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