481 | "Splitting or widening of non-power-of-2 MVTs is not implemented." ); |
| 1482 | |
| 1483 | // FIXME: We don't support non-power-of-2-sized vectors for now. |
| 1484 | // Ideally we could break down into LHS/RHS like LegalizeDAG does. |
| 1485 | if (!isPowerOf2_32(Value: EC.getKnownMinValue())) { |
| 1486 | // Split EC to unit size (scalable property is preserved). |
| 1487 | NumVectorRegs = EC.getKnownMinValue(); |
| 1488 | EC = ElementCount::getFixed(MinVal: 1); |
| 1489 | } |
| 1490 | |
| 1491 | // Divide the input until we get to a supported size. This will |
| 1492 | // always end up with an EC that represent a scalar or a scalable |
| 1493 | // scalar. |
| 1494 | while (EC.getKnownMinValue() > 1 && |
| 1495 | !TLI->isTypeLegal(VT: MVT::getVectorVT(VT: EltTy, EC))) { |
| 1496 | EC = EC.divideCoefficientBy(RHS: 2); |
| 1497 | NumVectorRegs <<= 1; |
| 1498 | } |
| 1499 | |
| 1500 | NumIntermediates = NumVectorRegs; |
| 1501 | |
| 1502 | MVT NewVT = MVT::getVectorVT(VT: EltTy, EC); |
| 1503 | if (!TLI->isTypeLegal(VT: NewVT)) |
| 1504 | NewVT = EltTy; |
| 1505 | IntermediateVT = NewVT; |
| 1506 | |
| 1507 | unsigned LaneSizeInBits = NewVT.getScalarSizeInBits(); |
| 1508 | |
| 1509 | // Convert sizes such as i33 to i64. |
| 1510 | LaneSizeInBits = llvm::bit_ceil(Value: LaneSizeInBits); |
| 1511 | |
| 1512 | MVT DestVT = TLI->getRegisterType(VT: NewVT); |
| 1513 | RegisterVT = DestVT; |
| 1514 | if (EVT(DestVT).bitsLT(VT: NewVT)) // Value is expanded, e.g. i64 -> i16. |
| 1515 | return NumVectorRegs * (LaneSizeInBits / DestVT.getScalarSizeInBits()); |
| 1516 | |
| 1517 | // Otherwise, promotion or legal types use the same number of registers as |
| 1518 | // the vector decimated to the appropriate level. |
| 1519 | return NumVectorRegs; |
| 1520 | } |
| 1521 | |
| 1522 | /// isLegalRC - Return true if the value types that can be represented by the |
| 1523 | /// specified register class are all legal. |
| 1524 | bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, |
| 1525 | const TargetRegisterClass &RC) const { |
| 1526 | for (const auto *I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) |
| 1527 | if (isTypeLegal(VT: *I)) |
| 1528 | return true; |
| 1529 | return false; |
| 1530 | } |
| 1531 | |
| 1532 | /// Replace/modify any TargetFrameIndex operands with a targte-dependent |
| 1533 | /// sequence of memory operands that is recognized by PrologEpilogInserter. |
| 1534 | MachineBasicBlock * |
| 1535 | TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, |
| 1536 | MachineBasicBlock *MBB) const { |
| 1537 | MachineInstr *MI = &InitialMI; |
| 1538 | MachineFunction &MF = *MI->getMF(); |
| 1539 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1540 | |
| 1541 | // We're handling multiple types of operands here: |
| 1542 | // PATCHPOINT MetaArgs - live-in, read only, direct |
| 1543 | // STATEPOINT Deopt Spill - live-through, read only, indirect |
| 1544 | // STATEPOINT Deopt Alloca - live-through, read only, direct |
| 1545 | // (We're currently conservative and mark the deopt slots read/write in |
| 1546 | // practice.) |
| 1547 | // STATEPOINT GC Spill - live-through, read/write, indirect |
| 1548 | // STATEPOINT GC Alloca - live-through, read/write, direct |
| 1549 | // The live-in vs live-through is handled already (the live through ones are |
| 1550 | // all stack slots), but we need to handle the different type of stackmap |
| 1551 | // operands and memory effects here. |
| 1552 | |
| 1553 | if (llvm::none_of(Range: MI->operands(), |
| 1554 | P: [](MachineOperand &Operand) { return Operand.isFI(); })) |
| 1555 | return MBB; |
| 1556 | |
| 1557 | MachineInstrBuilder MIB = BuildMI(MF, MIMD: MI->getDebugLoc(), MCID: MI->getDesc()); |
| 1558 | |
| 1559 | // Inherit previous memory operands. |
| 1560 | MIB.cloneMemRefs(OtherMI: *MI); |
| 1561 | |
| 1562 | for (unsigned i = 0; i < MI->getNumOperands(); ++i) { |
| 1563 | MachineOperand &MO = MI->getOperand(i); |
| 1564 | if (!MO.isFI()) { |
| 1565 | // Index of Def operand this Use it tied to. |
| 1566 | // Since Defs are coming before Uses, if Use is tied, then |
| 1567 | // index of Def must be smaller that index of that Use. |
| 1568 | // Also, Defs preserve their position in new MI. |
| 1569 | unsigned TiedTo = i; |
| 1570 | if (MO.isReg() && MO.isTied()) |
| 1571 | TiedTo = MI->findTiedOperandIdx(OpIdx: i); |
| 1572 | MIB.add(MO); |
| 1573 | if (TiedTo < i) |
| 1574 | MIB->tieOperands(DefIdx: TiedTo, UseIdx: MIB->getNumOperands() - 1); |
| 1575 | continue; |
| 1576 | } |
| 1577 | |
| 1578 | // foldMemoryOperand builds a new MI after replacing a single FI operand |
| 1579 | // with the canonical set of five x86 addressing-mode operands. |
| 1580 | int FI = MO.getIndex(); |
| 1581 | |
| 1582 | // Add frame index operands recognized by stackmaps.cpp |
| 1583 | if (MFI.isStatepointSpillSlotObjectIndex(ObjectIdx: FI)) { |
| 1584 | // indirect-mem-ref tag, size, #FI, offset. |
| 1585 | // Used for spills inserted by StatepointLowering. This codepath is not |
| 1586 | // used for patchpoints/stackmaps at all, for these spilling is done via |
| 1587 | // foldMemoryOperand callback only. |
| 1588 | assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity" ); |
| 1589 | MIB.addImm(Val: StackMaps::IndirectMemRefOp); |
| 1590 | MIB.addImm(Val: MFI.getObjectSize(ObjectIdx: FI)); |
| 1591 | MIB.add(MO); |
| 1592 | MIB.addImm(Val: 0); |
| 1593 | } else { |
| 1594 | // direct-mem-ref tag, #FI, offset. |
| 1595 | // Used by patchpoint, and direct alloca arguments to statepoints |
| 1596 | MIB.addImm(Val: StackMaps::DirectMemRefOp); |
| 1597 | MIB.add(MO); |
| 1598 | MIB.addImm(Val: 0); |
| 1599 | } |
| 1600 | |
| 1601 | assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!" ); |
| 1602 | |
| 1603 | // Add a new memory operand for this FI. |
| 1604 | assert(MFI.getObjectOffset(FI) != -1); |
| 1605 | |
| 1606 | // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and |
| 1607 | // PATCHPOINT should be updated to do the same. (TODO) |
| 1608 | if (MI->getOpcode() != TargetOpcode::STATEPOINT) { |
| 1609 | auto Flags = MachineMemOperand::MOLoad; |
| 1610 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 1611 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI), F: Flags, |
| 1612 | Size: MF.getDataLayout().getPointerSize(), BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
| 1613 | MIB->addMemOperand(MF, MO: MMO); |
| 1614 | } |
| 1615 | } |
| 1616 | MBB->insert(I: MachineBasicBlock::iterator(MI), MI: MIB); |
| 1617 | MI->eraseFromParent(); |
| 1618 | return MBB; |
| 1619 | } |
| 1620 | |
| 1621 | /// findRepresentativeClass - Return the largest legal super-reg register class |
| 1622 | /// of the register class for the specified type and its associated "cost". |
| 1623 | // This function is in TargetLowering because it uses RegClassForVT which would |
| 1624 | // need to be moved to TargetRegisterInfo and would necessitate moving |
| 1625 | // isTypeLegal over as well - a massive change that would just require |
| 1626 | // TargetLowering having a TargetRegisterInfo class member that it would use. |
| 1627 | std::pair<const TargetRegisterClass *, uint8_t> |
| 1628 | TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, |
| 1629 | MVT VT) const { |
| 1630 | const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; |
| 1631 | if (!RC) |
| 1632 | return std::make_pair(x&: RC, y: 0); |
| 1633 | |
| 1634 | // Compute the set of all super-register classes. |
| 1635 | BitVector SuperRegRC(TRI->getNumRegClasses()); |
| 1636 | for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) |
| 1637 | SuperRegRC.setBitsInMask(Mask: RCI.getMask()); |
| 1638 | |
| 1639 | // Find the first legal register class with the largest spill size. |
| 1640 | const TargetRegisterClass *BestRC = RC; |
| 1641 | for (unsigned i : SuperRegRC.set_bits()) { |
| 1642 | const TargetRegisterClass *SuperRC = TRI->getRegClass(i); |
| 1643 | // We want the largest possible spill size. |
| 1644 | if (TRI->getSpillSize(RC: *SuperRC) <= TRI->getSpillSize(RC: *BestRC)) |
| 1645 | continue; |
| 1646 | if (!isLegalRC(TRI: *TRI, RC: *SuperRC)) |
| 1647 | continue; |
| 1648 | BestRC = SuperRC; |
| 1649 | } |
| 1650 | return std::make_pair(x&: BestRC, y: 1); |
| 1651 | } |
| 1652 | |
| 1653 | /// computeRegisterProperties - Once all of the register classes are added, |
| 1654 | /// this allows us to compute derived properties we expose. |
| 1655 | void TargetLoweringBase::computeRegisterProperties( |
| 1656 | const TargetRegisterInfo *TRI) { |
| 1657 | // Everything defaults to needing one register. |
| 1658 | for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) { |
| 1659 | NumRegistersForVT[i] = 1; |
| 1660 | RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; |
| 1661 | } |
| 1662 | // ...except isVoid, which doesn't need any registers. |
| 1663 | NumRegistersForVT[MVT::isVoid] = 0; |
| 1664 | |
| 1665 | // Find the largest integer register class. |
| 1666 | unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; |
| 1667 | for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) |
| 1668 | assert(LargestIntReg != MVT::i1 && "No integer registers defined!" ); |
| 1669 | |
| 1670 | // Every integer value type larger than this largest register takes twice as |
| 1671 | // many registers to represent as the previous ValueType. |
| 1672 | for (unsigned ExpandedReg = LargestIntReg + 1; |
| 1673 | ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { |
| 1674 | NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; |
| 1675 | RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; |
| 1676 | TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); |
| 1677 | ValueTypeActions.setTypeAction(VT: (MVT::SimpleValueType)ExpandedReg, |
| 1678 | Action: TypeExpandInteger); |
| 1679 | } |
| 1680 | |
| 1681 | // Inspect all of the ValueType's smaller than the largest integer |
| 1682 | // register to see which ones need promotion. |
| 1683 | unsigned LegalIntReg = LargestIntReg; |
| 1684 | for (unsigned IntReg = LargestIntReg - 1; |
| 1685 | IntReg >= (unsigned)MVT::i1; --IntReg) { |
| 1686 | MVT IVT = (MVT::SimpleValueType)IntReg; |
| 1687 | if (isTypeLegal(VT: IVT)) { |
| 1688 | LegalIntReg = IntReg; |
| 1689 | } else { |
| 1690 | RegisterTypeForVT[IntReg] = TransformToType[IntReg] = |
| 1691 | (MVT::SimpleValueType)LegalIntReg; |
| 1692 | ValueTypeActions.setTypeAction(VT: IVT, Action: TypePromoteInteger); |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | // ppcf128 type is really two f64's. |
| 1697 | if (!isTypeLegal(VT: MVT::ppcf128)) { |
| 1698 | if (isTypeLegal(VT: MVT::f64)) { |
| 1699 | NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; |
| 1700 | RegisterTypeForVT[MVT::ppcf128] = MVT::f64; |
| 1701 | TransformToType[MVT::ppcf128] = MVT::f64; |
| 1702 | ValueTypeActions.setTypeAction(VT: MVT::ppcf128, Action: TypeExpandFloat); |
| 1703 | } else { |
| 1704 | NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; |
| 1705 | RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; |
| 1706 | TransformToType[MVT::ppcf128] = MVT::i128; |
| 1707 | ValueTypeActions.setTypeAction(VT: MVT::ppcf128, Action: TypeSoftenFloat); |
| 1708 | } |
| 1709 | } |
| 1710 | |
| 1711 | // Decide how to handle f128. If the target does not have native f128 support, |
| 1712 | // expand it to i128 and we will be generating soft float library calls. |
| 1713 | if (!isTypeLegal(VT: MVT::f128)) { |
| 1714 | NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; |
| 1715 | RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; |
| 1716 | TransformToType[MVT::f128] = MVT::i128; |
| 1717 | ValueTypeActions.setTypeAction(VT: MVT::f128, Action: TypeSoftenFloat); |
| 1718 | } |
| 1719 | |
| 1720 | // Decide how to handle f80. If the target does not have native f80 support, |
| 1721 | // expand it to i96 and we will be generating soft float library calls. |
| 1722 | if (!isTypeLegal(VT: MVT::f80)) { |
| 1723 | NumRegistersForVT[MVT::f80] = 3*NumRegistersForVT[MVT::i32]; |
| 1724 | RegisterTypeForVT[MVT::f80] = RegisterTypeForVT[MVT::i32]; |
| 1725 | TransformToType[MVT::f80] = MVT::i32; |
| 1726 | ValueTypeActions.setTypeAction(VT: MVT::f80, Action: TypeSoftenFloat); |
| 1727 | } |
| 1728 | |
| 1729 | // Decide how to handle f64. If the target does not have native f64 support, |
| 1730 | // expand it to i64 and we will be generating soft float library calls. |
| 1731 | if (!isTypeLegal(VT: MVT::f64)) { |
| 1732 | NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; |
| 1733 | RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; |
| 1734 | TransformToType[MVT::f64] = MVT::i64; |
| 1735 | ValueTypeActions.setTypeAction(VT: MVT::f64, Action: TypeSoftenFloat); |
| 1736 | } |
| 1737 | |
| 1738 | // Decide how to handle f32. If the target does not have native f32 support, |
| 1739 | // expand it to i32 and we will be generating soft float library calls. |
| 1740 | if (!isTypeLegal(VT: MVT::f32)) { |
| 1741 | NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; |
| 1742 | RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; |
| 1743 | TransformToType[MVT::f32] = MVT::i32; |
| 1744 | ValueTypeActions.setTypeAction(VT: MVT::f32, Action: TypeSoftenFloat); |
| 1745 | } |
| 1746 | |
| 1747 | // Decide how to handle f16. If the target does not have native f16 support, |
| 1748 | // promote it to f32, because there are no f16 library calls (except for |
| 1749 | // conversions). |
| 1750 | if (!isTypeLegal(VT: MVT::f16)) { |
| 1751 | // Allow targets to control how we legalize half. |
| 1752 | bool UseFPRegsForHalfType = useFPRegsForHalfType(); |
| 1753 | |
| 1754 | if (!UseFPRegsForHalfType) { |
| 1755 | NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::i16]; |
| 1756 | RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::i16]; |
| 1757 | } else { |
| 1758 | NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; |
| 1759 | RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; |
| 1760 | } |
| 1761 | TransformToType[MVT::f16] = MVT::f32; |
| 1762 | ValueTypeActions.setTypeAction(VT: MVT::f16, Action: TypeSoftPromoteHalf); |
| 1763 | } |
| 1764 | |
| 1765 | // Decide how to handle bf16. If the target does not have native bf16 support, |
| 1766 | // promote it to f32, because there are no bf16 library calls (except for |
| 1767 | // converting from f32 to bf16). |
| 1768 | if (!isTypeLegal(VT: MVT::bf16)) { |
| 1769 | NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32]; |
| 1770 | RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32]; |
| 1771 | TransformToType[MVT::bf16] = MVT::f32; |
| 1772 | ValueTypeActions.setTypeAction(VT: MVT::bf16, Action: TypeSoftPromoteHalf); |
| 1773 | } |
| 1774 | |
| 1775 | // Loop over all of the vector value types to see which need transformations. |
| 1776 | for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; |
| 1777 | i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { |
| 1778 | MVT VT = (MVT::SimpleValueType) i; |
| 1779 | if (isTypeLegal(VT)) |
| 1780 | continue; |
| 1781 | |
| 1782 | MVT EltVT = VT.getVectorElementType(); |
| 1783 | ElementCount EC = VT.getVectorElementCount(); |
| 1784 | bool IsLegalWiderType = false; |
| 1785 | bool IsScalable = VT.isScalableVector(); |
| 1786 | LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); |
| 1787 | switch (PreferredAction) { |
| 1788 | case TypePromoteInteger: { |
| 1789 | MVT::SimpleValueType EndVT = IsScalable ? |
| 1790 | MVT::LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE : |
| 1791 | MVT::LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE; |
| 1792 | // Try to promote the elements of integer vectors. If no legal |
| 1793 | // promotion was found, fall through to the widen-vector method. |
| 1794 | for (unsigned nVT = i + 1; |
| 1795 | (MVT::SimpleValueType)nVT <= EndVT; ++nVT) { |
| 1796 | MVT SVT = (MVT::SimpleValueType) nVT; |
| 1797 | // Promote vectors of integers to vectors with the same number |
| 1798 | // of elements, with a wider element type. |
| 1799 | if (SVT.getScalarSizeInBits() > EltVT.getFixedSizeInBits() && |
| 1800 | SVT.getVectorElementCount() == EC && isTypeLegal(VT: SVT)) { |
| 1801 | TransformToType[i] = SVT; |
| 1802 | RegisterTypeForVT[i] = SVT; |
| 1803 | NumRegistersForVT[i] = 1; |
| 1804 | ValueTypeActions.setTypeAction(VT, Action: TypePromoteInteger); |
| 1805 | IsLegalWiderType = true; |
| 1806 | break; |
| 1807 | } |
| 1808 | } |
| 1809 | if (IsLegalWiderType) |
| 1810 | break; |
| 1811 | [[fallthrough]]; |
| 1812 | } |
| 1813 | |
| 1814 | case TypeWidenVector: |
| 1815 | if (isPowerOf2_32(Value: EC.getKnownMinValue())) { |
| 1816 | // Try to widen the vector. |
| 1817 | for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { |
| 1818 | MVT SVT = (MVT::SimpleValueType) nVT; |
| 1819 | if (SVT.getVectorElementType() == EltVT && |
| 1820 | SVT.isScalableVector() == IsScalable && |
| 1821 | SVT.getVectorElementCount().getKnownMinValue() > |
| 1822 | EC.getKnownMinValue() && |
| 1823 | isTypeLegal(VT: SVT)) { |
| 1824 | TransformToType[i] = SVT; |
| 1825 | RegisterTypeForVT[i] = SVT; |
| 1826 | NumRegistersForVT[i] = 1; |
| 1827 | ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector); |
| 1828 | IsLegalWiderType = true; |
| 1829 | break; |
| 1830 | } |
| 1831 | } |
| 1832 | if (IsLegalWiderType) |
| 1833 | break; |
| 1834 | } else { |
| 1835 | // Only widen to the next power of 2 to keep consistency with EVT. |
| 1836 | MVT NVT = VT.getPow2VectorType(); |
| 1837 | if (isTypeLegal(VT: NVT)) { |
| 1838 | TransformToType[i] = NVT; |
| 1839 | ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector); |
| 1840 | RegisterTypeForVT[i] = NVT; |
| 1841 | NumRegistersForVT[i] = 1; |
| 1842 | break; |
| 1843 | } |
| 1844 | } |
| 1845 | [[fallthrough]]; |
| 1846 | |
| 1847 | case TypeSplitVector: |
| 1848 | case TypeScalarizeVector: { |
| 1849 | MVT IntermediateVT; |
| 1850 | MVT RegisterVT; |
| 1851 | unsigned NumIntermediates; |
| 1852 | unsigned NumRegisters = getVectorTypeBreakdownMVT(VT, IntermediateVT, |
| 1853 | NumIntermediates, RegisterVT, TLI: this); |
| 1854 | NumRegistersForVT[i] = NumRegisters; |
| 1855 | assert(NumRegistersForVT[i] == NumRegisters && |
| 1856 | "NumRegistersForVT size cannot represent NumRegisters!" ); |
| 1857 | RegisterTypeForVT[i] = RegisterVT; |
| 1858 | |
| 1859 | MVT NVT = VT.getPow2VectorType(); |
| 1860 | if (NVT == VT) { |
| 1861 | // Type is already a power of 2. The default action is to split. |
| 1862 | TransformToType[i] = MVT::Other; |
| 1863 | if (PreferredAction == TypeScalarizeVector) |
| 1864 | ValueTypeActions.setTypeAction(VT, Action: TypeScalarizeVector); |
| 1865 | else if (PreferredAction == TypeSplitVector) |
| 1866 | ValueTypeActions.setTypeAction(VT, Action: TypeSplitVector); |
| 1867 | else if (EC.getKnownMinValue() > 1) |
| 1868 | ValueTypeActions.setTypeAction(VT, Action: TypeSplitVector); |
| 1869 | else |
| 1870 | ValueTypeActions.setTypeAction(VT, Action: EC.isScalable() |
| 1871 | ? TypeScalarizeScalableVector |
| 1872 | : TypeScalarizeVector); |
| 1873 | } else { |
| 1874 | TransformToType[i] = NVT; |
| 1875 | ValueTypeActions.setTypeAction(VT, Action: TypeWidenVector); |
| 1876 | } |
| 1877 | break; |
| 1878 | } |
| 1879 | default: |
| 1880 | llvm_unreachable("Unknown vector legalization action!" ); |
| 1881 | } |
| 1882 | } |
| 1883 | |
| 1884 | // Determine the 'representative' register class for each value type. |
| 1885 | // An representative register class is the largest (meaning one which is |
| 1886 | // not a sub-register class / subreg register class) legal register class for |
| 1887 | // a group of value types. For example, on i386, i8, i16, and i32 |
| 1888 | // representative would be GR32; while on x86_64 it's GR64. |
| 1889 | for (unsigned i = 0; i != MVT::VALUETYPE_SIZE; ++i) { |
| 1890 | const TargetRegisterClass* RRC; |
| 1891 | uint8_t Cost; |
| 1892 | std::tie(args&: RRC, args&: Cost) = findRepresentativeClass(TRI, VT: (MVT::SimpleValueType)i); |
| 1893 | RepRegClassForVT[i] = RRC; |
| 1894 | RepRegClassCostForVT[i] = Cost; |
| 1895 | } |
| 1896 | } |
| 1897 | |
| 1898 | EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, |
| 1899 | EVT VT) const { |
| 1900 | assert(!VT.isVector() && "No default SetCC type for vectors!" ); |
| 1901 | return getPointerTy(DL).SimpleTy; |
| 1902 | } |
| 1903 | |
| 1904 | MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { |
| 1905 | return MVT::i32; // return the default value |
| 1906 | } |
| 1907 | |
| 1908 | /// getVectorTypeBreakdown - Vector types are broken down into some number of |
| 1909 | /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 |
| 1910 | /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. |
| 1911 | /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. |
| 1912 | /// |
| 1913 | /// This method returns the number of registers needed, and the VT for each |
| 1914 | /// register. It also returns the VT and quantity of the intermediate values |
| 1915 | /// before they are promoted/expanded. |
| 1916 | unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, |
| 1917 | EVT VT, EVT &IntermediateVT, |
| 1918 | unsigned &NumIntermediates, |
| 1919 | MVT &RegisterVT) const { |
| 1920 | ElementCount EltCnt = VT.getVectorElementCount(); |
| 1921 | |
| 1922 | // If there is a wider vector type with the same element type as this one, |
| 1923 | // or a promoted vector type that has the same number of elements which |
| 1924 | // are wider, then we should convert to that legal vector type. |
| 1925 | // This handles things like <2 x float> -> <4 x float> and |
| 1926 | // <4 x i1> -> <4 x i32>. |
| 1927 | LegalizeTypeAction TA = getTypeAction(Context, VT); |
| 1928 | if (!EltCnt.isScalar() && |
| 1929 | (TA == TypeWidenVector || TA == TypePromoteInteger)) { |
| 1930 | EVT RegisterEVT = getTypeToTransformTo(Context, VT); |
| 1931 | if (isTypeLegal(VT: RegisterEVT)) { |
| 1932 | IntermediateVT = RegisterEVT; |
| 1933 | RegisterVT = RegisterEVT.getSimpleVT(); |
| 1934 | NumIntermediates = 1; |
| 1935 | return 1; |
| 1936 | } |
| 1937 | } |
| 1938 | |
| 1939 | // Figure out the right, legal destination reg to copy into. |
| 1940 | EVT EltTy = VT.getVectorElementType(); |
| 1941 | |
| 1942 | unsigned NumVectorRegs = 1; |
| 1943 | |
| 1944 | // Scalable vectors cannot be scalarized, so handle the legalisation of the |
| 1945 | // types like done elsewhere in SelectionDAG. |
| 1946 | if (EltCnt.isScalable()) { |
| 1947 | LegalizeKind LK; |
| 1948 | EVT PartVT = VT; |
| 1949 | do { |
| 1950 | // Iterate until we've found a legal (part) type to hold VT. |
| 1951 | LK = getTypeConversion(Context, VT: PartVT); |
| 1952 | PartVT = LK.second; |
| 1953 | } while (LK.first != TypeLegal); |
| 1954 | |
| 1955 | if (!PartVT.isVector()) { |
| 1956 | report_fatal_error( |
| 1957 | reason: "Don't know how to legalize this scalable vector type" ); |
| 1958 | } |
| 1959 | |
| 1960 | NumIntermediates = |
| 1961 | divideCeil(Numerator: VT.getVectorElementCount().getKnownMinValue(), |
| 1962 | Denominator: PartVT.getVectorElementCount().getKnownMinValue()); |
| 1963 | IntermediateVT = PartVT; |
| 1964 | RegisterVT = getRegisterType(Context, VT: IntermediateVT); |
| 1965 | return NumIntermediates; |
| 1966 | } |
| 1967 | |
| 1968 | // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally |
| 1969 | // we could break down into LHS/RHS like LegalizeDAG does. |
| 1970 | if (!isPowerOf2_32(Value: EltCnt.getKnownMinValue())) { |
| 1971 | NumVectorRegs = EltCnt.getKnownMinValue(); |
| 1972 | EltCnt = ElementCount::getFixed(MinVal: 1); |
| 1973 | } |
| 1974 | |
| 1975 | // Divide the input until we get to a supported size. This will always |
| 1976 | // end with a scalar if the target doesn't support vectors. |
| 1977 | while (EltCnt.getKnownMinValue() > 1 && |
| 1978 | !isTypeLegal(VT: EVT::getVectorVT(Context, VT: EltTy, EC: EltCnt))) { |
| 1979 | EltCnt = EltCnt.divideCoefficientBy(RHS: 2); |
| 1980 | NumVectorRegs <<= 1; |
| 1981 | } |
| 1982 | |
| 1983 | NumIntermediates = NumVectorRegs; |
| 1984 | |
| 1985 | EVT NewVT = EVT::getVectorVT(Context, VT: EltTy, EC: EltCnt); |
| 1986 | if (!isTypeLegal(VT: NewVT)) |
| 1987 | NewVT = EltTy; |
| 1988 | IntermediateVT = NewVT; |
| 1989 | |
| 1990 | MVT DestVT = getRegisterType(Context, VT: NewVT); |
| 1991 | RegisterVT = DestVT; |
| 1992 | |
| 1993 | if (EVT(DestVT).bitsLT(VT: NewVT)) { // Value is expanded, e.g. i64 -> i16. |
| 1994 | TypeSize NewVTSize = NewVT.getSizeInBits(); |
| 1995 | // Convert sizes such as i33 to i64. |
| 1996 | if (!llvm::has_single_bit<uint32_t>(Value: NewVTSize.getKnownMinValue())) |
| 1997 | NewVTSize = NewVTSize.coefficientNextPowerOf2(); |
| 1998 | return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); |
| 1999 | } |
| 2000 | |
| 2001 | // Otherwise, promotion or legal types use the same number of registers as |
| 2002 | // the vector decimated to the appropriate level. |
| 2003 | return NumVectorRegs; |
| 2004 | } |
| 2005 | |
| 2006 | bool TargetLoweringBase::isSuitableForJumpTable(const SwitchInst *SI, |
| 2007 | uint64_t NumCases, |
| 2008 | uint64_t Range, |
| 2009 | ProfileSummaryInfo *PSI, |
| 2010 | BlockFrequencyInfo *BFI) const { |
| 2011 | // FIXME: This function check the maximum table size and density, but the |
| 2012 | // minimum size is not checked. It would be nice if the minimum size is |
| 2013 | // also combined within this function. Currently, the minimum size check is |
| 2014 | // performed in findJumpTable() in SelectionDAGBuiler and |
| 2015 | // getEstimatedNumberOfCaseClusters() in BasicTTIImpl. |
| 2016 | const bool OptForSize = |
| 2017 | llvm::shouldOptimizeForSize(BB: SI->getParent(), PSI, BFI); |
| 2018 | const unsigned MinDensity = getMinimumJumpTableDensity(OptForSize); |
| 2019 | const unsigned MaxJumpTableSize = getMaximumJumpTableSize(); |
| 2020 | |
| 2021 | // Check whether the number of cases is small enough and |
| 2022 | // the range is dense enough for a jump table. |
| 2023 | return (OptForSize || Range <= MaxJumpTableSize) && |
| 2024 | (NumCases * 100 >= Range * MinDensity); |
| 2025 | } |
| 2026 | |
| 2027 | MVT TargetLoweringBase::getPreferredSwitchConditionType(LLVMContext &Context, |
| 2028 | EVT ConditionVT) const { |
| 2029 | return getRegisterType(Context, VT: ConditionVT); |
| 2030 | } |
| 2031 | |
| 2032 | /// Get the EVTs and ArgFlags collections that represent the legalized return |
| 2033 | /// type of the given function. This does not require a DAG or a return value, |
| 2034 | /// and is suitable for use before any DAGs for the function are constructed. |
| 2035 | /// TODO: Move this out of TargetLowering.cpp. |
| 2036 | void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, |
| 2037 | AttributeList attr, |
| 2038 | SmallVectorImpl<ISD::OutputArg> &Outs, |
| 2039 | const TargetLowering &TLI, const DataLayout &DL) { |
| 2040 | SmallVector<Type *, 4> Types; |
| 2041 | ComputeValueTypes(DL, Ty: ReturnType, Types); |
| 2042 | unsigned NumValues = Types.size(); |
| 2043 | if (NumValues == 0) return; |
| 2044 | |
| 2045 | for (Type *Ty : Types) { |
| 2046 | EVT VT = TLI.getValueType(DL, Ty); |
| 2047 | ISD::NodeType ExtendKind = ISD::ANY_EXTEND; |
| 2048 | |
| 2049 | if (attr.hasRetAttr(Kind: Attribute::SExt)) |
| 2050 | ExtendKind = ISD::SIGN_EXTEND; |
| 2051 | else if (attr.hasRetAttr(Kind: Attribute::ZExt)) |
| 2052 | ExtendKind = ISD::ZERO_EXTEND; |
| 2053 | |
| 2054 | if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) |
| 2055 | VT = TLI.getTypeForExtReturn(Context&: ReturnType->getContext(), VT, ExtendKind); |
| 2056 | |
| 2057 | unsigned NumParts = |
| 2058 | TLI.getNumRegistersForCallingConv(Context&: ReturnType->getContext(), CC, VT); |
| 2059 | MVT PartVT = |
| 2060 | TLI.getRegisterTypeForCallingConv(Context&: ReturnType->getContext(), CC, VT); |
| 2061 | |
| 2062 | // 'inreg' on function refers to return value |
| 2063 | ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); |
| 2064 | if (attr.hasRetAttr(Kind: Attribute::InReg)) |
| 2065 | Flags.setInReg(); |
| 2066 | |
| 2067 | // Propagate extension type if any |
| 2068 | if (attr.hasRetAttr(Kind: Attribute::SExt)) |
| 2069 | Flags.setSExt(); |
| 2070 | else if (attr.hasRetAttr(Kind: Attribute::ZExt)) |
| 2071 | Flags.setZExt(); |
| 2072 | |
| 2073 | for (unsigned i = 0; i < NumParts; ++i) |
| 2074 | Outs.push_back(Elt: ISD::OutputArg(Flags, PartVT, VT, Ty, 0, 0)); |
| 2075 | } |
| 2076 | } |
| 2077 | |
| 2078 | Align TargetLoweringBase::getByValTypeAlignment(Type *Ty, |
| 2079 | const DataLayout &DL) const { |
| 2080 | return DL.getABITypeAlign(Ty); |
| 2081 | } |
| 2082 | |
| 2083 | bool TargetLoweringBase::allowsMemoryAccessForAlignment( |
| 2084 | LLVMContext &Context, const DataLayout &DL, EVT VT, unsigned AddrSpace, |
| 2085 | Align Alignment, MachineMemOperand::Flags Flags, unsigned *Fast) const { |
| 2086 | // Check if the specified alignment is sufficient based on the data layout. |
| 2087 | // TODO: While using the data layout works in practice, a better solution |
| 2088 | // would be to implement this check directly (make this a virtual function). |
| 2089 | // For example, the ABI alignment may change based on software platform while |
| 2090 | // this function should only be affected by hardware implementation. |
| 2091 | Type *Ty = VT.getTypeForEVT(Context); |
| 2092 | if (VT.isZeroSized() || Alignment >= DL.getABITypeAlign(Ty)) { |
| 2093 | // Assume that an access that meets the ABI-specified alignment is fast. |
| 2094 | if (Fast != nullptr) |
| 2095 | *Fast = 1; |
| 2096 | return true; |
| 2097 | } |
| 2098 | |
| 2099 | // This is a misaligned access. |
| 2100 | return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Flags, Fast); |
| 2101 | } |
| 2102 | |
| 2103 | bool TargetLoweringBase::allowsMemoryAccessForAlignment( |
| 2104 | LLVMContext &Context, const DataLayout &DL, EVT VT, |
| 2105 | const MachineMemOperand &MMO, unsigned *Fast) const { |
| 2106 | return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace: MMO.getAddrSpace(), |
| 2107 | Alignment: MMO.getAlign(), Flags: MMO.getFlags(), Fast); |
| 2108 | } |
| 2109 | |
| 2110 | bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, |
| 2111 | const DataLayout &DL, EVT VT, |
| 2112 | unsigned AddrSpace, Align Alignment, |
| 2113 | MachineMemOperand::Flags Flags, |
| 2114 | unsigned *Fast) const { |
| 2115 | return allowsMemoryAccessForAlignment(Context, DL, VT, AddrSpace, Alignment, |
| 2116 | Flags, Fast); |
| 2117 | } |
| 2118 | |
| 2119 | bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, |
| 2120 | const DataLayout &DL, EVT VT, |
| 2121 | const MachineMemOperand &MMO, |
| 2122 | unsigned *Fast) const { |
| 2123 | return allowsMemoryAccess(Context, DL, VT, AddrSpace: MMO.getAddrSpace(), Alignment: MMO.getAlign(), |
| 2124 | Flags: MMO.getFlags(), Fast); |
| 2125 | } |
| 2126 | |
| 2127 | bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, |
| 2128 | const DataLayout &DL, LLT Ty, |
| 2129 | const MachineMemOperand &MMO, |
| 2130 | unsigned *Fast) const { |
| 2131 | EVT VT = getApproximateEVTForLLT(Ty, Ctx&: Context); |
| 2132 | return allowsMemoryAccess(Context, DL, VT, AddrSpace: MMO.getAddrSpace(), Alignment: MMO.getAlign(), |
| 2133 | Flags: MMO.getFlags(), Fast); |
| 2134 | } |
| 2135 | |
| 2136 | unsigned TargetLoweringBase::getMaxStoresPerMemset(bool OptSize) const { |
| 2137 | if (MaxStoresPerMemsetOverride > 0) |
| 2138 | return MaxStoresPerMemsetOverride; |
| 2139 | |
| 2140 | return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset; |
| 2141 | } |
| 2142 | |
| 2143 | unsigned TargetLoweringBase::getMaxStoresPerMemcpy(bool OptSize) const { |
| 2144 | if (MaxStoresPerMemcpyOverride > 0) |
| 2145 | return MaxStoresPerMemcpyOverride; |
| 2146 | |
| 2147 | return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy; |
| 2148 | } |
| 2149 | |
| 2150 | unsigned TargetLoweringBase::getMaxStoresPerMemmove(bool OptSize) const { |
| 2151 | if (MaxStoresPerMemmoveOverride > 0) |
| 2152 | return MaxStoresPerMemmoveOverride; |
| 2153 | |
| 2154 | return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove; |
| 2155 | } |
| 2156 | |
| 2157 | //===----------------------------------------------------------------------===// |
| 2158 | // TargetTransformInfo Helpers |
| 2159 | //===----------------------------------------------------------------------===// |
| 2160 | |
| 2161 | int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { |
| 2162 | enum InstructionOpcodes { |
| 2163 | #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, |
| 2164 | #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM |
| 2165 | #include "llvm/IR/Instruction.def" |
| 2166 | }; |
| 2167 | switch (static_cast<InstructionOpcodes>(Opcode)) { |
| 2168 | case Ret: return 0; |
| 2169 | case Br: return 0; |
| 2170 | case Switch: return 0; |
| 2171 | case IndirectBr: return 0; |
| 2172 | case Invoke: return 0; |
| 2173 | case CallBr: return 0; |
| 2174 | case Resume: return 0; |
| 2175 | case Unreachable: return 0; |
| 2176 | case CleanupRet: return 0; |
| 2177 | case CatchRet: return 0; |
| 2178 | case CatchPad: return 0; |
| 2179 | case CatchSwitch: return 0; |
| 2180 | case CleanupPad: return 0; |
| 2181 | case FNeg: return ISD::FNEG; |
| 2182 | case Add: return ISD::ADD; |
| 2183 | case FAdd: return ISD::FADD; |
| 2184 | case Sub: return ISD::SUB; |
| 2185 | case FSub: return ISD::FSUB; |
| 2186 | case Mul: return ISD::MUL; |
| 2187 | case FMul: return ISD::FMUL; |
| 2188 | case UDiv: return ISD::UDIV; |
| 2189 | case SDiv: return ISD::SDIV; |
| 2190 | case FDiv: return ISD::FDIV; |
| 2191 | case URem: return ISD::UREM; |
| 2192 | case SRem: return ISD::SREM; |
| 2193 | case FRem: return ISD::FREM; |
| 2194 | case Shl: return ISD::SHL; |
| 2195 | case LShr: return ISD::SRL; |
| 2196 | case AShr: return ISD::SRA; |
| 2197 | case And: return ISD::AND; |
| 2198 | case Or: return ISD::OR; |
| 2199 | case Xor: return ISD::XOR; |
| 2200 | case Alloca: return 0; |
| 2201 | case Load: return ISD::LOAD; |
| 2202 | case Store: return ISD::STORE; |
| 2203 | case GetElementPtr: return 0; |
| 2204 | case Fence: return 0; |
| 2205 | case AtomicCmpXchg: return 0; |
| 2206 | case AtomicRMW: return 0; |
| 2207 | case Trunc: return ISD::TRUNCATE; |
| 2208 | case ZExt: return ISD::ZERO_EXTEND; |
| 2209 | case SExt: return ISD::SIGN_EXTEND; |
| 2210 | case FPToUI: return ISD::FP_TO_UINT; |
| 2211 | case FPToSI: return ISD::FP_TO_SINT; |
| 2212 | case UIToFP: return ISD::UINT_TO_FP; |
| 2213 | case SIToFP: return ISD::SINT_TO_FP; |
| 2214 | case FPTrunc: return ISD::FP_ROUND; |
| 2215 | case FPExt: return ISD::FP_EXTEND; |
| 2216 | case PtrToAddr: return ISD::BITCAST; |
| 2217 | case PtrToInt: return ISD::BITCAST; |
| 2218 | case IntToPtr: return ISD::BITCAST; |
| 2219 | case BitCast: return ISD::BITCAST; |
| 2220 | case AddrSpaceCast: return ISD::ADDRSPACECAST; |
| 2221 | case ICmp: return ISD::SETCC; |
| 2222 | case FCmp: return ISD::SETCC; |
| 2223 | case PHI: return 0; |
| 2224 | case Call: return 0; |
| 2225 | case Select: return ISD::SELECT; |
| 2226 | case UserOp1: return 0; |
| 2227 | case UserOp2: return 0; |
| 2228 | case VAArg: return 0; |
| 2229 | case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; |
| 2230 | case InsertElement: return ISD::INSERT_VECTOR_ELT; |
| 2231 | case ShuffleVector: return ISD::VECTOR_SHUFFLE; |
| 2232 | case ExtractValue: return ISD::MERGE_VALUES; |
| 2233 | case InsertValue: return ISD::MERGE_VALUES; |
| 2234 | case LandingPad: return 0; |
| 2235 | case Freeze: return ISD::FREEZE; |
| 2236 | } |
| 2237 | |
| 2238 | llvm_unreachable("Unknown instruction type encountered!" ); |
| 2239 | } |
| 2240 | |
| 2241 | int TargetLoweringBase::IntrinsicIDToISD(Intrinsic::ID ID) const { |
| 2242 | switch (ID) { |
| 2243 | case Intrinsic::exp: |
| 2244 | return ISD::FEXP; |
| 2245 | case Intrinsic::exp2: |
| 2246 | return ISD::FEXP2; |
| 2247 | case Intrinsic::log: |
| 2248 | return ISD::FLOG; |
| 2249 | default: |
| 2250 | return ISD::DELETED_NODE; |
| 2251 | } |
| 2252 | } |
| 2253 | |
| 2254 | Value * |
| 2255 | TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilderBase &IRB, |
| 2256 | bool UseTLS) const { |
| 2257 | // compiler-rt provides a variable with a magic name. Targets that do not |
| 2258 | // link with compiler-rt may also provide such a variable. |
| 2259 | Module *M = IRB.GetInsertBlock()->getParent()->getParent(); |
| 2260 | const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr" ; |
| 2261 | auto UnsafeStackPtr = |
| 2262 | dyn_cast_or_null<GlobalVariable>(Val: M->getNamedValue(Name: UnsafeStackPtrVar)); |
| 2263 | |
| 2264 | const DataLayout &DL = M->getDataLayout(); |
| 2265 | PointerType *StackPtrTy = DL.getAllocaPtrType(Ctx&: M->getContext()); |
| 2266 | |
| 2267 | if (!UnsafeStackPtr) { |
| 2268 | auto TLSModel = UseTLS ? |
| 2269 | GlobalValue::InitialExecTLSModel : |
| 2270 | GlobalValue::NotThreadLocal; |
| 2271 | // The global variable is not defined yet, define it ourselves. |
| 2272 | // We use the initial-exec TLS model because we do not support the |
| 2273 | // variable living anywhere other than in the main executable. |
| 2274 | UnsafeStackPtr = new GlobalVariable( |
| 2275 | *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, |
| 2276 | UnsafeStackPtrVar, nullptr, TLSModel); |
| 2277 | } else { |
| 2278 | // The variable exists, check its type and attributes. |
| 2279 | // |
| 2280 | // FIXME: Move to IR verifier. |
| 2281 | if (UnsafeStackPtr->getValueType() != StackPtrTy) |
| 2282 | report_fatal_error(reason: Twine(UnsafeStackPtrVar) + " must have void* type" ); |
| 2283 | if (UseTLS != UnsafeStackPtr->isThreadLocal()) |
| 2284 | report_fatal_error(reason: Twine(UnsafeStackPtrVar) + " must " + |
| 2285 | (UseTLS ? "" : "not " ) + "be thread-local" ); |
| 2286 | } |
| 2287 | return UnsafeStackPtr; |
| 2288 | } |
| 2289 | |
| 2290 | Value *TargetLoweringBase::getSafeStackPointerLocation( |
| 2291 | IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const { |
| 2292 | // FIXME: Can this triple check be replaced with SAFESTACK_POINTER_ADDRESS |
| 2293 | // being available? |
| 2294 | if (!TM.getTargetTriple().isAndroid()) |
| 2295 | return getDefaultSafeStackPointerLocation(IRB, UseTLS: true); |
| 2296 | |
| 2297 | Module *M = IRB.GetInsertBlock()->getParent()->getParent(); |
| 2298 | auto *PtrTy = PointerType::getUnqual(C&: M->getContext()); |
| 2299 | |
| 2300 | RTLIB::LibcallImpl SafestackPointerAddressImpl = |
| 2301 | Libcalls.getLibcallImpl(Call: RTLIB::SAFESTACK_POINTER_ADDRESS); |
| 2302 | if (SafestackPointerAddressImpl == RTLIB::Unsupported) { |
| 2303 | M->getContext().emitError( |
| 2304 | ErrorStr: "no libcall available for safestack pointer address" ); |
| 2305 | return PoisonValue::get(T: PtrTy); |
| 2306 | } |
| 2307 | |
| 2308 | // Android provides a libc function to retrieve the address of the current |
| 2309 | // thread's unsafe stack pointer. |
| 2310 | FunctionCallee Fn = |
| 2311 | M->getOrInsertFunction(Name: RTLIB::RuntimeLibcallsInfo::getLibcallImplName( |
| 2312 | CallImpl: SafestackPointerAddressImpl), |
| 2313 | RetTy: PtrTy); |
| 2314 | return IRB.CreateCall(Callee: Fn); |
| 2315 | } |
| 2316 | |
| 2317 | //===----------------------------------------------------------------------===// |
| 2318 | // Loop Strength Reduction hooks |
| 2319 | //===----------------------------------------------------------------------===// |
| 2320 | |
| 2321 | /// isLegalAddressingMode - Return true if the addressing mode represented |
| 2322 | /// by AM is legal for this target, for a load/store of the specified type. |
| 2323 | bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, |
| 2324 | const AddrMode &AM, Type *Ty, |
| 2325 | unsigned AS, Instruction *I) const { |
| 2326 | // The default implementation of this implements a conservative RISCy, r+r and |
| 2327 | // r+i addr mode. |
| 2328 | |
| 2329 | // Scalable offsets not supported |
| 2330 | if (AM.ScalableOffset) |
| 2331 | return false; |
| 2332 | |
| 2333 | // Allows a sign-extended 16-bit immediate field. |
| 2334 | if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) |
| 2335 | return false; |
| 2336 | |
| 2337 | // No global is ever allowed as a base. |
| 2338 | if (AM.BaseGV) |
| 2339 | return false; |
| 2340 | |
| 2341 | // Only support r+r, |
| 2342 | switch (AM.Scale) { |
| 2343 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
| 2344 | break; |
| 2345 | case 1: |
| 2346 | if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. |
| 2347 | return false; |
| 2348 | // Otherwise we have r+r or r+i. |
| 2349 | break; |
| 2350 | case 2: |
| 2351 | if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. |
| 2352 | return false; |
| 2353 | // Allow 2*r as r+r. |
| 2354 | break; |
| 2355 | default: // Don't allow n * r |
| 2356 | return false; |
| 2357 | } |
| 2358 | |
| 2359 | return true; |
| 2360 | } |
| 2361 | |
| 2362 | //===----------------------------------------------------------------------===// |
| 2363 | // Stack Protector |
| 2364 | //===----------------------------------------------------------------------===// |
| 2365 | |
| 2366 | // For OpenBSD return its special guard variable. Otherwise return nullptr, |
| 2367 | // so that SelectionDAG handle SSP. |
| 2368 | Value * |
| 2369 | TargetLoweringBase::getIRStackGuard(IRBuilderBase &IRB, |
| 2370 | const LibcallLoweringInfo &Libcalls) const { |
| 2371 | RTLIB::LibcallImpl GuardLocalImpl = |
| 2372 | Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD); |
| 2373 | if (GuardLocalImpl != RTLIB::impl___guard_local) |
| 2374 | return nullptr; |
| 2375 | |
| 2376 | Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); |
| 2377 | const DataLayout &DL = M.getDataLayout(); |
| 2378 | PointerType *PtrTy = |
| 2379 | PointerType::get(C&: M.getContext(), AddressSpace: DL.getDefaultGlobalsAddressSpace()); |
| 2380 | GlobalVariable *G = |
| 2381 | M.getOrInsertGlobal(Name: getLibcallImplName(Call: GuardLocalImpl), Ty: PtrTy); |
| 2382 | G->setVisibility(GlobalValue::HiddenVisibility); |
| 2383 | return G; |
| 2384 | } |
| 2385 | |
| 2386 | // Currently only support "standard" __stack_chk_guard. |
| 2387 | // TODO: add LOAD_STACK_GUARD support. |
| 2388 | void TargetLoweringBase::insertSSPDeclarations( |
| 2389 | Module &M, const LibcallLoweringInfo &Libcalls) const { |
| 2390 | RTLIB::LibcallImpl StackGuardImpl = |
| 2391 | Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD); |
| 2392 | if (StackGuardImpl == RTLIB::Unsupported) |
| 2393 | return; |
| 2394 | |
| 2395 | StringRef StackGuardVarName = getLibcallImplName(Call: StackGuardImpl); |
| 2396 | M.getOrInsertGlobal( |
| 2397 | Name: StackGuardVarName, Ty: PointerType::getUnqual(C&: M.getContext()), CreateGlobalCallback: [=, &M]() { |
| 2398 | auto *GV = new GlobalVariable(M, PointerType::getUnqual(C&: M.getContext()), |
| 2399 | false, GlobalVariable::ExternalLinkage, |
| 2400 | nullptr, StackGuardVarName); |
| 2401 | |
| 2402 | // FreeBSD has "__stack_chk_guard" defined externally on libc.so |
| 2403 | if (M.getDirectAccessExternalData() && |
| 2404 | !TM.getTargetTriple().isOSCygMing() && |
| 2405 | !(TM.getTargetTriple().isPPC64() && |
| 2406 | TM.getTargetTriple().isOSFreeBSD()) && |
| 2407 | (!TM.getTargetTriple().isOSDarwin() || |
| 2408 | TM.getRelocationModel() == Reloc::Static)) |
| 2409 | GV->setDSOLocal(true); |
| 2410 | |
| 2411 | return GV; |
| 2412 | }); |
| 2413 | } |
| 2414 | |
| 2415 | // Currently only support "standard" __stack_chk_guard. |
| 2416 | // TODO: add LOAD_STACK_GUARD support. |
| 2417 | Value *TargetLoweringBase::getSDagStackGuard( |
| 2418 | const Module &M, const LibcallLoweringInfo &Libcalls) const { |
| 2419 | RTLIB::LibcallImpl GuardVarImpl = |
| 2420 | Libcalls.getLibcallImpl(Call: RTLIB::STACK_CHECK_GUARD); |
| 2421 | if (GuardVarImpl == RTLIB::Unsupported) |
| 2422 | return nullptr; |
| 2423 | return M.getNamedValue(Name: getLibcallImplName(Call: GuardVarImpl)); |
| 2424 | } |
| 2425 | |
| 2426 | Function *TargetLoweringBase::getSSPStackGuardCheck( |
| 2427 | const Module &M, const LibcallLoweringInfo &Libcalls) const { |
| 2428 | // MSVC CRT has a function to validate security cookie. |
| 2429 | RTLIB::LibcallImpl SecurityCheckCookieLibcall = |
| 2430 | Libcalls.getLibcallImpl(Call: RTLIB::SECURITY_CHECK_COOKIE); |
| 2431 | if (SecurityCheckCookieLibcall != RTLIB::Unsupported) |
| 2432 | return M.getFunction(Name: getLibcallImplName(Call: SecurityCheckCookieLibcall)); |
| 2433 | return nullptr; |
| 2434 | } |
| 2435 | |
| 2436 | unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { |
| 2437 | return MinimumJumpTableEntries; |
| 2438 | } |
| 2439 | |
| 2440 | void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { |
| 2441 | MinimumJumpTableEntries = Val; |
| 2442 | } |
| 2443 | |
| 2444 | unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { |
| 2445 | return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; |
| 2446 | } |
| 2447 | |
| 2448 | unsigned TargetLoweringBase::getMaximumJumpTableSize() const { |
| 2449 | return MaximumJumpTableSize; |
| 2450 | } |
| 2451 | |
| 2452 | void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { |
| 2453 | MaximumJumpTableSize = Val; |
| 2454 | } |
| 2455 | |
| 2456 | bool TargetLoweringBase::isJumpTableRelative() const { |
| 2457 | return getTargetMachine().isPositionIndependent(); |
| 2458 | } |
| 2459 | |
| 2460 | unsigned TargetLoweringBase::getMinimumBitTestCmps() const { |
| 2461 | return MinimumBitTestCmps; |
| 2462 | } |
| 2463 | |
| 2464 | void TargetLoweringBase::setMinimumBitTestCmps(unsigned Val) { |
| 2465 | MinimumBitTestCmps = Val; |
| 2466 | } |
| 2467 | |
| 2468 | Align TargetLoweringBase::getPrefLoopAlignment(MachineLoop *ML) const { |
| 2469 | if (TM.Options.LoopAlignment) |
| 2470 | return Align(TM.Options.LoopAlignment); |
| 2471 | return PrefLoopAlignment; |
| 2472 | } |
| 2473 | |
| 2474 | unsigned TargetLoweringBase::getMaxPermittedBytesForAlignment( |
| 2475 | MachineBasicBlock *MBB) const { |
| 2476 | return MaxBytesForAlignment; |
| 2477 | } |
| 2478 | |
| 2479 | //===----------------------------------------------------------------------===// |
| 2480 | // Reciprocal Estimates |
| 2481 | //===----------------------------------------------------------------------===// |
| 2482 | |
| 2483 | /// Get the reciprocal estimate attribute string for a function that will |
| 2484 | /// override the target defaults. |
| 2485 | static StringRef getRecipEstimateForFunc(MachineFunction &MF) { |
| 2486 | const Function &F = MF.getFunction(); |
| 2487 | return F.getFnAttribute(Kind: "reciprocal-estimates" ).getValueAsString(); |
| 2488 | } |
| 2489 | |
| 2490 | /// Construct a string for the given reciprocal operation of the given type. |
| 2491 | /// This string should match the corresponding option to the front-end's |
| 2492 | /// "-mrecip" flag assuming those strings have been passed through in an |
| 2493 | /// attribute string. For example, "vec-divf" for a division of a vXf32. |
| 2494 | static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { |
| 2495 | std::string Name = VT.isVector() ? "vec-" : "" ; |
| 2496 | |
| 2497 | Name += IsSqrt ? "sqrt" : "div" ; |
| 2498 | |
| 2499 | // TODO: Handle other float types? |
| 2500 | if (VT.getScalarType() == MVT::f64) { |
| 2501 | Name += "d" ; |
| 2502 | } else if (VT.getScalarType() == MVT::f16) { |
| 2503 | Name += "h" ; |
| 2504 | } else { |
| 2505 | assert(VT.getScalarType() == MVT::f32 && |
| 2506 | "Unexpected FP type for reciprocal estimate" ); |
| 2507 | Name += "f" ; |
| 2508 | } |
| 2509 | |
| 2510 | return Name; |
| 2511 | } |
| 2512 | |
| 2513 | /// Return the character position and value (a single numeric character) of a |
| 2514 | /// customized refinement operation in the input string if it exists. Return |
| 2515 | /// false if there is no customized refinement step count. |
| 2516 | static bool parseRefinementStep(StringRef In, size_t &Position, |
| 2517 | uint8_t &Value) { |
| 2518 | const char RefStepToken = ':'; |
| 2519 | Position = In.find(C: RefStepToken); |
| 2520 | if (Position == StringRef::npos) |
| 2521 | return false; |
| 2522 | |
| 2523 | StringRef RefStepString = In.substr(Start: Position + 1); |
| 2524 | // Allow exactly one numeric character for the additional refinement |
| 2525 | // step parameter. |
| 2526 | if (RefStepString.size() == 1) { |
| 2527 | char RefStepChar = RefStepString[0]; |
| 2528 | if (isDigit(C: RefStepChar)) { |
| 2529 | Value = RefStepChar - '0'; |
| 2530 | return true; |
| 2531 | } |
| 2532 | } |
| 2533 | report_fatal_error(reason: "Invalid refinement step for -recip." ); |
| 2534 | } |
| 2535 | |
| 2536 | /// For the input attribute string, return one of the ReciprocalEstimate enum |
| 2537 | /// status values (enabled, disabled, or not specified) for this operation on |
| 2538 | /// the specified data type. |
| 2539 | static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { |
| 2540 | if (Override.empty()) |
| 2541 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2542 | |
| 2543 | SmallVector<StringRef, 4> OverrideVector; |
| 2544 | Override.split(A&: OverrideVector, Separator: ','); |
| 2545 | unsigned NumArgs = OverrideVector.size(); |
| 2546 | |
| 2547 | // Check if "all", "none", or "default" was specified. |
| 2548 | if (NumArgs == 1) { |
| 2549 | // Look for an optional setting of the number of refinement steps needed |
| 2550 | // for this type of reciprocal operation. |
| 2551 | size_t RefPos; |
| 2552 | uint8_t RefSteps; |
| 2553 | if (parseRefinementStep(In: Override, Position&: RefPos, Value&: RefSteps)) { |
| 2554 | // Split the string for further processing. |
| 2555 | Override = Override.substr(Start: 0, N: RefPos); |
| 2556 | } |
| 2557 | |
| 2558 | // All reciprocal types are enabled. |
| 2559 | if (Override == "all" ) |
| 2560 | return TargetLoweringBase::ReciprocalEstimate::Enabled; |
| 2561 | |
| 2562 | // All reciprocal types are disabled. |
| 2563 | if (Override == "none" ) |
| 2564 | return TargetLoweringBase::ReciprocalEstimate::Disabled; |
| 2565 | |
| 2566 | // Target defaults for enablement are used. |
| 2567 | if (Override == "default" ) |
| 2568 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2569 | } |
| 2570 | |
| 2571 | // The attribute string may omit the size suffix ('f'/'d'). |
| 2572 | std::string VTName = getReciprocalOpName(IsSqrt, VT); |
| 2573 | std::string VTNameNoSize = VTName; |
| 2574 | VTNameNoSize.pop_back(); |
| 2575 | static const char DisabledPrefix = '!'; |
| 2576 | |
| 2577 | for (StringRef RecipType : OverrideVector) { |
| 2578 | size_t RefPos; |
| 2579 | uint8_t RefSteps; |
| 2580 | if (parseRefinementStep(In: RecipType, Position&: RefPos, Value&: RefSteps)) |
| 2581 | RecipType = RecipType.substr(Start: 0, N: RefPos); |
| 2582 | |
| 2583 | // Ignore the disablement token for string matching. |
| 2584 | bool IsDisabled = RecipType[0] == DisabledPrefix; |
| 2585 | if (IsDisabled) |
| 2586 | RecipType = RecipType.substr(Start: 1); |
| 2587 | |
| 2588 | if (RecipType == VTName || RecipType == VTNameNoSize) |
| 2589 | return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled |
| 2590 | : TargetLoweringBase::ReciprocalEstimate::Enabled; |
| 2591 | } |
| 2592 | |
| 2593 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2594 | } |
| 2595 | |
| 2596 | /// For the input attribute string, return the customized refinement step count |
| 2597 | /// for this operation on the specified data type. If the step count does not |
| 2598 | /// exist, return the ReciprocalEstimate enum value for unspecified. |
| 2599 | static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { |
| 2600 | if (Override.empty()) |
| 2601 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2602 | |
| 2603 | SmallVector<StringRef, 4> OverrideVector; |
| 2604 | Override.split(A&: OverrideVector, Separator: ','); |
| 2605 | unsigned NumArgs = OverrideVector.size(); |
| 2606 | |
| 2607 | // Check if "all", "default", or "none" was specified. |
| 2608 | if (NumArgs == 1) { |
| 2609 | // Look for an optional setting of the number of refinement steps needed |
| 2610 | // for this type of reciprocal operation. |
| 2611 | size_t RefPos; |
| 2612 | uint8_t RefSteps; |
| 2613 | if (!parseRefinementStep(In: Override, Position&: RefPos, Value&: RefSteps)) |
| 2614 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2615 | |
| 2616 | // Split the string for further processing. |
| 2617 | Override = Override.substr(Start: 0, N: RefPos); |
| 2618 | assert(Override != "none" && |
| 2619 | "Disabled reciprocals, but specifed refinement steps?" ); |
| 2620 | |
| 2621 | // If this is a general override, return the specified number of steps. |
| 2622 | if (Override == "all" || Override == "default" ) |
| 2623 | return RefSteps; |
| 2624 | } |
| 2625 | |
| 2626 | // The attribute string may omit the size suffix ('f'/'d'). |
| 2627 | std::string VTName = getReciprocalOpName(IsSqrt, VT); |
| 2628 | std::string VTNameNoSize = VTName; |
| 2629 | VTNameNoSize.pop_back(); |
| 2630 | |
| 2631 | for (StringRef RecipType : OverrideVector) { |
| 2632 | size_t RefPos; |
| 2633 | uint8_t RefSteps; |
| 2634 | if (!parseRefinementStep(In: RecipType, Position&: RefPos, Value&: RefSteps)) |
| 2635 | continue; |
| 2636 | |
| 2637 | RecipType = RecipType.substr(Start: 0, N: RefPos); |
| 2638 | if (RecipType == VTName || RecipType == VTNameNoSize) |
| 2639 | return RefSteps; |
| 2640 | } |
| 2641 | |
| 2642 | return TargetLoweringBase::ReciprocalEstimate::Unspecified; |
| 2643 | } |
| 2644 | |
| 2645 | int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, |
| 2646 | MachineFunction &MF) const { |
| 2647 | return getOpEnabled(IsSqrt: true, VT, Override: getRecipEstimateForFunc(MF)); |
| 2648 | } |
| 2649 | |
| 2650 | int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, |
| 2651 | MachineFunction &MF) const { |
| 2652 | return getOpEnabled(IsSqrt: false, VT, Override: getRecipEstimateForFunc(MF)); |
| 2653 | } |
| 2654 | |
| 2655 | int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, |
| 2656 | MachineFunction &MF) const { |
| 2657 | return getOpRefinementSteps(IsSqrt: true, VT, Override: getRecipEstimateForFunc(MF)); |
| 2658 | } |
| 2659 | |
| 2660 | int TargetLoweringBase::getDivRefinementSteps(EVT VT, |
| 2661 | MachineFunction &MF) const { |
| 2662 | return getOpRefinementSteps(IsSqrt: false, VT, Override: getRecipEstimateForFunc(MF)); |
| 2663 | } |
| 2664 | |
| 2665 | bool TargetLoweringBase::isLoadBitCastBeneficial( |
| 2666 | EVT LoadVT, EVT BitcastVT, const SelectionDAG &DAG, |
| 2667 | const MachineMemOperand &MMO) const { |
| 2668 | // Single-element vectors are scalarized, so we should generally avoid having |
| 2669 | // any memory operations on such types, as they would get scalarized too. |
| 2670 | if (LoadVT.isFixedLengthVector() && BitcastVT.isFixedLengthVector() && |
| 2671 | BitcastVT.getVectorNumElements() == 1) |
| 2672 | return false; |
| 2673 | |
| 2674 | // Don't do if we could do an indexed load on the original type, but not on |
| 2675 | // the new one. |
| 2676 | if (!LoadVT.isSimple() || !BitcastVT.isSimple()) |
| 2677 | return true; |
| 2678 | |
| 2679 | MVT LoadMVT = LoadVT.getSimpleVT(); |
| 2680 | |
| 2681 | // Don't bother doing this if it's just going to be promoted again later, as |
| 2682 | // doing so might interfere with other combines. |
| 2683 | if (getOperationAction(Op: ISD::LOAD, VT: LoadMVT) == Promote && |
| 2684 | getTypeToPromoteTo(Op: ISD::LOAD, VT: LoadMVT) == BitcastVT.getSimpleVT()) |
| 2685 | return false; |
| 2686 | |
| 2687 | unsigned Fast = 0; |
| 2688 | return allowsMemoryAccess(Context&: *DAG.getContext(), DL: DAG.getDataLayout(), VT: BitcastVT, |
| 2689 | MMO, Fast: &Fast) && |
| 2690 | Fast; |
| 2691 | } |
| 2692 | |
| 2693 | void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { |
| 2694 | MF.getRegInfo().freezeReservedRegs(); |
| 2695 | } |
| 2696 | |
| 2697 | MachineMemOperand::Flags TargetLoweringBase::getLoadMemOperandFlags( |
| 2698 | const LoadInst &LI, const DataLayout &DL, AssumptionCache *AC, |
| 2699 | const TargetLibraryInfo *LibInfo) const { |
| 2700 | MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad; |
| 2701 | if (LI.isVolatile()) |
| 2702 | Flags |= MachineMemOperand::MOVolatile; |
| 2703 | |
| 2704 | if (LI.hasMetadata(KindID: LLVMContext::MD_nontemporal)) |
| 2705 | Flags |= MachineMemOperand::MONonTemporal; |
| 2706 | |
| 2707 | if (LI.hasMetadata(KindID: LLVMContext::MD_invariant_load)) |
| 2708 | Flags |= MachineMemOperand::MOInvariant; |
| 2709 | |
| 2710 | if (isDereferenceableAndAlignedPointer(V: LI.getPointerOperand(), Ty: LI.getType(), |
| 2711 | Alignment: LI.getAlign(), DL, CtxI: &LI, AC, |
| 2712 | /*DT=*/nullptr, TLI: LibInfo)) |
| 2713 | Flags |= MachineMemOperand::MODereferenceable; |
| 2714 | |
| 2715 | Flags |= getTargetMMOFlags(I: LI); |
| 2716 | return Flags; |
| 2717 | } |
| 2718 | |
| 2719 | MachineMemOperand::Flags |
| 2720 | TargetLoweringBase::getStoreMemOperandFlags(const StoreInst &SI, |
| 2721 | const DataLayout &DL) const { |
| 2722 | MachineMemOperand::Flags Flags = MachineMemOperand::MOStore; |
| 2723 | |
| 2724 | if (SI.isVolatile()) |
| 2725 | Flags |= MachineMemOperand::MOVolatile; |
| 2726 | |
| 2727 | if (SI.hasMetadata(KindID: LLVMContext::MD_nontemporal)) |
| 2728 | Flags |= MachineMemOperand::MONonTemporal; |
| 2729 | |
| 2730 | // FIXME: Not preserving dereferenceable |
| 2731 | Flags |= getTargetMMOFlags(I: SI); |
| 2732 | return Flags; |
| 2733 | } |
| 2734 | |
| 2735 | MachineMemOperand::Flags |
| 2736 | TargetLoweringBase::getAtomicMemOperandFlags(const Instruction &AI, |
| 2737 | const DataLayout &DL) const { |
| 2738 | auto Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; |
| 2739 | |
| 2740 | if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: &AI)) { |
| 2741 | if (RMW->isVolatile()) |
| 2742 | Flags |= MachineMemOperand::MOVolatile; |
| 2743 | } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: &AI)) { |
| 2744 | if (CmpX->isVolatile()) |
| 2745 | Flags |= MachineMemOperand::MOVolatile; |
| 2746 | } else |
| 2747 | llvm_unreachable("not an atomic instruction" ); |
| 2748 | |
| 2749 | // FIXME: Not preserving dereferenceable |
| 2750 | Flags |= getTargetMMOFlags(I: AI); |
| 2751 | return Flags; |
| 2752 | } |
| 2753 | |
| 2754 | MachineMemOperand::Flags TargetLoweringBase::getVPIntrinsicMemOperandFlags( |
| 2755 | const VPIntrinsic &VPIntrin) const { |
| 2756 | MachineMemOperand::Flags Flags = MachineMemOperand::MONone; |
| 2757 | Intrinsic::ID IntrinID = VPIntrin.getIntrinsicID(); |
| 2758 | |
| 2759 | switch (IntrinID) { |
| 2760 | default: |
| 2761 | llvm_unreachable("unexpected intrinsic. Existing code may be appropriate " |
| 2762 | "for it, but support must be explicitly enabled" ); |
| 2763 | case Intrinsic::vp_load: |
| 2764 | case Intrinsic::vp_gather: |
| 2765 | case Intrinsic::experimental_vp_strided_load: |
| 2766 | Flags = MachineMemOperand::MOLoad; |
| 2767 | break; |
| 2768 | case Intrinsic::vp_store: |
| 2769 | case Intrinsic::vp_scatter: |
| 2770 | case Intrinsic::experimental_vp_strided_store: |
| 2771 | Flags = MachineMemOperand::MOStore; |
| 2772 | break; |
| 2773 | } |
| 2774 | |
| 2775 | if (VPIntrin.hasMetadata(KindID: LLVMContext::MD_nontemporal)) |
| 2776 | Flags |= MachineMemOperand::MONonTemporal; |
| 2777 | |
| 2778 | Flags |= getTargetMMOFlags(I: VPIntrin); |
| 2779 | return Flags; |
| 2780 | } |
| 2781 | |
| 2782 | Instruction *TargetLoweringBase::emitLeadingFence(IRBuilderBase &Builder, |
| 2783 | Instruction *Inst, |
| 2784 | AtomicOrdering Ord) const { |
| 2785 | if (isReleaseOrStronger(AO: Ord) && Inst->hasAtomicStore()) |
| 2786 | return Builder.CreateFence(Ordering: Ord); |
| 2787 | else |
| 2788 | return nullptr; |
| 2789 | } |
| 2790 | |
| 2791 | Instruction *TargetLoweringBase::emitTrailingFence(IRBuilderBase &Builder, |
| 2792 | Instruction *Inst, |
| 2793 | AtomicOrdering Ord) const { |
| 2794 | if (isAcquireOrStronger(AO: Ord)) |
| 2795 | return Builder.CreateFence(Ordering: Ord); |
| 2796 | else |
| 2797 | return nullptr; |
| 2798 | } |
| 2799 | |
| 2800 | //===----------------------------------------------------------------------===// |
| 2801 | // GlobalISel Hooks |
| 2802 | //===----------------------------------------------------------------------===// |
| 2803 | |
| 2804 | bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI, |
| 2805 | const TargetTransformInfo *TTI) const { |
| 2806 | auto &MF = *MI.getMF(); |
| 2807 | auto &MRI = MF.getRegInfo(); |
| 2808 | // Assuming a spill and reload of a value has a cost of 1 instruction each, |
| 2809 | // this helper function computes the maximum number of uses we should consider |
| 2810 | // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We |
| 2811 | // break even in terms of code size when the original MI has 2 users vs |
| 2812 | // choosing to potentially spill. Any more than 2 users we we have a net code |
| 2813 | // size increase. This doesn't take into account register pressure though. |
| 2814 | auto maxUses = [](unsigned RematCost) { |
| 2815 | // A cost of 1 means remats are basically free. |
| 2816 | if (RematCost == 1) |
| 2817 | return std::numeric_limits<unsigned>::max(); |
| 2818 | if (RematCost == 2) |
| 2819 | return 2U; |
| 2820 | |
| 2821 | // Remat is too expensive, only sink if there's one user. |
| 2822 | if (RematCost > 2) |
| 2823 | return 1U; |
| 2824 | llvm_unreachable("Unexpected remat cost" ); |
| 2825 | }; |
| 2826 | |
| 2827 | switch (MI.getOpcode()) { |
| 2828 | default: |
| 2829 | return false; |
| 2830 | // Constants-like instructions should be close to their users. |
| 2831 | // We don't want long live-ranges for them. |
| 2832 | case TargetOpcode::G_CONSTANT: |
| 2833 | case TargetOpcode::G_FCONSTANT: |
| 2834 | case TargetOpcode::G_FRAME_INDEX: |
| 2835 | case TargetOpcode::G_INTTOPTR: |
| 2836 | return true; |
| 2837 | case TargetOpcode::G_GLOBAL_VALUE: { |
| 2838 | unsigned RematCost = TTI->getGISelRematGlobalCost(); |
| 2839 | Register Reg = MI.getOperand(i: 0).getReg(); |
| 2840 | unsigned MaxUses = maxUses(RematCost); |
| 2841 | if (MaxUses == UINT_MAX) |
| 2842 | return true; // Remats are "free" so always localize. |
| 2843 | return MRI.hasAtMostUserInstrs(Reg, MaxUsers: MaxUses); |
| 2844 | } |
| 2845 | } |
| 2846 | } |
| 2847 | |