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 "PPCMachineScheduler.h" 17 #include "PPCSubtarget.h" 18 #include "PPCTargetObjectFile.h" 19 #include "PPCTargetTransformInfo.h" 20 #include "llvm/ADT/Optional.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/ADT/Triple.h" 24 #include "llvm/Analysis/TargetTransformInfo.h" 25 #include "llvm/CodeGen/Passes.h" 26 #include "llvm/CodeGen/TargetPassConfig.h" 27 #include "llvm/CodeGen/MachineScheduler.h" 28 #include "llvm/IR/Attributes.h" 29 #include "llvm/IR/DataLayout.h" 30 #include "llvm/IR/Function.h" 31 #include "llvm/Pass.h" 32 #include "llvm/Support/CodeGen.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/TargetRegistry.h" 35 #include "llvm/Target/TargetLoweringObjectFile.h" 36 #include "llvm/Target/TargetOptions.h" 37 #include "llvm/Transforms/Scalar.h" 38 #include <cassert> 39 #include <memory> 40 #include <string> 41 42 using namespace llvm; 43 44 45 static cl::opt<bool> 46 EnableBranchCoalescing("enable-ppc-branch-coalesce", cl::Hidden, 47 cl::desc("enable coalescing of duplicate branches for PPC")); 48 static cl:: 49 opt<bool> DisableCTRLoops("disable-ppc-ctrloops", cl::Hidden, 50 cl::desc("Disable CTR loops for PPC")); 51 52 static cl:: 53 opt<bool> DisablePreIncPrep("disable-ppc-preinc-prep", cl::Hidden, 54 cl::desc("Disable PPC loop preinc prep")); 55 56 static cl::opt<bool> 57 VSXFMAMutateEarly("schedule-ppc-vsx-fma-mutation-early", 58 cl::Hidden, cl::desc("Schedule VSX FMA instruction mutation early")); 59 60 static cl:: 61 opt<bool> DisableVSXSwapRemoval("disable-ppc-vsx-swap-removal", cl::Hidden, 62 cl::desc("Disable VSX Swap Removal for PPC")); 63 64 static cl:: 65 opt<bool> DisableQPXLoadSplat("disable-ppc-qpx-load-splat", cl::Hidden, 66 cl::desc("Disable QPX load splat simplification")); 67 68 static cl:: 69 opt<bool> DisableMIPeephole("disable-ppc-peephole", cl::Hidden, 70 cl::desc("Disable machine peepholes for PPC")); 71 72 static cl::opt<bool> 73 EnableGEPOpt("ppc-gep-opt", cl::Hidden, 74 cl::desc("Enable optimizations on complex GEPs"), 75 cl::init(true)); 76 77 static cl::opt<bool> 78 EnablePrefetch("enable-ppc-prefetching", 79 cl::desc("disable software prefetching on PPC"), 80 cl::init(false), cl::Hidden); 81 82 static cl::opt<bool> 83 EnableExtraTOCRegDeps("enable-ppc-extra-toc-reg-deps", 84 cl::desc("Add extra TOC register dependencies"), 85 cl::init(true), cl::Hidden); 86 87 static cl::opt<bool> 88 EnableMachineCombinerPass("ppc-machine-combiner", 89 cl::desc("Enable the machine combiner pass"), 90 cl::init(true), cl::Hidden); 91 92 static cl::opt<bool> 93 ReduceCRLogical("ppc-reduce-cr-logicals", 94 cl::desc("Expand eligible cr-logical binary ops to branches"), 95 cl::init(false), cl::Hidden); 96 extern "C" void LLVMInitializePowerPCTarget() { 97 // Register the targets 98 RegisterTargetMachine<PPCTargetMachine> A(getThePPC32Target()); 99 RegisterTargetMachine<PPCTargetMachine> B(getThePPC64Target()); 100 RegisterTargetMachine<PPCTargetMachine> C(getThePPC64LETarget()); 101 102 PassRegistry &PR = *PassRegistry::getPassRegistry(); 103 initializePPCBoolRetToIntPass(PR); 104 initializePPCExpandISELPass(PR); 105 initializePPCPreEmitPeepholePass(PR); 106 initializePPCTLSDynamicCallPass(PR); 107 initializePPCMIPeepholePass(PR); 108 } 109 110 /// Return the datalayout string of a subtarget. 111 static std::string getDataLayoutString(const Triple &T) { 112 bool is64Bit = T.getArch() == Triple::ppc64 || T.getArch() == Triple::ppc64le; 113 std::string Ret; 114 115 // Most PPC* platforms are big endian, PPC64LE is little endian. 116 if (T.getArch() == Triple::ppc64le) 117 Ret = "e"; 118 else 119 Ret = "E"; 120 121 Ret += DataLayout::getManglingComponent(T); 122 123 // PPC32 has 32 bit pointers. The PS3 (OS Lv2) is a PPC64 machine with 32 bit 124 // pointers. 125 if (!is64Bit || T.getOS() == Triple::Lv2) 126 Ret += "-p:32:32"; 127 128 // Note, the alignment values for f64 and i64 on ppc64 in Darwin 129 // documentation are wrong; these are correct (i.e. "what gcc does"). 130 if (is64Bit || !T.isOSDarwin()) 131 Ret += "-i64:64"; 132 else 133 Ret += "-f64:32:64"; 134 135 // PPC64 has 32 and 64 bit registers, PPC32 has only 32 bit ones. 136 if (is64Bit) 137 Ret += "-n32:64"; 138 else 139 Ret += "-n32"; 140 141 return Ret; 142 } 143 144 static std::string computeFSAdditions(StringRef FS, CodeGenOpt::Level OL, 145 const Triple &TT) { 146 std::string FullFS = FS; 147 148 // Make sure 64-bit features are available when CPUname is generic 149 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) { 150 if (!FullFS.empty()) 151 FullFS = "+64bit," + FullFS; 152 else 153 FullFS = "+64bit"; 154 } 155 156 if (OL >= CodeGenOpt::Default) { 157 if (!FullFS.empty()) 158 FullFS = "+crbits," + FullFS; 159 else 160 FullFS = "+crbits"; 161 } 162 163 if (OL != CodeGenOpt::None) { 164 if (!FullFS.empty()) 165 FullFS = "+invariant-function-descriptors," + FullFS; 166 else 167 FullFS = "+invariant-function-descriptors"; 168 } 169 170 return FullFS; 171 } 172 173 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 174 // If it isn't a Mach-O file then it's going to be a linux ELF 175 // object file. 176 if (TT.isOSDarwin()) 177 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 178 179 return llvm::make_unique<PPC64LinuxTargetObjectFile>(); 180 } 181 182 static PPCTargetMachine::PPCABI computeTargetABI(const Triple &TT, 183 const TargetOptions &Options) { 184 if (TT.isOSDarwin()) 185 report_fatal_error("Darwin is no longer supported for PowerPC"); 186 187 if (Options.MCOptions.getABIName().startswith("elfv1")) 188 return PPCTargetMachine::PPC_ABI_ELFv1; 189 else if (Options.MCOptions.getABIName().startswith("elfv2")) 190 return PPCTargetMachine::PPC_ABI_ELFv2; 191 192 assert(Options.MCOptions.getABIName().empty() && 193 "Unknown target-abi option!"); 194 195 if (TT.isMacOSX()) 196 return PPCTargetMachine::PPC_ABI_UNKNOWN; 197 198 switch (TT.getArch()) { 199 case Triple::ppc64le: 200 return PPCTargetMachine::PPC_ABI_ELFv2; 201 case Triple::ppc64: 202 return PPCTargetMachine::PPC_ABI_ELFv1; 203 default: 204 return PPCTargetMachine::PPC_ABI_UNKNOWN; 205 } 206 } 207 208 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 209 Optional<Reloc::Model> RM) { 210 if (RM.hasValue()) 211 return *RM; 212 213 // Darwin defaults to dynamic-no-pic. 214 if (TT.isOSDarwin()) 215 return Reloc::DynamicNoPIC; 216 217 // Big Endian PPC is PIC by default. 218 if (TT.getArch() == Triple::ppc64) 219 return Reloc::PIC_; 220 221 // Rest are static by default. 222 return Reloc::Static; 223 } 224 225 static CodeModel::Model getEffectivePPCCodeModel(const Triple &TT, 226 Optional<CodeModel::Model> CM, 227 bool JIT) { 228 if (CM) { 229 if (*CM == CodeModel::Tiny) 230 report_fatal_error("Target does not support the tiny CodeModel"); 231 if (*CM == CodeModel::Kernel) 232 report_fatal_error("Target does not support the kernel CodeModel"); 233 return *CM; 234 } 235 if (!TT.isOSDarwin() && !JIT && 236 (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le)) 237 return CodeModel::Medium; 238 return CodeModel::Small; 239 } 240 241 242 static ScheduleDAGInstrs *createPPCMachineScheduler(MachineSchedContext *C) { 243 ScheduleDAGMILive *DAG = 244 new ScheduleDAGMILive(C, llvm::make_unique<PPCPreRASchedStrategy>(C)); 245 // add DAG Mutations here. 246 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI)); 247 return DAG; 248 } 249 250 static ScheduleDAGInstrs *createPPCPostMachineScheduler( 251 MachineSchedContext *C) { 252 ScheduleDAGMI *DAG = 253 new ScheduleDAGMI(C, llvm::make_unique<PPCPostRASchedStrategy>(C), true); 254 // add DAG Mutations here. 255 return DAG; 256 } 257 258 // The FeatureString here is a little subtle. We are modifying the feature 259 // string with what are (currently) non-function specific overrides as it goes 260 // into the LLVMTargetMachine constructor and then using the stored value in the 261 // Subtarget constructor below it. 262 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT, 263 StringRef CPU, StringRef FS, 264 const TargetOptions &Options, 265 Optional<Reloc::Model> RM, 266 Optional<CodeModel::Model> CM, 267 CodeGenOpt::Level OL, bool JIT) 268 : LLVMTargetMachine(T, getDataLayoutString(TT), TT, CPU, 269 computeFSAdditions(FS, OL, TT), Options, 270 getEffectiveRelocModel(TT, RM), 271 getEffectivePPCCodeModel(TT, CM, JIT), OL), 272 TLOF(createTLOF(getTargetTriple())), 273 TargetABI(computeTargetABI(TT, Options)) { 274 initAsmInfo(); 275 } 276 277 PPCTargetMachine::~PPCTargetMachine() = default; 278 279 const PPCSubtarget * 280 PPCTargetMachine::getSubtargetImpl(const Function &F) const { 281 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 282 Attribute FSAttr = F.getFnAttribute("target-features"); 283 284 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 285 ? CPUAttr.getValueAsString().str() 286 : TargetCPU; 287 std::string FS = !FSAttr.hasAttribute(Attribute::None) 288 ? FSAttr.getValueAsString().str() 289 : TargetFS; 290 291 // FIXME: This is related to the code below to reset the target options, 292 // we need to know whether or not the soft float flag is set on the 293 // function before we can generate a subtarget. We also need to use 294 // it as a key for the subtarget since that can be the only difference 295 // between two functions. 296 bool SoftFloat = 297 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 298 // If the soft float attribute is set on the function turn on the soft float 299 // subtarget feature. 300 if (SoftFloat) 301 FS += FS.empty() ? "-hard-float" : ",-hard-float"; 302 303 auto &I = SubtargetMap[CPU + FS]; 304 if (!I) { 305 // This needs to be done before we create a new subtarget since any 306 // creation will depend on the TM and the code generation flags on the 307 // function that reside in TargetOptions. 308 resetTargetOptions(F); 309 I = llvm::make_unique<PPCSubtarget>( 310 TargetTriple, CPU, 311 // FIXME: It would be good to have the subtarget additions here 312 // not necessary. Anything that turns them on/off (overrides) ends 313 // up being put at the end of the feature string, but the defaults 314 // shouldn't require adding them. Fixing this means pulling Feature64Bit 315 // out of most of the target cpus in the .td file and making it set only 316 // as part of initialization via the TargetTriple. 317 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this); 318 } 319 return I.get(); 320 } 321 322 //===----------------------------------------------------------------------===// 323 // Pass Pipeline Configuration 324 //===----------------------------------------------------------------------===// 325 326 namespace { 327 328 /// PPC Code Generator Pass Configuration Options. 329 class PPCPassConfig : public TargetPassConfig { 330 public: 331 PPCPassConfig(PPCTargetMachine &TM, PassManagerBase &PM) 332 : TargetPassConfig(TM, PM) { 333 // At any optimization level above -O0 we use the Machine Scheduler and not 334 // the default Post RA List Scheduler. 335 if (TM.getOptLevel() != CodeGenOpt::None) 336 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 337 } 338 339 PPCTargetMachine &getPPCTargetMachine() const { 340 return getTM<PPCTargetMachine>(); 341 } 342 343 void addIRPasses() override; 344 bool addPreISel() override; 345 bool addILPOpts() override; 346 bool addInstSelector() override; 347 void addMachineSSAOptimization() override; 348 void addPreRegAlloc() override; 349 void addPreSched2() override; 350 void addPreEmitPass() override; 351 ScheduleDAGInstrs * 352 createMachineScheduler(MachineSchedContext *C) const override { 353 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>(); 354 if (ST.usePPCPreRASchedStrategy()) 355 return createPPCMachineScheduler(C); 356 return nullptr; 357 } 358 ScheduleDAGInstrs * 359 createPostMachineScheduler(MachineSchedContext *C) const override { 360 const PPCSubtarget &ST = C->MF->getSubtarget<PPCSubtarget>(); 361 if (ST.usePPCPostRASchedStrategy()) 362 return createPPCPostMachineScheduler(C); 363 return nullptr; 364 } 365 }; 366 367 } // end anonymous namespace 368 369 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) { 370 return new PPCPassConfig(*this, PM); 371 } 372 373 void PPCPassConfig::addIRPasses() { 374 if (TM->getOptLevel() != CodeGenOpt::None) 375 addPass(createPPCBoolRetToIntPass()); 376 addPass(createAtomicExpandPass()); 377 378 // For the BG/Q (or if explicitly requested), add explicit data prefetch 379 // intrinsics. 380 bool UsePrefetching = TM->getTargetTriple().getVendor() == Triple::BGQ && 381 getOptLevel() != CodeGenOpt::None; 382 if (EnablePrefetch.getNumOccurrences() > 0) 383 UsePrefetching = EnablePrefetch; 384 if (UsePrefetching) 385 addPass(createLoopDataPrefetchPass()); 386 387 if (TM->getOptLevel() >= CodeGenOpt::Default && EnableGEPOpt) { 388 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 389 // and lower a GEP with multiple indices to either arithmetic operations or 390 // multiple GEPs with single index. 391 addPass(createSeparateConstOffsetFromGEPPass(true)); 392 // Call EarlyCSE pass to find and remove subexpressions in the lowered 393 // result. 394 addPass(createEarlyCSEPass()); 395 // Do loop invariant code motion in case part of the lowered result is 396 // invariant. 397 addPass(createLICMPass()); 398 } 399 400 TargetPassConfig::addIRPasses(); 401 } 402 403 bool PPCPassConfig::addPreISel() { 404 if (!DisablePreIncPrep && getOptLevel() != CodeGenOpt::None) 405 addPass(createPPCLoopPreIncPrepPass(getPPCTargetMachine())); 406 407 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 408 addPass(createPPCCTRLoops()); 409 410 return false; 411 } 412 413 bool PPCPassConfig::addILPOpts() { 414 addPass(&EarlyIfConverterID); 415 416 if (EnableMachineCombinerPass) 417 addPass(&MachineCombinerID); 418 419 return true; 420 } 421 422 bool PPCPassConfig::addInstSelector() { 423 // Install an instruction selector. 424 addPass(createPPCISelDag(getPPCTargetMachine(), getOptLevel())); 425 426 #ifndef NDEBUG 427 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 428 addPass(createPPCCTRLoopsVerify()); 429 #endif 430 431 addPass(createPPCVSXCopyPass()); 432 return false; 433 } 434 435 void PPCPassConfig::addMachineSSAOptimization() { 436 // PPCBranchCoalescingPass need to be done before machine sinking 437 // since it merges empty blocks. 438 if (EnableBranchCoalescing && getOptLevel() != CodeGenOpt::None) 439 addPass(createPPCBranchCoalescingPass()); 440 TargetPassConfig::addMachineSSAOptimization(); 441 // For little endian, remove where possible the vector swap instructions 442 // introduced at code generation to normalize vector element order. 443 if (TM->getTargetTriple().getArch() == Triple::ppc64le && 444 !DisableVSXSwapRemoval) 445 addPass(createPPCVSXSwapRemovalPass()); 446 // Reduce the number of cr-logical ops. 447 if (ReduceCRLogical && getOptLevel() != CodeGenOpt::None) 448 addPass(createPPCReduceCRLogicalsPass()); 449 // Target-specific peephole cleanups performed after instruction 450 // selection. 451 if (!DisableMIPeephole) { 452 addPass(createPPCMIPeepholePass()); 453 addPass(&DeadMachineInstructionElimID); 454 } 455 } 456 457 void PPCPassConfig::addPreRegAlloc() { 458 if (getOptLevel() != CodeGenOpt::None) { 459 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); 460 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID, 461 &PPCVSXFMAMutateID); 462 } 463 464 // FIXME: We probably don't need to run these for -fPIE. 465 if (getPPCTargetMachine().isPositionIndependent()) { 466 // FIXME: LiveVariables should not be necessary here! 467 // PPCTLSDynamicCallPass uses LiveIntervals which previously dependent on 468 // LiveVariables. This (unnecessary) dependency has been removed now, 469 // however a stage-2 clang build fails without LiveVariables computed here. 470 addPass(&LiveVariablesID, false); 471 addPass(createPPCTLSDynamicCallPass()); 472 } 473 if (EnableExtraTOCRegDeps) 474 addPass(createPPCTOCRegDepsPass()); 475 } 476 477 void PPCPassConfig::addPreSched2() { 478 if (getOptLevel() != CodeGenOpt::None) { 479 addPass(&IfConverterID); 480 481 // This optimization must happen after anything that might do store-to-load 482 // forwarding. Here we're after RA (and, thus, when spills are inserted) 483 // but before post-RA scheduling. 484 if (!DisableQPXLoadSplat) 485 addPass(createPPCQPXLoadSplatPass()); 486 } 487 } 488 489 void PPCPassConfig::addPreEmitPass() { 490 addPass(createPPCPreEmitPeepholePass()); 491 addPass(createPPCExpandISELPass()); 492 493 if (getOptLevel() != CodeGenOpt::None) 494 addPass(createPPCEarlyReturnPass(), false); 495 // Must run branch selection immediately preceding the asm printer. 496 addPass(createPPCBranchSelectionPass(), false); 497 } 498 499 TargetTransformInfo 500 PPCTargetMachine::getTargetTransformInfo(const Function &F) { 501 return TargetTransformInfo(PPCTTIImpl(this, F)); 502 } 503 504 static MachineSchedRegistry 505 PPCPreRASchedRegistry("ppc-prera", 506 "Run PowerPC PreRA specific scheduler", 507 createPPCMachineScheduler); 508 509 static MachineSchedRegistry 510 PPCPostRASchedRegistry("ppc-postra", 511 "Run PowerPC PostRA specific scheduler", 512 createPPCPostMachineScheduler); 513