1 //===- TargetPassConfig.cpp - Target independent code generation passes ---===// 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 // This file defines interfaces to access the target independent code 10 // generation passes provided by the LLVM backend. 11 // 12 //===---------------------------------------------------------------------===// 13 14 #include "llvm/CodeGen/TargetPassConfig.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/Analysis/BasicAliasAnalysis.h" 19 #include "llvm/Analysis/CallGraphSCCPass.h" 20 #include "llvm/Analysis/ScopedNoAliasAA.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 23 #include "llvm/CodeGen/BasicBlockSectionsProfileReader.h" 24 #include "llvm/CodeGen/CSEConfigBase.h" 25 #include "llvm/CodeGen/MachineFunctionPass.h" 26 #include "llvm/CodeGen/MachinePassRegistry.h" 27 #include "llvm/CodeGen/Passes.h" 28 #include "llvm/CodeGen/RegAllocRegistry.h" 29 #include "llvm/IR/IRPrintingPasses.h" 30 #include "llvm/IR/LegacyPassManager.h" 31 #include "llvm/IR/PassInstrumentation.h" 32 #include "llvm/IR/Verifier.h" 33 #include "llvm/InitializePasses.h" 34 #include "llvm/MC/MCAsmInfo.h" 35 #include "llvm/MC/MCTargetOptions.h" 36 #include "llvm/Pass.h" 37 #include "llvm/Support/CodeGen.h" 38 #include "llvm/Support/CommandLine.h" 39 #include "llvm/Support/Compiler.h" 40 #include "llvm/Support/Debug.h" 41 #include "llvm/Support/Discriminator.h" 42 #include "llvm/Support/ErrorHandling.h" 43 #include "llvm/Support/SaveAndRestore.h" 44 #include "llvm/Support/Threading.h" 45 #include "llvm/Support/VirtualFileSystem.h" 46 #include "llvm/Support/WithColor.h" 47 #include "llvm/Target/CGPassBuilderOption.h" 48 #include "llvm/Target/TargetMachine.h" 49 #include "llvm/Transforms/Scalar.h" 50 #include "llvm/Transforms/Utils.h" 51 #include <cassert> 52 #include <optional> 53 #include <string> 54 55 using namespace llvm; 56 57 static cl::opt<bool> 58 EnableIPRA("enable-ipra", cl::init(false), cl::Hidden, 59 cl::desc("Enable interprocedural register allocation " 60 "to reduce load/store at procedure calls.")); 61 static cl::opt<bool> DisablePostRASched("disable-post-ra", cl::Hidden, 62 cl::desc("Disable Post Regalloc Scheduler")); 63 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden, 64 cl::desc("Disable branch folding")); 65 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden, 66 cl::desc("Disable tail duplication")); 67 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden, 68 cl::desc("Disable pre-register allocation tail duplication")); 69 static cl::opt<bool> DisableBlockPlacement("disable-block-placement", 70 cl::Hidden, cl::desc("Disable probability-driven block placement")); 71 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats", 72 cl::Hidden, cl::desc("Collect probability-driven block placement stats")); 73 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden, 74 cl::desc("Disable Stack Slot Coloring")); 75 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden, 76 cl::desc("Disable Machine Dead Code Elimination")); 77 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden, 78 cl::desc("Disable Early If-conversion")); 79 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden, 80 cl::desc("Disable Machine LICM")); 81 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden, 82 cl::desc("Disable Machine Common Subexpression Elimination")); 83 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc( 84 "optimize-regalloc", cl::Hidden, 85 cl::desc("Enable optimized register allocation compilation path.")); 86 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm", 87 cl::Hidden, 88 cl::desc("Disable Machine LICM")); 89 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden, 90 cl::desc("Disable Machine Sinking")); 91 static cl::opt<bool> DisablePostRAMachineSink("disable-postra-machine-sink", 92 cl::Hidden, 93 cl::desc("Disable PostRA Machine Sinking")); 94 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden, 95 cl::desc("Disable Loop Strength Reduction Pass")); 96 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting", 97 cl::Hidden, cl::desc("Disable ConstantHoisting")); 98 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden, 99 cl::desc("Disable Codegen Prepare")); 100 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden, 101 cl::desc("Disable Copy Propagation pass")); 102 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining", 103 cl::Hidden, cl::desc("Disable Partial Libcall Inlining")); 104 static cl::opt<bool> DisableAtExitBasedGlobalDtorLowering( 105 "disable-atexit-based-global-dtor-lowering", cl::Hidden, 106 cl::desc("For MachO, disable atexit()-based global destructor lowering")); 107 static cl::opt<bool> EnableImplicitNullChecks( 108 "enable-implicit-null-checks", 109 cl::desc("Fold null checks into faulting memory operations"), 110 cl::init(false), cl::Hidden); 111 static cl::opt<bool> DisableMergeICmps("disable-mergeicmps", 112 cl::desc("Disable MergeICmps Pass"), 113 cl::init(false), cl::Hidden); 114 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden, 115 cl::desc("Print LLVM IR produced by the loop-reduce pass")); 116 static cl::opt<bool> 117 PrintISelInput("print-isel-input", cl::Hidden, 118 cl::desc("Print LLVM IR input to isel pass")); 119 static cl::opt<cl::boolOrDefault> 120 VerifyMachineCode("verify-machineinstrs", cl::Hidden, 121 cl::desc("Verify generated machine code")); 122 static cl::opt<cl::boolOrDefault> 123 DebugifyAndStripAll("debugify-and-strip-all-safe", cl::Hidden, 124 cl::desc("Debugify MIR before and Strip debug after " 125 "each pass except those known to be unsafe " 126 "when debug info is present")); 127 static cl::opt<cl::boolOrDefault> DebugifyCheckAndStripAll( 128 "debugify-check-and-strip-all-safe", cl::Hidden, 129 cl::desc( 130 "Debugify MIR before, by checking and stripping the debug info after, " 131 "each pass except those known to be unsafe when debug info is " 132 "present")); 133 // Enable or disable the MachineOutliner. 134 static cl::opt<RunOutliner> EnableMachineOutliner( 135 "enable-machine-outliner", cl::desc("Enable the machine outliner"), 136 cl::Hidden, cl::ValueOptional, cl::init(RunOutliner::TargetDefault), 137 cl::values(clEnumValN(RunOutliner::AlwaysOutline, "always", 138 "Run on all functions guaranteed to be beneficial"), 139 clEnumValN(RunOutliner::NeverOutline, "never", 140 "Disable all outlining"), 141 // Sentinel value for unspecified option. 142 clEnumValN(RunOutliner::AlwaysOutline, "", ""))); 143 // Disable the pass to fix unwind information. Whether the pass is included in 144 // the pipeline is controlled via the target options, this option serves as 145 // manual override. 146 static cl::opt<bool> DisableCFIFixup("disable-cfi-fixup", cl::Hidden, 147 cl::desc("Disable the CFI fixup pass")); 148 // Enable or disable FastISel. Both options are needed, because 149 // FastISel is enabled by default with -fast, and we wish to be 150 // able to enable or disable fast-isel independently from -O0. 151 static cl::opt<cl::boolOrDefault> 152 EnableFastISelOption("fast-isel", cl::Hidden, 153 cl::desc("Enable the \"fast\" instruction selector")); 154 155 static cl::opt<cl::boolOrDefault> EnableGlobalISelOption( 156 "global-isel", cl::Hidden, 157 cl::desc("Enable the \"global\" instruction selector")); 158 159 // FIXME: remove this after switching to NPM or GlobalISel, whichever gets there 160 // first... 161 static cl::opt<bool> 162 PrintAfterISel("print-after-isel", cl::init(false), cl::Hidden, 163 cl::desc("Print machine instrs after ISel")); 164 165 static cl::opt<GlobalISelAbortMode> EnableGlobalISelAbort( 166 "global-isel-abort", cl::Hidden, 167 cl::desc("Enable abort calls when \"global\" instruction selection " 168 "fails to lower/select an instruction"), 169 cl::values( 170 clEnumValN(GlobalISelAbortMode::Disable, "0", "Disable the abort"), 171 clEnumValN(GlobalISelAbortMode::Enable, "1", "Enable the abort"), 172 clEnumValN(GlobalISelAbortMode::DisableWithDiag, "2", 173 "Disable the abort but emit a diagnostic on failure"))); 174 175 // Disable MIRProfileLoader before RegAlloc. This is for for debugging and 176 // tuning purpose. 177 static cl::opt<bool> DisableRAFSProfileLoader( 178 "disable-ra-fsprofile-loader", cl::init(false), cl::Hidden, 179 cl::desc("Disable MIRProfileLoader before RegAlloc")); 180 // Disable MIRProfileLoader before BloackPlacement. This is for for debugging 181 // and tuning purpose. 182 static cl::opt<bool> DisableLayoutFSProfileLoader( 183 "disable-layout-fsprofile-loader", cl::init(false), cl::Hidden, 184 cl::desc("Disable MIRProfileLoader before BlockPlacement")); 185 // Specify FSProfile file name. 186 static cl::opt<std::string> 187 FSProfileFile("fs-profile-file", cl::init(""), cl::value_desc("filename"), 188 cl::desc("Flow Sensitive profile file name."), cl::Hidden); 189 // Specify Remapping file for FSProfile. 190 static cl::opt<std::string> FSRemappingFile( 191 "fs-remapping-file", cl::init(""), cl::value_desc("filename"), 192 cl::desc("Flow Sensitive profile remapping file name."), cl::Hidden); 193 194 // Temporary option to allow experimenting with MachineScheduler as a post-RA 195 // scheduler. Targets can "properly" enable this with 196 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID). 197 // Targets can return true in targetSchedulesPostRAScheduling() and 198 // insert a PostRA scheduling pass wherever it wants. 199 static cl::opt<bool> MISchedPostRA( 200 "misched-postra", cl::Hidden, 201 cl::desc( 202 "Run MachineScheduler post regalloc (independent of preRA sched)")); 203 204 // Experimental option to run live interval analysis early. 205 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden, 206 cl::desc("Run live interval analysis earlier in the pipeline")); 207 208 static cl::opt<bool> DisableReplaceWithVecLib( 209 "disable-replace-with-vec-lib", cl::Hidden, 210 cl::desc("Disable replace with vector math call pass")); 211 212 /// Option names for limiting the codegen pipeline. 213 /// Those are used in error reporting and we didn't want 214 /// to duplicate their names all over the place. 215 static const char StartAfterOptName[] = "start-after"; 216 static const char StartBeforeOptName[] = "start-before"; 217 static const char StopAfterOptName[] = "stop-after"; 218 static const char StopBeforeOptName[] = "stop-before"; 219 220 static cl::opt<std::string> 221 StartAfterOpt(StringRef(StartAfterOptName), 222 cl::desc("Resume compilation after a specific pass"), 223 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 224 225 static cl::opt<std::string> 226 StartBeforeOpt(StringRef(StartBeforeOptName), 227 cl::desc("Resume compilation before a specific pass"), 228 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 229 230 static cl::opt<std::string> 231 StopAfterOpt(StringRef(StopAfterOptName), 232 cl::desc("Stop compilation after a specific pass"), 233 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 234 235 static cl::opt<std::string> 236 StopBeforeOpt(StringRef(StopBeforeOptName), 237 cl::desc("Stop compilation before a specific pass"), 238 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 239 240 /// Enable the machine function splitter pass. 241 static cl::opt<bool> EnableMachineFunctionSplitter( 242 "enable-split-machine-functions", cl::Hidden, 243 cl::desc("Split out cold blocks from machine functions based on profile " 244 "information.")); 245 246 /// Disable the expand reductions pass for testing. 247 static cl::opt<bool> DisableExpandReductions( 248 "disable-expand-reductions", cl::init(false), cl::Hidden, 249 cl::desc("Disable the expand reduction intrinsics pass from running")); 250 251 /// Disable the select optimization pass. 252 static cl::opt<bool> DisableSelectOptimize( 253 "disable-select-optimize", cl::init(true), cl::Hidden, 254 cl::desc("Disable the select-optimization pass from running")); 255 256 /// Enable garbage-collecting empty basic blocks. 257 static cl::opt<bool> 258 GCEmptyBlocks("gc-empty-basic-blocks", cl::init(false), cl::Hidden, 259 cl::desc("Enable garbage-collecting empty basic blocks")); 260 261 /// Allow standard passes to be disabled by command line options. This supports 262 /// simple binary flags that either suppress the pass or do nothing. 263 /// i.e. -disable-mypass=false has no effect. 264 /// These should be converted to boolOrDefault in order to use applyOverride. 265 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID, 266 bool Override) { 267 if (Override) 268 return IdentifyingPassPtr(); 269 return PassID; 270 } 271 272 /// Allow standard passes to be disabled by the command line, regardless of who 273 /// is adding the pass. 274 /// 275 /// StandardID is the pass identified in the standard pass pipeline and provided 276 /// to addPass(). It may be a target-specific ID in the case that the target 277 /// directly adds its own pass, but in that case we harmlessly fall through. 278 /// 279 /// TargetID is the pass that the target has configured to override StandardID. 280 /// 281 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real 282 /// pass to run. This allows multiple options to control a single pass depending 283 /// on where in the pipeline that pass is added. 284 static IdentifyingPassPtr overridePass(AnalysisID StandardID, 285 IdentifyingPassPtr TargetID) { 286 if (StandardID == &PostRASchedulerID) 287 return applyDisable(TargetID, DisablePostRASched); 288 289 if (StandardID == &BranchFolderPassID) 290 return applyDisable(TargetID, DisableBranchFold); 291 292 if (StandardID == &TailDuplicateID) 293 return applyDisable(TargetID, DisableTailDuplicate); 294 295 if (StandardID == &EarlyTailDuplicateID) 296 return applyDisable(TargetID, DisableEarlyTailDup); 297 298 if (StandardID == &MachineBlockPlacementID) 299 return applyDisable(TargetID, DisableBlockPlacement); 300 301 if (StandardID == &StackSlotColoringID) 302 return applyDisable(TargetID, DisableSSC); 303 304 if (StandardID == &DeadMachineInstructionElimID) 305 return applyDisable(TargetID, DisableMachineDCE); 306 307 if (StandardID == &EarlyIfConverterID) 308 return applyDisable(TargetID, DisableEarlyIfConversion); 309 310 if (StandardID == &EarlyMachineLICMID) 311 return applyDisable(TargetID, DisableMachineLICM); 312 313 if (StandardID == &MachineCSEID) 314 return applyDisable(TargetID, DisableMachineCSE); 315 316 if (StandardID == &MachineLICMID) 317 return applyDisable(TargetID, DisablePostRAMachineLICM); 318 319 if (StandardID == &MachineSinkingID) 320 return applyDisable(TargetID, DisableMachineSink); 321 322 if (StandardID == &PostRAMachineSinkingID) 323 return applyDisable(TargetID, DisablePostRAMachineSink); 324 325 if (StandardID == &MachineCopyPropagationID) 326 return applyDisable(TargetID, DisableCopyProp); 327 328 return TargetID; 329 } 330 331 // Find the FSProfile file name. The internal option takes the precedence 332 // before getting from TargetMachine. 333 static std::string getFSProfileFile(const TargetMachine *TM) { 334 if (!FSProfileFile.empty()) 335 return FSProfileFile.getValue(); 336 const std::optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 337 if (PGOOpt == std::nullopt || PGOOpt->Action != PGOOptions::SampleUse) 338 return std::string(); 339 return PGOOpt->ProfileFile; 340 } 341 342 // Find the Profile remapping file name. The internal option takes the 343 // precedence before getting from TargetMachine. 344 static std::string getFSRemappingFile(const TargetMachine *TM) { 345 if (!FSRemappingFile.empty()) 346 return FSRemappingFile.getValue(); 347 const std::optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 348 if (PGOOpt == std::nullopt || PGOOpt->Action != PGOOptions::SampleUse) 349 return std::string(); 350 return PGOOpt->ProfileRemappingFile; 351 } 352 353 //===---------------------------------------------------------------------===// 354 /// TargetPassConfig 355 //===---------------------------------------------------------------------===// 356 357 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig", 358 "Target Pass Configuration", false, false) 359 char TargetPassConfig::ID = 0; 360 361 namespace { 362 363 struct InsertedPass { 364 AnalysisID TargetPassID; 365 IdentifyingPassPtr InsertedPassID; 366 367 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID) 368 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID) {} 369 370 Pass *getInsertedPass() const { 371 assert(InsertedPassID.isValid() && "Illegal Pass ID!"); 372 if (InsertedPassID.isInstance()) 373 return InsertedPassID.getInstance(); 374 Pass *NP = Pass::createPass(InsertedPassID.getID()); 375 assert(NP && "Pass ID not registered"); 376 return NP; 377 } 378 }; 379 380 } // end anonymous namespace 381 382 namespace llvm { 383 384 extern cl::opt<bool> EnableFSDiscriminator; 385 386 class PassConfigImpl { 387 public: 388 // List of passes explicitly substituted by this target. Normally this is 389 // empty, but it is a convenient way to suppress or replace specific passes 390 // that are part of a standard pass pipeline without overridding the entire 391 // pipeline. This mechanism allows target options to inherit a standard pass's 392 // user interface. For example, a target may disable a standard pass by 393 // default by substituting a pass ID of zero, and the user may still enable 394 // that standard pass with an explicit command line option. 395 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses; 396 397 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass 398 /// is inserted after each instance of the first one. 399 SmallVector<InsertedPass, 4> InsertedPasses; 400 }; 401 402 } // end namespace llvm 403 404 // Out of line virtual method. 405 TargetPassConfig::~TargetPassConfig() { 406 delete Impl; 407 } 408 409 static const PassInfo *getPassInfo(StringRef PassName) { 410 if (PassName.empty()) 411 return nullptr; 412 413 const PassRegistry &PR = *PassRegistry::getPassRegistry(); 414 const PassInfo *PI = PR.getPassInfo(PassName); 415 if (!PI) 416 report_fatal_error(Twine('\"') + Twine(PassName) + 417 Twine("\" pass is not registered.")); 418 return PI; 419 } 420 421 static AnalysisID getPassIDFromName(StringRef PassName) { 422 const PassInfo *PI = getPassInfo(PassName); 423 return PI ? PI->getTypeInfo() : nullptr; 424 } 425 426 static std::pair<StringRef, unsigned> 427 getPassNameAndInstanceNum(StringRef PassName) { 428 StringRef Name, InstanceNumStr; 429 std::tie(Name, InstanceNumStr) = PassName.split(','); 430 431 unsigned InstanceNum = 0; 432 if (!InstanceNumStr.empty() && InstanceNumStr.getAsInteger(10, InstanceNum)) 433 report_fatal_error("invalid pass instance specifier " + PassName); 434 435 return std::make_pair(Name, InstanceNum); 436 } 437 438 void TargetPassConfig::setStartStopPasses() { 439 StringRef StartBeforeName; 440 std::tie(StartBeforeName, StartBeforeInstanceNum) = 441 getPassNameAndInstanceNum(StartBeforeOpt); 442 443 StringRef StartAfterName; 444 std::tie(StartAfterName, StartAfterInstanceNum) = 445 getPassNameAndInstanceNum(StartAfterOpt); 446 447 StringRef StopBeforeName; 448 std::tie(StopBeforeName, StopBeforeInstanceNum) 449 = getPassNameAndInstanceNum(StopBeforeOpt); 450 451 StringRef StopAfterName; 452 std::tie(StopAfterName, StopAfterInstanceNum) 453 = getPassNameAndInstanceNum(StopAfterOpt); 454 455 StartBefore = getPassIDFromName(StartBeforeName); 456 StartAfter = getPassIDFromName(StartAfterName); 457 StopBefore = getPassIDFromName(StopBeforeName); 458 StopAfter = getPassIDFromName(StopAfterName); 459 if (StartBefore && StartAfter) 460 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") + 461 Twine(StartAfterOptName) + Twine(" specified!")); 462 if (StopBefore && StopAfter) 463 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") + 464 Twine(StopAfterOptName) + Twine(" specified!")); 465 Started = (StartAfter == nullptr) && (StartBefore == nullptr); 466 } 467 468 CGPassBuilderOption llvm::getCGPassBuilderOption() { 469 CGPassBuilderOption Opt; 470 471 #define SET_OPTION(Option) \ 472 if (Option.getNumOccurrences()) \ 473 Opt.Option = Option; 474 475 SET_OPTION(EnableFastISelOption) 476 SET_OPTION(EnableGlobalISelAbort) 477 SET_OPTION(EnableGlobalISelOption) 478 SET_OPTION(EnableIPRA) 479 SET_OPTION(OptimizeRegAlloc) 480 SET_OPTION(VerifyMachineCode) 481 SET_OPTION(DisableAtExitBasedGlobalDtorLowering) 482 SET_OPTION(DisableExpandReductions) 483 SET_OPTION(PrintAfterISel) 484 SET_OPTION(FSProfileFile) 485 SET_OPTION(GCEmptyBlocks) 486 487 #define SET_BOOLEAN_OPTION(Option) Opt.Option = Option; 488 489 SET_BOOLEAN_OPTION(EarlyLiveIntervals) 490 SET_BOOLEAN_OPTION(EnableBlockPlacementStats) 491 SET_BOOLEAN_OPTION(EnableImplicitNullChecks) 492 SET_BOOLEAN_OPTION(EnableMachineOutliner) 493 SET_BOOLEAN_OPTION(MISchedPostRA) 494 SET_BOOLEAN_OPTION(DisableMergeICmps) 495 SET_BOOLEAN_OPTION(DisableLSR) 496 SET_BOOLEAN_OPTION(DisableConstantHoisting) 497 SET_BOOLEAN_OPTION(DisableCGP) 498 SET_BOOLEAN_OPTION(DisablePartialLibcallInlining) 499 SET_BOOLEAN_OPTION(DisableSelectOptimize) 500 SET_BOOLEAN_OPTION(PrintLSR) 501 SET_BOOLEAN_OPTION(PrintISelInput) 502 SET_BOOLEAN_OPTION(DebugifyAndStripAll) 503 SET_BOOLEAN_OPTION(DebugifyCheckAndStripAll) 504 SET_BOOLEAN_OPTION(DisableRAFSProfileLoader) 505 SET_BOOLEAN_OPTION(DisableCFIFixup) 506 SET_BOOLEAN_OPTION(EnableMachineFunctionSplitter) 507 508 return Opt; 509 } 510 511 void llvm::registerCodeGenCallback(PassInstrumentationCallbacks &PIC, 512 LLVMTargetMachine &LLVMTM) { 513 514 // Register a callback for disabling passes. 515 PIC.registerShouldRunOptionalPassCallback([](StringRef P, Any) { 516 517 #define DISABLE_PASS(Option, Name) \ 518 if (Option && P.contains(#Name)) \ 519 return false; 520 DISABLE_PASS(DisableBlockPlacement, MachineBlockPlacementPass) 521 DISABLE_PASS(DisableBranchFold, BranchFolderPass) 522 DISABLE_PASS(DisableCopyProp, MachineCopyPropagationPass) 523 DISABLE_PASS(DisableEarlyIfConversion, EarlyIfConverterPass) 524 DISABLE_PASS(DisableEarlyTailDup, EarlyTailDuplicatePass) 525 DISABLE_PASS(DisableMachineCSE, MachineCSEPass) 526 DISABLE_PASS(DisableMachineDCE, DeadMachineInstructionElimPass) 527 DISABLE_PASS(DisableMachineLICM, EarlyMachineLICMPass) 528 DISABLE_PASS(DisableMachineSink, MachineSinkingPass) 529 DISABLE_PASS(DisablePostRAMachineLICM, MachineLICMPass) 530 DISABLE_PASS(DisablePostRAMachineSink, PostRAMachineSinkingPass) 531 DISABLE_PASS(DisablePostRASched, PostRASchedulerPass) 532 DISABLE_PASS(DisableSSC, StackSlotColoringPass) 533 DISABLE_PASS(DisableTailDuplicate, TailDuplicatePass) 534 535 return true; 536 }); 537 } 538 539 Expected<TargetPassConfig::StartStopInfo> 540 TargetPassConfig::getStartStopInfo(PassInstrumentationCallbacks &PIC) { 541 auto [StartBefore, StartBeforeInstanceNum] = 542 getPassNameAndInstanceNum(StartBeforeOpt); 543 auto [StartAfter, StartAfterInstanceNum] = 544 getPassNameAndInstanceNum(StartAfterOpt); 545 auto [StopBefore, StopBeforeInstanceNum] = 546 getPassNameAndInstanceNum(StopBeforeOpt); 547 auto [StopAfter, StopAfterInstanceNum] = 548 getPassNameAndInstanceNum(StopAfterOpt); 549 550 if (!StartBefore.empty() && !StartAfter.empty()) 551 return make_error<StringError>( 552 Twine(StartBeforeOptName) + " and " + StartAfterOptName + " specified!", 553 std::make_error_code(std::errc::invalid_argument)); 554 if (!StopBefore.empty() && !StopAfter.empty()) 555 return make_error<StringError>( 556 Twine(StopBeforeOptName) + " and " + StopAfterOptName + " specified!", 557 std::make_error_code(std::errc::invalid_argument)); 558 559 StartStopInfo Result; 560 Result.StartPass = StartBefore.empty() ? StartAfter : StartBefore; 561 Result.StopPass = StopBefore.empty() ? StopAfter : StopBefore; 562 Result.StartInstanceNum = 563 StartBefore.empty() ? StartAfterInstanceNum : StartBeforeInstanceNum; 564 Result.StopInstanceNum = 565 StopBefore.empty() ? StopAfterInstanceNum : StopBeforeInstanceNum; 566 Result.StartAfter = !StartAfter.empty(); 567 Result.StopAfter = !StopAfter.empty(); 568 Result.StartInstanceNum += Result.StartInstanceNum == 0; 569 Result.StopInstanceNum += Result.StopInstanceNum == 0; 570 return Result; 571 } 572 573 // Out of line constructor provides default values for pass options and 574 // registers all common codegen passes. 575 TargetPassConfig::TargetPassConfig(LLVMTargetMachine &TM, PassManagerBase &pm) 576 : ImmutablePass(ID), PM(&pm), TM(&TM) { 577 Impl = new PassConfigImpl(); 578 579 // Register all target independent codegen passes to activate their PassIDs, 580 // including this pass itself. 581 initializeCodeGen(*PassRegistry::getPassRegistry()); 582 583 // Also register alias analysis passes required by codegen passes. 584 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry()); 585 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry()); 586 587 if (EnableIPRA.getNumOccurrences()) 588 TM.Options.EnableIPRA = EnableIPRA; 589 else { 590 // If not explicitly specified, use target default. 591 TM.Options.EnableIPRA |= TM.useIPRA(); 592 } 593 594 if (TM.Options.EnableIPRA) 595 setRequiresCodeGenSCCOrder(); 596 597 if (EnableGlobalISelAbort.getNumOccurrences()) 598 TM.Options.GlobalISelAbort = EnableGlobalISelAbort; 599 600 setStartStopPasses(); 601 } 602 603 CodeGenOptLevel TargetPassConfig::getOptLevel() const { 604 return TM->getOptLevel(); 605 } 606 607 /// Insert InsertedPassID pass after TargetPassID. 608 void TargetPassConfig::insertPass(AnalysisID TargetPassID, 609 IdentifyingPassPtr InsertedPassID) { 610 assert(((!InsertedPassID.isInstance() && 611 TargetPassID != InsertedPassID.getID()) || 612 (InsertedPassID.isInstance() && 613 TargetPassID != InsertedPassID.getInstance()->getPassID())) && 614 "Insert a pass after itself!"); 615 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID); 616 } 617 618 /// createPassConfig - Create a pass configuration object to be used by 619 /// addPassToEmitX methods for generating a pipeline of CodeGen passes. 620 /// 621 /// Targets may override this to extend TargetPassConfig. 622 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) { 623 return new TargetPassConfig(*this, PM); 624 } 625 626 TargetPassConfig::TargetPassConfig() 627 : ImmutablePass(ID) { 628 report_fatal_error("Trying to construct TargetPassConfig without a target " 629 "machine. Scheduling a CodeGen pass without a target " 630 "triple set?"); 631 } 632 633 bool TargetPassConfig::willCompleteCodeGenPipeline() { 634 return StopBeforeOpt.empty() && StopAfterOpt.empty(); 635 } 636 637 bool TargetPassConfig::hasLimitedCodeGenPipeline() { 638 return !StartBeforeOpt.empty() || !StartAfterOpt.empty() || 639 !willCompleteCodeGenPipeline(); 640 } 641 642 std::string TargetPassConfig::getLimitedCodeGenPipelineReason() { 643 if (!hasLimitedCodeGenPipeline()) 644 return std::string(); 645 std::string Res; 646 static cl::opt<std::string> *PassNames[] = {&StartAfterOpt, &StartBeforeOpt, 647 &StopAfterOpt, &StopBeforeOpt}; 648 static const char *OptNames[] = {StartAfterOptName, StartBeforeOptName, 649 StopAfterOptName, StopBeforeOptName}; 650 bool IsFirst = true; 651 for (int Idx = 0; Idx < 4; ++Idx) 652 if (!PassNames[Idx]->empty()) { 653 if (!IsFirst) 654 Res += " and "; 655 IsFirst = false; 656 Res += OptNames[Idx]; 657 } 658 return Res; 659 } 660 661 // Helper to verify the analysis is really immutable. 662 void TargetPassConfig::setOpt(bool &Opt, bool Val) { 663 assert(!Initialized && "PassConfig is immutable"); 664 Opt = Val; 665 } 666 667 void TargetPassConfig::substitutePass(AnalysisID StandardID, 668 IdentifyingPassPtr TargetID) { 669 Impl->TargetPasses[StandardID] = TargetID; 670 } 671 672 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const { 673 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator 674 I = Impl->TargetPasses.find(ID); 675 if (I == Impl->TargetPasses.end()) 676 return ID; 677 return I->second; 678 } 679 680 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const { 681 IdentifyingPassPtr TargetID = getPassSubstitution(ID); 682 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID); 683 return !FinalPtr.isValid() || FinalPtr.isInstance() || 684 FinalPtr.getID() != ID; 685 } 686 687 /// Add a pass to the PassManager if that pass is supposed to be run. If the 688 /// Started/Stopped flags indicate either that the compilation should start at 689 /// a later pass or that it should stop after an earlier pass, then do not add 690 /// the pass. Finally, compare the current pass against the StartAfter 691 /// and StopAfter options and change the Started/Stopped flags accordingly. 692 void TargetPassConfig::addPass(Pass *P) { 693 assert(!Initialized && "PassConfig is immutable"); 694 695 // Cache the Pass ID here in case the pass manager finds this pass is 696 // redundant with ones already scheduled / available, and deletes it. 697 // Fundamentally, once we add the pass to the manager, we no longer own it 698 // and shouldn't reference it. 699 AnalysisID PassID = P->getPassID(); 700 701 if (StartBefore == PassID && StartBeforeCount++ == StartBeforeInstanceNum) 702 Started = true; 703 if (StopBefore == PassID && StopBeforeCount++ == StopBeforeInstanceNum) 704 Stopped = true; 705 if (Started && !Stopped) { 706 if (AddingMachinePasses) { 707 // Construct banner message before PM->add() as that may delete the pass. 708 std::string Banner = 709 std::string("After ") + std::string(P->getPassName()); 710 addMachinePrePasses(); 711 PM->add(P); 712 addMachinePostPasses(Banner); 713 } else { 714 PM->add(P); 715 } 716 717 // Add the passes after the pass P if there is any. 718 for (const auto &IP : Impl->InsertedPasses) 719 if (IP.TargetPassID == PassID) 720 addPass(IP.getInsertedPass()); 721 } else { 722 delete P; 723 } 724 725 if (StopAfter == PassID && StopAfterCount++ == StopAfterInstanceNum) 726 Stopped = true; 727 728 if (StartAfter == PassID && StartAfterCount++ == StartAfterInstanceNum) 729 Started = true; 730 if (Stopped && !Started) 731 report_fatal_error("Cannot stop compilation after pass that is not run"); 732 } 733 734 /// Add a CodeGen pass at this point in the pipeline after checking for target 735 /// and command line overrides. 736 /// 737 /// addPass cannot return a pointer to the pass instance because is internal the 738 /// PassManager and the instance we create here may already be freed. 739 AnalysisID TargetPassConfig::addPass(AnalysisID PassID) { 740 IdentifyingPassPtr TargetID = getPassSubstitution(PassID); 741 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID); 742 if (!FinalPtr.isValid()) 743 return nullptr; 744 745 Pass *P; 746 if (FinalPtr.isInstance()) 747 P = FinalPtr.getInstance(); 748 else { 749 P = Pass::createPass(FinalPtr.getID()); 750 if (!P) 751 llvm_unreachable("Pass ID not registered"); 752 } 753 AnalysisID FinalID = P->getPassID(); 754 addPass(P); // Ends the lifetime of P. 755 756 return FinalID; 757 } 758 759 void TargetPassConfig::printAndVerify(const std::string &Banner) { 760 addPrintPass(Banner); 761 addVerifyPass(Banner); 762 } 763 764 void TargetPassConfig::addPrintPass(const std::string &Banner) { 765 if (PrintAfterISel) 766 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner)); 767 } 768 769 void TargetPassConfig::addVerifyPass(const std::string &Banner) { 770 bool Verify = VerifyMachineCode == cl::BOU_TRUE; 771 #ifdef EXPENSIVE_CHECKS 772 if (VerifyMachineCode == cl::BOU_UNSET) 773 Verify = TM->isMachineVerifierClean(); 774 #endif 775 if (Verify) 776 PM->add(createMachineVerifierPass(Banner)); 777 } 778 779 void TargetPassConfig::addDebugifyPass() { 780 PM->add(createDebugifyMachineModulePass()); 781 } 782 783 void TargetPassConfig::addStripDebugPass() { 784 PM->add(createStripDebugMachineModulePass(/*OnlyDebugified=*/true)); 785 } 786 787 void TargetPassConfig::addCheckDebugPass() { 788 PM->add(createCheckDebugMachineModulePass()); 789 } 790 791 void TargetPassConfig::addMachinePrePasses(bool AllowDebugify) { 792 if (AllowDebugify && DebugifyIsSafe && 793 (DebugifyAndStripAll == cl::BOU_TRUE || 794 DebugifyCheckAndStripAll == cl::BOU_TRUE)) 795 addDebugifyPass(); 796 } 797 798 void TargetPassConfig::addMachinePostPasses(const std::string &Banner) { 799 if (DebugifyIsSafe) { 800 if (DebugifyCheckAndStripAll == cl::BOU_TRUE) { 801 addCheckDebugPass(); 802 addStripDebugPass(); 803 } else if (DebugifyAndStripAll == cl::BOU_TRUE) 804 addStripDebugPass(); 805 } 806 addVerifyPass(Banner); 807 } 808 809 /// Add common target configurable passes that perform LLVM IR to IR transforms 810 /// following machine independent optimization. 811 void TargetPassConfig::addIRPasses() { 812 // Before running any passes, run the verifier to determine if the input 813 // coming from the front-end and/or optimizer is valid. 814 if (!DisableVerify) 815 addPass(createVerifierPass()); 816 817 if (getOptLevel() != CodeGenOptLevel::None) { 818 // Basic AliasAnalysis support. 819 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 820 // BasicAliasAnalysis wins if they disagree. This is intended to help 821 // support "obvious" type-punning idioms. 822 addPass(createTypeBasedAAWrapperPass()); 823 addPass(createScopedNoAliasAAWrapperPass()); 824 addPass(createBasicAAWrapperPass()); 825 826 // Run loop strength reduction before anything else. 827 if (!DisableLSR) { 828 addPass(createCanonicalizeFreezeInLoopsPass()); 829 addPass(createLoopStrengthReducePass()); 830 if (PrintLSR) 831 addPass(createPrintFunctionPass(dbgs(), 832 "\n\n*** Code after LSR ***\n")); 833 } 834 835 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of 836 // loads and compares. ExpandMemCmpPass then tries to expand those calls 837 // into optimally-sized loads and compares. The transforms are enabled by a 838 // target lowering hook. 839 if (!DisableMergeICmps) 840 addPass(createMergeICmpsLegacyPass()); 841 addPass(createExpandMemCmpLegacyPass()); 842 } 843 844 // Run GC lowering passes for builtin collectors 845 // TODO: add a pass insertion point here 846 addPass(&GCLoweringID); 847 addPass(&ShadowStackGCLoweringID); 848 addPass(createLowerConstantIntrinsicsPass()); 849 850 // For MachO, lower @llvm.global_dtors into @llvm.global_ctors with 851 // __cxa_atexit() calls to avoid emitting the deprecated __mod_term_func. 852 if (TM->getTargetTriple().isOSBinFormatMachO() && 853 !DisableAtExitBasedGlobalDtorLowering) 854 addPass(createLowerGlobalDtorsLegacyPass()); 855 856 // Make sure that no unreachable blocks are instruction selected. 857 addPass(createUnreachableBlockEliminationPass()); 858 859 // Prepare expensive constants for SelectionDAG. 860 if (getOptLevel() != CodeGenOptLevel::None && !DisableConstantHoisting) 861 addPass(createConstantHoistingPass()); 862 863 if (getOptLevel() != CodeGenOptLevel::None && !DisableReplaceWithVecLib) 864 addPass(createReplaceWithVeclibLegacyPass()); 865 866 if (getOptLevel() != CodeGenOptLevel::None && !DisablePartialLibcallInlining) 867 addPass(createPartiallyInlineLibCallsPass()); 868 869 // Expand vector predication intrinsics into standard IR instructions. 870 // This pass has to run before ScalarizeMaskedMemIntrin and ExpandReduction 871 // passes since it emits those kinds of intrinsics. 872 addPass(createExpandVectorPredicationPass()); 873 874 // Add scalarization of target's unsupported masked memory intrinsics pass. 875 // the unsupported intrinsic will be replaced with a chain of basic blocks, 876 // that stores/loads element one-by-one if the appropriate mask bit is set. 877 addPass(createScalarizeMaskedMemIntrinLegacyPass()); 878 879 // Expand reduction intrinsics into shuffle sequences if the target wants to. 880 // Allow disabling it for testing purposes. 881 if (!DisableExpandReductions) 882 addPass(createExpandReductionsPass()); 883 884 if (getOptLevel() != CodeGenOptLevel::None) 885 addPass(createTLSVariableHoistPass()); 886 887 // Convert conditional moves to conditional jumps when profitable. 888 if (getOptLevel() != CodeGenOptLevel::None && !DisableSelectOptimize) 889 addPass(createSelectOptimizePass()); 890 } 891 892 /// Turn exception handling constructs into something the code generators can 893 /// handle. 894 void TargetPassConfig::addPassesToHandleExceptions() { 895 const MCAsmInfo *MCAI = TM->getMCAsmInfo(); 896 assert(MCAI && "No MCAsmInfo"); 897 switch (MCAI->getExceptionHandlingType()) { 898 case ExceptionHandling::SjLj: 899 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both 900 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise, 901 // catch info can get misplaced when a selector ends up more than one block 902 // removed from the parent invoke(s). This could happen when a landing 903 // pad is shared by multiple invokes and is also a target of a normal 904 // edge from elsewhere. 905 addPass(createSjLjEHPreparePass(TM)); 906 [[fallthrough]]; 907 case ExceptionHandling::DwarfCFI: 908 case ExceptionHandling::ARM: 909 case ExceptionHandling::AIX: 910 case ExceptionHandling::ZOS: 911 addPass(createDwarfEHPass(getOptLevel())); 912 break; 913 case ExceptionHandling::WinEH: 914 // We support using both GCC-style and MSVC-style exceptions on Windows, so 915 // add both preparation passes. Each pass will only actually run if it 916 // recognizes the personality function. 917 addPass(createWinEHPass()); 918 addPass(createDwarfEHPass(getOptLevel())); 919 break; 920 case ExceptionHandling::Wasm: 921 // Wasm EH uses Windows EH instructions, but it does not need to demote PHIs 922 // on catchpads and cleanuppads because it does not outline them into 923 // funclets. Catchswitch blocks are not lowered in SelectionDAG, so we 924 // should remove PHIs there. 925 addPass(createWinEHPass(/*DemoteCatchSwitchPHIOnly=*/true)); 926 addPass(createWasmEHPass()); 927 break; 928 case ExceptionHandling::None: 929 addPass(createLowerInvokePass()); 930 931 // The lower invoke pass may create unreachable code. Remove it. 932 addPass(createUnreachableBlockEliminationPass()); 933 break; 934 } 935 } 936 937 /// Add pass to prepare the LLVM IR for code generation. This should be done 938 /// before exception handling preparation passes. 939 void TargetPassConfig::addCodeGenPrepare() { 940 if (getOptLevel() != CodeGenOptLevel::None && !DisableCGP) 941 addPass(createCodeGenPrepareLegacyPass()); 942 } 943 944 /// Add common passes that perform LLVM IR to IR transforms in preparation for 945 /// instruction selection. 946 void TargetPassConfig::addISelPrepare() { 947 addPreISel(); 948 949 // Force codegen to run according to the callgraph. 950 if (requiresCodeGenSCCOrder()) 951 addPass(new DummyCGSCCPass); 952 953 addPass(createCallBrPass()); 954 955 // Add both the safe stack and the stack protection passes: each of them will 956 // only protect functions that have corresponding attributes. 957 addPass(createSafeStackPass()); 958 addPass(createStackProtectorPass()); 959 960 if (PrintISelInput) 961 addPass(createPrintFunctionPass( 962 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n")); 963 964 // All passes which modify the LLVM IR are now complete; run the verifier 965 // to ensure that the IR is valid. 966 if (!DisableVerify) 967 addPass(createVerifierPass()); 968 } 969 970 bool TargetPassConfig::addCoreISelPasses() { 971 // Enable FastISel with -fast-isel, but allow that to be overridden. 972 TM->setO0WantsFastISel(EnableFastISelOption != cl::BOU_FALSE); 973 974 // Determine an instruction selector. 975 enum class SelectorType { SelectionDAG, FastISel, GlobalISel }; 976 SelectorType Selector; 977 978 if (EnableFastISelOption == cl::BOU_TRUE) 979 Selector = SelectorType::FastISel; 980 else if (EnableGlobalISelOption == cl::BOU_TRUE || 981 (TM->Options.EnableGlobalISel && 982 EnableGlobalISelOption != cl::BOU_FALSE)) 983 Selector = SelectorType::GlobalISel; 984 else if (TM->getOptLevel() == CodeGenOptLevel::None && 985 TM->getO0WantsFastISel()) 986 Selector = SelectorType::FastISel; 987 else 988 Selector = SelectorType::SelectionDAG; 989 990 // Set consistently TM->Options.EnableFastISel and EnableGlobalISel. 991 if (Selector == SelectorType::FastISel) { 992 TM->setFastISel(true); 993 TM->setGlobalISel(false); 994 } else if (Selector == SelectorType::GlobalISel) { 995 TM->setFastISel(false); 996 TM->setGlobalISel(true); 997 } 998 999 // FIXME: Injecting into the DAGISel pipeline seems to cause issues with 1000 // analyses needing to be re-run. This can result in being unable to 1001 // schedule passes (particularly with 'Function Alias Analysis 1002 // Results'). It's not entirely clear why but AFAICT this seems to be 1003 // due to one FunctionPassManager not being able to use analyses from a 1004 // previous one. As we're injecting a ModulePass we break the usual 1005 // pass manager into two. GlobalISel with the fallback path disabled 1006 // and -run-pass seem to be unaffected. The majority of GlobalISel 1007 // testing uses -run-pass so this probably isn't too bad. 1008 SaveAndRestore SavedDebugifyIsSafe(DebugifyIsSafe); 1009 if (Selector != SelectorType::GlobalISel || !isGlobalISelAbortEnabled()) 1010 DebugifyIsSafe = false; 1011 1012 // Add instruction selector passes. 1013 if (Selector == SelectorType::GlobalISel) { 1014 SaveAndRestore SavedAddingMachinePasses(AddingMachinePasses, true); 1015 if (addIRTranslator()) 1016 return true; 1017 1018 addPreLegalizeMachineIR(); 1019 1020 if (addLegalizeMachineIR()) 1021 return true; 1022 1023 // Before running the register bank selector, ask the target if it 1024 // wants to run some passes. 1025 addPreRegBankSelect(); 1026 1027 if (addRegBankSelect()) 1028 return true; 1029 1030 addPreGlobalInstructionSelect(); 1031 1032 if (addGlobalInstructionSelect()) 1033 return true; 1034 1035 // Pass to reset the MachineFunction if the ISel failed. 1036 addPass(createResetMachineFunctionPass( 1037 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled())); 1038 1039 // Provide a fallback path when we do not want to abort on 1040 // not-yet-supported input. 1041 if (!isGlobalISelAbortEnabled() && addInstSelector()) 1042 return true; 1043 1044 } else if (addInstSelector()) 1045 return true; 1046 1047 // Expand pseudo-instructions emitted by ISel. Don't run the verifier before 1048 // FinalizeISel. 1049 addPass(&FinalizeISelID); 1050 1051 // Print the instruction selected machine code... 1052 printAndVerify("After Instruction Selection"); 1053 1054 return false; 1055 } 1056 1057 bool TargetPassConfig::addISelPasses() { 1058 if (TM->useEmulatedTLS()) 1059 addPass(createLowerEmuTLSPass()); 1060 1061 PM->add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis())); 1062 addPass(createPreISelIntrinsicLoweringPass()); 1063 addPass(createExpandLargeDivRemPass()); 1064 addPass(createExpandLargeFpConvertPass()); 1065 addIRPasses(); 1066 addCodeGenPrepare(); 1067 addPassesToHandleExceptions(); 1068 addISelPrepare(); 1069 1070 return addCoreISelPasses(); 1071 } 1072 1073 /// -regalloc=... command line option. 1074 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } 1075 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false, 1076 RegisterPassParser<RegisterRegAlloc>> 1077 RegAlloc("regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator), 1078 cl::desc("Register allocator to use")); 1079 1080 /// Add the complete set of target-independent postISel code generator passes. 1081 /// 1082 /// This can be read as the standard order of major LLVM CodeGen stages. Stages 1083 /// with nontrivial configuration or multiple passes are broken out below in 1084 /// add%Stage routines. 1085 /// 1086 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The 1087 /// addPre/Post methods with empty header implementations allow injecting 1088 /// target-specific fixups just before or after major stages. Additionally, 1089 /// targets have the flexibility to change pass order within a stage by 1090 /// overriding default implementation of add%Stage routines below. Each 1091 /// technique has maintainability tradeoffs because alternate pass orders are 1092 /// not well supported. addPre/Post works better if the target pass is easily 1093 /// tied to a common pass. But if it has subtle dependencies on multiple passes, 1094 /// the target should override the stage instead. 1095 /// 1096 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection 1097 /// before/after any target-independent pass. But it's currently overkill. 1098 void TargetPassConfig::addMachinePasses() { 1099 AddingMachinePasses = true; 1100 1101 // Add passes that optimize machine instructions in SSA form. 1102 if (getOptLevel() != CodeGenOptLevel::None) { 1103 addMachineSSAOptimization(); 1104 } else { 1105 // If the target requests it, assign local variables to stack slots relative 1106 // to one another and simplify frame index references where possible. 1107 addPass(&LocalStackSlotAllocationID); 1108 } 1109 1110 if (TM->Options.EnableIPRA) 1111 addPass(createRegUsageInfoPropPass()); 1112 1113 // Run pre-ra passes. 1114 addPreRegAlloc(); 1115 1116 // Debugifying the register allocator passes seems to provoke some 1117 // non-determinism that affects CodeGen and there doesn't seem to be a point 1118 // where it becomes safe again so stop debugifying here. 1119 DebugifyIsSafe = false; 1120 1121 // Add a FSDiscriminator pass right before RA, so that we could get 1122 // more precise SampleFDO profile for RA. 1123 if (EnableFSDiscriminator) { 1124 addPass(createMIRAddFSDiscriminatorsPass( 1125 sampleprof::FSDiscriminatorPass::Pass1)); 1126 const std::string ProfileFile = getFSProfileFile(TM); 1127 if (!ProfileFile.empty() && !DisableRAFSProfileLoader) 1128 addPass(createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1129 sampleprof::FSDiscriminatorPass::Pass1, 1130 nullptr)); 1131 } 1132 1133 // Run register allocation and passes that are tightly coupled with it, 1134 // including phi elimination and scheduling. 1135 if (getOptimizeRegAlloc()) 1136 addOptimizedRegAlloc(); 1137 else 1138 addFastRegAlloc(); 1139 1140 // Run post-ra passes. 1141 addPostRegAlloc(); 1142 1143 addPass(&RemoveRedundantDebugValuesID); 1144 1145 addPass(&FixupStatepointCallerSavedID); 1146 1147 // Insert prolog/epilog code. Eliminate abstract frame index references... 1148 if (getOptLevel() != CodeGenOptLevel::None) { 1149 addPass(&PostRAMachineSinkingID); 1150 addPass(&ShrinkWrapID); 1151 } 1152 1153 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only 1154 // do so if it hasn't been disabled, substituted, or overridden. 1155 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID)) 1156 addPass(createPrologEpilogInserterPass()); 1157 1158 /// Add passes that optimize machine instructions after register allocation. 1159 if (getOptLevel() != CodeGenOptLevel::None) 1160 addMachineLateOptimization(); 1161 1162 // Expand pseudo instructions before second scheduling pass. 1163 addPass(&ExpandPostRAPseudosID); 1164 1165 // Run pre-sched2 passes. 1166 addPreSched2(); 1167 1168 if (EnableImplicitNullChecks) 1169 addPass(&ImplicitNullChecksID); 1170 1171 // Second pass scheduler. 1172 // Let Target optionally insert this pass by itself at some other 1173 // point. 1174 if (getOptLevel() != CodeGenOptLevel::None && 1175 !TM->targetSchedulesPostRAScheduling()) { 1176 if (MISchedPostRA) 1177 addPass(&PostMachineSchedulerID); 1178 else 1179 addPass(&PostRASchedulerID); 1180 } 1181 1182 // GC 1183 addGCPasses(); 1184 1185 // Basic block placement. 1186 if (getOptLevel() != CodeGenOptLevel::None) 1187 addBlockPlacement(); 1188 1189 // Insert before XRay Instrumentation. 1190 addPass(&FEntryInserterID); 1191 1192 addPass(&XRayInstrumentationID); 1193 addPass(&PatchableFunctionID); 1194 1195 addPreEmitPass(); 1196 1197 if (TM->Options.EnableIPRA) 1198 // Collect register usage information and produce a register mask of 1199 // clobbered registers, to be used to optimize call sites. 1200 addPass(createRegUsageInfoCollector()); 1201 1202 // FIXME: Some backends are incompatible with running the verifier after 1203 // addPreEmitPass. Maybe only pass "false" here for those targets? 1204 addPass(&FuncletLayoutID); 1205 1206 addPass(&StackMapLivenessID); 1207 addPass(&LiveDebugValuesID); 1208 addPass(&MachineSanitizerBinaryMetadataID); 1209 1210 if (TM->Options.EnableMachineOutliner && 1211 getOptLevel() != CodeGenOptLevel::None && 1212 EnableMachineOutliner != RunOutliner::NeverOutline) { 1213 bool RunOnAllFunctions = 1214 (EnableMachineOutliner == RunOutliner::AlwaysOutline); 1215 bool AddOutliner = 1216 RunOnAllFunctions || TM->Options.SupportsDefaultOutlining; 1217 if (AddOutliner) 1218 addPass(createMachineOutlinerPass(RunOnAllFunctions)); 1219 } 1220 1221 if (GCEmptyBlocks) 1222 addPass(llvm::createGCEmptyBasicBlocksPass()); 1223 1224 if (EnableFSDiscriminator) 1225 addPass(createMIRAddFSDiscriminatorsPass( 1226 sampleprof::FSDiscriminatorPass::PassLast)); 1227 1228 bool NeedsBBSections = 1229 TM->getBBSectionsType() != llvm::BasicBlockSection::None; 1230 // Machine function splitter uses the basic block sections feature. Both 1231 // cannot be enabled at the same time. We do not apply machine function 1232 // splitter if -basic-block-sections is requested. 1233 if (!NeedsBBSections && (TM->Options.EnableMachineFunctionSplitter || 1234 EnableMachineFunctionSplitter)) { 1235 const std::string ProfileFile = getFSProfileFile(TM); 1236 if (!ProfileFile.empty()) { 1237 if (EnableFSDiscriminator) { 1238 addPass(createMIRProfileLoaderPass( 1239 ProfileFile, getFSRemappingFile(TM), 1240 sampleprof::FSDiscriminatorPass::PassLast, nullptr)); 1241 } else { 1242 // Sample profile is given, but FSDiscriminator is not 1243 // enabled, this may result in performance regression. 1244 WithColor::warning() 1245 << "Using AutoFDO without FSDiscriminator for MFS may regress " 1246 "performance.\n"; 1247 } 1248 } 1249 addPass(createMachineFunctionSplitterPass()); 1250 } 1251 // We run the BasicBlockSections pass if either we need BB sections or BB 1252 // address map (or both). 1253 if (NeedsBBSections || TM->Options.BBAddrMap) { 1254 if (TM->getBBSectionsType() == llvm::BasicBlockSection::List) { 1255 addPass(llvm::createBasicBlockSectionsProfileReaderWrapperPass( 1256 TM->getBBSectionsFuncListBuf())); 1257 addPass(llvm::createBasicBlockPathCloningPass()); 1258 } 1259 addPass(llvm::createBasicBlockSectionsPass()); 1260 } 1261 1262 addPostBBSections(); 1263 1264 if (!DisableCFIFixup && TM->Options.EnableCFIFixup) 1265 addPass(createCFIFixup()); 1266 1267 PM->add(createStackFrameLayoutAnalysisPass()); 1268 1269 // Add passes that directly emit MI after all other MI passes. 1270 addPreEmitPass2(); 1271 1272 AddingMachinePasses = false; 1273 } 1274 1275 /// Add passes that optimize machine instructions in SSA form. 1276 void TargetPassConfig::addMachineSSAOptimization() { 1277 // Pre-ra tail duplication. 1278 addPass(&EarlyTailDuplicateID); 1279 1280 // Optimize PHIs before DCE: removing dead PHI cycles may make more 1281 // instructions dead. 1282 addPass(&OptimizePHIsID); 1283 1284 // This pass merges large allocas. StackSlotColoring is a different pass 1285 // which merges spill slots. 1286 addPass(&StackColoringID); 1287 1288 // If the target requests it, assign local variables to stack slots relative 1289 // to one another and simplify frame index references where possible. 1290 addPass(&LocalStackSlotAllocationID); 1291 1292 // With optimization, dead code should already be eliminated. However 1293 // there is one known exception: lowered code for arguments that are only 1294 // used by tail calls, where the tail calls reuse the incoming stack 1295 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll). 1296 addPass(&DeadMachineInstructionElimID); 1297 1298 // Allow targets to insert passes that improve instruction level parallelism, 1299 // like if-conversion. Such passes will typically need dominator trees and 1300 // loop info, just like LICM and CSE below. 1301 addILPOpts(); 1302 1303 addPass(&EarlyMachineLICMID); 1304 addPass(&MachineCSEID); 1305 1306 addPass(&MachineSinkingID); 1307 1308 addPass(&PeepholeOptimizerID); 1309 // Clean-up the dead code that may have been generated by peephole 1310 // rewriting. 1311 addPass(&DeadMachineInstructionElimID); 1312 } 1313 1314 //===---------------------------------------------------------------------===// 1315 /// Register Allocation Pass Configuration 1316 //===---------------------------------------------------------------------===// 1317 1318 bool TargetPassConfig::getOptimizeRegAlloc() const { 1319 switch (OptimizeRegAlloc) { 1320 case cl::BOU_UNSET: 1321 return getOptLevel() != CodeGenOptLevel::None; 1322 case cl::BOU_TRUE: return true; 1323 case cl::BOU_FALSE: return false; 1324 } 1325 llvm_unreachable("Invalid optimize-regalloc state"); 1326 } 1327 1328 /// A dummy default pass factory indicates whether the register allocator is 1329 /// overridden on the command line. 1330 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag; 1331 1332 static RegisterRegAlloc 1333 defaultRegAlloc("default", 1334 "pick register allocator based on -O option", 1335 useDefaultRegisterAllocator); 1336 1337 static void initializeDefaultRegisterAllocatorOnce() { 1338 if (!RegisterRegAlloc::getDefault()) 1339 RegisterRegAlloc::setDefault(RegAlloc); 1340 } 1341 1342 /// Instantiate the default register allocator pass for this target for either 1343 /// the optimized or unoptimized allocation path. This will be added to the pass 1344 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc 1345 /// in the optimized case. 1346 /// 1347 /// A target that uses the standard regalloc pass order for fast or optimized 1348 /// allocation may still override this for per-target regalloc 1349 /// selection. But -regalloc=... always takes precedence. 1350 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) { 1351 if (Optimized) 1352 return createGreedyRegisterAllocator(); 1353 else 1354 return createFastRegisterAllocator(); 1355 } 1356 1357 /// Find and instantiate the register allocation pass requested by this target 1358 /// at the current optimization level. Different register allocators are 1359 /// defined as separate passes because they may require different analysis. 1360 /// 1361 /// This helper ensures that the regalloc= option is always available, 1362 /// even for targets that override the default allocator. 1363 /// 1364 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs, 1365 /// this can be folded into addPass. 1366 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) { 1367 // Initialize the global default. 1368 llvm::call_once(InitializeDefaultRegisterAllocatorFlag, 1369 initializeDefaultRegisterAllocatorOnce); 1370 1371 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault(); 1372 if (Ctor != useDefaultRegisterAllocator) 1373 return Ctor(); 1374 1375 // With no -regalloc= override, ask the target for a regalloc pass. 1376 return createTargetRegisterAllocator(Optimized); 1377 } 1378 1379 bool TargetPassConfig::isCustomizedRegAlloc() { 1380 return RegAlloc != 1381 (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator; 1382 } 1383 1384 bool TargetPassConfig::addRegAssignAndRewriteFast() { 1385 if (RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator && 1386 RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&createFastRegisterAllocator) 1387 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc."); 1388 1389 addPass(createRegAllocPass(false)); 1390 1391 // Allow targets to change the register assignments after 1392 // fast register allocation. 1393 addPostFastRegAllocRewrite(); 1394 return true; 1395 } 1396 1397 bool TargetPassConfig::addRegAssignAndRewriteOptimized() { 1398 // Add the selected register allocation pass. 1399 addPass(createRegAllocPass(true)); 1400 1401 // Allow targets to change the register assignments before rewriting. 1402 addPreRewrite(); 1403 1404 // Finally rewrite virtual registers. 1405 addPass(&VirtRegRewriterID); 1406 1407 // Regalloc scoring for ML-driven eviction - noop except when learning a new 1408 // eviction policy. 1409 addPass(createRegAllocScoringPass()); 1410 return true; 1411 } 1412 1413 /// Return true if the default global register allocator is in use and 1414 /// has not be overriden on the command line with '-regalloc=...' 1415 bool TargetPassConfig::usingDefaultRegAlloc() const { 1416 return RegAlloc.getNumOccurrences() == 0; 1417 } 1418 1419 /// Add the minimum set of target-independent passes that are required for 1420 /// register allocation. No coalescing or scheduling. 1421 void TargetPassConfig::addFastRegAlloc() { 1422 addPass(&PHIEliminationID); 1423 addPass(&TwoAddressInstructionPassID); 1424 1425 addRegAssignAndRewriteFast(); 1426 } 1427 1428 /// Add standard target-independent passes that are tightly coupled with 1429 /// optimized register allocation, including coalescing, machine instruction 1430 /// scheduling, and register allocation itself. 1431 void TargetPassConfig::addOptimizedRegAlloc() { 1432 addPass(&DetectDeadLanesID); 1433 1434 addPass(&InitUndefID); 1435 1436 addPass(&ProcessImplicitDefsID); 1437 1438 // LiveVariables currently requires pure SSA form. 1439 // 1440 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags, 1441 // LiveVariables can be removed completely, and LiveIntervals can be directly 1442 // computed. (We still either need to regenerate kill flags after regalloc, or 1443 // preferably fix the scavenger to not depend on them). 1444 // FIXME: UnreachableMachineBlockElim is a dependant pass of LiveVariables. 1445 // When LiveVariables is removed this has to be removed/moved either. 1446 // Explicit addition of UnreachableMachineBlockElim allows stopping before or 1447 // after it with -stop-before/-stop-after. 1448 addPass(&UnreachableMachineBlockElimID); 1449 addPass(&LiveVariablesID); 1450 1451 // Edge splitting is smarter with machine loop info. 1452 addPass(&MachineLoopInfoID); 1453 addPass(&PHIEliminationID); 1454 1455 // Eventually, we want to run LiveIntervals before PHI elimination. 1456 if (EarlyLiveIntervals) 1457 addPass(&LiveIntervalsID); 1458 1459 addPass(&TwoAddressInstructionPassID); 1460 addPass(&RegisterCoalescerID); 1461 1462 // The machine scheduler may accidentally create disconnected components 1463 // when moving subregister definitions around, avoid this by splitting them to 1464 // separate vregs before. Splitting can also improve reg. allocation quality. 1465 addPass(&RenameIndependentSubregsID); 1466 1467 // PreRA instruction scheduling. 1468 addPass(&MachineSchedulerID); 1469 1470 if (addRegAssignAndRewriteOptimized()) { 1471 // Perform stack slot coloring and post-ra machine LICM. 1472 addPass(&StackSlotColoringID); 1473 1474 // Allow targets to expand pseudo instructions depending on the choice of 1475 // registers before MachineCopyPropagation. 1476 addPostRewrite(); 1477 1478 // Copy propagate to forward register uses and try to eliminate COPYs that 1479 // were not coalesced. 1480 addPass(&MachineCopyPropagationID); 1481 1482 // Run post-ra machine LICM to hoist reloads / remats. 1483 // 1484 // FIXME: can this move into MachineLateOptimization? 1485 addPass(&MachineLICMID); 1486 } 1487 } 1488 1489 //===---------------------------------------------------------------------===// 1490 /// Post RegAlloc Pass Configuration 1491 //===---------------------------------------------------------------------===// 1492 1493 /// Add passes that optimize machine instructions after register allocation. 1494 void TargetPassConfig::addMachineLateOptimization() { 1495 // Cleanup of redundant immediate/address loads. 1496 addPass(&MachineLateInstrsCleanupID); 1497 1498 // Branch folding must be run after regalloc and prolog/epilog insertion. 1499 addPass(&BranchFolderPassID); 1500 1501 // Tail duplication. 1502 // Note that duplicating tail just increases code size and degrades 1503 // performance for targets that require Structured Control Flow. 1504 // In addition it can also make CFG irreducible. Thus we disable it. 1505 if (!TM->requiresStructuredCFG()) 1506 addPass(&TailDuplicateID); 1507 1508 // Copy propagation. 1509 addPass(&MachineCopyPropagationID); 1510 } 1511 1512 /// Add standard GC passes. 1513 bool TargetPassConfig::addGCPasses() { 1514 addPass(&GCMachineCodeAnalysisID); 1515 return true; 1516 } 1517 1518 /// Add standard basic block placement passes. 1519 void TargetPassConfig::addBlockPlacement() { 1520 if (EnableFSDiscriminator) { 1521 addPass(createMIRAddFSDiscriminatorsPass( 1522 sampleprof::FSDiscriminatorPass::Pass2)); 1523 const std::string ProfileFile = getFSProfileFile(TM); 1524 if (!ProfileFile.empty() && !DisableLayoutFSProfileLoader) 1525 addPass(createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1526 sampleprof::FSDiscriminatorPass::Pass2, 1527 nullptr)); 1528 } 1529 if (addPass(&MachineBlockPlacementID)) { 1530 // Run a separate pass to collect block placement statistics. 1531 if (EnableBlockPlacementStats) 1532 addPass(&MachineBlockPlacementStatsID); 1533 } 1534 } 1535 1536 //===---------------------------------------------------------------------===// 1537 /// GlobalISel Configuration 1538 //===---------------------------------------------------------------------===// 1539 bool TargetPassConfig::isGlobalISelAbortEnabled() const { 1540 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::Enable; 1541 } 1542 1543 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const { 1544 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::DisableWithDiag; 1545 } 1546 1547 bool TargetPassConfig::isGISelCSEEnabled() const { 1548 return true; 1549 } 1550 1551 std::unique_ptr<CSEConfigBase> TargetPassConfig::getCSEConfig() const { 1552 return std::make_unique<CSEConfigBase>(); 1553 } 1554