1 //===-- PPCTargetMachine.cpp - Define TargetMachine for PowerPC -----------===// 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 // Top-level implementation for the PowerPC target. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "PPCTargetMachine.h" 14 #include "MCTargetDesc/PPCMCTargetDesc.h" 15 #include "PPC.h" 16 #include "PPCMachineFunctionInfo.h" 17 #include "PPCMachineScheduler.h" 18 #include "PPCMacroFusion.h" 19 #include "PPCSubtarget.h" 20 #include "PPCTargetObjectFile.h" 21 #include "PPCTargetTransformInfo.h" 22 #include "TargetInfo/PowerPCTargetInfo.h" 23 #include "llvm/ADT/StringRef.h" 24 #include "llvm/Analysis/TargetTransformInfo.h" 25 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 26 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 27 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 28 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 29 #include "llvm/CodeGen/GlobalISel/Localizer.h" 30 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 31 #include "llvm/CodeGen/MachineScheduler.h" 32 #include "llvm/CodeGen/Passes.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/IR/Attributes.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/Function.h" 37 #include "llvm/InitializePasses.h" 38 #include "llvm/MC/TargetRegistry.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/CodeGen.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Target/TargetLoweringObjectFile.h" 43 #include "llvm/Target/TargetOptions.h" 44 #include "llvm/TargetParser/Triple.h" 45 #include "llvm/Transforms/Scalar.h" 46 #include <cassert> 47 #include <memory> 48 #include <optional> 49 #include <string> 50 51 using namespace llvm; 52 53 54 static cl::opt<bool> 55 EnableBranchCoalescing("enable-ppc-branch-coalesce", cl::Hidden, 56 cl::desc("enable coalescing of duplicate branches for PPC")); 57 static cl:: 58 opt<bool> DisableCTRLoops("disable-ppc-ctrloops", cl::Hidden, 59 cl::desc("Disable CTR loops for PPC")); 60 61 static cl:: 62 opt<bool> DisableInstrFormPrep("disable-ppc-instr-form-prep", cl::Hidden, 63 cl::desc("Disable PPC loop instr form prep")); 64 65 static cl::opt<bool> 66 VSXFMAMutateEarly("schedule-ppc-vsx-fma-mutation-early", 67 cl::Hidden, cl::desc("Schedule VSX FMA instruction mutation early")); 68 69 static cl:: 70 opt<bool> DisableVSXSwapRemoval("disable-ppc-vsx-swap-removal", cl::Hidden, 71 cl::desc("Disable VSX Swap Removal for PPC")); 72 73 static cl:: 74 opt<bool> DisableMIPeephole("disable-ppc-peephole", cl::Hidden, 75 cl::desc("Disable machine peepholes for PPC")); 76 77 static cl::opt<bool> 78 EnableGEPOpt("ppc-gep-opt", cl::Hidden, 79 cl::desc("Enable optimizations on complex GEPs"), 80 cl::init(true)); 81 82 static cl::opt<bool> 83 EnablePrefetch("enable-ppc-prefetching", 84 cl::desc("enable software prefetching on PPC"), 85 cl::init(false), cl::Hidden); 86 87 static cl::opt<bool> 88 EnableExtraTOCRegDeps("enable-ppc-extra-toc-reg-deps", 89 cl::desc("Add extra TOC register dependencies"), 90 cl::init(true), cl::Hidden); 91 92 static cl::opt<bool> 93 EnableMachineCombinerPass("ppc-machine-combiner", 94 cl::desc("Enable the machine combiner pass"), 95 cl::init(true), cl::Hidden); 96 97 static cl::opt<bool> 98 ReduceCRLogical("ppc-reduce-cr-logicals", 99 cl::desc("Expand eligible cr-logical binary ops to branches"), 100 cl::init(true), cl::Hidden); 101 102 static cl::opt<bool> EnablePPCGenScalarMASSEntries( 103 "enable-ppc-gen-scalar-mass", cl::init(false), 104 cl::desc("Enable lowering math functions to their corresponding MASS " 105 "(scalar) entries"), 106 cl::Hidden); 107 108 static cl::opt<bool> 109 EnableGlobalMerge("ppc-global-merge", cl::Hidden, cl::init(false), 110 cl::desc("Enable the global merge pass")); 111 112 static cl::opt<unsigned> 113 GlobalMergeMaxOffset("ppc-global-merge-max-offset", cl::Hidden, 114 cl::init(0x7fff), 115 cl::desc("Maximum global merge offset")); 116 117 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializePowerPCTarget() { 118 // Register the targets 119 RegisterTargetMachine<PPCTargetMachine> A(getThePPC32Target()); 120 RegisterTargetMachine<PPCTargetMachine> B(getThePPC32LETarget()); 121 RegisterTargetMachine<PPCTargetMachine> C(getThePPC64Target()); 122 RegisterTargetMachine<PPCTargetMachine> D(getThePPC64LETarget()); 123 124 PassRegistry &PR = *PassRegistry::getPassRegistry(); 125 #ifndef NDEBUG 126 initializePPCCTRLoopsVerifyPass(PR); 127 #endif 128 initializePPCLoopInstrFormPrepPass(PR); 129 initializePPCTOCRegDepsPass(PR); 130 initializePPCEarlyReturnPass(PR); 131 initializePPCVSXCopyPass(PR); 132 initializePPCVSXFMAMutatePass(PR); 133 initializePPCVSXSwapRemovalPass(PR); 134 initializePPCReduceCRLogicalsPass(PR); 135 initializePPCBSelPass(PR); 136 initializePPCBranchCoalescingPass(PR); 137 initializePPCBoolRetToIntPass(PR); 138 initializePPCPreEmitPeepholePass(PR); 139 initializePPCTLSDynamicCallPass(PR); 140 initializePPCMIPeepholePass(PR); 141 initializePPCLowerMASSVEntriesPass(PR); 142 initializePPCGenScalarMASSEntriesPass(PR); 143 initializePPCExpandAtomicPseudoPass(PR); 144 initializeGlobalISel(PR); 145 initializePPCCTRLoopsPass(PR); 146 initializePPCDAGToDAGISelLegacyPass(PR); 147 } 148 149 static bool isLittleEndianTriple(const Triple &T) { 150 return T.getArch() == Triple::ppc64le || T.getArch() == Triple::ppcle; 151 } 152 153 /// Return the datalayout string of a subtarget. 154 static std::string getDataLayoutString(const Triple &T) { 155 bool is64Bit = T.getArch() == Triple::ppc64 || T.getArch() == Triple::ppc64le; 156 std::string Ret; 157 158 // Most PPC* platforms are big endian, PPC(64)LE is little endian. 159 if (isLittleEndianTriple(T)) 160 Ret = "e"; 161 else 162 Ret = "E"; 163 164 Ret += DataLayout::getManglingComponent(T); 165 166 // PPC32 has 32 bit pointers. The PS3 (OS Lv2) is a PPC64 machine with 32 bit 167 // pointers. 168 if (!is64Bit || T.getOS() == Triple::Lv2) 169 Ret += "-p:32:32"; 170 171 // If the target ABI uses function descriptors, then the alignment of function 172 // pointers depends on the alignment used to emit the descriptor. Otherwise, 173 // function pointers are aligned to 32 bits because the instructions must be. 174 if ((T.getArch() == Triple::ppc64 && !T.isPPC64ELFv2ABI())) { 175 Ret += "-Fi64"; 176 } else if (T.isOSAIX()) { 177 Ret += is64Bit ? "-Fi64" : "-Fi32"; 178 } else { 179 Ret += "-Fn32"; 180 } 181 182 // Note, the alignment values for f64 and i64 on ppc64 in Darwin 183 // documentation are wrong; these are correct (i.e. "what gcc does"). 184 Ret += "-i64:64"; 185 186 // PPC64 has 32 and 64 bit registers, PPC32 has only 32 bit ones. 187 if (is64Bit) 188 Ret += "-i128:128-n32:64"; 189 else 190 Ret += "-n32"; 191 192 // Specify the vector alignment explicitly. For v256i1 and v512i1, the 193 // calculated alignment would be 256*alignment(i1) and 512*alignment(i1), 194 // which is 256 and 512 bytes - way over aligned. 195 if (is64Bit && (T.isOSAIX() || T.isOSLinux())) 196 Ret += "-S128-v256:256:256-v512:512:512"; 197 198 return Ret; 199 } 200 201 static std::string computeFSAdditions(StringRef FS, CodeGenOptLevel OL, 202 const Triple &TT) { 203 std::string FullFS = std::string(FS); 204 205 // Make sure 64-bit features are available when CPUname is generic 206 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) { 207 if (!FullFS.empty()) 208 FullFS = "+64bit," + FullFS; 209 else 210 FullFS = "+64bit"; 211 } 212 213 if (OL >= CodeGenOptLevel::Default) { 214 if (!FullFS.empty()) 215 FullFS = "+crbits," + FullFS; 216 else 217 FullFS = "+crbits"; 218 } 219 220 if (OL != CodeGenOptLevel::None) { 221 if (!FullFS.empty()) 222 FullFS = "+invariant-function-descriptors," + FullFS; 223 else 224 FullFS = "+invariant-function-descriptors"; 225 } 226 227 if (TT.isOSAIX()) { 228 if (!FullFS.empty()) 229 FullFS = "+aix," + FullFS; 230 else 231 FullFS = "+aix"; 232 } 233 234 return FullFS; 235 } 236 237 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 238 if (TT.isOSAIX()) 239 return std::make_unique<TargetLoweringObjectFileXCOFF>(); 240 241 return std::make_unique<PPC64LinuxTargetObjectFile>(); 242 } 243 244 static PPCTargetMachine::PPCABI computeTargetABI(const Triple &TT, 245 const TargetOptions &Options) { 246 if (Options.MCOptions.getABIName().starts_with("elfv1")) 247 return PPCTargetMachine::PPC_ABI_ELFv1; 248 else if (Options.MCOptions.getABIName().starts_with("elfv2")) 249 return PPCTargetMachine::PPC_ABI_ELFv2; 250 251 assert(Options.MCOptions.getABIName().empty() && 252 "Unknown target-abi option!"); 253 254 switch (TT.getArch()) { 255 case Triple::ppc64le: 256 return PPCTargetMachine::PPC_ABI_ELFv2; 257 case Triple::ppc64: 258 if (TT.isPPC64ELFv2ABI()) 259 return PPCTargetMachine::PPC_ABI_ELFv2; 260 else 261 return PPCTargetMachine::PPC_ABI_ELFv1; 262 default: 263 return PPCTargetMachine::PPC_ABI_UNKNOWN; 264 } 265 } 266 267 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 268 std::optional<Reloc::Model> RM) { 269 if (TT.isOSAIX() && RM && *RM != Reloc::PIC_) 270 report_fatal_error("invalid relocation model, AIX only supports PIC", 271 false); 272 273 if (RM) 274 return *RM; 275 276 // Big Endian PPC and AIX default to PIC. 277 if (TT.getArch() == Triple::ppc64 || TT.isOSAIX()) 278 return Reloc::PIC_; 279 280 // Rest are static by default. 281 return Reloc::Static; 282 } 283 284 static CodeModel::Model 285 getEffectivePPCCodeModel(const Triple &TT, std::optional<CodeModel::Model> CM, 286 bool JIT) { 287 if (CM) { 288 if (*CM == CodeModel::Tiny) 289 report_fatal_error("Target does not support the tiny CodeModel", false); 290 if (*CM == CodeModel::Kernel) 291 report_fatal_error("Target does not support the kernel CodeModel", false); 292 return *CM; 293 } 294 295 if (JIT) 296 return CodeModel::Small; 297 if (TT.isOSAIX()) 298 return CodeModel::Small; 299 300 assert(TT.isOSBinFormatELF() && "All remaining PPC OSes are ELF based."); 301 302 if (TT.isArch32Bit()) 303 return CodeModel::Small; 304 305 assert(TT.isArch64Bit() && "Unsupported PPC architecture."); 306 return CodeModel::Medium; 307 } 308 309 310 static ScheduleDAGInstrs *createPPCMachineScheduler(MachineSchedContext *C) { 311 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>(); 312 ScheduleDAGMILive *DAG = 313 new ScheduleDAGMILive(C, ST.usePPCPreRASchedStrategy() ? 314 std::make_unique<PPCPreRASchedStrategy>(C) : 315 std::make_unique<GenericScheduler>(C)); 316 // add DAG Mutations here. 317 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI)); 318 if (ST.hasStoreFusion()) 319 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 320 if (ST.hasFusion()) 321 DAG->addMutation(createPowerPCMacroFusionDAGMutation()); 322 323 return DAG; 324 } 325 326 static ScheduleDAGInstrs *createPPCPostMachineScheduler( 327 MachineSchedContext *C) { 328 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>(); 329 ScheduleDAGMI *DAG = 330 new ScheduleDAGMI(C, ST.usePPCPostRASchedStrategy() ? 331 std::make_unique<PPCPostRASchedStrategy>(C) : 332 std::make_unique<PostGenericScheduler>(C), true); 333 // add DAG Mutations here. 334 if (ST.hasStoreFusion()) 335 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 336 if (ST.hasFusion()) 337 DAG->addMutation(createPowerPCMacroFusionDAGMutation()); 338 return DAG; 339 } 340 341 // The FeatureString here is a little subtle. We are modifying the feature 342 // string with what are (currently) non-function specific overrides as it goes 343 // into the CodeGenTargetMachineImpl constructor and then using the stored value 344 // in the Subtarget constructor below it. 345 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT, 346 StringRef CPU, StringRef FS, 347 const TargetOptions &Options, 348 std::optional<Reloc::Model> RM, 349 std::optional<CodeModel::Model> CM, 350 CodeGenOptLevel OL, bool JIT) 351 : CodeGenTargetMachineImpl(T, getDataLayoutString(TT), TT, CPU, 352 computeFSAdditions(FS, OL, TT), Options, 353 getEffectiveRelocModel(TT, RM), 354 getEffectivePPCCodeModel(TT, CM, JIT), OL), 355 TLOF(createTLOF(getTargetTriple())), 356 TargetABI(computeTargetABI(TT, Options)), 357 Endianness(isLittleEndianTriple(TT) ? Endian::LITTLE : Endian::BIG) { 358 initAsmInfo(); 359 } 360 361 PPCTargetMachine::~PPCTargetMachine() = default; 362 363 const PPCSubtarget * 364 PPCTargetMachine::getSubtargetImpl(const Function &F) const { 365 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 366 Attribute TuneAttr = F.getFnAttribute("tune-cpu"); 367 Attribute FSAttr = F.getFnAttribute("target-features"); 368 369 std::string CPU = 370 CPUAttr.isValid() ? CPUAttr.getValueAsString().str() : TargetCPU; 371 std::string TuneCPU = 372 TuneAttr.isValid() ? TuneAttr.getValueAsString().str() : CPU; 373 std::string FS = 374 FSAttr.isValid() ? FSAttr.getValueAsString().str() : TargetFS; 375 376 // FIXME: This is related to the code below to reset the target options, 377 // we need to know whether or not the soft float flag is set on the 378 // function before we can generate a subtarget. We also need to use 379 // it as a key for the subtarget since that can be the only difference 380 // between two functions. 381 bool SoftFloat = F.getFnAttribute("use-soft-float").getValueAsBool(); 382 // If the soft float attribute is set on the function turn on the soft float 383 // subtarget feature. 384 if (SoftFloat) 385 FS += FS.empty() ? "-hard-float" : ",-hard-float"; 386 387 auto &I = SubtargetMap[CPU + TuneCPU + FS]; 388 if (!I) { 389 // This needs to be done before we create a new subtarget since any 390 // creation will depend on the TM and the code generation flags on the 391 // function that reside in TargetOptions. 392 resetTargetOptions(F); 393 I = std::make_unique<PPCSubtarget>( 394 TargetTriple, CPU, TuneCPU, 395 // FIXME: It would be good to have the subtarget additions here 396 // not necessary. Anything that turns them on/off (overrides) ends 397 // up being put at the end of the feature string, but the defaults 398 // shouldn't require adding them. Fixing this means pulling Feature64Bit 399 // out of most of the target cpus in the .td file and making it set only 400 // as part of initialization via the TargetTriple. 401 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this); 402 } 403 return I.get(); 404 } 405 406 //===----------------------------------------------------------------------===// 407 // Pass Pipeline Configuration 408 //===----------------------------------------------------------------------===// 409 410 namespace { 411 412 /// PPC Code Generator Pass Configuration Options. 413 class PPCPassConfig : public TargetPassConfig { 414 public: 415 PPCPassConfig(PPCTargetMachine &TM, PassManagerBase &PM) 416 : TargetPassConfig(TM, PM) { 417 // At any optimization level above -O0 we use the Machine Scheduler and not 418 // the default Post RA List Scheduler. 419 if (TM.getOptLevel() != CodeGenOptLevel::None) 420 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 421 } 422 423 PPCTargetMachine &getPPCTargetMachine() const { 424 return getTM<PPCTargetMachine>(); 425 } 426 427 void addIRPasses() override; 428 bool addPreISel() override; 429 bool addILPOpts() override; 430 bool addInstSelector() override; 431 void addMachineSSAOptimization() override; 432 void addPreRegAlloc() override; 433 void addPreSched2() override; 434 void addPreEmitPass() override; 435 void addPreEmitPass2() override; 436 // GlobalISEL 437 bool addIRTranslator() override; 438 bool addLegalizeMachineIR() override; 439 bool addRegBankSelect() override; 440 bool addGlobalInstructionSelect() override; 441 442 ScheduleDAGInstrs * 443 createMachineScheduler(MachineSchedContext *C) const override { 444 return createPPCMachineScheduler(C); 445 } 446 ScheduleDAGInstrs * 447 createPostMachineScheduler(MachineSchedContext *C) const override { 448 return createPPCPostMachineScheduler(C); 449 } 450 }; 451 452 } // end anonymous namespace 453 454 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) { 455 return new PPCPassConfig(*this, PM); 456 } 457 458 void PPCPassConfig::addIRPasses() { 459 if (TM->getOptLevel() != CodeGenOptLevel::None) 460 addPass(createPPCBoolRetToIntPass()); 461 addPass(createAtomicExpandLegacyPass()); 462 463 // Lower generic MASSV routines to PowerPC subtarget-specific entries. 464 addPass(createPPCLowerMASSVEntriesPass()); 465 466 // Generate PowerPC target-specific entries for scalar math functions 467 // that are available in IBM MASS (scalar) library. 468 if (TM->getOptLevel() == CodeGenOptLevel::Aggressive && 469 EnablePPCGenScalarMASSEntries) { 470 TM->Options.PPCGenScalarMASSEntries = EnablePPCGenScalarMASSEntries; 471 addPass(createPPCGenScalarMASSEntriesPass()); 472 } 473 474 // If explicitly requested, add explicit data prefetch intrinsics. 475 if (EnablePrefetch.getNumOccurrences() > 0) 476 addPass(createLoopDataPrefetchPass()); 477 478 if (TM->getOptLevel() >= CodeGenOptLevel::Default && EnableGEPOpt) { 479 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 480 // and lower a GEP with multiple indices to either arithmetic operations or 481 // multiple GEPs with single index. 482 addPass(createSeparateConstOffsetFromGEPPass(true)); 483 // Call EarlyCSE pass to find and remove subexpressions in the lowered 484 // result. 485 addPass(createEarlyCSEPass()); 486 // Do loop invariant code motion in case part of the lowered result is 487 // invariant. 488 addPass(createLICMPass()); 489 } 490 491 TargetPassConfig::addIRPasses(); 492 } 493 494 bool PPCPassConfig::addPreISel() { 495 // The GlobalMerge pass is intended to be on by default on AIX. 496 // Specifying the command line option overrides the AIX default. 497 if ((EnableGlobalMerge.getNumOccurrences() > 0) 498 ? EnableGlobalMerge 499 : getOptLevel() != CodeGenOptLevel::None) 500 addPass(createGlobalMergePass(TM, GlobalMergeMaxOffset, false, false, true, 501 true)); 502 503 if (!DisableInstrFormPrep && getOptLevel() != CodeGenOptLevel::None) 504 addPass(createPPCLoopInstrFormPrepPass(getPPCTargetMachine())); 505 506 if (!DisableCTRLoops && getOptLevel() != CodeGenOptLevel::None) 507 addPass(createHardwareLoopsLegacyPass()); 508 509 return false; 510 } 511 512 bool PPCPassConfig::addILPOpts() { 513 addPass(&EarlyIfConverterLegacyID); 514 515 if (EnableMachineCombinerPass) 516 addPass(&MachineCombinerID); 517 518 return true; 519 } 520 521 bool PPCPassConfig::addInstSelector() { 522 // Install an instruction selector. 523 addPass(createPPCISelDag(getPPCTargetMachine(), getOptLevel())); 524 525 #ifndef NDEBUG 526 if (!DisableCTRLoops && getOptLevel() != CodeGenOptLevel::None) 527 addPass(createPPCCTRLoopsVerify()); 528 #endif 529 530 addPass(createPPCVSXCopyPass()); 531 return false; 532 } 533 534 void PPCPassConfig::addMachineSSAOptimization() { 535 // Run CTR loops pass before any cfg modification pass to prevent the 536 // canonical form of hardware loop from being destroied. 537 if (!DisableCTRLoops && getOptLevel() != CodeGenOptLevel::None) 538 addPass(createPPCCTRLoopsPass()); 539 540 // PPCBranchCoalescingPass need to be done before machine sinking 541 // since it merges empty blocks. 542 if (EnableBranchCoalescing && getOptLevel() != CodeGenOptLevel::None) 543 addPass(createPPCBranchCoalescingPass()); 544 TargetPassConfig::addMachineSSAOptimization(); 545 // For little endian, remove where possible the vector swap instructions 546 // introduced at code generation to normalize vector element order. 547 if (TM->getTargetTriple().getArch() == Triple::ppc64le && 548 !DisableVSXSwapRemoval) 549 addPass(createPPCVSXSwapRemovalPass()); 550 // Reduce the number of cr-logical ops. 551 if (ReduceCRLogical && getOptLevel() != CodeGenOptLevel::None) 552 addPass(createPPCReduceCRLogicalsPass()); 553 // Target-specific peephole cleanups performed after instruction 554 // selection. 555 if (!DisableMIPeephole) { 556 addPass(createPPCMIPeepholePass()); 557 addPass(&DeadMachineInstructionElimID); 558 } 559 } 560 561 void PPCPassConfig::addPreRegAlloc() { 562 if (getOptLevel() != CodeGenOptLevel::None) { 563 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); 564 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID, 565 &PPCVSXFMAMutateID); 566 } 567 568 // FIXME: We probably don't need to run these for -fPIE. 569 if (getPPCTargetMachine().isPositionIndependent()) { 570 // FIXME: LiveVariables should not be necessary here! 571 // PPCTLSDynamicCallPass uses LiveIntervals which previously dependent on 572 // LiveVariables. This (unnecessary) dependency has been removed now, 573 // however a stage-2 clang build fails without LiveVariables computed here. 574 addPass(&LiveVariablesID); 575 addPass(createPPCTLSDynamicCallPass()); 576 } 577 if (EnableExtraTOCRegDeps) 578 addPass(createPPCTOCRegDepsPass()); 579 580 if (getOptLevel() != CodeGenOptLevel::None) 581 addPass(&MachinePipelinerID); 582 } 583 584 void PPCPassConfig::addPreSched2() { 585 if (getOptLevel() != CodeGenOptLevel::None) 586 addPass(&IfConverterID); 587 } 588 589 void PPCPassConfig::addPreEmitPass() { 590 addPass(createPPCPreEmitPeepholePass()); 591 592 if (getOptLevel() != CodeGenOptLevel::None) 593 addPass(createPPCEarlyReturnPass()); 594 } 595 596 void PPCPassConfig::addPreEmitPass2() { 597 // Schedule the expansion of AMOs at the last possible moment, avoiding the 598 // possibility for other passes to break the requirements for forward 599 // progress in the LL/SC block. 600 addPass(createPPCExpandAtomicPseudoPass()); 601 // Must run branch selection immediately preceding the asm printer. 602 addPass(createPPCBranchSelectionPass()); 603 } 604 605 TargetTransformInfo 606 PPCTargetMachine::getTargetTransformInfo(const Function &F) const { 607 return TargetTransformInfo(PPCTTIImpl(this, F)); 608 } 609 610 bool PPCTargetMachine::isLittleEndian() const { 611 assert(Endianness != Endian::NOT_DETECTED && 612 "Unable to determine endianness"); 613 return Endianness == Endian::LITTLE; 614 } 615 616 MachineFunctionInfo *PPCTargetMachine::createMachineFunctionInfo( 617 BumpPtrAllocator &Allocator, const Function &F, 618 const TargetSubtargetInfo *STI) const { 619 return PPCFunctionInfo::create<PPCFunctionInfo>(Allocator, F, STI); 620 } 621 622 static MachineSchedRegistry 623 PPCPreRASchedRegistry("ppc-prera", 624 "Run PowerPC PreRA specific scheduler", 625 createPPCMachineScheduler); 626 627 static MachineSchedRegistry 628 PPCPostRASchedRegistry("ppc-postra", 629 "Run PowerPC PostRA specific scheduler", 630 createPPCPostMachineScheduler); 631 632 // Global ISEL 633 bool PPCPassConfig::addIRTranslator() { 634 addPass(new IRTranslator()); 635 return false; 636 } 637 638 bool PPCPassConfig::addLegalizeMachineIR() { 639 addPass(new Legalizer()); 640 return false; 641 } 642 643 bool PPCPassConfig::addRegBankSelect() { 644 addPass(new RegBankSelect()); 645 return false; 646 } 647 648 bool PPCPassConfig::addGlobalInstructionSelect() { 649 addPass(new InstructionSelect(getOptLevel())); 650 return false; 651 } 652