1 //===- ARMLegalizerInfo.cpp --------------------------------------*- C++ -*-==// 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 /// \file 9 /// This file implements the targeting of the Machinelegalizer class for ARM. 10 /// \todo This should be generated by TableGen. 11 //===----------------------------------------------------------------------===// 12 13 #include "ARMLegalizerInfo.h" 14 #include "ARMCallLowering.h" 15 #include "ARMSubtarget.h" 16 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h" 17 #include "llvm/CodeGen/LowLevelType.h" 18 #include "llvm/CodeGen/MachineRegisterInfo.h" 19 #include "llvm/CodeGen/TargetOpcodes.h" 20 #include "llvm/CodeGen/ValueTypes.h" 21 #include "llvm/IR/DerivedTypes.h" 22 #include "llvm/IR/Type.h" 23 24 using namespace llvm; 25 using namespace LegalizeActions; 26 27 /// FIXME: The following static functions are SizeChangeStrategy functions 28 /// that are meant to temporarily mimic the behaviour of the old legalization 29 /// based on doubling/halving non-legal types as closely as possible. This is 30 /// not entirly possible as only legalizing the types that are exactly a power 31 /// of 2 times the size of the legal types would require specifying all those 32 /// sizes explicitly. 33 /// In practice, not specifying those isn't a problem, and the below functions 34 /// should disappear quickly as we add support for legalizing non-power-of-2 35 /// sized types further. 36 static void 37 addAndInterleaveWithUnsupported(LegalizerInfo::SizeAndActionsVec &result, 38 const LegalizerInfo::SizeAndActionsVec &v) { 39 for (unsigned i = 0; i < v.size(); ++i) { 40 result.push_back(v[i]); 41 if (i + 1 < v[i].first && i + 1 < v.size() && 42 v[i + 1].first != v[i].first + 1) 43 result.push_back({v[i].first + 1, Unsupported}); 44 } 45 } 46 47 static LegalizerInfo::SizeAndActionsVec 48 widen_8_16(const LegalizerInfo::SizeAndActionsVec &v) { 49 assert(v.size() >= 1); 50 assert(v[0].first > 17); 51 LegalizerInfo::SizeAndActionsVec result = {{1, Unsupported}, 52 {8, WidenScalar}, 53 {9, Unsupported}, 54 {16, WidenScalar}, 55 {17, Unsupported}}; 56 addAndInterleaveWithUnsupported(result, v); 57 auto Largest = result.back().first; 58 result.push_back({Largest + 1, Unsupported}); 59 return result; 60 } 61 62 static bool AEABI(const ARMSubtarget &ST) { 63 return ST.isTargetAEABI() || ST.isTargetGNUAEABI() || ST.isTargetMuslAEABI(); 64 } 65 66 ARMLegalizerInfo::ARMLegalizerInfo(const ARMSubtarget &ST) { 67 using namespace TargetOpcode; 68 69 const LLT p0 = LLT::pointer(0, 32); 70 71 const LLT s1 = LLT::scalar(1); 72 const LLT s8 = LLT::scalar(8); 73 const LLT s16 = LLT::scalar(16); 74 const LLT s32 = LLT::scalar(32); 75 const LLT s64 = LLT::scalar(64); 76 77 if (ST.isThumb1Only()) { 78 // Thumb1 is not supported yet. 79 computeTables(); 80 verify(*ST.getInstrInfo()); 81 return; 82 } 83 84 getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT}) 85 .legalForCartesianProduct({s8, s16, s32}, {s1, s8, s16}); 86 87 getActionDefinitionsBuilder({G_MUL, G_AND, G_OR, G_XOR}) 88 .legalFor({s32}) 89 .minScalar(0, s32); 90 91 if (ST.hasNEON()) 92 getActionDefinitionsBuilder({G_ADD, G_SUB}) 93 .legalFor({s32, s64}) 94 .minScalar(0, s32); 95 else 96 getActionDefinitionsBuilder({G_ADD, G_SUB}) 97 .legalFor({s32}) 98 .minScalar(0, s32); 99 100 getActionDefinitionsBuilder({G_ASHR, G_LSHR, G_SHL}) 101 .legalFor({{s32, s32}}) 102 .minScalar(0, s32) 103 .clampScalar(1, s32, s32); 104 105 bool HasHWDivide = (!ST.isThumb() && ST.hasDivideInARMMode()) || 106 (ST.isThumb() && ST.hasDivideInThumbMode()); 107 if (HasHWDivide) 108 getActionDefinitionsBuilder({G_SDIV, G_UDIV}) 109 .legalFor({s32}) 110 .clampScalar(0, s32, s32); 111 else 112 getActionDefinitionsBuilder({G_SDIV, G_UDIV}) 113 .libcallFor({s32}) 114 .clampScalar(0, s32, s32); 115 116 for (unsigned Op : {G_SREM, G_UREM}) { 117 setLegalizeScalarToDifferentSizeStrategy(Op, 0, widen_8_16); 118 if (HasHWDivide) 119 setAction({Op, s32}, Lower); 120 else if (AEABI(ST)) 121 setAction({Op, s32}, Custom); 122 else 123 setAction({Op, s32}, Libcall); 124 } 125 126 getActionDefinitionsBuilder(G_INTTOPTR) 127 .legalFor({{p0, s32}}) 128 .minScalar(1, s32); 129 getActionDefinitionsBuilder(G_PTRTOINT) 130 .legalFor({{s32, p0}}) 131 .minScalar(0, s32); 132 133 getActionDefinitionsBuilder(G_CONSTANT) 134 .legalFor({s32, p0}) 135 .clampScalar(0, s32, s32); 136 137 getActionDefinitionsBuilder(G_ICMP) 138 .legalForCartesianProduct({s1}, {s32, p0}) 139 .minScalar(1, s32); 140 141 getActionDefinitionsBuilder(G_SELECT) 142 .legalForCartesianProduct({s32, p0}, {s1}) 143 .minScalar(0, s32); 144 145 // We're keeping these builders around because we'll want to add support for 146 // floating point to them. 147 auto &LoadStoreBuilder = getActionDefinitionsBuilder({G_LOAD, G_STORE}) 148 .legalForTypesWithMemDesc({{s1, p0, 8, 8}, 149 {s8, p0, 8, 8}, 150 {s16, p0, 16, 8}, 151 {s32, p0, 32, 8}, 152 {p0, p0, 32, 8}}) 153 .unsupportedIfMemSizeNotPow2(); 154 155 getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({p0}); 156 getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0}); 157 158 auto &PhiBuilder = 159 getActionDefinitionsBuilder(G_PHI) 160 .legalFor({s32, p0}) 161 .minScalar(0, s32); 162 163 getActionDefinitionsBuilder(G_GEP) 164 .legalFor({{p0, s32}}) 165 .minScalar(1, s32); 166 167 getActionDefinitionsBuilder(G_BRCOND).legalFor({s1}); 168 169 if (!ST.useSoftFloat() && ST.hasVFP2Base()) { 170 getActionDefinitionsBuilder( 171 {G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FCONSTANT, G_FNEG}) 172 .legalFor({s32, s64}); 173 174 LoadStoreBuilder 175 .legalForTypesWithMemDesc({{s64, p0, 64, 32}}) 176 .maxScalar(0, s32); 177 PhiBuilder.legalFor({s64}); 178 179 getActionDefinitionsBuilder(G_FCMP).legalForCartesianProduct({s1}, 180 {s32, s64}); 181 182 getActionDefinitionsBuilder(G_MERGE_VALUES).legalFor({{s64, s32}}); 183 getActionDefinitionsBuilder(G_UNMERGE_VALUES).legalFor({{s32, s64}}); 184 185 getActionDefinitionsBuilder(G_FPEXT).legalFor({{s64, s32}}); 186 getActionDefinitionsBuilder(G_FPTRUNC).legalFor({{s32, s64}}); 187 188 getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI}) 189 .legalForCartesianProduct({s32}, {s32, s64}); 190 getActionDefinitionsBuilder({G_SITOFP, G_UITOFP}) 191 .legalForCartesianProduct({s32, s64}, {s32}); 192 } else { 193 getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV}) 194 .libcallFor({s32, s64}); 195 196 LoadStoreBuilder.maxScalar(0, s32); 197 198 for (auto Ty : {s32, s64}) 199 setAction({G_FNEG, Ty}, Lower); 200 201 getActionDefinitionsBuilder(G_FCONSTANT).customFor({s32, s64}); 202 203 getActionDefinitionsBuilder(G_FCMP).customForCartesianProduct({s1}, 204 {s32, s64}); 205 206 if (AEABI(ST)) 207 setFCmpLibcallsAEABI(); 208 else 209 setFCmpLibcallsGNU(); 210 211 getActionDefinitionsBuilder(G_FPEXT).libcallFor({{s64, s32}}); 212 getActionDefinitionsBuilder(G_FPTRUNC).libcallFor({{s32, s64}}); 213 214 getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI}) 215 .libcallForCartesianProduct({s32}, {s32, s64}); 216 getActionDefinitionsBuilder({G_SITOFP, G_UITOFP}) 217 .libcallForCartesianProduct({s32, s64}, {s32}); 218 } 219 220 if (!ST.useSoftFloat() && ST.hasVFP4Base()) 221 getActionDefinitionsBuilder(G_FMA).legalFor({s32, s64}); 222 else 223 getActionDefinitionsBuilder(G_FMA).libcallFor({s32, s64}); 224 225 getActionDefinitionsBuilder({G_FREM, G_FPOW}).libcallFor({s32, s64}); 226 227 if (ST.hasV5TOps()) { 228 getActionDefinitionsBuilder(G_CTLZ) 229 .legalFor({s32, s32}) 230 .clampScalar(1, s32, s32) 231 .clampScalar(0, s32, s32); 232 getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF) 233 .lowerFor({s32, s32}) 234 .clampScalar(1, s32, s32) 235 .clampScalar(0, s32, s32); 236 } else { 237 getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF) 238 .libcallFor({s32, s32}) 239 .clampScalar(1, s32, s32) 240 .clampScalar(0, s32, s32); 241 getActionDefinitionsBuilder(G_CTLZ) 242 .lowerFor({s32, s32}) 243 .clampScalar(1, s32, s32) 244 .clampScalar(0, s32, s32); 245 } 246 247 computeTables(); 248 verify(*ST.getInstrInfo()); 249 } 250 251 void ARMLegalizerInfo::setFCmpLibcallsAEABI() { 252 // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be 253 // default-initialized. 254 FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 255 FCmp32Libcalls[CmpInst::FCMP_OEQ] = { 256 {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}}; 257 FCmp32Libcalls[CmpInst::FCMP_OGE] = { 258 {RTLIB::OGE_F32, CmpInst::BAD_ICMP_PREDICATE}}; 259 FCmp32Libcalls[CmpInst::FCMP_OGT] = { 260 {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 261 FCmp32Libcalls[CmpInst::FCMP_OLE] = { 262 {RTLIB::OLE_F32, CmpInst::BAD_ICMP_PREDICATE}}; 263 FCmp32Libcalls[CmpInst::FCMP_OLT] = { 264 {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 265 FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}}; 266 FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_EQ}}; 267 FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_EQ}}; 268 FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_EQ}}; 269 FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_EQ}}; 270 FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_EQ}}; 271 FCmp32Libcalls[CmpInst::FCMP_UNO] = { 272 {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}}; 273 FCmp32Libcalls[CmpInst::FCMP_ONE] = { 274 {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}, 275 {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 276 FCmp32Libcalls[CmpInst::FCMP_UEQ] = { 277 {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}, 278 {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}}; 279 280 FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 281 FCmp64Libcalls[CmpInst::FCMP_OEQ] = { 282 {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}}; 283 FCmp64Libcalls[CmpInst::FCMP_OGE] = { 284 {RTLIB::OGE_F64, CmpInst::BAD_ICMP_PREDICATE}}; 285 FCmp64Libcalls[CmpInst::FCMP_OGT] = { 286 {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 287 FCmp64Libcalls[CmpInst::FCMP_OLE] = { 288 {RTLIB::OLE_F64, CmpInst::BAD_ICMP_PREDICATE}}; 289 FCmp64Libcalls[CmpInst::FCMP_OLT] = { 290 {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 291 FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}}; 292 FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_EQ}}; 293 FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_EQ}}; 294 FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_EQ}}; 295 FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_EQ}}; 296 FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_EQ}}; 297 FCmp64Libcalls[CmpInst::FCMP_UNO] = { 298 {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}}; 299 FCmp64Libcalls[CmpInst::FCMP_ONE] = { 300 {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}, 301 {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 302 FCmp64Libcalls[CmpInst::FCMP_UEQ] = { 303 {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}, 304 {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}}; 305 } 306 307 void ARMLegalizerInfo::setFCmpLibcallsGNU() { 308 // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be 309 // default-initialized. 310 FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 311 FCmp32Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}}; 312 FCmp32Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F32, CmpInst::ICMP_SGE}}; 313 FCmp32Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}}; 314 FCmp32Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F32, CmpInst::ICMP_SLE}}; 315 FCmp32Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F32, CmpInst::ICMP_SLT}}; 316 FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}}; 317 FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_SGE}}; 318 FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_SGT}}; 319 FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SLE}}; 320 FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_SLT}}; 321 FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_NE}}; 322 FCmp32Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F32, CmpInst::ICMP_NE}}; 323 FCmp32Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}, 324 {RTLIB::OLT_F32, CmpInst::ICMP_SLT}}; 325 FCmp32Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}, 326 {RTLIB::UO_F32, CmpInst::ICMP_NE}}; 327 328 FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 329 FCmp64Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}}; 330 FCmp64Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F64, CmpInst::ICMP_SGE}}; 331 FCmp64Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}}; 332 FCmp64Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F64, CmpInst::ICMP_SLE}}; 333 FCmp64Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F64, CmpInst::ICMP_SLT}}; 334 FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}}; 335 FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_SGE}}; 336 FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_SGT}}; 337 FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SLE}}; 338 FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_SLT}}; 339 FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_NE}}; 340 FCmp64Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F64, CmpInst::ICMP_NE}}; 341 FCmp64Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}, 342 {RTLIB::OLT_F64, CmpInst::ICMP_SLT}}; 343 FCmp64Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}, 344 {RTLIB::UO_F64, CmpInst::ICMP_NE}}; 345 } 346 347 ARMLegalizerInfo::FCmpLibcallsList 348 ARMLegalizerInfo::getFCmpLibcalls(CmpInst::Predicate Predicate, 349 unsigned Size) const { 350 assert(CmpInst::isFPPredicate(Predicate) && "Unsupported FCmp predicate"); 351 if (Size == 32) 352 return FCmp32Libcalls[Predicate]; 353 if (Size == 64) 354 return FCmp64Libcalls[Predicate]; 355 llvm_unreachable("Unsupported size for FCmp predicate"); 356 } 357 358 bool ARMLegalizerInfo::legalizeCustom(MachineInstr &MI, 359 MachineRegisterInfo &MRI, 360 MachineIRBuilder &MIRBuilder, 361 GISelChangeObserver &Observer) const { 362 using namespace TargetOpcode; 363 364 MIRBuilder.setInstr(MI); 365 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 366 367 switch (MI.getOpcode()) { 368 default: 369 return false; 370 case G_SREM: 371 case G_UREM: { 372 unsigned OriginalResult = MI.getOperand(0).getReg(); 373 auto Size = MRI.getType(OriginalResult).getSizeInBits(); 374 if (Size != 32) 375 return false; 376 377 auto Libcall = 378 MI.getOpcode() == G_SREM ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; 379 380 // Our divmod libcalls return a struct containing the quotient and the 381 // remainder. We need to create a virtual register for it. 382 Type *ArgTy = Type::getInt32Ty(Ctx); 383 StructType *RetTy = StructType::get(Ctx, {ArgTy, ArgTy}, /* Packed */ true); 384 auto RetVal = MRI.createGenericVirtualRegister( 385 getLLTForType(*RetTy, MIRBuilder.getMF().getDataLayout())); 386 387 auto Status = createLibcall(MIRBuilder, Libcall, {RetVal, RetTy}, 388 {{MI.getOperand(1).getReg(), ArgTy}, 389 {MI.getOperand(2).getReg(), ArgTy}}); 390 if (Status != LegalizerHelper::Legalized) 391 return false; 392 393 // The remainder is the second result of divmod. Split the return value into 394 // a new, unused register for the quotient and the destination of the 395 // original instruction for the remainder. 396 MIRBuilder.buildUnmerge( 397 {MRI.createGenericVirtualRegister(LLT::scalar(32)), OriginalResult}, 398 RetVal); 399 break; 400 } 401 case G_FCMP: { 402 assert(MRI.getType(MI.getOperand(2).getReg()) == 403 MRI.getType(MI.getOperand(3).getReg()) && 404 "Mismatched operands for G_FCMP"); 405 auto OpSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits(); 406 407 auto OriginalResult = MI.getOperand(0).getReg(); 408 auto Predicate = 409 static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); 410 auto Libcalls = getFCmpLibcalls(Predicate, OpSize); 411 412 if (Libcalls.empty()) { 413 assert((Predicate == CmpInst::FCMP_TRUE || 414 Predicate == CmpInst::FCMP_FALSE) && 415 "Predicate needs libcalls, but none specified"); 416 MIRBuilder.buildConstant(OriginalResult, 417 Predicate == CmpInst::FCMP_TRUE ? 1 : 0); 418 MI.eraseFromParent(); 419 return true; 420 } 421 422 assert((OpSize == 32 || OpSize == 64) && "Unsupported operand size"); 423 auto *ArgTy = OpSize == 32 ? Type::getFloatTy(Ctx) : Type::getDoubleTy(Ctx); 424 auto *RetTy = Type::getInt32Ty(Ctx); 425 426 SmallVector<Register, 2> Results; 427 for (auto Libcall : Libcalls) { 428 auto LibcallResult = MRI.createGenericVirtualRegister(LLT::scalar(32)); 429 auto Status = 430 createLibcall(MIRBuilder, Libcall.LibcallID, {LibcallResult, RetTy}, 431 {{MI.getOperand(2).getReg(), ArgTy}, 432 {MI.getOperand(3).getReg(), ArgTy}}); 433 434 if (Status != LegalizerHelper::Legalized) 435 return false; 436 437 auto ProcessedResult = 438 Libcalls.size() == 1 439 ? OriginalResult 440 : MRI.createGenericVirtualRegister(MRI.getType(OriginalResult)); 441 442 // We have a result, but we need to transform it into a proper 1-bit 0 or 443 // 1, taking into account the different peculiarities of the values 444 // returned by the comparison functions. 445 CmpInst::Predicate ResultPred = Libcall.Predicate; 446 if (ResultPred == CmpInst::BAD_ICMP_PREDICATE) { 447 // We have a nice 0 or 1, and we just need to truncate it back to 1 bit 448 // to keep the types consistent. 449 MIRBuilder.buildTrunc(ProcessedResult, LibcallResult); 450 } else { 451 // We need to compare against 0. 452 assert(CmpInst::isIntPredicate(ResultPred) && "Unsupported predicate"); 453 auto Zero = MRI.createGenericVirtualRegister(LLT::scalar(32)); 454 MIRBuilder.buildConstant(Zero, 0); 455 MIRBuilder.buildICmp(ResultPred, ProcessedResult, LibcallResult, Zero); 456 } 457 Results.push_back(ProcessedResult); 458 } 459 460 if (Results.size() != 1) { 461 assert(Results.size() == 2 && "Unexpected number of results"); 462 MIRBuilder.buildOr(OriginalResult, Results[0], Results[1]); 463 } 464 break; 465 } 466 case G_FCONSTANT: { 467 // Convert to integer constants, while preserving the binary representation. 468 auto AsInteger = 469 MI.getOperand(1).getFPImm()->getValueAPF().bitcastToAPInt(); 470 MIRBuilder.buildConstant(MI.getOperand(0).getReg(), 471 *ConstantInt::get(Ctx, AsInteger)); 472 break; 473 } 474 } 475 476 MI.eraseFromParent(); 477 return true; 478 } 479