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_SEXT_INREG).lower(); 88 89 getActionDefinitionsBuilder({G_MUL, G_AND, G_OR, G_XOR}) 90 .legalFor({s32}) 91 .minScalar(0, s32); 92 93 if (ST.hasNEON()) 94 getActionDefinitionsBuilder({G_ADD, G_SUB}) 95 .legalFor({s32, s64}) 96 .minScalar(0, s32); 97 else 98 getActionDefinitionsBuilder({G_ADD, G_SUB}) 99 .legalFor({s32}) 100 .minScalar(0, s32); 101 102 getActionDefinitionsBuilder({G_ASHR, G_LSHR, G_SHL}) 103 .legalFor({{s32, s32}}) 104 .minScalar(0, s32) 105 .clampScalar(1, s32, s32); 106 107 bool HasHWDivide = (!ST.isThumb() && ST.hasDivideInARMMode()) || 108 (ST.isThumb() && ST.hasDivideInThumbMode()); 109 if (HasHWDivide) 110 getActionDefinitionsBuilder({G_SDIV, G_UDIV}) 111 .legalFor({s32}) 112 .clampScalar(0, s32, s32); 113 else 114 getActionDefinitionsBuilder({G_SDIV, G_UDIV}) 115 .libcallFor({s32}) 116 .clampScalar(0, s32, s32); 117 118 for (unsigned Op : {G_SREM, G_UREM}) { 119 setLegalizeScalarToDifferentSizeStrategy(Op, 0, widen_8_16); 120 if (HasHWDivide) 121 setAction({Op, s32}, Lower); 122 else if (AEABI(ST)) 123 setAction({Op, s32}, Custom); 124 else 125 setAction({Op, s32}, Libcall); 126 } 127 128 getActionDefinitionsBuilder(G_INTTOPTR) 129 .legalFor({{p0, s32}}) 130 .minScalar(1, s32); 131 getActionDefinitionsBuilder(G_PTRTOINT) 132 .legalFor({{s32, p0}}) 133 .minScalar(0, s32); 134 135 getActionDefinitionsBuilder(G_CONSTANT) 136 .legalFor({s32, p0}) 137 .clampScalar(0, s32, s32); 138 139 getActionDefinitionsBuilder(G_ICMP) 140 .legalForCartesianProduct({s1}, {s32, p0}) 141 .minScalar(1, s32); 142 143 getActionDefinitionsBuilder(G_SELECT) 144 .legalForCartesianProduct({s32, p0}, {s1}) 145 .minScalar(0, s32); 146 147 // We're keeping these builders around because we'll want to add support for 148 // floating point to them. 149 auto &LoadStoreBuilder = getActionDefinitionsBuilder({G_LOAD, G_STORE}) 150 .legalForTypesWithMemDesc({{s1, p0, 8, 8}, 151 {s8, p0, 8, 8}, 152 {s16, p0, 16, 8}, 153 {s32, p0, 32, 8}, 154 {p0, p0, 32, 8}}) 155 .unsupportedIfMemSizeNotPow2(); 156 157 getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({p0}); 158 getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0}); 159 160 auto &PhiBuilder = 161 getActionDefinitionsBuilder(G_PHI) 162 .legalFor({s32, p0}) 163 .minScalar(0, s32); 164 165 getActionDefinitionsBuilder(G_GEP) 166 .legalFor({{p0, s32}}) 167 .minScalar(1, s32); 168 169 getActionDefinitionsBuilder(G_BRCOND).legalFor({s1}); 170 171 if (!ST.useSoftFloat() && ST.hasVFP2Base()) { 172 getActionDefinitionsBuilder( 173 {G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FCONSTANT, G_FNEG}) 174 .legalFor({s32, s64}); 175 176 LoadStoreBuilder 177 .legalForTypesWithMemDesc({{s64, p0, 64, 32}}) 178 .maxScalar(0, s32); 179 PhiBuilder.legalFor({s64}); 180 181 getActionDefinitionsBuilder(G_FCMP).legalForCartesianProduct({s1}, 182 {s32, s64}); 183 184 getActionDefinitionsBuilder(G_MERGE_VALUES).legalFor({{s64, s32}}); 185 getActionDefinitionsBuilder(G_UNMERGE_VALUES).legalFor({{s32, s64}}); 186 187 getActionDefinitionsBuilder(G_FPEXT).legalFor({{s64, s32}}); 188 getActionDefinitionsBuilder(G_FPTRUNC).legalFor({{s32, s64}}); 189 190 getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI}) 191 .legalForCartesianProduct({s32}, {s32, s64}); 192 getActionDefinitionsBuilder({G_SITOFP, G_UITOFP}) 193 .legalForCartesianProduct({s32, s64}, {s32}); 194 } else { 195 getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV}) 196 .libcallFor({s32, s64}); 197 198 LoadStoreBuilder.maxScalar(0, s32); 199 200 for (auto Ty : {s32, s64}) 201 setAction({G_FNEG, Ty}, Lower); 202 203 getActionDefinitionsBuilder(G_FCONSTANT).customFor({s32, s64}); 204 205 getActionDefinitionsBuilder(G_FCMP).customForCartesianProduct({s1}, 206 {s32, s64}); 207 208 if (AEABI(ST)) 209 setFCmpLibcallsAEABI(); 210 else 211 setFCmpLibcallsGNU(); 212 213 getActionDefinitionsBuilder(G_FPEXT).libcallFor({{s64, s32}}); 214 getActionDefinitionsBuilder(G_FPTRUNC).libcallFor({{s32, s64}}); 215 216 getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI}) 217 .libcallForCartesianProduct({s32}, {s32, s64}); 218 getActionDefinitionsBuilder({G_SITOFP, G_UITOFP}) 219 .libcallForCartesianProduct({s32, s64}, {s32}); 220 } 221 222 if (!ST.useSoftFloat() && ST.hasVFP4Base()) 223 getActionDefinitionsBuilder(G_FMA).legalFor({s32, s64}); 224 else 225 getActionDefinitionsBuilder(G_FMA).libcallFor({s32, s64}); 226 227 getActionDefinitionsBuilder({G_FREM, G_FPOW}).libcallFor({s32, s64}); 228 229 if (ST.hasV5TOps()) { 230 getActionDefinitionsBuilder(G_CTLZ) 231 .legalFor({s32, s32}) 232 .clampScalar(1, s32, s32) 233 .clampScalar(0, s32, s32); 234 getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF) 235 .lowerFor({s32, s32}) 236 .clampScalar(1, s32, s32) 237 .clampScalar(0, s32, s32); 238 } else { 239 getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF) 240 .libcallFor({s32, s32}) 241 .clampScalar(1, s32, s32) 242 .clampScalar(0, s32, s32); 243 getActionDefinitionsBuilder(G_CTLZ) 244 .lowerFor({s32, s32}) 245 .clampScalar(1, s32, s32) 246 .clampScalar(0, s32, s32); 247 } 248 249 computeTables(); 250 verify(*ST.getInstrInfo()); 251 } 252 253 void ARMLegalizerInfo::setFCmpLibcallsAEABI() { 254 // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be 255 // default-initialized. 256 FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 257 FCmp32Libcalls[CmpInst::FCMP_OEQ] = { 258 {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}}; 259 FCmp32Libcalls[CmpInst::FCMP_OGE] = { 260 {RTLIB::OGE_F32, CmpInst::BAD_ICMP_PREDICATE}}; 261 FCmp32Libcalls[CmpInst::FCMP_OGT] = { 262 {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 263 FCmp32Libcalls[CmpInst::FCMP_OLE] = { 264 {RTLIB::OLE_F32, CmpInst::BAD_ICMP_PREDICATE}}; 265 FCmp32Libcalls[CmpInst::FCMP_OLT] = { 266 {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 267 FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}}; 268 FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_EQ}}; 269 FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_EQ}}; 270 FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_EQ}}; 271 FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_EQ}}; 272 FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_EQ}}; 273 FCmp32Libcalls[CmpInst::FCMP_UNO] = { 274 {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}}; 275 FCmp32Libcalls[CmpInst::FCMP_ONE] = { 276 {RTLIB::OGT_F32, CmpInst::BAD_ICMP_PREDICATE}, 277 {RTLIB::OLT_F32, CmpInst::BAD_ICMP_PREDICATE}}; 278 FCmp32Libcalls[CmpInst::FCMP_UEQ] = { 279 {RTLIB::OEQ_F32, CmpInst::BAD_ICMP_PREDICATE}, 280 {RTLIB::UO_F32, CmpInst::BAD_ICMP_PREDICATE}}; 281 282 FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 283 FCmp64Libcalls[CmpInst::FCMP_OEQ] = { 284 {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}}; 285 FCmp64Libcalls[CmpInst::FCMP_OGE] = { 286 {RTLIB::OGE_F64, CmpInst::BAD_ICMP_PREDICATE}}; 287 FCmp64Libcalls[CmpInst::FCMP_OGT] = { 288 {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 289 FCmp64Libcalls[CmpInst::FCMP_OLE] = { 290 {RTLIB::OLE_F64, CmpInst::BAD_ICMP_PREDICATE}}; 291 FCmp64Libcalls[CmpInst::FCMP_OLT] = { 292 {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 293 FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}}; 294 FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_EQ}}; 295 FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_EQ}}; 296 FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_EQ}}; 297 FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_EQ}}; 298 FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_EQ}}; 299 FCmp64Libcalls[CmpInst::FCMP_UNO] = { 300 {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}}; 301 FCmp64Libcalls[CmpInst::FCMP_ONE] = { 302 {RTLIB::OGT_F64, CmpInst::BAD_ICMP_PREDICATE}, 303 {RTLIB::OLT_F64, CmpInst::BAD_ICMP_PREDICATE}}; 304 FCmp64Libcalls[CmpInst::FCMP_UEQ] = { 305 {RTLIB::OEQ_F64, CmpInst::BAD_ICMP_PREDICATE}, 306 {RTLIB::UO_F64, CmpInst::BAD_ICMP_PREDICATE}}; 307 } 308 309 void ARMLegalizerInfo::setFCmpLibcallsGNU() { 310 // FCMP_TRUE and FCMP_FALSE don't need libcalls, they should be 311 // default-initialized. 312 FCmp32Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 313 FCmp32Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}}; 314 FCmp32Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F32, CmpInst::ICMP_SGE}}; 315 FCmp32Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}}; 316 FCmp32Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F32, CmpInst::ICMP_SLE}}; 317 FCmp32Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F32, CmpInst::ICMP_SLT}}; 318 FCmp32Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F32, CmpInst::ICMP_EQ}}; 319 FCmp32Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F32, CmpInst::ICMP_SGE}}; 320 FCmp32Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F32, CmpInst::ICMP_SGT}}; 321 FCmp32Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SLE}}; 322 FCmp32Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F32, CmpInst::ICMP_SLT}}; 323 FCmp32Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F32, CmpInst::ICMP_NE}}; 324 FCmp32Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F32, CmpInst::ICMP_NE}}; 325 FCmp32Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F32, CmpInst::ICMP_SGT}, 326 {RTLIB::OLT_F32, CmpInst::ICMP_SLT}}; 327 FCmp32Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F32, CmpInst::ICMP_EQ}, 328 {RTLIB::UO_F32, CmpInst::ICMP_NE}}; 329 330 FCmp64Libcalls.resize(CmpInst::LAST_FCMP_PREDICATE + 1); 331 FCmp64Libcalls[CmpInst::FCMP_OEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}}; 332 FCmp64Libcalls[CmpInst::FCMP_OGE] = {{RTLIB::OGE_F64, CmpInst::ICMP_SGE}}; 333 FCmp64Libcalls[CmpInst::FCMP_OGT] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}}; 334 FCmp64Libcalls[CmpInst::FCMP_OLE] = {{RTLIB::OLE_F64, CmpInst::ICMP_SLE}}; 335 FCmp64Libcalls[CmpInst::FCMP_OLT] = {{RTLIB::OLT_F64, CmpInst::ICMP_SLT}}; 336 FCmp64Libcalls[CmpInst::FCMP_ORD] = {{RTLIB::O_F64, CmpInst::ICMP_EQ}}; 337 FCmp64Libcalls[CmpInst::FCMP_UGE] = {{RTLIB::OLT_F64, CmpInst::ICMP_SGE}}; 338 FCmp64Libcalls[CmpInst::FCMP_UGT] = {{RTLIB::OLE_F64, CmpInst::ICMP_SGT}}; 339 FCmp64Libcalls[CmpInst::FCMP_ULE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SLE}}; 340 FCmp64Libcalls[CmpInst::FCMP_ULT] = {{RTLIB::OGE_F64, CmpInst::ICMP_SLT}}; 341 FCmp64Libcalls[CmpInst::FCMP_UNE] = {{RTLIB::UNE_F64, CmpInst::ICMP_NE}}; 342 FCmp64Libcalls[CmpInst::FCMP_UNO] = {{RTLIB::UO_F64, CmpInst::ICMP_NE}}; 343 FCmp64Libcalls[CmpInst::FCMP_ONE] = {{RTLIB::OGT_F64, CmpInst::ICMP_SGT}, 344 {RTLIB::OLT_F64, CmpInst::ICMP_SLT}}; 345 FCmp64Libcalls[CmpInst::FCMP_UEQ] = {{RTLIB::OEQ_F64, CmpInst::ICMP_EQ}, 346 {RTLIB::UO_F64, CmpInst::ICMP_NE}}; 347 } 348 349 ARMLegalizerInfo::FCmpLibcallsList 350 ARMLegalizerInfo::getFCmpLibcalls(CmpInst::Predicate Predicate, 351 unsigned Size) const { 352 assert(CmpInst::isFPPredicate(Predicate) && "Unsupported FCmp predicate"); 353 if (Size == 32) 354 return FCmp32Libcalls[Predicate]; 355 if (Size == 64) 356 return FCmp64Libcalls[Predicate]; 357 llvm_unreachable("Unsupported size for FCmp predicate"); 358 } 359 360 bool ARMLegalizerInfo::legalizeCustom(MachineInstr &MI, 361 MachineRegisterInfo &MRI, 362 MachineIRBuilder &MIRBuilder, 363 GISelChangeObserver &Observer) const { 364 using namespace TargetOpcode; 365 366 MIRBuilder.setInstr(MI); 367 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 368 369 switch (MI.getOpcode()) { 370 default: 371 return false; 372 case G_SREM: 373 case G_UREM: { 374 Register OriginalResult = MI.getOperand(0).getReg(); 375 auto Size = MRI.getType(OriginalResult).getSizeInBits(); 376 if (Size != 32) 377 return false; 378 379 auto Libcall = 380 MI.getOpcode() == G_SREM ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; 381 382 // Our divmod libcalls return a struct containing the quotient and the 383 // remainder. Create a new, unused register for the quotient and use the 384 // destination of the original instruction for the remainder. 385 Type *ArgTy = Type::getInt32Ty(Ctx); 386 StructType *RetTy = StructType::get(Ctx, {ArgTy, ArgTy}, /* Packed */ true); 387 Register RetRegs[] = {MRI.createGenericVirtualRegister(LLT::scalar(32)), 388 OriginalResult}; 389 auto Status = createLibcall(MIRBuilder, Libcall, {RetRegs, RetTy}, 390 {{MI.getOperand(1).getReg(), ArgTy}, 391 {MI.getOperand(2).getReg(), ArgTy}}); 392 if (Status != LegalizerHelper::Legalized) 393 return false; 394 break; 395 } 396 case G_FCMP: { 397 assert(MRI.getType(MI.getOperand(2).getReg()) == 398 MRI.getType(MI.getOperand(3).getReg()) && 399 "Mismatched operands for G_FCMP"); 400 auto OpSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits(); 401 402 auto OriginalResult = MI.getOperand(0).getReg(); 403 auto Predicate = 404 static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); 405 auto Libcalls = getFCmpLibcalls(Predicate, OpSize); 406 407 if (Libcalls.empty()) { 408 assert((Predicate == CmpInst::FCMP_TRUE || 409 Predicate == CmpInst::FCMP_FALSE) && 410 "Predicate needs libcalls, but none specified"); 411 MIRBuilder.buildConstant(OriginalResult, 412 Predicate == CmpInst::FCMP_TRUE ? 1 : 0); 413 MI.eraseFromParent(); 414 return true; 415 } 416 417 assert((OpSize == 32 || OpSize == 64) && "Unsupported operand size"); 418 auto *ArgTy = OpSize == 32 ? Type::getFloatTy(Ctx) : Type::getDoubleTy(Ctx); 419 auto *RetTy = Type::getInt32Ty(Ctx); 420 421 SmallVector<Register, 2> Results; 422 for (auto Libcall : Libcalls) { 423 auto LibcallResult = MRI.createGenericVirtualRegister(LLT::scalar(32)); 424 auto Status = 425 createLibcall(MIRBuilder, Libcall.LibcallID, {LibcallResult, RetTy}, 426 {{MI.getOperand(2).getReg(), ArgTy}, 427 {MI.getOperand(3).getReg(), ArgTy}}); 428 429 if (Status != LegalizerHelper::Legalized) 430 return false; 431 432 auto ProcessedResult = 433 Libcalls.size() == 1 434 ? OriginalResult 435 : MRI.createGenericVirtualRegister(MRI.getType(OriginalResult)); 436 437 // We have a result, but we need to transform it into a proper 1-bit 0 or 438 // 1, taking into account the different peculiarities of the values 439 // returned by the comparison functions. 440 CmpInst::Predicate ResultPred = Libcall.Predicate; 441 if (ResultPred == CmpInst::BAD_ICMP_PREDICATE) { 442 // We have a nice 0 or 1, and we just need to truncate it back to 1 bit 443 // to keep the types consistent. 444 MIRBuilder.buildTrunc(ProcessedResult, LibcallResult); 445 } else { 446 // We need to compare against 0. 447 assert(CmpInst::isIntPredicate(ResultPred) && "Unsupported predicate"); 448 auto Zero = MRI.createGenericVirtualRegister(LLT::scalar(32)); 449 MIRBuilder.buildConstant(Zero, 0); 450 MIRBuilder.buildICmp(ResultPred, ProcessedResult, LibcallResult, Zero); 451 } 452 Results.push_back(ProcessedResult); 453 } 454 455 if (Results.size() != 1) { 456 assert(Results.size() == 2 && "Unexpected number of results"); 457 MIRBuilder.buildOr(OriginalResult, Results[0], Results[1]); 458 } 459 break; 460 } 461 case G_FCONSTANT: { 462 // Convert to integer constants, while preserving the binary representation. 463 auto AsInteger = 464 MI.getOperand(1).getFPImm()->getValueAPF().bitcastToAPInt(); 465 MIRBuilder.buildConstant(MI.getOperand(0).getReg(), 466 *ConstantInt::get(Ctx, AsInteger)); 467 break; 468 } 469 } 470 471 MI.eraseFromParent(); 472 return true; 473 } 474