1 //===-- PPCTargetMachine.cpp - Define TargetMachine for PowerPC -----------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // Top-level implementation for the PowerPC target. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "PPCTargetMachine.h" 15 #include "MCTargetDesc/PPCMCTargetDesc.h" 16 #include "PPC.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 // Non-darwin 64-bit platforms are PIC by default. 218 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) 219 return Reloc::PIC_; 220 221 // 32-bit is static by default. 222 return Reloc::Static; 223 } 224 225 static CodeModel::Model getEffectiveCodeModel(const Triple &TT, 226 Optional<CodeModel::Model> CM, 227 bool JIT) { 228 if (CM) 229 return *CM; 230 if (!TT.isOSDarwin() && !JIT && 231 (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le)) 232 return CodeModel::Medium; 233 return CodeModel::Small; 234 } 235 236 // The FeatureString here is a little subtle. We are modifying the feature 237 // string with what are (currently) non-function specific overrides as it goes 238 // into the LLVMTargetMachine constructor and then using the stored value in the 239 // Subtarget constructor below it. 240 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT, 241 StringRef CPU, StringRef FS, 242 const TargetOptions &Options, 243 Optional<Reloc::Model> RM, 244 Optional<CodeModel::Model> CM, 245 CodeGenOpt::Level OL, bool JIT) 246 : LLVMTargetMachine(T, getDataLayoutString(TT), TT, CPU, 247 computeFSAdditions(FS, OL, TT), Options, 248 getEffectiveRelocModel(TT, RM), 249 getEffectiveCodeModel(TT, CM, JIT), OL), 250 TLOF(createTLOF(getTargetTriple())), 251 TargetABI(computeTargetABI(TT, Options)) { 252 initAsmInfo(); 253 } 254 255 PPCTargetMachine::~PPCTargetMachine() = default; 256 257 const PPCSubtarget * 258 PPCTargetMachine::getSubtargetImpl(const Function &F) const { 259 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 260 Attribute FSAttr = F.getFnAttribute("target-features"); 261 262 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 263 ? CPUAttr.getValueAsString().str() 264 : TargetCPU; 265 std::string FS = !FSAttr.hasAttribute(Attribute::None) 266 ? FSAttr.getValueAsString().str() 267 : TargetFS; 268 269 // FIXME: This is related to the code below to reset the target options, 270 // we need to know whether or not the soft float flag is set on the 271 // function before we can generate a subtarget. We also need to use 272 // it as a key for the subtarget since that can be the only difference 273 // between two functions. 274 bool SoftFloat = 275 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 276 // If the soft float attribute is set on the function turn on the soft float 277 // subtarget feature. 278 if (SoftFloat) 279 FS += FS.empty() ? "-hard-float" : ",-hard-float"; 280 281 auto &I = SubtargetMap[CPU + FS]; 282 if (!I) { 283 // This needs to be done before we create a new subtarget since any 284 // creation will depend on the TM and the code generation flags on the 285 // function that reside in TargetOptions. 286 resetTargetOptions(F); 287 I = llvm::make_unique<PPCSubtarget>( 288 TargetTriple, CPU, 289 // FIXME: It would be good to have the subtarget additions here 290 // not necessary. Anything that turns them on/off (overrides) ends 291 // up being put at the end of the feature string, but the defaults 292 // shouldn't require adding them. Fixing this means pulling Feature64Bit 293 // out of most of the target cpus in the .td file and making it set only 294 // as part of initialization via the TargetTriple. 295 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this); 296 } 297 return I.get(); 298 } 299 300 //===----------------------------------------------------------------------===// 301 // Pass Pipeline Configuration 302 //===----------------------------------------------------------------------===// 303 304 namespace { 305 306 /// PPC Code Generator Pass Configuration Options. 307 class PPCPassConfig : public TargetPassConfig { 308 public: 309 PPCPassConfig(PPCTargetMachine &TM, PassManagerBase &PM) 310 : TargetPassConfig(TM, PM) { 311 // At any optimization level above -O0 we use the Machine Scheduler and not 312 // the default Post RA List Scheduler. 313 if (TM.getOptLevel() != CodeGenOpt::None) 314 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 315 } 316 317 PPCTargetMachine &getPPCTargetMachine() const { 318 return getTM<PPCTargetMachine>(); 319 } 320 321 void addIRPasses() override; 322 bool addPreISel() override; 323 bool addILPOpts() override; 324 bool addInstSelector() override; 325 void addMachineSSAOptimization() override; 326 void addPreRegAlloc() override; 327 void addPreSched2() override; 328 void addPreEmitPass() override; 329 }; 330 331 } // end anonymous namespace 332 333 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) { 334 return new PPCPassConfig(*this, PM); 335 } 336 337 void PPCPassConfig::addIRPasses() { 338 if (TM->getOptLevel() != CodeGenOpt::None) 339 addPass(createPPCBoolRetToIntPass()); 340 addPass(createAtomicExpandPass()); 341 342 // For the BG/Q (or if explicitly requested), add explicit data prefetch 343 // intrinsics. 344 bool UsePrefetching = TM->getTargetTriple().getVendor() == Triple::BGQ && 345 getOptLevel() != CodeGenOpt::None; 346 if (EnablePrefetch.getNumOccurrences() > 0) 347 UsePrefetching = EnablePrefetch; 348 if (UsePrefetching) 349 addPass(createLoopDataPrefetchPass()); 350 351 if (TM->getOptLevel() >= CodeGenOpt::Default && EnableGEPOpt) { 352 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 353 // and lower a GEP with multiple indices to either arithmetic operations or 354 // multiple GEPs with single index. 355 addPass(createSeparateConstOffsetFromGEPPass(true)); 356 // Call EarlyCSE pass to find and remove subexpressions in the lowered 357 // result. 358 addPass(createEarlyCSEPass()); 359 // Do loop invariant code motion in case part of the lowered result is 360 // invariant. 361 addPass(createLICMPass()); 362 } 363 364 TargetPassConfig::addIRPasses(); 365 } 366 367 bool PPCPassConfig::addPreISel() { 368 if (!DisablePreIncPrep && getOptLevel() != CodeGenOpt::None) 369 addPass(createPPCLoopPreIncPrepPass(getPPCTargetMachine())); 370 371 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 372 addPass(createPPCCTRLoops()); 373 374 return false; 375 } 376 377 bool PPCPassConfig::addILPOpts() { 378 addPass(&EarlyIfConverterID); 379 380 if (EnableMachineCombinerPass) 381 addPass(&MachineCombinerID); 382 383 return true; 384 } 385 386 bool PPCPassConfig::addInstSelector() { 387 // Install an instruction selector. 388 addPass(createPPCISelDag(getPPCTargetMachine(), getOptLevel())); 389 390 #ifndef NDEBUG 391 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 392 addPass(createPPCCTRLoopsVerify()); 393 #endif 394 395 addPass(createPPCVSXCopyPass()); 396 return false; 397 } 398 399 void PPCPassConfig::addMachineSSAOptimization() { 400 // PPCBranchCoalescingPass need to be done before machine sinking 401 // since it merges empty blocks. 402 if (EnableBranchCoalescing && getOptLevel() != CodeGenOpt::None) 403 addPass(createPPCBranchCoalescingPass()); 404 TargetPassConfig::addMachineSSAOptimization(); 405 // For little endian, remove where possible the vector swap instructions 406 // introduced at code generation to normalize vector element order. 407 if (TM->getTargetTriple().getArch() == Triple::ppc64le && 408 !DisableVSXSwapRemoval) 409 addPass(createPPCVSXSwapRemovalPass()); 410 // Reduce the number of cr-logical ops. 411 if (ReduceCRLogical && getOptLevel() != CodeGenOpt::None) 412 addPass(createPPCReduceCRLogicalsPass()); 413 // Target-specific peephole cleanups performed after instruction 414 // selection. 415 if (!DisableMIPeephole) { 416 addPass(createPPCMIPeepholePass()); 417 addPass(&DeadMachineInstructionElimID); 418 } 419 } 420 421 void PPCPassConfig::addPreRegAlloc() { 422 if (getOptLevel() != CodeGenOpt::None) { 423 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); 424 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID, 425 &PPCVSXFMAMutateID); 426 } 427 428 // FIXME: We probably don't need to run these for -fPIE. 429 if (getPPCTargetMachine().isPositionIndependent()) { 430 // FIXME: LiveVariables should not be necessary here! 431 // PPCTLSDynamicCallPass uses LiveIntervals which previously dependent on 432 // LiveVariables. This (unnecessary) dependency has been removed now, 433 // however a stage-2 clang build fails without LiveVariables computed here. 434 addPass(&LiveVariablesID, false); 435 addPass(createPPCTLSDynamicCallPass()); 436 } 437 if (EnableExtraTOCRegDeps) 438 addPass(createPPCTOCRegDepsPass()); 439 } 440 441 void PPCPassConfig::addPreSched2() { 442 if (getOptLevel() != CodeGenOpt::None) { 443 addPass(&IfConverterID); 444 445 // This optimization must happen after anything that might do store-to-load 446 // forwarding. Here we're after RA (and, thus, when spills are inserted) 447 // but before post-RA scheduling. 448 if (!DisableQPXLoadSplat) 449 addPass(createPPCQPXLoadSplatPass()); 450 } 451 } 452 453 void PPCPassConfig::addPreEmitPass() { 454 addPass(createPPCPreEmitPeepholePass()); 455 addPass(createPPCExpandISELPass()); 456 457 if (getOptLevel() != CodeGenOpt::None) 458 addPass(createPPCEarlyReturnPass(), false); 459 // Must run branch selection immediately preceding the asm printer. 460 addPass(createPPCBranchSelectionPass(), false); 461 } 462 463 TargetTransformInfo 464 PPCTargetMachine::getTargetTransformInfo(const Function &F) { 465 return TargetTransformInfo(PPCTTIImpl(this, F)); 466 } 467