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/CodeGenTargetMachineImpl.h" 26 #include "llvm/CodeGen/MachineFunctionPass.h" 27 #include "llvm/CodeGen/MachinePassRegistry.h" 28 #include "llvm/CodeGen/Passes.h" 29 #include "llvm/CodeGen/RegAllocRegistry.h" 30 #include "llvm/IR/IRPrintingPasses.h" 31 #include "llvm/IR/LegacyPassManager.h" 32 #include "llvm/IR/PassInstrumentation.h" 33 #include "llvm/IR/Verifier.h" 34 #include "llvm/InitializePasses.h" 35 #include "llvm/MC/MCAsmInfo.h" 36 #include "llvm/MC/MCTargetOptions.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/CodeGen.h" 39 #include "llvm/Support/CommandLine.h" 40 #include "llvm/Support/Compiler.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/Discriminator.h" 43 #include "llvm/Support/ErrorHandling.h" 44 #include "llvm/Support/SaveAndRestore.h" 45 #include "llvm/Support/Threading.h" 46 #include "llvm/Support/VirtualFileSystem.h" 47 #include "llvm/Support/WithColor.h" 48 #include "llvm/Target/CGPassBuilderOption.h" 49 #include "llvm/Target/TargetMachine.h" 50 #include "llvm/Transforms/ObjCARC.h" 51 #include "llvm/Transforms/Scalar.h" 52 #include "llvm/Transforms/Utils.h" 53 #include <cassert> 54 #include <optional> 55 #include <string> 56 57 using namespace llvm; 58 59 static cl::opt<bool> 60 EnableIPRA("enable-ipra", cl::init(false), cl::Hidden, 61 cl::desc("Enable interprocedural register allocation " 62 "to reduce load/store at procedure calls.")); 63 static cl::opt<bool> DisablePostRASched("disable-post-ra", cl::Hidden, 64 cl::desc("Disable Post Regalloc Scheduler")); 65 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden, 66 cl::desc("Disable branch folding")); 67 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden, 68 cl::desc("Disable tail duplication")); 69 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden, 70 cl::desc("Disable pre-register allocation tail duplication")); 71 static cl::opt<bool> DisableBlockPlacement("disable-block-placement", 72 cl::Hidden, cl::desc("Disable probability-driven block placement")); 73 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats", 74 cl::Hidden, cl::desc("Collect probability-driven block placement stats")); 75 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden, 76 cl::desc("Disable Stack Slot Coloring")); 77 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden, 78 cl::desc("Disable Machine Dead Code Elimination")); 79 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden, 80 cl::desc("Disable Early If-conversion")); 81 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden, 82 cl::desc("Disable Machine LICM")); 83 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden, 84 cl::desc("Disable Machine Common Subexpression Elimination")); 85 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc( 86 "optimize-regalloc", cl::Hidden, 87 cl::desc("Enable optimized register allocation compilation path.")); 88 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm", 89 cl::Hidden, 90 cl::desc("Disable Machine LICM")); 91 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden, 92 cl::desc("Disable Machine Sinking")); 93 static cl::opt<bool> DisablePostRAMachineSink("disable-postra-machine-sink", 94 cl::Hidden, 95 cl::desc("Disable PostRA Machine Sinking")); 96 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden, 97 cl::desc("Disable Loop Strength Reduction Pass")); 98 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting", 99 cl::Hidden, cl::desc("Disable ConstantHoisting")); 100 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden, 101 cl::desc("Disable Codegen Prepare")); 102 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden, 103 cl::desc("Disable Copy Propagation pass")); 104 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining", 105 cl::Hidden, cl::desc("Disable Partial Libcall Inlining")); 106 static cl::opt<bool> DisableAtExitBasedGlobalDtorLowering( 107 "disable-atexit-based-global-dtor-lowering", cl::Hidden, 108 cl::desc("For MachO, disable atexit()-based global destructor lowering")); 109 static cl::opt<bool> EnableImplicitNullChecks( 110 "enable-implicit-null-checks", 111 cl::desc("Fold null checks into faulting memory operations"), 112 cl::init(false), cl::Hidden); 113 static cl::opt<bool> DisableMergeICmps("disable-mergeicmps", 114 cl::desc("Disable MergeICmps Pass"), 115 cl::init(false), cl::Hidden); 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 static cl::opt<bool> EnableGlobalMergeFunc( 144 "enable-global-merge-func", cl::Hidden, 145 cl::desc("Enable global merge functions that are based on hash function")); 146 // Disable the pass to fix unwind information. Whether the pass is included in 147 // the pipeline is controlled via the target options, this option serves as 148 // manual override. 149 static cl::opt<bool> DisableCFIFixup("disable-cfi-fixup", cl::Hidden, 150 cl::desc("Disable the CFI fixup pass")); 151 // Enable or disable FastISel. Both options are needed, because 152 // FastISel is enabled by default with -fast, and we wish to be 153 // able to enable or disable fast-isel independently from -O0. 154 static cl::opt<cl::boolOrDefault> 155 EnableFastISelOption("fast-isel", cl::Hidden, 156 cl::desc("Enable the \"fast\" instruction selector")); 157 158 static cl::opt<cl::boolOrDefault> EnableGlobalISelOption( 159 "global-isel", cl::Hidden, 160 cl::desc("Enable the \"global\" instruction selector")); 161 162 // FIXME: remove this after switching to NPM or GlobalISel, whichever gets there 163 // first... 164 static cl::opt<bool> 165 PrintAfterISel("print-after-isel", cl::init(false), cl::Hidden, 166 cl::desc("Print machine instrs after ISel")); 167 168 static cl::opt<GlobalISelAbortMode> EnableGlobalISelAbort( 169 "global-isel-abort", cl::Hidden, 170 cl::desc("Enable abort calls when \"global\" instruction selection " 171 "fails to lower/select an instruction"), 172 cl::values( 173 clEnumValN(GlobalISelAbortMode::Disable, "0", "Disable the abort"), 174 clEnumValN(GlobalISelAbortMode::Enable, "1", "Enable the abort"), 175 clEnumValN(GlobalISelAbortMode::DisableWithDiag, "2", 176 "Disable the abort but emit a diagnostic on failure"))); 177 178 // Disable MIRProfileLoader before RegAlloc. This is for for debugging and 179 // tuning purpose. 180 static cl::opt<bool> DisableRAFSProfileLoader( 181 "disable-ra-fsprofile-loader", cl::init(false), cl::Hidden, 182 cl::desc("Disable MIRProfileLoader before RegAlloc")); 183 // Disable MIRProfileLoader before BloackPlacement. This is for for debugging 184 // and tuning purpose. 185 static cl::opt<bool> DisableLayoutFSProfileLoader( 186 "disable-layout-fsprofile-loader", cl::init(false), cl::Hidden, 187 cl::desc("Disable MIRProfileLoader before BlockPlacement")); 188 // Specify FSProfile file name. 189 static cl::opt<std::string> 190 FSProfileFile("fs-profile-file", cl::init(""), cl::value_desc("filename"), 191 cl::desc("Flow Sensitive profile file name."), cl::Hidden); 192 // Specify Remapping file for FSProfile. 193 static cl::opt<std::string> FSRemappingFile( 194 "fs-remapping-file", cl::init(""), cl::value_desc("filename"), 195 cl::desc("Flow Sensitive profile remapping file name."), cl::Hidden); 196 197 // Temporary option to allow experimenting with MachineScheduler as a post-RA 198 // scheduler. Targets can "properly" enable this with 199 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID). 200 // Targets can return true in targetSchedulesPostRAScheduling() and 201 // insert a PostRA scheduling pass wherever it wants. 202 static cl::opt<bool> MISchedPostRA( 203 "misched-postra", cl::Hidden, 204 cl::desc( 205 "Run MachineScheduler post regalloc (independent of preRA sched)")); 206 207 // Experimental option to run live interval analysis early. 208 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden, 209 cl::desc("Run live interval analysis earlier in the pipeline")); 210 211 static cl::opt<bool> DisableReplaceWithVecLib( 212 "disable-replace-with-vec-lib", cl::Hidden, 213 cl::desc("Disable replace with vector math call pass")); 214 215 /// Option names for limiting the codegen pipeline. 216 /// Those are used in error reporting and we didn't want 217 /// to duplicate their names all over the place. 218 static const char StartAfterOptName[] = "start-after"; 219 static const char StartBeforeOptName[] = "start-before"; 220 static const char StopAfterOptName[] = "stop-after"; 221 static const char StopBeforeOptName[] = "stop-before"; 222 223 static cl::opt<std::string> 224 StartAfterOpt(StringRef(StartAfterOptName), 225 cl::desc("Resume compilation after a specific pass"), 226 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 227 228 static cl::opt<std::string> 229 StartBeforeOpt(StringRef(StartBeforeOptName), 230 cl::desc("Resume compilation before a specific pass"), 231 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 232 233 static cl::opt<std::string> 234 StopAfterOpt(StringRef(StopAfterOptName), 235 cl::desc("Stop compilation after a specific pass"), 236 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 237 238 static cl::opt<std::string> 239 StopBeforeOpt(StringRef(StopBeforeOptName), 240 cl::desc("Stop compilation before a specific pass"), 241 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 242 243 /// Enable the machine function splitter pass. 244 static cl::opt<bool> EnableMachineFunctionSplitter( 245 "enable-split-machine-functions", cl::Hidden, 246 cl::desc("Split out cold blocks from machine functions based on profile " 247 "information.")); 248 249 /// Disable the expand reductions pass for testing. 250 static cl::opt<bool> DisableExpandReductions( 251 "disable-expand-reductions", cl::init(false), cl::Hidden, 252 cl::desc("Disable the expand reduction intrinsics pass from running")); 253 254 /// Disable the select optimization pass. 255 static cl::opt<bool> DisableSelectOptimize( 256 "disable-select-optimize", cl::init(true), cl::Hidden, 257 cl::desc("Disable the select-optimization pass from running")); 258 259 /// Enable garbage-collecting empty basic blocks. 260 static cl::opt<bool> 261 GCEmptyBlocks("gc-empty-basic-blocks", cl::init(false), cl::Hidden, 262 cl::desc("Enable garbage-collecting empty basic blocks")); 263 264 static cl::opt<bool> 265 SplitStaticData("split-static-data", cl::Hidden, cl::init(false), 266 cl::desc("Split static data sections into hot and cold " 267 "sections using profile information")); 268 269 /// Allow standard passes to be disabled by command line options. This supports 270 /// simple binary flags that either suppress the pass or do nothing. 271 /// i.e. -disable-mypass=false has no effect. 272 /// These should be converted to boolOrDefault in order to use applyOverride. 273 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID, 274 bool Override) { 275 if (Override) 276 return IdentifyingPassPtr(); 277 return PassID; 278 } 279 280 /// Allow standard passes to be disabled by the command line, regardless of who 281 /// is adding the pass. 282 /// 283 /// StandardID is the pass identified in the standard pass pipeline and provided 284 /// to addPass(). It may be a target-specific ID in the case that the target 285 /// directly adds its own pass, but in that case we harmlessly fall through. 286 /// 287 /// TargetID is the pass that the target has configured to override StandardID. 288 /// 289 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real 290 /// pass to run. This allows multiple options to control a single pass depending 291 /// on where in the pipeline that pass is added. 292 static IdentifyingPassPtr overridePass(AnalysisID StandardID, 293 IdentifyingPassPtr TargetID) { 294 if (StandardID == &PostRASchedulerID) 295 return applyDisable(TargetID, DisablePostRASched); 296 297 if (StandardID == &BranchFolderPassID) 298 return applyDisable(TargetID, DisableBranchFold); 299 300 if (StandardID == &TailDuplicateLegacyID) 301 return applyDisable(TargetID, DisableTailDuplicate); 302 303 if (StandardID == &EarlyTailDuplicateLegacyID) 304 return applyDisable(TargetID, DisableEarlyTailDup); 305 306 if (StandardID == &MachineBlockPlacementID) 307 return applyDisable(TargetID, DisableBlockPlacement); 308 309 if (StandardID == &StackSlotColoringID) 310 return applyDisable(TargetID, DisableSSC); 311 312 if (StandardID == &DeadMachineInstructionElimID) 313 return applyDisable(TargetID, DisableMachineDCE); 314 315 if (StandardID == &EarlyIfConverterLegacyID) 316 return applyDisable(TargetID, DisableEarlyIfConversion); 317 318 if (StandardID == &EarlyMachineLICMID) 319 return applyDisable(TargetID, DisableMachineLICM); 320 321 if (StandardID == &MachineCSELegacyID) 322 return applyDisable(TargetID, DisableMachineCSE); 323 324 if (StandardID == &MachineLICMID) 325 return applyDisable(TargetID, DisablePostRAMachineLICM); 326 327 if (StandardID == &MachineSinkingID) 328 return applyDisable(TargetID, DisableMachineSink); 329 330 if (StandardID == &PostRAMachineSinkingID) 331 return applyDisable(TargetID, DisablePostRAMachineSink); 332 333 if (StandardID == &MachineCopyPropagationID) 334 return applyDisable(TargetID, DisableCopyProp); 335 336 return TargetID; 337 } 338 339 // Find the FSProfile file name. The internal option takes the precedence 340 // before getting from TargetMachine. 341 static std::string getFSProfileFile(const TargetMachine *TM) { 342 if (!FSProfileFile.empty()) 343 return FSProfileFile.getValue(); 344 const std::optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 345 if (PGOOpt == std::nullopt || PGOOpt->Action != PGOOptions::SampleUse) 346 return std::string(); 347 return PGOOpt->ProfileFile; 348 } 349 350 // Find the Profile remapping file name. The internal option takes the 351 // precedence before getting from TargetMachine. 352 static std::string getFSRemappingFile(const TargetMachine *TM) { 353 if (!FSRemappingFile.empty()) 354 return FSRemappingFile.getValue(); 355 const std::optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 356 if (PGOOpt == std::nullopt || PGOOpt->Action != PGOOptions::SampleUse) 357 return std::string(); 358 return PGOOpt->ProfileRemappingFile; 359 } 360 361 //===---------------------------------------------------------------------===// 362 /// TargetPassConfig 363 //===---------------------------------------------------------------------===// 364 365 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig", 366 "Target Pass Configuration", false, false) 367 char TargetPassConfig::ID = 0; 368 369 namespace { 370 371 struct InsertedPass { 372 AnalysisID TargetPassID; 373 IdentifyingPassPtr InsertedPassID; 374 375 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID) 376 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID) {} 377 378 Pass *getInsertedPass() const { 379 assert(InsertedPassID.isValid() && "Illegal Pass ID!"); 380 if (InsertedPassID.isInstance()) 381 return InsertedPassID.getInstance(); 382 Pass *NP = Pass::createPass(InsertedPassID.getID()); 383 assert(NP && "Pass ID not registered"); 384 return NP; 385 } 386 }; 387 388 } // end anonymous namespace 389 390 namespace llvm { 391 392 extern cl::opt<bool> EnableFSDiscriminator; 393 394 class PassConfigImpl { 395 public: 396 // List of passes explicitly substituted by this target. Normally this is 397 // empty, but it is a convenient way to suppress or replace specific passes 398 // that are part of a standard pass pipeline without overridding the entire 399 // pipeline. This mechanism allows target options to inherit a standard pass's 400 // user interface. For example, a target may disable a standard pass by 401 // default by substituting a pass ID of zero, and the user may still enable 402 // that standard pass with an explicit command line option. 403 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses; 404 405 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass 406 /// is inserted after each instance of the first one. 407 SmallVector<InsertedPass, 4> InsertedPasses; 408 }; 409 410 } // end namespace llvm 411 412 // Out of line virtual method. 413 TargetPassConfig::~TargetPassConfig() { 414 delete Impl; 415 } 416 417 static const PassInfo *getPassInfo(StringRef PassName) { 418 if (PassName.empty()) 419 return nullptr; 420 421 const PassRegistry &PR = *PassRegistry::getPassRegistry(); 422 const PassInfo *PI = PR.getPassInfo(PassName); 423 if (!PI) 424 report_fatal_error(Twine('\"') + Twine(PassName) + 425 Twine("\" pass is not registered.")); 426 return PI; 427 } 428 429 static AnalysisID getPassIDFromName(StringRef PassName) { 430 const PassInfo *PI = getPassInfo(PassName); 431 return PI ? PI->getTypeInfo() : nullptr; 432 } 433 434 static std::pair<StringRef, unsigned> 435 getPassNameAndInstanceNum(StringRef PassName) { 436 StringRef Name, InstanceNumStr; 437 std::tie(Name, InstanceNumStr) = PassName.split(','); 438 439 unsigned InstanceNum = 0; 440 if (!InstanceNumStr.empty() && InstanceNumStr.getAsInteger(10, InstanceNum)) 441 report_fatal_error("invalid pass instance specifier " + PassName); 442 443 return std::make_pair(Name, InstanceNum); 444 } 445 446 void TargetPassConfig::setStartStopPasses() { 447 StringRef StartBeforeName; 448 std::tie(StartBeforeName, StartBeforeInstanceNum) = 449 getPassNameAndInstanceNum(StartBeforeOpt); 450 451 StringRef StartAfterName; 452 std::tie(StartAfterName, StartAfterInstanceNum) = 453 getPassNameAndInstanceNum(StartAfterOpt); 454 455 StringRef StopBeforeName; 456 std::tie(StopBeforeName, StopBeforeInstanceNum) 457 = getPassNameAndInstanceNum(StopBeforeOpt); 458 459 StringRef StopAfterName; 460 std::tie(StopAfterName, StopAfterInstanceNum) 461 = getPassNameAndInstanceNum(StopAfterOpt); 462 463 StartBefore = getPassIDFromName(StartBeforeName); 464 StartAfter = getPassIDFromName(StartAfterName); 465 StopBefore = getPassIDFromName(StopBeforeName); 466 StopAfter = getPassIDFromName(StopAfterName); 467 if (StartBefore && StartAfter) 468 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") + 469 Twine(StartAfterOptName) + Twine(" specified!")); 470 if (StopBefore && StopAfter) 471 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") + 472 Twine(StopAfterOptName) + Twine(" specified!")); 473 Started = (StartAfter == nullptr) && (StartBefore == nullptr); 474 } 475 476 CGPassBuilderOption llvm::getCGPassBuilderOption() { 477 CGPassBuilderOption Opt; 478 479 #define SET_OPTION(Option) \ 480 if (Option.getNumOccurrences()) \ 481 Opt.Option = Option; 482 483 SET_OPTION(EnableFastISelOption) 484 SET_OPTION(EnableGlobalISelAbort) 485 SET_OPTION(EnableGlobalISelOption) 486 SET_OPTION(EnableIPRA) 487 SET_OPTION(OptimizeRegAlloc) 488 SET_OPTION(VerifyMachineCode) 489 SET_OPTION(DisableAtExitBasedGlobalDtorLowering) 490 SET_OPTION(DisableExpandReductions) 491 SET_OPTION(PrintAfterISel) 492 SET_OPTION(FSProfileFile) 493 SET_OPTION(GCEmptyBlocks) 494 495 #define SET_BOOLEAN_OPTION(Option) Opt.Option = Option; 496 497 SET_BOOLEAN_OPTION(EarlyLiveIntervals) 498 SET_BOOLEAN_OPTION(EnableBlockPlacementStats) 499 SET_BOOLEAN_OPTION(EnableGlobalMergeFunc) 500 SET_BOOLEAN_OPTION(EnableImplicitNullChecks) 501 SET_BOOLEAN_OPTION(EnableMachineOutliner) 502 SET_BOOLEAN_OPTION(MISchedPostRA) 503 SET_BOOLEAN_OPTION(DisableMergeICmps) 504 SET_BOOLEAN_OPTION(DisableLSR) 505 SET_BOOLEAN_OPTION(DisableConstantHoisting) 506 SET_BOOLEAN_OPTION(DisableCGP) 507 SET_BOOLEAN_OPTION(DisablePartialLibcallInlining) 508 SET_BOOLEAN_OPTION(DisableSelectOptimize) 509 SET_BOOLEAN_OPTION(PrintISelInput) 510 SET_BOOLEAN_OPTION(DebugifyAndStripAll) 511 SET_BOOLEAN_OPTION(DebugifyCheckAndStripAll) 512 SET_BOOLEAN_OPTION(DisableRAFSProfileLoader) 513 SET_BOOLEAN_OPTION(DisableCFIFixup) 514 SET_BOOLEAN_OPTION(EnableMachineFunctionSplitter) 515 516 return Opt; 517 } 518 519 void llvm::registerCodeGenCallback(PassInstrumentationCallbacks &PIC, 520 TargetMachine &TM) { 521 522 // Register a callback for disabling passes. 523 PIC.registerShouldRunOptionalPassCallback([](StringRef P, Any) { 524 525 #define DISABLE_PASS(Option, Name) \ 526 if (Option && P.contains(#Name)) \ 527 return false; 528 DISABLE_PASS(DisableBlockPlacement, MachineBlockPlacementPass) 529 DISABLE_PASS(DisableBranchFold, BranchFolderPass) 530 DISABLE_PASS(DisableCopyProp, MachineCopyPropagationPass) 531 DISABLE_PASS(DisableEarlyIfConversion, EarlyIfConverterLegacyPass) 532 DISABLE_PASS(DisableEarlyTailDup, EarlyTailDuplicatePass) 533 DISABLE_PASS(DisableMachineCSE, MachineCSELegacyPass) 534 DISABLE_PASS(DisableMachineDCE, DeadMachineInstructionElimPass) 535 DISABLE_PASS(DisableMachineLICM, EarlyMachineLICMPass) 536 DISABLE_PASS(DisableMachineSink, MachineSinkingPass) 537 DISABLE_PASS(DisablePostRAMachineLICM, MachineLICMPass) 538 DISABLE_PASS(DisablePostRAMachineSink, PostRAMachineSinkingPass) 539 DISABLE_PASS(DisablePostRASched, PostRASchedulerPass) 540 DISABLE_PASS(DisableSSC, StackSlotColoringPass) 541 DISABLE_PASS(DisableTailDuplicate, TailDuplicatePass) 542 543 return true; 544 }); 545 } 546 547 Expected<TargetPassConfig::StartStopInfo> 548 TargetPassConfig::getStartStopInfo(PassInstrumentationCallbacks &PIC) { 549 auto [StartBefore, StartBeforeInstanceNum] = 550 getPassNameAndInstanceNum(StartBeforeOpt); 551 auto [StartAfter, StartAfterInstanceNum] = 552 getPassNameAndInstanceNum(StartAfterOpt); 553 auto [StopBefore, StopBeforeInstanceNum] = 554 getPassNameAndInstanceNum(StopBeforeOpt); 555 auto [StopAfter, StopAfterInstanceNum] = 556 getPassNameAndInstanceNum(StopAfterOpt); 557 558 if (!StartBefore.empty() && !StartAfter.empty()) 559 return make_error<StringError>( 560 Twine(StartBeforeOptName) + " and " + StartAfterOptName + " specified!", 561 std::make_error_code(std::errc::invalid_argument)); 562 if (!StopBefore.empty() && !StopAfter.empty()) 563 return make_error<StringError>( 564 Twine(StopBeforeOptName) + " and " + StopAfterOptName + " specified!", 565 std::make_error_code(std::errc::invalid_argument)); 566 567 StartStopInfo Result; 568 Result.StartPass = StartBefore.empty() ? StartAfter : StartBefore; 569 Result.StopPass = StopBefore.empty() ? StopAfter : StopBefore; 570 Result.StartInstanceNum = 571 StartBefore.empty() ? StartAfterInstanceNum : StartBeforeInstanceNum; 572 Result.StopInstanceNum = 573 StopBefore.empty() ? StopAfterInstanceNum : StopBeforeInstanceNum; 574 Result.StartAfter = !StartAfter.empty(); 575 Result.StopAfter = !StopAfter.empty(); 576 Result.StartInstanceNum += Result.StartInstanceNum == 0; 577 Result.StopInstanceNum += Result.StopInstanceNum == 0; 578 return Result; 579 } 580 581 // Out of line constructor provides default values for pass options and 582 // registers all common codegen passes. 583 TargetPassConfig::TargetPassConfig(TargetMachine &TM, PassManagerBase &PM) 584 : ImmutablePass(ID), PM(&PM), TM(&TM) { 585 Impl = new PassConfigImpl(); 586 587 // Register all target independent codegen passes to activate their PassIDs, 588 // including this pass itself. 589 initializeCodeGen(*PassRegistry::getPassRegistry()); 590 591 // Also register alias analysis passes required by codegen passes. 592 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry()); 593 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry()); 594 595 if (EnableIPRA.getNumOccurrences()) 596 TM.Options.EnableIPRA = EnableIPRA; 597 else { 598 // If not explicitly specified, use target default. 599 TM.Options.EnableIPRA |= TM.useIPRA(); 600 } 601 602 if (TM.Options.EnableIPRA) 603 setRequiresCodeGenSCCOrder(); 604 605 if (EnableGlobalISelAbort.getNumOccurrences()) 606 TM.Options.GlobalISelAbort = EnableGlobalISelAbort; 607 608 setStartStopPasses(); 609 } 610 611 CodeGenOptLevel TargetPassConfig::getOptLevel() const { 612 return TM->getOptLevel(); 613 } 614 615 /// Insert InsertedPassID pass after TargetPassID. 616 void TargetPassConfig::insertPass(AnalysisID TargetPassID, 617 IdentifyingPassPtr InsertedPassID) { 618 assert(((!InsertedPassID.isInstance() && 619 TargetPassID != InsertedPassID.getID()) || 620 (InsertedPassID.isInstance() && 621 TargetPassID != InsertedPassID.getInstance()->getPassID())) && 622 "Insert a pass after itself!"); 623 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID); 624 } 625 626 /// createPassConfig - Create a pass configuration object to be used by 627 /// addPassToEmitX methods for generating a pipeline of CodeGen passes. 628 /// 629 /// Targets may override this to extend TargetPassConfig. 630 TargetPassConfig * 631 CodeGenTargetMachineImpl::createPassConfig(PassManagerBase &PM) { 632 return new TargetPassConfig(*this, PM); 633 } 634 635 TargetPassConfig::TargetPassConfig() 636 : ImmutablePass(ID) { 637 report_fatal_error("Trying to construct TargetPassConfig without a target " 638 "machine. Scheduling a CodeGen pass without a target " 639 "triple set?"); 640 } 641 642 bool TargetPassConfig::willCompleteCodeGenPipeline() { 643 return StopBeforeOpt.empty() && StopAfterOpt.empty(); 644 } 645 646 bool TargetPassConfig::hasLimitedCodeGenPipeline() { 647 return !StartBeforeOpt.empty() || !StartAfterOpt.empty() || 648 !willCompleteCodeGenPipeline(); 649 } 650 651 std::string TargetPassConfig::getLimitedCodeGenPipelineReason() { 652 if (!hasLimitedCodeGenPipeline()) 653 return std::string(); 654 std::string Res; 655 static cl::opt<std::string> *PassNames[] = {&StartAfterOpt, &StartBeforeOpt, 656 &StopAfterOpt, &StopBeforeOpt}; 657 static const char *OptNames[] = {StartAfterOptName, StartBeforeOptName, 658 StopAfterOptName, StopBeforeOptName}; 659 bool IsFirst = true; 660 for (int Idx = 0; Idx < 4; ++Idx) 661 if (!PassNames[Idx]->empty()) { 662 if (!IsFirst) 663 Res += " and "; 664 IsFirst = false; 665 Res += OptNames[Idx]; 666 } 667 return Res; 668 } 669 670 // Helper to verify the analysis is really immutable. 671 void TargetPassConfig::setOpt(bool &Opt, bool Val) { 672 assert(!Initialized && "PassConfig is immutable"); 673 Opt = Val; 674 } 675 676 void TargetPassConfig::substitutePass(AnalysisID StandardID, 677 IdentifyingPassPtr TargetID) { 678 Impl->TargetPasses[StandardID] = TargetID; 679 } 680 681 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const { 682 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator 683 I = Impl->TargetPasses.find(ID); 684 if (I == Impl->TargetPasses.end()) 685 return ID; 686 return I->second; 687 } 688 689 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const { 690 IdentifyingPassPtr TargetID = getPassSubstitution(ID); 691 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID); 692 return !FinalPtr.isValid() || FinalPtr.isInstance() || 693 FinalPtr.getID() != ID; 694 } 695 696 /// Add a pass to the PassManager if that pass is supposed to be run. If the 697 /// Started/Stopped flags indicate either that the compilation should start at 698 /// a later pass or that it should stop after an earlier pass, then do not add 699 /// the pass. Finally, compare the current pass against the StartAfter 700 /// and StopAfter options and change the Started/Stopped flags accordingly. 701 void TargetPassConfig::addPass(Pass *P) { 702 assert(!Initialized && "PassConfig is immutable"); 703 704 // Cache the Pass ID here in case the pass manager finds this pass is 705 // redundant with ones already scheduled / available, and deletes it. 706 // Fundamentally, once we add the pass to the manager, we no longer own it 707 // and shouldn't reference it. 708 AnalysisID PassID = P->getPassID(); 709 710 if (StartBefore == PassID && StartBeforeCount++ == StartBeforeInstanceNum) 711 Started = true; 712 if (StopBefore == PassID && StopBeforeCount++ == StopBeforeInstanceNum) 713 Stopped = true; 714 if (Started && !Stopped) { 715 if (AddingMachinePasses) { 716 // Construct banner message before PM->add() as that may delete the pass. 717 std::string Banner = 718 std::string("After ") + std::string(P->getPassName()); 719 addMachinePrePasses(); 720 PM->add(P); 721 addMachinePostPasses(Banner); 722 } else { 723 PM->add(P); 724 } 725 726 // Add the passes after the pass P if there is any. 727 for (const auto &IP : Impl->InsertedPasses) 728 if (IP.TargetPassID == PassID) 729 addPass(IP.getInsertedPass()); 730 } else { 731 delete P; 732 } 733 734 if (StopAfter == PassID && StopAfterCount++ == StopAfterInstanceNum) 735 Stopped = true; 736 737 if (StartAfter == PassID && StartAfterCount++ == StartAfterInstanceNum) 738 Started = true; 739 if (Stopped && !Started) 740 report_fatal_error("Cannot stop compilation after pass that is not run"); 741 } 742 743 /// Add a CodeGen pass at this point in the pipeline after checking for target 744 /// and command line overrides. 745 /// 746 /// addPass cannot return a pointer to the pass instance because is internal the 747 /// PassManager and the instance we create here may already be freed. 748 AnalysisID TargetPassConfig::addPass(AnalysisID PassID) { 749 IdentifyingPassPtr TargetID = getPassSubstitution(PassID); 750 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID); 751 if (!FinalPtr.isValid()) 752 return nullptr; 753 754 Pass *P; 755 if (FinalPtr.isInstance()) 756 P = FinalPtr.getInstance(); 757 else { 758 P = Pass::createPass(FinalPtr.getID()); 759 if (!P) 760 llvm_unreachable("Pass ID not registered"); 761 } 762 AnalysisID FinalID = P->getPassID(); 763 addPass(P); // Ends the lifetime of P. 764 765 return FinalID; 766 } 767 768 void TargetPassConfig::printAndVerify(const std::string &Banner) { 769 addPrintPass(Banner); 770 addVerifyPass(Banner); 771 } 772 773 void TargetPassConfig::addPrintPass(const std::string &Banner) { 774 if (PrintAfterISel) 775 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner)); 776 } 777 778 void TargetPassConfig::addVerifyPass(const std::string &Banner) { 779 bool Verify = VerifyMachineCode == cl::BOU_TRUE; 780 #ifdef EXPENSIVE_CHECKS 781 if (VerifyMachineCode == cl::BOU_UNSET) 782 Verify = TM->isMachineVerifierClean(); 783 #endif 784 if (Verify) 785 PM->add(createMachineVerifierPass(Banner)); 786 } 787 788 void TargetPassConfig::addDebugifyPass() { 789 PM->add(createDebugifyMachineModulePass()); 790 } 791 792 void TargetPassConfig::addStripDebugPass() { 793 PM->add(createStripDebugMachineModulePass(/*OnlyDebugified=*/true)); 794 } 795 796 void TargetPassConfig::addCheckDebugPass() { 797 PM->add(createCheckDebugMachineModulePass()); 798 } 799 800 void TargetPassConfig::addMachinePrePasses(bool AllowDebugify) { 801 if (AllowDebugify && DebugifyIsSafe && 802 (DebugifyAndStripAll == cl::BOU_TRUE || 803 DebugifyCheckAndStripAll == cl::BOU_TRUE)) 804 addDebugifyPass(); 805 } 806 807 void TargetPassConfig::addMachinePostPasses(const std::string &Banner) { 808 if (DebugifyIsSafe) { 809 if (DebugifyCheckAndStripAll == cl::BOU_TRUE) { 810 addCheckDebugPass(); 811 addStripDebugPass(); 812 } else if (DebugifyAndStripAll == cl::BOU_TRUE) 813 addStripDebugPass(); 814 } 815 addVerifyPass(Banner); 816 } 817 818 /// Add common target configurable passes that perform LLVM IR to IR transforms 819 /// following machine independent optimization. 820 void TargetPassConfig::addIRPasses() { 821 // Before running any passes, run the verifier to determine if the input 822 // coming from the front-end and/or optimizer is valid. 823 if (!DisableVerify) 824 addPass(createVerifierPass()); 825 826 if (getOptLevel() != CodeGenOptLevel::None) { 827 // Basic AliasAnalysis support. 828 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 829 // BasicAliasAnalysis wins if they disagree. This is intended to help 830 // support "obvious" type-punning idioms. 831 addPass(createTypeBasedAAWrapperPass()); 832 addPass(createScopedNoAliasAAWrapperPass()); 833 addPass(createBasicAAWrapperPass()); 834 835 // Run loop strength reduction before anything else. 836 if (!DisableLSR) { 837 addPass(createCanonicalizeFreezeInLoopsPass()); 838 addPass(createLoopStrengthReducePass()); 839 if (EnableLoopTermFold) 840 addPass(createLoopTermFoldPass()); 841 } 842 843 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of 844 // loads and compares. ExpandMemCmpPass then tries to expand those calls 845 // into optimally-sized loads and compares. The transforms are enabled by a 846 // target lowering hook. 847 if (!DisableMergeICmps) 848 addPass(createMergeICmpsLegacyPass()); 849 addPass(createExpandMemCmpLegacyPass()); 850 } 851 852 // Run GC lowering passes for builtin collectors 853 // TODO: add a pass insertion point here 854 addPass(&GCLoweringID); 855 addPass(&ShadowStackGCLoweringID); 856 857 // For MachO, lower @llvm.global_dtors into @llvm.global_ctors with 858 // __cxa_atexit() calls to avoid emitting the deprecated __mod_term_func. 859 if (TM->getTargetTriple().isOSBinFormatMachO() && 860 !DisableAtExitBasedGlobalDtorLowering) 861 addPass(createLowerGlobalDtorsLegacyPass()); 862 863 // Make sure that no unreachable blocks are instruction selected. 864 addPass(createUnreachableBlockEliminationPass()); 865 866 // Prepare expensive constants for SelectionDAG. 867 if (getOptLevel() != CodeGenOptLevel::None && !DisableConstantHoisting) 868 addPass(createConstantHoistingPass()); 869 870 if (getOptLevel() != CodeGenOptLevel::None && !DisableReplaceWithVecLib) 871 addPass(createReplaceWithVeclibLegacyPass()); 872 873 if (getOptLevel() != CodeGenOptLevel::None && !DisablePartialLibcallInlining) 874 addPass(createPartiallyInlineLibCallsPass()); 875 876 // Instrument function entry after all inlining. 877 addPass(createPostInlineEntryExitInstrumenterPass()); 878 879 // Add scalarization of target's unsupported masked memory intrinsics pass. 880 // the unsupported intrinsic will be replaced with a chain of basic blocks, 881 // that stores/loads element one-by-one if the appropriate mask bit is set. 882 addPass(createScalarizeMaskedMemIntrinLegacyPass()); 883 884 // Expand reduction intrinsics into shuffle sequences if the target wants to. 885 // Allow disabling it for testing purposes. 886 if (!DisableExpandReductions) 887 addPass(createExpandReductionsPass()); 888 889 // Convert conditional moves to conditional jumps when profitable. 890 if (getOptLevel() != CodeGenOptLevel::None && !DisableSelectOptimize) 891 addPass(createSelectOptimizePass()); 892 893 if (EnableGlobalMergeFunc) 894 addPass(createGlobalMergeFuncPass()); 895 } 896 897 /// Turn exception handling constructs into something the code generators can 898 /// handle. 899 void TargetPassConfig::addPassesToHandleExceptions() { 900 const MCAsmInfo *MCAI = TM->getMCAsmInfo(); 901 assert(MCAI && "No MCAsmInfo"); 902 switch (MCAI->getExceptionHandlingType()) { 903 case ExceptionHandling::SjLj: 904 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both 905 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise, 906 // catch info can get misplaced when a selector ends up more than one block 907 // removed from the parent invoke(s). This could happen when a landing 908 // pad is shared by multiple invokes and is also a target of a normal 909 // edge from elsewhere. 910 addPass(createSjLjEHPreparePass(TM)); 911 [[fallthrough]]; 912 case ExceptionHandling::DwarfCFI: 913 case ExceptionHandling::ARM: 914 case ExceptionHandling::AIX: 915 case ExceptionHandling::ZOS: 916 addPass(createDwarfEHPass(getOptLevel())); 917 break; 918 case ExceptionHandling::WinEH: 919 // We support using both GCC-style and MSVC-style exceptions on Windows, so 920 // add both preparation passes. Each pass will only actually run if it 921 // recognizes the personality function. 922 addPass(createWinEHPass()); 923 addPass(createDwarfEHPass(getOptLevel())); 924 break; 925 case ExceptionHandling::Wasm: 926 // Wasm EH uses Windows EH instructions, but it does not need to demote PHIs 927 // on catchpads and cleanuppads because it does not outline them into 928 // funclets. Catchswitch blocks are not lowered in SelectionDAG, so we 929 // should remove PHIs there. 930 addPass(createWinEHPass(/*DemoteCatchSwitchPHIOnly=*/true)); 931 addPass(createWasmEHPass()); 932 break; 933 case ExceptionHandling::None: 934 addPass(createLowerInvokePass()); 935 936 // The lower invoke pass may create unreachable code. Remove it. 937 addPass(createUnreachableBlockEliminationPass()); 938 break; 939 } 940 } 941 942 /// Add pass to prepare the LLVM IR for code generation. This should be done 943 /// before exception handling preparation passes. 944 void TargetPassConfig::addCodeGenPrepare() { 945 if (getOptLevel() != CodeGenOptLevel::None && !DisableCGP) 946 addPass(createCodeGenPrepareLegacyPass()); 947 } 948 949 /// Add common passes that perform LLVM IR to IR transforms in preparation for 950 /// instruction selection. 951 void TargetPassConfig::addISelPrepare() { 952 addPreISel(); 953 954 // Force codegen to run according to the callgraph. 955 if (requiresCodeGenSCCOrder()) 956 addPass(new DummyCGSCCPass); 957 958 if (getOptLevel() != CodeGenOptLevel::None) 959 addPass(createObjCARCContractPass()); 960 961 addPass(createCallBrPass()); 962 963 // Add both the safe stack and the stack protection passes: each of them will 964 // only protect functions that have corresponding attributes. 965 addPass(createSafeStackPass()); 966 addPass(createStackProtectorPass()); 967 968 if (PrintISelInput) 969 addPass(createPrintFunctionPass( 970 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n")); 971 972 // All passes which modify the LLVM IR are now complete; run the verifier 973 // to ensure that the IR is valid. 974 if (!DisableVerify) 975 addPass(createVerifierPass()); 976 } 977 978 bool TargetPassConfig::addCoreISelPasses() { 979 // Enable FastISel with -fast-isel, but allow that to be overridden. 980 TM->setO0WantsFastISel(EnableFastISelOption != cl::BOU_FALSE); 981 982 // Determine an instruction selector. 983 enum class SelectorType { SelectionDAG, FastISel, GlobalISel }; 984 SelectorType Selector; 985 986 if (EnableFastISelOption == cl::BOU_TRUE) 987 Selector = SelectorType::FastISel; 988 else if (EnableGlobalISelOption == cl::BOU_TRUE || 989 (TM->Options.EnableGlobalISel && 990 EnableGlobalISelOption != cl::BOU_FALSE)) 991 Selector = SelectorType::GlobalISel; 992 else if (TM->getOptLevel() == CodeGenOptLevel::None && 993 TM->getO0WantsFastISel()) 994 Selector = SelectorType::FastISel; 995 else 996 Selector = SelectorType::SelectionDAG; 997 998 // Set consistently TM->Options.EnableFastISel and EnableGlobalISel. 999 if (Selector == SelectorType::FastISel) { 1000 TM->setFastISel(true); 1001 TM->setGlobalISel(false); 1002 } else if (Selector == SelectorType::GlobalISel) { 1003 TM->setFastISel(false); 1004 TM->setGlobalISel(true); 1005 } 1006 1007 // FIXME: Injecting into the DAGISel pipeline seems to cause issues with 1008 // analyses needing to be re-run. This can result in being unable to 1009 // schedule passes (particularly with 'Function Alias Analysis 1010 // Results'). It's not entirely clear why but AFAICT this seems to be 1011 // due to one FunctionPassManager not being able to use analyses from a 1012 // previous one. As we're injecting a ModulePass we break the usual 1013 // pass manager into two. GlobalISel with the fallback path disabled 1014 // and -run-pass seem to be unaffected. The majority of GlobalISel 1015 // testing uses -run-pass so this probably isn't too bad. 1016 SaveAndRestore SavedDebugifyIsSafe(DebugifyIsSafe); 1017 if (Selector != SelectorType::GlobalISel || !isGlobalISelAbortEnabled()) 1018 DebugifyIsSafe = false; 1019 1020 // Add instruction selector passes for global isel if enabled. 1021 if (Selector == SelectorType::GlobalISel) { 1022 SaveAndRestore SavedAddingMachinePasses(AddingMachinePasses, true); 1023 if (addIRTranslator()) 1024 return true; 1025 1026 addPreLegalizeMachineIR(); 1027 1028 if (addLegalizeMachineIR()) 1029 return true; 1030 1031 // Before running the register bank selector, ask the target if it 1032 // wants to run some passes. 1033 addPreRegBankSelect(); 1034 1035 if (addRegBankSelect()) 1036 return true; 1037 1038 addPreGlobalInstructionSelect(); 1039 1040 if (addGlobalInstructionSelect()) 1041 return true; 1042 } 1043 1044 // Pass to reset the MachineFunction if the ISel failed. Outside of the above 1045 // if so that the verifier is not added to it. 1046 if (Selector == SelectorType::GlobalISel) 1047 addPass(createResetMachineFunctionPass( 1048 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled())); 1049 1050 // Run the SDAG InstSelector, providing a fallback path when we do not want to 1051 // abort on not-yet-supported input. 1052 if (Selector != SelectorType::GlobalISel || !isGlobalISelAbortEnabled()) 1053 if (addInstSelector()) 1054 return true; 1055 1056 // Expand pseudo-instructions emitted by ISel. Don't run the verifier before 1057 // FinalizeISel. 1058 addPass(&FinalizeISelID); 1059 1060 // Print the instruction selected machine code... 1061 printAndVerify("After Instruction Selection"); 1062 1063 return false; 1064 } 1065 1066 bool TargetPassConfig::addISelPasses() { 1067 if (TM->useEmulatedTLS()) 1068 addPass(createLowerEmuTLSPass()); 1069 1070 PM->add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis())); 1071 addPass(createPreISelIntrinsicLoweringPass()); 1072 addPass(createExpandLargeDivRemPass()); 1073 addPass(createExpandLargeFpConvertPass()); 1074 addIRPasses(); 1075 addCodeGenPrepare(); 1076 addPassesToHandleExceptions(); 1077 addISelPrepare(); 1078 1079 return addCoreISelPasses(); 1080 } 1081 1082 /// -regalloc=... command line option. 1083 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } 1084 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false, 1085 RegisterPassParser<RegisterRegAlloc>> 1086 RegAlloc("regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator), 1087 cl::desc("Register allocator to use")); 1088 1089 /// Add the complete set of target-independent postISel code generator passes. 1090 /// 1091 /// This can be read as the standard order of major LLVM CodeGen stages. Stages 1092 /// with nontrivial configuration or multiple passes are broken out below in 1093 /// add%Stage routines. 1094 /// 1095 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The 1096 /// addPre/Post methods with empty header implementations allow injecting 1097 /// target-specific fixups just before or after major stages. Additionally, 1098 /// targets have the flexibility to change pass order within a stage by 1099 /// overriding default implementation of add%Stage routines below. Each 1100 /// technique has maintainability tradeoffs because alternate pass orders are 1101 /// not well supported. addPre/Post works better if the target pass is easily 1102 /// tied to a common pass. But if it has subtle dependencies on multiple passes, 1103 /// the target should override the stage instead. 1104 /// 1105 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection 1106 /// before/after any target-independent pass. But it's currently overkill. 1107 void TargetPassConfig::addMachinePasses() { 1108 AddingMachinePasses = true; 1109 1110 // Add passes that optimize machine instructions in SSA form. 1111 if (getOptLevel() != CodeGenOptLevel::None) { 1112 addMachineSSAOptimization(); 1113 } else { 1114 // If the target requests it, assign local variables to stack slots relative 1115 // to one another and simplify frame index references where possible. 1116 addPass(&LocalStackSlotAllocationID); 1117 } 1118 1119 if (TM->Options.EnableIPRA) 1120 addPass(createRegUsageInfoPropPass()); 1121 1122 // Run pre-ra passes. 1123 addPreRegAlloc(); 1124 1125 // Debugifying the register allocator passes seems to provoke some 1126 // non-determinism that affects CodeGen and there doesn't seem to be a point 1127 // where it becomes safe again so stop debugifying here. 1128 DebugifyIsSafe = false; 1129 1130 // Add a FSDiscriminator pass right before RA, so that we could get 1131 // more precise SampleFDO profile for RA. 1132 if (EnableFSDiscriminator) { 1133 addPass(createMIRAddFSDiscriminatorsPass( 1134 sampleprof::FSDiscriminatorPass::Pass1)); 1135 const std::string ProfileFile = getFSProfileFile(TM); 1136 if (!ProfileFile.empty() && !DisableRAFSProfileLoader) 1137 addPass(createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1138 sampleprof::FSDiscriminatorPass::Pass1, 1139 nullptr)); 1140 } 1141 1142 // Run register allocation and passes that are tightly coupled with it, 1143 // including phi elimination and scheduling. 1144 if (getOptimizeRegAlloc()) 1145 addOptimizedRegAlloc(); 1146 else 1147 addFastRegAlloc(); 1148 1149 // Run post-ra passes. 1150 addPostRegAlloc(); 1151 1152 addPass(&RemoveRedundantDebugValuesID); 1153 1154 addPass(&FixupStatepointCallerSavedID); 1155 1156 // Insert prolog/epilog code. Eliminate abstract frame index references... 1157 if (getOptLevel() != CodeGenOptLevel::None) { 1158 addPass(&PostRAMachineSinkingID); 1159 addPass(&ShrinkWrapID); 1160 } 1161 1162 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only 1163 // do so if it hasn't been disabled, substituted, or overridden. 1164 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID)) 1165 addPass(createPrologEpilogInserterPass()); 1166 1167 /// Add passes that optimize machine instructions after register allocation. 1168 if (getOptLevel() != CodeGenOptLevel::None) 1169 addMachineLateOptimization(); 1170 1171 // Expand pseudo instructions before second scheduling pass. 1172 addPass(&ExpandPostRAPseudosID); 1173 1174 // Run pre-sched2 passes. 1175 addPreSched2(); 1176 1177 if (EnableImplicitNullChecks) 1178 addPass(&ImplicitNullChecksID); 1179 1180 // Second pass scheduler. 1181 // Let Target optionally insert this pass by itself at some other 1182 // point. 1183 if (getOptLevel() != CodeGenOptLevel::None && 1184 !TM->targetSchedulesPostRAScheduling()) { 1185 if (MISchedPostRA) 1186 addPass(&PostMachineSchedulerID); 1187 else 1188 addPass(&PostRASchedulerID); 1189 } 1190 1191 // GC 1192 addGCPasses(); 1193 1194 // Basic block placement. 1195 if (getOptLevel() != CodeGenOptLevel::None) 1196 addBlockPlacement(); 1197 1198 // Insert before XRay Instrumentation. 1199 addPass(&FEntryInserterID); 1200 1201 addPass(&XRayInstrumentationID); 1202 addPass(&PatchableFunctionID); 1203 1204 addPreEmitPass(); 1205 1206 if (TM->Options.EnableIPRA) 1207 // Collect register usage information and produce a register mask of 1208 // clobbered registers, to be used to optimize call sites. 1209 addPass(createRegUsageInfoCollector()); 1210 1211 // FIXME: Some backends are incompatible with running the verifier after 1212 // addPreEmitPass. Maybe only pass "false" here for those targets? 1213 addPass(&FuncletLayoutID); 1214 1215 addPass(&RemoveLoadsIntoFakeUsesID); 1216 addPass(&StackMapLivenessID); 1217 addPass(&LiveDebugValuesID); 1218 addPass(&MachineSanitizerBinaryMetadataID); 1219 1220 if (TM->Options.EnableMachineOutliner && 1221 getOptLevel() != CodeGenOptLevel::None && 1222 EnableMachineOutliner != RunOutliner::NeverOutline) { 1223 bool RunOnAllFunctions = 1224 (EnableMachineOutliner == RunOutliner::AlwaysOutline); 1225 bool AddOutliner = 1226 RunOnAllFunctions || TM->Options.SupportsDefaultOutlining; 1227 if (AddOutliner) 1228 addPass(createMachineOutlinerPass(RunOnAllFunctions)); 1229 } 1230 1231 if (GCEmptyBlocks) 1232 addPass(llvm::createGCEmptyBasicBlocksPass()); 1233 1234 if (EnableFSDiscriminator) 1235 addPass(createMIRAddFSDiscriminatorsPass( 1236 sampleprof::FSDiscriminatorPass::PassLast)); 1237 1238 // Machine function splitter uses the basic block sections feature. 1239 // When used along with `-basic-block-sections=`, the basic-block-sections 1240 // feature takes precedence. This means functions eligible for 1241 // basic-block-sections optimizations (`=all`, or `=list=` with function 1242 // included in the list profile) will get that optimization instead. 1243 if (TM->Options.EnableMachineFunctionSplitter || 1244 EnableMachineFunctionSplitter) { 1245 const std::string ProfileFile = getFSProfileFile(TM); 1246 if (!ProfileFile.empty()) { 1247 if (EnableFSDiscriminator) { 1248 addPass(createMIRProfileLoaderPass( 1249 ProfileFile, getFSRemappingFile(TM), 1250 sampleprof::FSDiscriminatorPass::PassLast, nullptr)); 1251 } else { 1252 // Sample profile is given, but FSDiscriminator is not 1253 // enabled, this may result in performance regression. 1254 WithColor::warning() 1255 << "Using AutoFDO without FSDiscriminator for MFS may regress " 1256 "performance.\n"; 1257 } 1258 } 1259 addPass(createMachineFunctionSplitterPass()); 1260 if (SplitStaticData || TM->Options.EnableStaticDataPartitioning) 1261 addPass(createStaticDataSplitterPass()); 1262 } 1263 // We run the BasicBlockSections pass if either we need BB sections or BB 1264 // address map (or both). 1265 if (TM->getBBSectionsType() != llvm::BasicBlockSection::None || 1266 TM->Options.BBAddrMap) { 1267 if (TM->getBBSectionsType() == llvm::BasicBlockSection::List) { 1268 addPass(llvm::createBasicBlockSectionsProfileReaderWrapperPass( 1269 TM->getBBSectionsFuncListBuf())); 1270 addPass(llvm::createBasicBlockPathCloningPass()); 1271 } 1272 addPass(llvm::createBasicBlockSectionsPass()); 1273 } 1274 1275 addPostBBSections(); 1276 1277 if (!DisableCFIFixup && TM->Options.EnableCFIFixup) 1278 addPass(createCFIFixup()); 1279 1280 PM->add(createStackFrameLayoutAnalysisPass()); 1281 1282 // Add passes that directly emit MI after all other MI passes. 1283 addPreEmitPass2(); 1284 1285 AddingMachinePasses = false; 1286 } 1287 1288 /// Add passes that optimize machine instructions in SSA form. 1289 void TargetPassConfig::addMachineSSAOptimization() { 1290 // Pre-ra tail duplication. 1291 addPass(&EarlyTailDuplicateLegacyID); 1292 1293 // Optimize PHIs before DCE: removing dead PHI cycles may make more 1294 // instructions dead. 1295 addPass(&OptimizePHIsLegacyID); 1296 1297 // This pass merges large allocas. StackSlotColoring is a different pass 1298 // which merges spill slots. 1299 addPass(&StackColoringLegacyID); 1300 1301 // If the target requests it, assign local variables to stack slots relative 1302 // to one another and simplify frame index references where possible. 1303 addPass(&LocalStackSlotAllocationID); 1304 1305 // With optimization, dead code should already be eliminated. However 1306 // there is one known exception: lowered code for arguments that are only 1307 // used by tail calls, where the tail calls reuse the incoming stack 1308 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll). 1309 addPass(&DeadMachineInstructionElimID); 1310 1311 // Allow targets to insert passes that improve instruction level parallelism, 1312 // like if-conversion. Such passes will typically need dominator trees and 1313 // loop info, just like LICM and CSE below. 1314 addILPOpts(); 1315 1316 addPass(&EarlyMachineLICMID); 1317 addPass(&MachineCSELegacyID); 1318 1319 addPass(&MachineSinkingID); 1320 1321 addPass(&PeepholeOptimizerLegacyID); 1322 // Clean-up the dead code that may have been generated by peephole 1323 // rewriting. 1324 addPass(&DeadMachineInstructionElimID); 1325 } 1326 1327 //===---------------------------------------------------------------------===// 1328 /// Register Allocation Pass Configuration 1329 //===---------------------------------------------------------------------===// 1330 1331 bool TargetPassConfig::getOptimizeRegAlloc() const { 1332 switch (OptimizeRegAlloc) { 1333 case cl::BOU_UNSET: 1334 return getOptLevel() != CodeGenOptLevel::None; 1335 case cl::BOU_TRUE: return true; 1336 case cl::BOU_FALSE: return false; 1337 } 1338 llvm_unreachable("Invalid optimize-regalloc state"); 1339 } 1340 1341 /// A dummy default pass factory indicates whether the register allocator is 1342 /// overridden on the command line. 1343 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag; 1344 1345 static RegisterRegAlloc 1346 defaultRegAlloc("default", 1347 "pick register allocator based on -O option", 1348 useDefaultRegisterAllocator); 1349 1350 static void initializeDefaultRegisterAllocatorOnce() { 1351 if (!RegisterRegAlloc::getDefault()) 1352 RegisterRegAlloc::setDefault(RegAlloc); 1353 } 1354 1355 /// Instantiate the default register allocator pass for this target for either 1356 /// the optimized or unoptimized allocation path. This will be added to the pass 1357 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc 1358 /// in the optimized case. 1359 /// 1360 /// A target that uses the standard regalloc pass order for fast or optimized 1361 /// allocation may still override this for per-target regalloc 1362 /// selection. But -regalloc=... always takes precedence. 1363 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) { 1364 if (Optimized) 1365 return createGreedyRegisterAllocator(); 1366 else 1367 return createFastRegisterAllocator(); 1368 } 1369 1370 /// Find and instantiate the register allocation pass requested by this target 1371 /// at the current optimization level. Different register allocators are 1372 /// defined as separate passes because they may require different analysis. 1373 /// 1374 /// This helper ensures that the regalloc= option is always available, 1375 /// even for targets that override the default allocator. 1376 /// 1377 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs, 1378 /// this can be folded into addPass. 1379 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) { 1380 // Initialize the global default. 1381 llvm::call_once(InitializeDefaultRegisterAllocatorFlag, 1382 initializeDefaultRegisterAllocatorOnce); 1383 1384 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault(); 1385 if (Ctor != useDefaultRegisterAllocator) 1386 return Ctor(); 1387 1388 // With no -regalloc= override, ask the target for a regalloc pass. 1389 return createTargetRegisterAllocator(Optimized); 1390 } 1391 1392 bool TargetPassConfig::isCustomizedRegAlloc() { 1393 return RegAlloc != 1394 (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator; 1395 } 1396 1397 bool TargetPassConfig::addRegAssignAndRewriteFast() { 1398 if (RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator && 1399 RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&createFastRegisterAllocator) 1400 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc."); 1401 1402 addPass(createRegAllocPass(false)); 1403 1404 // Allow targets to change the register assignments after 1405 // fast register allocation. 1406 addPostFastRegAllocRewrite(); 1407 return true; 1408 } 1409 1410 bool TargetPassConfig::addRegAssignAndRewriteOptimized() { 1411 // Add the selected register allocation pass. 1412 addPass(createRegAllocPass(true)); 1413 1414 // Allow targets to change the register assignments before rewriting. 1415 addPreRewrite(); 1416 1417 // Finally rewrite virtual registers. 1418 addPass(&VirtRegRewriterID); 1419 1420 // Regalloc scoring for ML-driven eviction - noop except when learning a new 1421 // eviction policy. 1422 addPass(createRegAllocScoringPass()); 1423 return true; 1424 } 1425 1426 /// Return true if the default global register allocator is in use and 1427 /// has not be overriden on the command line with '-regalloc=...' 1428 bool TargetPassConfig::usingDefaultRegAlloc() const { 1429 return RegAlloc.getNumOccurrences() == 0; 1430 } 1431 1432 /// Add the minimum set of target-independent passes that are required for 1433 /// register allocation. No coalescing or scheduling. 1434 void TargetPassConfig::addFastRegAlloc() { 1435 addPass(&PHIEliminationID); 1436 addPass(&TwoAddressInstructionPassID); 1437 1438 addRegAssignAndRewriteFast(); 1439 } 1440 1441 /// Add standard target-independent passes that are tightly coupled with 1442 /// optimized register allocation, including coalescing, machine instruction 1443 /// scheduling, and register allocation itself. 1444 void TargetPassConfig::addOptimizedRegAlloc() { 1445 addPass(&DetectDeadLanesID); 1446 1447 addPass(&InitUndefID); 1448 1449 addPass(&ProcessImplicitDefsID); 1450 1451 // LiveVariables currently requires pure SSA form. 1452 // 1453 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags, 1454 // LiveVariables can be removed completely, and LiveIntervals can be directly 1455 // computed. (We still either need to regenerate kill flags after regalloc, or 1456 // preferably fix the scavenger to not depend on them). 1457 // FIXME: UnreachableMachineBlockElim is a dependant pass of LiveVariables. 1458 // When LiveVariables is removed this has to be removed/moved either. 1459 // Explicit addition of UnreachableMachineBlockElim allows stopping before or 1460 // after it with -stop-before/-stop-after. 1461 addPass(&UnreachableMachineBlockElimID); 1462 addPass(&LiveVariablesID); 1463 1464 // Edge splitting is smarter with machine loop info. 1465 addPass(&MachineLoopInfoID); 1466 addPass(&PHIEliminationID); 1467 1468 // Eventually, we want to run LiveIntervals before PHI elimination. 1469 if (EarlyLiveIntervals) 1470 addPass(&LiveIntervalsID); 1471 1472 addPass(&TwoAddressInstructionPassID); 1473 addPass(&RegisterCoalescerID); 1474 1475 // The machine scheduler may accidentally create disconnected components 1476 // when moving subregister definitions around, avoid this by splitting them to 1477 // separate vregs before. Splitting can also improve reg. allocation quality. 1478 addPass(&RenameIndependentSubregsID); 1479 1480 // PreRA instruction scheduling. 1481 addPass(&MachineSchedulerID); 1482 1483 if (addRegAssignAndRewriteOptimized()) { 1484 // Perform stack slot coloring and post-ra machine LICM. 1485 addPass(&StackSlotColoringID); 1486 1487 // Allow targets to expand pseudo instructions depending on the choice of 1488 // registers before MachineCopyPropagation. 1489 addPostRewrite(); 1490 1491 // Copy propagate to forward register uses and try to eliminate COPYs that 1492 // were not coalesced. 1493 addPass(&MachineCopyPropagationID); 1494 1495 // Run post-ra machine LICM to hoist reloads / remats. 1496 // 1497 // FIXME: can this move into MachineLateOptimization? 1498 addPass(&MachineLICMID); 1499 } 1500 } 1501 1502 //===---------------------------------------------------------------------===// 1503 /// Post RegAlloc Pass Configuration 1504 //===---------------------------------------------------------------------===// 1505 1506 /// Add passes that optimize machine instructions after register allocation. 1507 void TargetPassConfig::addMachineLateOptimization() { 1508 // Cleanup of redundant immediate/address loads. 1509 addPass(&MachineLateInstrsCleanupID); 1510 1511 // Branch folding must be run after regalloc and prolog/epilog insertion. 1512 addPass(&BranchFolderPassID); 1513 1514 // Tail duplication. 1515 // Note that duplicating tail just increases code size and degrades 1516 // performance for targets that require Structured Control Flow. 1517 // In addition it can also make CFG irreducible. Thus we disable it. 1518 if (!TM->requiresStructuredCFG()) 1519 addPass(&TailDuplicateLegacyID); 1520 1521 // Copy propagation. 1522 addPass(&MachineCopyPropagationID); 1523 } 1524 1525 /// Add standard GC passes. 1526 bool TargetPassConfig::addGCPasses() { 1527 addPass(&GCMachineCodeAnalysisID); 1528 return true; 1529 } 1530 1531 /// Add standard basic block placement passes. 1532 void TargetPassConfig::addBlockPlacement() { 1533 if (EnableFSDiscriminator) { 1534 addPass(createMIRAddFSDiscriminatorsPass( 1535 sampleprof::FSDiscriminatorPass::Pass2)); 1536 const std::string ProfileFile = getFSProfileFile(TM); 1537 if (!ProfileFile.empty() && !DisableLayoutFSProfileLoader) 1538 addPass(createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1539 sampleprof::FSDiscriminatorPass::Pass2, 1540 nullptr)); 1541 } 1542 if (addPass(&MachineBlockPlacementID)) { 1543 // Run a separate pass to collect block placement statistics. 1544 if (EnableBlockPlacementStats) 1545 addPass(&MachineBlockPlacementStatsID); 1546 } 1547 } 1548 1549 //===---------------------------------------------------------------------===// 1550 /// GlobalISel Configuration 1551 //===---------------------------------------------------------------------===// 1552 bool TargetPassConfig::isGlobalISelAbortEnabled() const { 1553 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::Enable; 1554 } 1555 1556 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const { 1557 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::DisableWithDiag; 1558 } 1559 1560 bool TargetPassConfig::isGISelCSEEnabled() const { 1561 return true; 1562 } 1563 1564 std::unique_ptr<CSEConfigBase> TargetPassConfig::getCSEConfig() const { 1565 return std::make_unique<CSEConfigBase>(); 1566 } 1567