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/CFLAndersAliasAnalysis.h" 20 #include "llvm/Analysis/CFLSteensAliasAnalysis.h" 21 #include "llvm/Analysis/CallGraphSCCPass.h" 22 #include "llvm/Analysis/ScopedNoAliasAA.h" 23 #include "llvm/Analysis/TargetTransformInfo.h" 24 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 25 #include "llvm/CodeGen/BasicBlockSectionsProfileReader.h" 26 #include "llvm/CodeGen/CSEConfigBase.h" 27 #include "llvm/CodeGen/MachineFunctionPass.h" 28 #include "llvm/CodeGen/MachinePassRegistry.h" 29 #include "llvm/CodeGen/Passes.h" 30 #include "llvm/CodeGen/RegAllocRegistry.h" 31 #include "llvm/IR/IRPrintingPasses.h" 32 #include "llvm/IR/LegacyPassManager.h" 33 #include "llvm/IR/PassInstrumentation.h" 34 #include "llvm/IR/Verifier.h" 35 #include "llvm/InitializePasses.h" 36 #include "llvm/MC/MCAsmInfo.h" 37 #include "llvm/MC/MCTargetOptions.h" 38 #include "llvm/Pass.h" 39 #include "llvm/Support/CodeGen.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/Support/Compiler.h" 42 #include "llvm/Support/Debug.h" 43 #include "llvm/Support/Discriminator.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/SaveAndRestore.h" 46 #include "llvm/Support/Threading.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 <string> 53 54 using namespace llvm; 55 56 static cl::opt<bool> 57 EnableIPRA("enable-ipra", cl::init(false), cl::Hidden, 58 cl::desc("Enable interprocedural register allocation " 59 "to reduce load/store at procedure calls.")); 60 static cl::opt<bool> DisablePostRASched("disable-post-ra", cl::Hidden, 61 cl::desc("Disable Post Regalloc Scheduler")); 62 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden, 63 cl::desc("Disable branch folding")); 64 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden, 65 cl::desc("Disable tail duplication")); 66 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden, 67 cl::desc("Disable pre-register allocation tail duplication")); 68 static cl::opt<bool> DisableBlockPlacement("disable-block-placement", 69 cl::Hidden, cl::desc("Disable probability-driven block placement")); 70 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats", 71 cl::Hidden, cl::desc("Collect probability-driven block placement stats")); 72 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden, 73 cl::desc("Disable Stack Slot Coloring")); 74 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden, 75 cl::desc("Disable Machine Dead Code Elimination")); 76 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden, 77 cl::desc("Disable Early If-conversion")); 78 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden, 79 cl::desc("Disable Machine LICM")); 80 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden, 81 cl::desc("Disable Machine Common Subexpression Elimination")); 82 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc( 83 "optimize-regalloc", cl::Hidden, 84 cl::desc("Enable optimized register allocation compilation path.")); 85 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm", 86 cl::Hidden, 87 cl::desc("Disable Machine LICM")); 88 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden, 89 cl::desc("Disable Machine Sinking")); 90 static cl::opt<bool> DisablePostRAMachineSink("disable-postra-machine-sink", 91 cl::Hidden, 92 cl::desc("Disable PostRA Machine Sinking")); 93 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden, 94 cl::desc("Disable Loop Strength Reduction Pass")); 95 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting", 96 cl::Hidden, cl::desc("Disable ConstantHoisting")); 97 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden, 98 cl::desc("Disable Codegen Prepare")); 99 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden, 100 cl::desc("Disable Copy Propagation pass")); 101 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining", 102 cl::Hidden, cl::desc("Disable Partial Libcall Inlining")); 103 static cl::opt<bool> EnableImplicitNullChecks( 104 "enable-implicit-null-checks", 105 cl::desc("Fold null checks into faulting memory operations"), 106 cl::init(false), cl::Hidden); 107 static cl::opt<bool> DisableMergeICmps("disable-mergeicmps", 108 cl::desc("Disable MergeICmps Pass"), 109 cl::init(false), cl::Hidden); 110 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden, 111 cl::desc("Print LLVM IR produced by the loop-reduce pass")); 112 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden, 113 cl::desc("Print LLVM IR input to isel pass")); 114 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden, 115 cl::desc("Dump garbage collector data")); 116 static cl::opt<cl::boolOrDefault> 117 VerifyMachineCode("verify-machineinstrs", cl::Hidden, 118 cl::desc("Verify generated machine code")); 119 static cl::opt<cl::boolOrDefault> 120 DebugifyAndStripAll("debugify-and-strip-all-safe", cl::Hidden, 121 cl::desc("Debugify MIR before and Strip debug after " 122 "each pass except those known to be unsafe " 123 "when debug info is present")); 124 static cl::opt<cl::boolOrDefault> DebugifyCheckAndStripAll( 125 "debugify-check-and-strip-all-safe", cl::Hidden, 126 cl::desc( 127 "Debugify MIR before, by checking and stripping the debug info after, " 128 "each pass except those known to be unsafe when debug info is " 129 "present")); 130 // Enable or disable the MachineOutliner. 131 static cl::opt<RunOutliner> EnableMachineOutliner( 132 "enable-machine-outliner", cl::desc("Enable the machine outliner"), 133 cl::Hidden, cl::ValueOptional, cl::init(RunOutliner::TargetDefault), 134 cl::values(clEnumValN(RunOutliner::AlwaysOutline, "always", 135 "Run on all functions guaranteed to be beneficial"), 136 clEnumValN(RunOutliner::NeverOutline, "never", 137 "Disable all outlining"), 138 // Sentinel value for unspecified option. 139 clEnumValN(RunOutliner::AlwaysOutline, "", ""))); 140 // Disable the pass to fix unwind information. Whether the pass is included in 141 // the pipeline is controlled via the target options, this option serves as 142 // manual override. 143 static cl::opt<bool> DisableCFIFixup("disable-cfi-fixup", cl::Hidden, 144 cl::desc("Disable the CFI fixup pass")); 145 // Enable or disable FastISel. Both options are needed, because 146 // FastISel is enabled by default with -fast, and we wish to be 147 // able to enable or disable fast-isel independently from -O0. 148 static cl::opt<cl::boolOrDefault> 149 EnableFastISelOption("fast-isel", cl::Hidden, 150 cl::desc("Enable the \"fast\" instruction selector")); 151 152 static cl::opt<cl::boolOrDefault> EnableGlobalISelOption( 153 "global-isel", cl::Hidden, 154 cl::desc("Enable the \"global\" instruction selector")); 155 156 // FIXME: remove this after switching to NPM or GlobalISel, whichever gets there 157 // first... 158 static cl::opt<bool> 159 PrintAfterISel("print-after-isel", cl::init(false), cl::Hidden, 160 cl::desc("Print machine instrs after ISel")); 161 162 static cl::opt<GlobalISelAbortMode> EnableGlobalISelAbort( 163 "global-isel-abort", cl::Hidden, 164 cl::desc("Enable abort calls when \"global\" instruction selection " 165 "fails to lower/select an instruction"), 166 cl::values( 167 clEnumValN(GlobalISelAbortMode::Disable, "0", "Disable the abort"), 168 clEnumValN(GlobalISelAbortMode::Enable, "1", "Enable the abort"), 169 clEnumValN(GlobalISelAbortMode::DisableWithDiag, "2", 170 "Disable the abort but emit a diagnostic on failure"))); 171 172 // An option that disables inserting FS-AFDO discriminators before emit. 173 // This is mainly for debugging and tuning purpose. 174 static cl::opt<bool> 175 FSNoFinalDiscrim("fs-no-final-discrim", cl::init(false), cl::Hidden, 176 cl::desc("Do not insert FS-AFDO discriminators before " 177 "emit.")); 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 // Experimental option to use CFL-AA in codegen 212 static cl::opt<CFLAAType> UseCFLAA( 213 "use-cfl-aa-in-codegen", cl::init(CFLAAType::None), cl::Hidden, 214 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"), 215 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"), 216 clEnumValN(CFLAAType::Steensgaard, "steens", 217 "Enable unification-based CFL-AA"), 218 clEnumValN(CFLAAType::Andersen, "anders", 219 "Enable inclusion-based CFL-AA"), 220 clEnumValN(CFLAAType::Both, "both", 221 "Enable both variants of CFL-AA"))); 222 223 /// Option names for limiting the codegen pipeline. 224 /// Those are used in error reporting and we didn't want 225 /// to duplicate their names all over the place. 226 static const char StartAfterOptName[] = "start-after"; 227 static const char StartBeforeOptName[] = "start-before"; 228 static const char StopAfterOptName[] = "stop-after"; 229 static const char StopBeforeOptName[] = "stop-before"; 230 231 static cl::opt<std::string> 232 StartAfterOpt(StringRef(StartAfterOptName), 233 cl::desc("Resume compilation after a specific pass"), 234 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 235 236 static cl::opt<std::string> 237 StartBeforeOpt(StringRef(StartBeforeOptName), 238 cl::desc("Resume compilation before a specific pass"), 239 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 240 241 static cl::opt<std::string> 242 StopAfterOpt(StringRef(StopAfterOptName), 243 cl::desc("Stop compilation after a specific pass"), 244 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 245 246 static cl::opt<std::string> 247 StopBeforeOpt(StringRef(StopBeforeOptName), 248 cl::desc("Stop compilation before a specific pass"), 249 cl::value_desc("pass-name"), cl::init(""), cl::Hidden); 250 251 /// Enable the machine function splitter pass. 252 static cl::opt<bool> EnableMachineFunctionSplitter( 253 "enable-split-machine-functions", cl::Hidden, 254 cl::desc("Split out cold blocks from machine functions based on profile " 255 "information.")); 256 257 /// Disable the expand reductions pass for testing. 258 static cl::opt<bool> DisableExpandReductions( 259 "disable-expand-reductions", cl::init(false), cl::Hidden, 260 cl::desc("Disable the expand reduction intrinsics pass from running")); 261 262 /// Disable the select optimization pass. 263 static cl::opt<bool> DisableSelectOptimize( 264 "disable-select-optimize", cl::init(true), cl::Hidden, 265 cl::desc("Disable the select-optimization pass from running")); 266 267 /// Allow standard passes to be disabled by command line options. This supports 268 /// simple binary flags that either suppress the pass or do nothing. 269 /// i.e. -disable-mypass=false has no effect. 270 /// These should be converted to boolOrDefault in order to use applyOverride. 271 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID, 272 bool Override) { 273 if (Override) 274 return IdentifyingPassPtr(); 275 return PassID; 276 } 277 278 /// Allow standard passes to be disabled by the command line, regardless of who 279 /// is adding the pass. 280 /// 281 /// StandardID is the pass identified in the standard pass pipeline and provided 282 /// to addPass(). It may be a target-specific ID in the case that the target 283 /// directly adds its own pass, but in that case we harmlessly fall through. 284 /// 285 /// TargetID is the pass that the target has configured to override StandardID. 286 /// 287 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real 288 /// pass to run. This allows multiple options to control a single pass depending 289 /// on where in the pipeline that pass is added. 290 static IdentifyingPassPtr overridePass(AnalysisID StandardID, 291 IdentifyingPassPtr TargetID) { 292 if (StandardID == &PostRASchedulerID) 293 return applyDisable(TargetID, DisablePostRASched); 294 295 if (StandardID == &BranchFolderPassID) 296 return applyDisable(TargetID, DisableBranchFold); 297 298 if (StandardID == &TailDuplicateID) 299 return applyDisable(TargetID, DisableTailDuplicate); 300 301 if (StandardID == &EarlyTailDuplicateID) 302 return applyDisable(TargetID, DisableEarlyTailDup); 303 304 if (StandardID == &MachineBlockPlacementID) 305 return applyDisable(TargetID, DisableBlockPlacement); 306 307 if (StandardID == &StackSlotColoringID) 308 return applyDisable(TargetID, DisableSSC); 309 310 if (StandardID == &DeadMachineInstructionElimID) 311 return applyDisable(TargetID, DisableMachineDCE); 312 313 if (StandardID == &EarlyIfConverterID) 314 return applyDisable(TargetID, DisableEarlyIfConversion); 315 316 if (StandardID == &EarlyMachineLICMID) 317 return applyDisable(TargetID, DisableMachineLICM); 318 319 if (StandardID == &MachineCSEID) 320 return applyDisable(TargetID, DisableMachineCSE); 321 322 if (StandardID == &MachineLICMID) 323 return applyDisable(TargetID, DisablePostRAMachineLICM); 324 325 if (StandardID == &MachineSinkingID) 326 return applyDisable(TargetID, DisableMachineSink); 327 328 if (StandardID == &PostRAMachineSinkingID) 329 return applyDisable(TargetID, DisablePostRAMachineSink); 330 331 if (StandardID == &MachineCopyPropagationID) 332 return applyDisable(TargetID, DisableCopyProp); 333 334 return TargetID; 335 } 336 337 // Find the FSProfile file name. The internal option takes the precedence 338 // before getting from TargetMachine. 339 static std::string getFSProfileFile(const TargetMachine *TM) { 340 if (!FSProfileFile.empty()) 341 return FSProfileFile.getValue(); 342 const Optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 343 if (PGOOpt == None || PGOOpt->Action != PGOOptions::SampleUse) 344 return std::string(); 345 return PGOOpt->ProfileFile; 346 } 347 348 // Find the Profile remapping file name. The internal option takes the 349 // precedence before getting from TargetMachine. 350 static std::string getFSRemappingFile(const TargetMachine *TM) { 351 if (!FSRemappingFile.empty()) 352 return FSRemappingFile.getValue(); 353 const Optional<PGOOptions> &PGOOpt = TM->getPGOOption(); 354 if (PGOOpt == None || PGOOpt->Action != PGOOptions::SampleUse) 355 return std::string(); 356 return PGOOpt->ProfileRemappingFile; 357 } 358 359 //===---------------------------------------------------------------------===// 360 /// TargetPassConfig 361 //===---------------------------------------------------------------------===// 362 363 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig", 364 "Target Pass Configuration", false, false) 365 char TargetPassConfig::ID = 0; 366 367 namespace { 368 369 struct InsertedPass { 370 AnalysisID TargetPassID; 371 IdentifyingPassPtr InsertedPassID; 372 373 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID) 374 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID) {} 375 376 Pass *getInsertedPass() const { 377 assert(InsertedPassID.isValid() && "Illegal Pass ID!"); 378 if (InsertedPassID.isInstance()) 379 return InsertedPassID.getInstance(); 380 Pass *NP = Pass::createPass(InsertedPassID.getID()); 381 assert(NP && "Pass ID not registered"); 382 return NP; 383 } 384 }; 385 386 } // end anonymous namespace 387 388 namespace llvm { 389 390 extern cl::opt<bool> EnableFSDiscriminator; 391 392 class PassConfigImpl { 393 public: 394 // List of passes explicitly substituted by this target. Normally this is 395 // empty, but it is a convenient way to suppress or replace specific passes 396 // that are part of a standard pass pipeline without overridding the entire 397 // pipeline. This mechanism allows target options to inherit a standard pass's 398 // user interface. For example, a target may disable a standard pass by 399 // default by substituting a pass ID of zero, and the user may still enable 400 // that standard pass with an explicit command line option. 401 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses; 402 403 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass 404 /// is inserted after each instance of the first one. 405 SmallVector<InsertedPass, 4> InsertedPasses; 406 }; 407 408 } // end namespace llvm 409 410 // Out of line virtual method. 411 TargetPassConfig::~TargetPassConfig() { 412 delete Impl; 413 } 414 415 static const PassInfo *getPassInfo(StringRef PassName) { 416 if (PassName.empty()) 417 return nullptr; 418 419 const PassRegistry &PR = *PassRegistry::getPassRegistry(); 420 const PassInfo *PI = PR.getPassInfo(PassName); 421 if (!PI) 422 report_fatal_error(Twine('\"') + Twine(PassName) + 423 Twine("\" pass is not registered.")); 424 return PI; 425 } 426 427 static AnalysisID getPassIDFromName(StringRef PassName) { 428 const PassInfo *PI = getPassInfo(PassName); 429 return PI ? PI->getTypeInfo() : nullptr; 430 } 431 432 static std::pair<StringRef, unsigned> 433 getPassNameAndInstanceNum(StringRef PassName) { 434 StringRef Name, InstanceNumStr; 435 std::tie(Name, InstanceNumStr) = PassName.split(','); 436 437 unsigned InstanceNum = 0; 438 if (!InstanceNumStr.empty() && InstanceNumStr.getAsInteger(10, InstanceNum)) 439 report_fatal_error("invalid pass instance specifier " + PassName); 440 441 return std::make_pair(Name, InstanceNum); 442 } 443 444 void TargetPassConfig::setStartStopPasses() { 445 StringRef StartBeforeName; 446 std::tie(StartBeforeName, StartBeforeInstanceNum) = 447 getPassNameAndInstanceNum(StartBeforeOpt); 448 449 StringRef StartAfterName; 450 std::tie(StartAfterName, StartAfterInstanceNum) = 451 getPassNameAndInstanceNum(StartAfterOpt); 452 453 StringRef StopBeforeName; 454 std::tie(StopBeforeName, StopBeforeInstanceNum) 455 = getPassNameAndInstanceNum(StopBeforeOpt); 456 457 StringRef StopAfterName; 458 std::tie(StopAfterName, StopAfterInstanceNum) 459 = getPassNameAndInstanceNum(StopAfterOpt); 460 461 StartBefore = getPassIDFromName(StartBeforeName); 462 StartAfter = getPassIDFromName(StartAfterName); 463 StopBefore = getPassIDFromName(StopBeforeName); 464 StopAfter = getPassIDFromName(StopAfterName); 465 if (StartBefore && StartAfter) 466 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") + 467 Twine(StartAfterOptName) + Twine(" specified!")); 468 if (StopBefore && StopAfter) 469 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") + 470 Twine(StopAfterOptName) + Twine(" specified!")); 471 Started = (StartAfter == nullptr) && (StartBefore == nullptr); 472 } 473 474 CGPassBuilderOption llvm::getCGPassBuilderOption() { 475 CGPassBuilderOption Opt; 476 477 #define SET_OPTION(Option) \ 478 if (Option.getNumOccurrences()) \ 479 Opt.Option = Option; 480 481 SET_OPTION(EnableFastISelOption) 482 SET_OPTION(EnableGlobalISelAbort) 483 SET_OPTION(EnableGlobalISelOption) 484 SET_OPTION(EnableIPRA) 485 SET_OPTION(OptimizeRegAlloc) 486 SET_OPTION(VerifyMachineCode) 487 488 #define SET_BOOLEAN_OPTION(Option) Opt.Option = Option; 489 490 SET_BOOLEAN_OPTION(EarlyLiveIntervals) 491 SET_BOOLEAN_OPTION(EnableBlockPlacementStats) 492 SET_BOOLEAN_OPTION(EnableImplicitNullChecks) 493 SET_BOOLEAN_OPTION(EnableMachineOutliner) 494 SET_BOOLEAN_OPTION(MISchedPostRA) 495 SET_BOOLEAN_OPTION(UseCFLAA) 496 SET_BOOLEAN_OPTION(DisableMergeICmps) 497 SET_BOOLEAN_OPTION(DisableLSR) 498 SET_BOOLEAN_OPTION(DisableConstantHoisting) 499 SET_BOOLEAN_OPTION(DisableCGP) 500 SET_BOOLEAN_OPTION(DisablePartialLibcallInlining) 501 SET_BOOLEAN_OPTION(DisableSelectOptimize) 502 SET_BOOLEAN_OPTION(PrintLSR) 503 SET_BOOLEAN_OPTION(PrintISelInput) 504 SET_BOOLEAN_OPTION(PrintGCInfo) 505 506 return Opt; 507 } 508 509 static void registerPartialPipelineCallback(PassInstrumentationCallbacks &PIC, 510 LLVMTargetMachine &LLVMTM) { 511 StringRef StartBefore; 512 StringRef StartAfter; 513 StringRef StopBefore; 514 StringRef StopAfter; 515 516 unsigned StartBeforeInstanceNum = 0; 517 unsigned StartAfterInstanceNum = 0; 518 unsigned StopBeforeInstanceNum = 0; 519 unsigned StopAfterInstanceNum = 0; 520 521 std::tie(StartBefore, StartBeforeInstanceNum) = 522 getPassNameAndInstanceNum(StartBeforeOpt); 523 std::tie(StartAfter, StartAfterInstanceNum) = 524 getPassNameAndInstanceNum(StartAfterOpt); 525 std::tie(StopBefore, StopBeforeInstanceNum) = 526 getPassNameAndInstanceNum(StopBeforeOpt); 527 std::tie(StopAfter, StopAfterInstanceNum) = 528 getPassNameAndInstanceNum(StopAfterOpt); 529 530 if (StartBefore.empty() && StartAfter.empty() && StopBefore.empty() && 531 StopAfter.empty()) 532 return; 533 534 std::tie(StartBefore, std::ignore) = 535 LLVMTM.getPassNameFromLegacyName(StartBefore); 536 std::tie(StartAfter, std::ignore) = 537 LLVMTM.getPassNameFromLegacyName(StartAfter); 538 std::tie(StopBefore, std::ignore) = 539 LLVMTM.getPassNameFromLegacyName(StopBefore); 540 std::tie(StopAfter, std::ignore) = 541 LLVMTM.getPassNameFromLegacyName(StopAfter); 542 if (!StartBefore.empty() && !StartAfter.empty()) 543 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") + 544 Twine(StartAfterOptName) + Twine(" specified!")); 545 if (!StopBefore.empty() && !StopAfter.empty()) 546 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") + 547 Twine(StopAfterOptName) + Twine(" specified!")); 548 549 PIC.registerShouldRunOptionalPassCallback( 550 [=, EnableCurrent = StartBefore.empty() && StartAfter.empty(), 551 EnableNext = Optional<bool>(), StartBeforeCount = 0u, 552 StartAfterCount = 0u, StopBeforeCount = 0u, 553 StopAfterCount = 0u](StringRef P, Any) mutable { 554 bool StartBeforePass = !StartBefore.empty() && P.contains(StartBefore); 555 bool StartAfterPass = !StartAfter.empty() && P.contains(StartAfter); 556 bool StopBeforePass = !StopBefore.empty() && P.contains(StopBefore); 557 bool StopAfterPass = !StopAfter.empty() && P.contains(StopAfter); 558 559 // Implement -start-after/-stop-after 560 if (EnableNext) { 561 EnableCurrent = *EnableNext; 562 EnableNext.reset(); 563 } 564 565 // Using PIC.registerAfterPassCallback won't work because if this 566 // callback returns false, AfterPassCallback is also skipped. 567 if (StartAfterPass && StartAfterCount++ == StartAfterInstanceNum) { 568 assert(!EnableNext && "Error: assign to EnableNext more than once"); 569 EnableNext = true; 570 } 571 if (StopAfterPass && StopAfterCount++ == StopAfterInstanceNum) { 572 assert(!EnableNext && "Error: assign to EnableNext more than once"); 573 EnableNext = false; 574 } 575 576 if (StartBeforePass && StartBeforeCount++ == StartBeforeInstanceNum) 577 EnableCurrent = true; 578 if (StopBeforePass && StopBeforeCount++ == StopBeforeInstanceNum) 579 EnableCurrent = false; 580 return EnableCurrent; 581 }); 582 } 583 584 void llvm::registerCodeGenCallback(PassInstrumentationCallbacks &PIC, 585 LLVMTargetMachine &LLVMTM) { 586 587 // Register a callback for disabling passes. 588 PIC.registerShouldRunOptionalPassCallback([](StringRef P, Any) { 589 590 #define DISABLE_PASS(Option, Name) \ 591 if (Option && P.contains(#Name)) \ 592 return false; 593 DISABLE_PASS(DisableBlockPlacement, MachineBlockPlacementPass) 594 DISABLE_PASS(DisableBranchFold, BranchFolderPass) 595 DISABLE_PASS(DisableCopyProp, MachineCopyPropagationPass) 596 DISABLE_PASS(DisableEarlyIfConversion, EarlyIfConverterPass) 597 DISABLE_PASS(DisableEarlyTailDup, EarlyTailDuplicatePass) 598 DISABLE_PASS(DisableMachineCSE, MachineCSEPass) 599 DISABLE_PASS(DisableMachineDCE, DeadMachineInstructionElimPass) 600 DISABLE_PASS(DisableMachineLICM, EarlyMachineLICMPass) 601 DISABLE_PASS(DisableMachineSink, MachineSinkingPass) 602 DISABLE_PASS(DisablePostRAMachineLICM, MachineLICMPass) 603 DISABLE_PASS(DisablePostRAMachineSink, PostRAMachineSinkingPass) 604 DISABLE_PASS(DisablePostRASched, PostRASchedulerPass) 605 DISABLE_PASS(DisableSSC, StackSlotColoringPass) 606 DISABLE_PASS(DisableTailDuplicate, TailDuplicatePass) 607 608 return true; 609 }); 610 611 registerPartialPipelineCallback(PIC, LLVMTM); 612 } 613 614 // Out of line constructor provides default values for pass options and 615 // registers all common codegen passes. 616 TargetPassConfig::TargetPassConfig(LLVMTargetMachine &TM, PassManagerBase &pm) 617 : ImmutablePass(ID), PM(&pm), TM(&TM) { 618 Impl = new PassConfigImpl(); 619 620 // Register all target independent codegen passes to activate their PassIDs, 621 // including this pass itself. 622 initializeCodeGen(*PassRegistry::getPassRegistry()); 623 624 // Also register alias analysis passes required by codegen passes. 625 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry()); 626 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry()); 627 628 if (EnableIPRA.getNumOccurrences()) 629 TM.Options.EnableIPRA = EnableIPRA; 630 else { 631 // If not explicitly specified, use target default. 632 TM.Options.EnableIPRA |= TM.useIPRA(); 633 } 634 635 if (TM.Options.EnableIPRA) 636 setRequiresCodeGenSCCOrder(); 637 638 if (EnableGlobalISelAbort.getNumOccurrences()) 639 TM.Options.GlobalISelAbort = EnableGlobalISelAbort; 640 641 setStartStopPasses(); 642 } 643 644 CodeGenOpt::Level TargetPassConfig::getOptLevel() const { 645 return TM->getOptLevel(); 646 } 647 648 /// Insert InsertedPassID pass after TargetPassID. 649 void TargetPassConfig::insertPass(AnalysisID TargetPassID, 650 IdentifyingPassPtr InsertedPassID) { 651 assert(((!InsertedPassID.isInstance() && 652 TargetPassID != InsertedPassID.getID()) || 653 (InsertedPassID.isInstance() && 654 TargetPassID != InsertedPassID.getInstance()->getPassID())) && 655 "Insert a pass after itself!"); 656 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID); 657 } 658 659 /// createPassConfig - Create a pass configuration object to be used by 660 /// addPassToEmitX methods for generating a pipeline of CodeGen passes. 661 /// 662 /// Targets may override this to extend TargetPassConfig. 663 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) { 664 return new TargetPassConfig(*this, PM); 665 } 666 667 TargetPassConfig::TargetPassConfig() 668 : ImmutablePass(ID) { 669 report_fatal_error("Trying to construct TargetPassConfig without a target " 670 "machine. Scheduling a CodeGen pass without a target " 671 "triple set?"); 672 } 673 674 bool TargetPassConfig::willCompleteCodeGenPipeline() { 675 return StopBeforeOpt.empty() && StopAfterOpt.empty(); 676 } 677 678 bool TargetPassConfig::hasLimitedCodeGenPipeline() { 679 return !StartBeforeOpt.empty() || !StartAfterOpt.empty() || 680 !willCompleteCodeGenPipeline(); 681 } 682 683 std::string 684 TargetPassConfig::getLimitedCodeGenPipelineReason(const char *Separator) { 685 if (!hasLimitedCodeGenPipeline()) 686 return std::string(); 687 std::string Res; 688 static cl::opt<std::string> *PassNames[] = {&StartAfterOpt, &StartBeforeOpt, 689 &StopAfterOpt, &StopBeforeOpt}; 690 static const char *OptNames[] = {StartAfterOptName, StartBeforeOptName, 691 StopAfterOptName, StopBeforeOptName}; 692 bool IsFirst = true; 693 for (int Idx = 0; Idx < 4; ++Idx) 694 if (!PassNames[Idx]->empty()) { 695 if (!IsFirst) 696 Res += Separator; 697 IsFirst = false; 698 Res += OptNames[Idx]; 699 } 700 return Res; 701 } 702 703 // Helper to verify the analysis is really immutable. 704 void TargetPassConfig::setOpt(bool &Opt, bool Val) { 705 assert(!Initialized && "PassConfig is immutable"); 706 Opt = Val; 707 } 708 709 void TargetPassConfig::substitutePass(AnalysisID StandardID, 710 IdentifyingPassPtr TargetID) { 711 Impl->TargetPasses[StandardID] = TargetID; 712 } 713 714 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const { 715 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator 716 I = Impl->TargetPasses.find(ID); 717 if (I == Impl->TargetPasses.end()) 718 return ID; 719 return I->second; 720 } 721 722 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const { 723 IdentifyingPassPtr TargetID = getPassSubstitution(ID); 724 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID); 725 return !FinalPtr.isValid() || FinalPtr.isInstance() || 726 FinalPtr.getID() != ID; 727 } 728 729 /// Add a pass to the PassManager if that pass is supposed to be run. If the 730 /// Started/Stopped flags indicate either that the compilation should start at 731 /// a later pass or that it should stop after an earlier pass, then do not add 732 /// the pass. Finally, compare the current pass against the StartAfter 733 /// and StopAfter options and change the Started/Stopped flags accordingly. 734 void TargetPassConfig::addPass(Pass *P) { 735 assert(!Initialized && "PassConfig is immutable"); 736 737 // Cache the Pass ID here in case the pass manager finds this pass is 738 // redundant with ones already scheduled / available, and deletes it. 739 // Fundamentally, once we add the pass to the manager, we no longer own it 740 // and shouldn't reference it. 741 AnalysisID PassID = P->getPassID(); 742 743 if (StartBefore == PassID && StartBeforeCount++ == StartBeforeInstanceNum) 744 Started = true; 745 if (StopBefore == PassID && StopBeforeCount++ == StopBeforeInstanceNum) 746 Stopped = true; 747 if (Started && !Stopped) { 748 if (AddingMachinePasses) { 749 // Construct banner message before PM->add() as that may delete the pass. 750 std::string Banner = 751 std::string("After ") + std::string(P->getPassName()); 752 addMachinePrePasses(); 753 PM->add(P); 754 addMachinePostPasses(Banner); 755 } else { 756 PM->add(P); 757 } 758 759 // Add the passes after the pass P if there is any. 760 for (const auto &IP : Impl->InsertedPasses) 761 if (IP.TargetPassID == PassID) 762 addPass(IP.getInsertedPass()); 763 } else { 764 delete P; 765 } 766 767 if (StopAfter == PassID && StopAfterCount++ == StopAfterInstanceNum) 768 Stopped = true; 769 770 if (StartAfter == PassID && StartAfterCount++ == StartAfterInstanceNum) 771 Started = true; 772 if (Stopped && !Started) 773 report_fatal_error("Cannot stop compilation after pass that is not run"); 774 } 775 776 /// Add a CodeGen pass at this point in the pipeline after checking for target 777 /// and command line overrides. 778 /// 779 /// addPass cannot return a pointer to the pass instance because is internal the 780 /// PassManager and the instance we create here may already be freed. 781 AnalysisID TargetPassConfig::addPass(AnalysisID PassID) { 782 IdentifyingPassPtr TargetID = getPassSubstitution(PassID); 783 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID); 784 if (!FinalPtr.isValid()) 785 return nullptr; 786 787 Pass *P; 788 if (FinalPtr.isInstance()) 789 P = FinalPtr.getInstance(); 790 else { 791 P = Pass::createPass(FinalPtr.getID()); 792 if (!P) 793 llvm_unreachable("Pass ID not registered"); 794 } 795 AnalysisID FinalID = P->getPassID(); 796 addPass(P); // Ends the lifetime of P. 797 798 return FinalID; 799 } 800 801 void TargetPassConfig::printAndVerify(const std::string &Banner) { 802 addPrintPass(Banner); 803 addVerifyPass(Banner); 804 } 805 806 void TargetPassConfig::addPrintPass(const std::string &Banner) { 807 if (PrintAfterISel) 808 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner)); 809 } 810 811 void TargetPassConfig::addVerifyPass(const std::string &Banner) { 812 bool Verify = VerifyMachineCode == cl::BOU_TRUE; 813 #ifdef EXPENSIVE_CHECKS 814 if (VerifyMachineCode == cl::BOU_UNSET) 815 Verify = TM->isMachineVerifierClean(); 816 #endif 817 if (Verify) 818 PM->add(createMachineVerifierPass(Banner)); 819 } 820 821 void TargetPassConfig::addDebugifyPass() { 822 PM->add(createDebugifyMachineModulePass()); 823 } 824 825 void TargetPassConfig::addStripDebugPass() { 826 PM->add(createStripDebugMachineModulePass(/*OnlyDebugified=*/true)); 827 } 828 829 void TargetPassConfig::addCheckDebugPass() { 830 PM->add(createCheckDebugMachineModulePass()); 831 } 832 833 void TargetPassConfig::addMachinePrePasses(bool AllowDebugify) { 834 if (AllowDebugify && DebugifyIsSafe && 835 (DebugifyAndStripAll == cl::BOU_TRUE || 836 DebugifyCheckAndStripAll == cl::BOU_TRUE)) 837 addDebugifyPass(); 838 } 839 840 void TargetPassConfig::addMachinePostPasses(const std::string &Banner) { 841 if (DebugifyIsSafe) { 842 if (DebugifyCheckAndStripAll == cl::BOU_TRUE) { 843 addCheckDebugPass(); 844 addStripDebugPass(); 845 } else if (DebugifyAndStripAll == cl::BOU_TRUE) 846 addStripDebugPass(); 847 } 848 addVerifyPass(Banner); 849 } 850 851 /// Add common target configurable passes that perform LLVM IR to IR transforms 852 /// following machine independent optimization. 853 void TargetPassConfig::addIRPasses() { 854 // Before running any passes, run the verifier to determine if the input 855 // coming from the front-end and/or optimizer is valid. 856 if (!DisableVerify) 857 addPass(createVerifierPass()); 858 859 if (getOptLevel() != CodeGenOpt::None) { 860 switch (UseCFLAA) { 861 case CFLAAType::Steensgaard: 862 addPass(createCFLSteensAAWrapperPass()); 863 break; 864 case CFLAAType::Andersen: 865 addPass(createCFLAndersAAWrapperPass()); 866 break; 867 case CFLAAType::Both: 868 addPass(createCFLAndersAAWrapperPass()); 869 addPass(createCFLSteensAAWrapperPass()); 870 break; 871 default: 872 break; 873 } 874 875 // Basic AliasAnalysis support. 876 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 877 // BasicAliasAnalysis wins if they disagree. This is intended to help 878 // support "obvious" type-punning idioms. 879 addPass(createTypeBasedAAWrapperPass()); 880 addPass(createScopedNoAliasAAWrapperPass()); 881 addPass(createBasicAAWrapperPass()); 882 883 // Run loop strength reduction before anything else. 884 if (!DisableLSR) { 885 addPass(createCanonicalizeFreezeInLoopsPass()); 886 addPass(createLoopStrengthReducePass()); 887 if (PrintLSR) 888 addPass(createPrintFunctionPass(dbgs(), 889 "\n\n*** Code after LSR ***\n")); 890 } 891 892 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of 893 // loads and compares. ExpandMemCmpPass then tries to expand those calls 894 // into optimally-sized loads and compares. The transforms are enabled by a 895 // target lowering hook. 896 if (!DisableMergeICmps) 897 addPass(createMergeICmpsLegacyPass()); 898 addPass(createExpandMemCmpPass()); 899 } 900 901 // Run GC lowering passes for builtin collectors 902 // TODO: add a pass insertion point here 903 addPass(&GCLoweringID); 904 addPass(&ShadowStackGCLoweringID); 905 addPass(createLowerConstantIntrinsicsPass()); 906 907 // For MachO, lower @llvm.global_dtors into @llvm_global_ctors with 908 // __cxa_atexit() calls to avoid emitting the deprecated __mod_term_func. 909 if (TM->getTargetTriple().isOSBinFormatMachO() && 910 TM->Options.LowerGlobalDtorsViaCxaAtExit) 911 addPass(createLowerGlobalDtorsLegacyPass()); 912 913 // Make sure that no unreachable blocks are instruction selected. 914 addPass(createUnreachableBlockEliminationPass()); 915 916 // Prepare expensive constants for SelectionDAG. 917 if (getOptLevel() != CodeGenOpt::None && !DisableConstantHoisting) 918 addPass(createConstantHoistingPass()); 919 920 if (getOptLevel() != CodeGenOpt::None) 921 addPass(createReplaceWithVeclibLegacyPass()); 922 923 if (getOptLevel() != CodeGenOpt::None && !DisablePartialLibcallInlining) 924 addPass(createPartiallyInlineLibCallsPass()); 925 926 // Expand vector predication intrinsics into standard IR instructions. 927 // This pass has to run before ScalarizeMaskedMemIntrin and ExpandReduction 928 // passes since it emits those kinds of intrinsics. 929 addPass(createExpandVectorPredicationPass()); 930 931 // Add scalarization of target's unsupported masked memory intrinsics pass. 932 // the unsupported intrinsic will be replaced with a chain of basic blocks, 933 // that stores/loads element one-by-one if the appropriate mask bit is set. 934 addPass(createScalarizeMaskedMemIntrinLegacyPass()); 935 936 // Expand reduction intrinsics into shuffle sequences if the target wants to. 937 // Allow disabling it for testing purposes. 938 if (!DisableExpandReductions) 939 addPass(createExpandReductionsPass()); 940 941 if (getOptLevel() != CodeGenOpt::None) 942 addPass(createTLSVariableHoistPass()); 943 944 // Convert conditional moves to conditional jumps when profitable. 945 if (getOptLevel() != CodeGenOpt::None && !DisableSelectOptimize) 946 addPass(createSelectOptimizePass()); 947 } 948 949 /// Turn exception handling constructs into something the code generators can 950 /// handle. 951 void TargetPassConfig::addPassesToHandleExceptions() { 952 const MCAsmInfo *MCAI = TM->getMCAsmInfo(); 953 assert(MCAI && "No MCAsmInfo"); 954 switch (MCAI->getExceptionHandlingType()) { 955 case ExceptionHandling::SjLj: 956 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both 957 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise, 958 // catch info can get misplaced when a selector ends up more than one block 959 // removed from the parent invoke(s). This could happen when a landing 960 // pad is shared by multiple invokes and is also a target of a normal 961 // edge from elsewhere. 962 addPass(createSjLjEHPreparePass(TM)); 963 LLVM_FALLTHROUGH; 964 case ExceptionHandling::DwarfCFI: 965 case ExceptionHandling::ARM: 966 case ExceptionHandling::AIX: 967 addPass(createDwarfEHPass(getOptLevel())); 968 break; 969 case ExceptionHandling::WinEH: 970 // We support using both GCC-style and MSVC-style exceptions on Windows, so 971 // add both preparation passes. Each pass will only actually run if it 972 // recognizes the personality function. 973 addPass(createWinEHPass()); 974 addPass(createDwarfEHPass(getOptLevel())); 975 break; 976 case ExceptionHandling::Wasm: 977 // Wasm EH uses Windows EH instructions, but it does not need to demote PHIs 978 // on catchpads and cleanuppads because it does not outline them into 979 // funclets. Catchswitch blocks are not lowered in SelectionDAG, so we 980 // should remove PHIs there. 981 addPass(createWinEHPass(/*DemoteCatchSwitchPHIOnly=*/false)); 982 addPass(createWasmEHPass()); 983 break; 984 case ExceptionHandling::None: 985 addPass(createLowerInvokePass()); 986 987 // The lower invoke pass may create unreachable code. Remove it. 988 addPass(createUnreachableBlockEliminationPass()); 989 break; 990 } 991 } 992 993 /// Add pass to prepare the LLVM IR for code generation. This should be done 994 /// before exception handling preparation passes. 995 void TargetPassConfig::addCodeGenPrepare() { 996 if (getOptLevel() != CodeGenOpt::None && !DisableCGP) 997 addPass(createCodeGenPreparePass()); 998 } 999 1000 /// Add common passes that perform LLVM IR to IR transforms in preparation for 1001 /// instruction selection. 1002 void TargetPassConfig::addISelPrepare() { 1003 addPreISel(); 1004 1005 // Force codegen to run according to the callgraph. 1006 if (requiresCodeGenSCCOrder()) 1007 addPass(new DummyCGSCCPass); 1008 1009 // Add both the safe stack and the stack protection passes: each of them will 1010 // only protect functions that have corresponding attributes. 1011 addPass(createSafeStackPass()); 1012 addPass(createStackProtectorPass()); 1013 1014 if (PrintISelInput) 1015 addPass(createPrintFunctionPass( 1016 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n")); 1017 1018 // All passes which modify the LLVM IR are now complete; run the verifier 1019 // to ensure that the IR is valid. 1020 if (!DisableVerify) 1021 addPass(createVerifierPass()); 1022 } 1023 1024 bool TargetPassConfig::addCoreISelPasses() { 1025 // Enable FastISel with -fast-isel, but allow that to be overridden. 1026 TM->setO0WantsFastISel(EnableFastISelOption != cl::BOU_FALSE); 1027 1028 // Determine an instruction selector. 1029 enum class SelectorType { SelectionDAG, FastISel, GlobalISel }; 1030 SelectorType Selector; 1031 1032 if (EnableFastISelOption == cl::BOU_TRUE) 1033 Selector = SelectorType::FastISel; 1034 else if (EnableGlobalISelOption == cl::BOU_TRUE || 1035 (TM->Options.EnableGlobalISel && 1036 EnableGlobalISelOption != cl::BOU_FALSE)) 1037 Selector = SelectorType::GlobalISel; 1038 else if (TM->getOptLevel() == CodeGenOpt::None && TM->getO0WantsFastISel()) 1039 Selector = SelectorType::FastISel; 1040 else 1041 Selector = SelectorType::SelectionDAG; 1042 1043 // Set consistently TM->Options.EnableFastISel and EnableGlobalISel. 1044 if (Selector == SelectorType::FastISel) { 1045 TM->setFastISel(true); 1046 TM->setGlobalISel(false); 1047 } else if (Selector == SelectorType::GlobalISel) { 1048 TM->setFastISel(false); 1049 TM->setGlobalISel(true); 1050 } 1051 1052 // FIXME: Injecting into the DAGISel pipeline seems to cause issues with 1053 // analyses needing to be re-run. This can result in being unable to 1054 // schedule passes (particularly with 'Function Alias Analysis 1055 // Results'). It's not entirely clear why but AFAICT this seems to be 1056 // due to one FunctionPassManager not being able to use analyses from a 1057 // previous one. As we're injecting a ModulePass we break the usual 1058 // pass manager into two. GlobalISel with the fallback path disabled 1059 // and -run-pass seem to be unaffected. The majority of GlobalISel 1060 // testing uses -run-pass so this probably isn't too bad. 1061 SaveAndRestore<bool> SavedDebugifyIsSafe(DebugifyIsSafe); 1062 if (Selector != SelectorType::GlobalISel || !isGlobalISelAbortEnabled()) 1063 DebugifyIsSafe = false; 1064 1065 // Add instruction selector passes. 1066 if (Selector == SelectorType::GlobalISel) { 1067 SaveAndRestore<bool> SavedAddingMachinePasses(AddingMachinePasses, true); 1068 if (addIRTranslator()) 1069 return true; 1070 1071 addPreLegalizeMachineIR(); 1072 1073 if (addLegalizeMachineIR()) 1074 return true; 1075 1076 // Before running the register bank selector, ask the target if it 1077 // wants to run some passes. 1078 addPreRegBankSelect(); 1079 1080 if (addRegBankSelect()) 1081 return true; 1082 1083 addPreGlobalInstructionSelect(); 1084 1085 if (addGlobalInstructionSelect()) 1086 return true; 1087 1088 // Pass to reset the MachineFunction if the ISel failed. 1089 addPass(createResetMachineFunctionPass( 1090 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled())); 1091 1092 // Provide a fallback path when we do not want to abort on 1093 // not-yet-supported input. 1094 if (!isGlobalISelAbortEnabled() && addInstSelector()) 1095 return true; 1096 1097 } else if (addInstSelector()) 1098 return true; 1099 1100 // Expand pseudo-instructions emitted by ISel. Don't run the verifier before 1101 // FinalizeISel. 1102 addPass(&FinalizeISelID); 1103 1104 // Print the instruction selected machine code... 1105 printAndVerify("After Instruction Selection"); 1106 1107 return false; 1108 } 1109 1110 bool TargetPassConfig::addISelPasses() { 1111 if (TM->useEmulatedTLS()) 1112 addPass(createLowerEmuTLSPass()); 1113 1114 addPass(createPreISelIntrinsicLoweringPass()); 1115 PM->add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis())); 1116 addIRPasses(); 1117 addCodeGenPrepare(); 1118 addPassesToHandleExceptions(); 1119 addISelPrepare(); 1120 1121 return addCoreISelPasses(); 1122 } 1123 1124 /// -regalloc=... command line option. 1125 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } 1126 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false, 1127 RegisterPassParser<RegisterRegAlloc>> 1128 RegAlloc("regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator), 1129 cl::desc("Register allocator to use")); 1130 1131 /// Add the complete set of target-independent postISel code generator passes. 1132 /// 1133 /// This can be read as the standard order of major LLVM CodeGen stages. Stages 1134 /// with nontrivial configuration or multiple passes are broken out below in 1135 /// add%Stage routines. 1136 /// 1137 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The 1138 /// addPre/Post methods with empty header implementations allow injecting 1139 /// target-specific fixups just before or after major stages. Additionally, 1140 /// targets have the flexibility to change pass order within a stage by 1141 /// overriding default implementation of add%Stage routines below. Each 1142 /// technique has maintainability tradeoffs because alternate pass orders are 1143 /// not well supported. addPre/Post works better if the target pass is easily 1144 /// tied to a common pass. But if it has subtle dependencies on multiple passes, 1145 /// the target should override the stage instead. 1146 /// 1147 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection 1148 /// before/after any target-independent pass. But it's currently overkill. 1149 void TargetPassConfig::addMachinePasses() { 1150 AddingMachinePasses = true; 1151 1152 // Add passes that optimize machine instructions in SSA form. 1153 if (getOptLevel() != CodeGenOpt::None) { 1154 addMachineSSAOptimization(); 1155 } else { 1156 // If the target requests it, assign local variables to stack slots relative 1157 // to one another and simplify frame index references where possible. 1158 addPass(&LocalStackSlotAllocationID); 1159 } 1160 1161 if (TM->Options.EnableIPRA) 1162 addPass(createRegUsageInfoPropPass()); 1163 1164 // Run pre-ra passes. 1165 addPreRegAlloc(); 1166 1167 // Debugifying the register allocator passes seems to provoke some 1168 // non-determinism that affects CodeGen and there doesn't seem to be a point 1169 // where it becomes safe again so stop debugifying here. 1170 DebugifyIsSafe = false; 1171 1172 // Add a FSDiscriminator pass right before RA, so that we could get 1173 // more precise SampleFDO profile for RA. 1174 if (EnableFSDiscriminator) { 1175 addPass(createMIRAddFSDiscriminatorsPass( 1176 sampleprof::FSDiscriminatorPass::Pass1)); 1177 const std::string ProfileFile = getFSProfileFile(TM); 1178 if (!ProfileFile.empty() && !DisableRAFSProfileLoader) 1179 addPass( 1180 createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1181 sampleprof::FSDiscriminatorPass::Pass1)); 1182 } 1183 1184 // Run register allocation and passes that are tightly coupled with it, 1185 // including phi elimination and scheduling. 1186 if (getOptimizeRegAlloc()) 1187 addOptimizedRegAlloc(); 1188 else 1189 addFastRegAlloc(); 1190 1191 // Run post-ra passes. 1192 addPostRegAlloc(); 1193 1194 addPass(&RemoveRedundantDebugValuesID); 1195 1196 addPass(&FixupStatepointCallerSavedID); 1197 1198 // Insert prolog/epilog code. Eliminate abstract frame index references... 1199 if (getOptLevel() != CodeGenOpt::None) { 1200 addPass(&PostRAMachineSinkingID); 1201 addPass(&ShrinkWrapID); 1202 } 1203 1204 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only 1205 // do so if it hasn't been disabled, substituted, or overridden. 1206 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID)) 1207 addPass(createPrologEpilogInserterPass()); 1208 1209 /// Add passes that optimize machine instructions after register allocation. 1210 if (getOptLevel() != CodeGenOpt::None) 1211 addMachineLateOptimization(); 1212 1213 // Expand pseudo instructions before second scheduling pass. 1214 addPass(&ExpandPostRAPseudosID); 1215 1216 // Run pre-sched2 passes. 1217 addPreSched2(); 1218 1219 if (EnableImplicitNullChecks) 1220 addPass(&ImplicitNullChecksID); 1221 1222 // Second pass scheduler. 1223 // Let Target optionally insert this pass by itself at some other 1224 // point. 1225 if (getOptLevel() != CodeGenOpt::None && 1226 !TM->targetSchedulesPostRAScheduling()) { 1227 if (MISchedPostRA) 1228 addPass(&PostMachineSchedulerID); 1229 else 1230 addPass(&PostRASchedulerID); 1231 } 1232 1233 // GC 1234 if (addGCPasses()) { 1235 if (PrintGCInfo) 1236 addPass(createGCInfoPrinter(dbgs())); 1237 } 1238 1239 // Basic block placement. 1240 if (getOptLevel() != CodeGenOpt::None) 1241 addBlockPlacement(); 1242 1243 // Insert before XRay Instrumentation. 1244 addPass(&FEntryInserterID); 1245 1246 addPass(&XRayInstrumentationID); 1247 addPass(&PatchableFunctionID); 1248 1249 if (EnableFSDiscriminator && !FSNoFinalDiscrim) 1250 // Add FS discriminators here so that all the instruction duplicates 1251 // in different BBs get their own discriminators. With this, we can "sum" 1252 // the SampleFDO counters instead of using MAX. This will improve the 1253 // SampleFDO profile quality. 1254 addPass(createMIRAddFSDiscriminatorsPass( 1255 sampleprof::FSDiscriminatorPass::PassLast)); 1256 1257 addPreEmitPass(); 1258 1259 if (TM->Options.EnableIPRA) 1260 // Collect register usage information and produce a register mask of 1261 // clobbered registers, to be used to optimize call sites. 1262 addPass(createRegUsageInfoCollector()); 1263 1264 // FIXME: Some backends are incompatible with running the verifier after 1265 // addPreEmitPass. Maybe only pass "false" here for those targets? 1266 addPass(&FuncletLayoutID); 1267 1268 addPass(&StackMapLivenessID); 1269 addPass(&LiveDebugValuesID); 1270 1271 if (TM->Options.EnableMachineOutliner && getOptLevel() != CodeGenOpt::None && 1272 EnableMachineOutliner != RunOutliner::NeverOutline) { 1273 bool RunOnAllFunctions = 1274 (EnableMachineOutliner == RunOutliner::AlwaysOutline); 1275 bool AddOutliner = 1276 RunOnAllFunctions || TM->Options.SupportsDefaultOutlining; 1277 if (AddOutliner) 1278 addPass(createMachineOutlinerPass(RunOnAllFunctions)); 1279 } 1280 1281 // Machine function splitter uses the basic block sections feature. Both 1282 // cannot be enabled at the same time. Basic block sections takes precedence. 1283 // FIXME: In principle, BasicBlockSection::Labels and splitting can used 1284 // together. Update this check once we have addressed any issues. 1285 if (TM->getBBSectionsType() != llvm::BasicBlockSection::None) { 1286 if (TM->getBBSectionsType() == llvm::BasicBlockSection::List) { 1287 addPass(llvm::createBasicBlockSectionsProfileReaderPass( 1288 TM->getBBSectionsFuncListBuf())); 1289 } 1290 addPass(llvm::createBasicBlockSectionsPass()); 1291 } else if (TM->Options.EnableMachineFunctionSplitter || 1292 EnableMachineFunctionSplitter) { 1293 addPass(createMachineFunctionSplitterPass()); 1294 } 1295 1296 if (!DisableCFIFixup && TM->Options.EnableCFIFixup) 1297 addPass(createCFIFixup()); 1298 1299 // Add passes that directly emit MI after all other MI passes. 1300 addPreEmitPass2(); 1301 1302 AddingMachinePasses = false; 1303 } 1304 1305 /// Add passes that optimize machine instructions in SSA form. 1306 void TargetPassConfig::addMachineSSAOptimization() { 1307 // Pre-ra tail duplication. 1308 addPass(&EarlyTailDuplicateID); 1309 1310 // Optimize PHIs before DCE: removing dead PHI cycles may make more 1311 // instructions dead. 1312 addPass(&OptimizePHIsID); 1313 1314 // This pass merges large allocas. StackSlotColoring is a different pass 1315 // which merges spill slots. 1316 addPass(&StackColoringID); 1317 1318 // If the target requests it, assign local variables to stack slots relative 1319 // to one another and simplify frame index references where possible. 1320 addPass(&LocalStackSlotAllocationID); 1321 1322 // With optimization, dead code should already be eliminated. However 1323 // there is one known exception: lowered code for arguments that are only 1324 // used by tail calls, where the tail calls reuse the incoming stack 1325 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll). 1326 addPass(&DeadMachineInstructionElimID); 1327 1328 // Allow targets to insert passes that improve instruction level parallelism, 1329 // like if-conversion. Such passes will typically need dominator trees and 1330 // loop info, just like LICM and CSE below. 1331 addILPOpts(); 1332 1333 addPass(&EarlyMachineLICMID); 1334 addPass(&MachineCSEID); 1335 1336 addPass(&MachineSinkingID); 1337 1338 addPass(&PeepholeOptimizerID); 1339 // Clean-up the dead code that may have been generated by peephole 1340 // rewriting. 1341 addPass(&DeadMachineInstructionElimID); 1342 } 1343 1344 //===---------------------------------------------------------------------===// 1345 /// Register Allocation Pass Configuration 1346 //===---------------------------------------------------------------------===// 1347 1348 bool TargetPassConfig::getOptimizeRegAlloc() const { 1349 switch (OptimizeRegAlloc) { 1350 case cl::BOU_UNSET: return getOptLevel() != CodeGenOpt::None; 1351 case cl::BOU_TRUE: return true; 1352 case cl::BOU_FALSE: return false; 1353 } 1354 llvm_unreachable("Invalid optimize-regalloc state"); 1355 } 1356 1357 /// A dummy default pass factory indicates whether the register allocator is 1358 /// overridden on the command line. 1359 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag; 1360 1361 static RegisterRegAlloc 1362 defaultRegAlloc("default", 1363 "pick register allocator based on -O option", 1364 useDefaultRegisterAllocator); 1365 1366 static void initializeDefaultRegisterAllocatorOnce() { 1367 if (!RegisterRegAlloc::getDefault()) 1368 RegisterRegAlloc::setDefault(RegAlloc); 1369 } 1370 1371 /// Instantiate the default register allocator pass for this target for either 1372 /// the optimized or unoptimized allocation path. This will be added to the pass 1373 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc 1374 /// in the optimized case. 1375 /// 1376 /// A target that uses the standard regalloc pass order for fast or optimized 1377 /// allocation may still override this for per-target regalloc 1378 /// selection. But -regalloc=... always takes precedence. 1379 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) { 1380 if (Optimized) 1381 return createGreedyRegisterAllocator(); 1382 else 1383 return createFastRegisterAllocator(); 1384 } 1385 1386 /// Find and instantiate the register allocation pass requested by this target 1387 /// at the current optimization level. Different register allocators are 1388 /// defined as separate passes because they may require different analysis. 1389 /// 1390 /// This helper ensures that the regalloc= option is always available, 1391 /// even for targets that override the default allocator. 1392 /// 1393 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs, 1394 /// this can be folded into addPass. 1395 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) { 1396 // Initialize the global default. 1397 llvm::call_once(InitializeDefaultRegisterAllocatorFlag, 1398 initializeDefaultRegisterAllocatorOnce); 1399 1400 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault(); 1401 if (Ctor != useDefaultRegisterAllocator) 1402 return Ctor(); 1403 1404 // With no -regalloc= override, ask the target for a regalloc pass. 1405 return createTargetRegisterAllocator(Optimized); 1406 } 1407 1408 bool TargetPassConfig::isCustomizedRegAlloc() { 1409 return RegAlloc != 1410 (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator; 1411 } 1412 1413 bool TargetPassConfig::addRegAssignAndRewriteFast() { 1414 if (RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator && 1415 RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&createFastRegisterAllocator) 1416 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc."); 1417 1418 addPass(createRegAllocPass(false)); 1419 1420 // Allow targets to change the register assignments after 1421 // fast register allocation. 1422 addPostFastRegAllocRewrite(); 1423 return true; 1424 } 1425 1426 bool TargetPassConfig::addRegAssignAndRewriteOptimized() { 1427 // Add the selected register allocation pass. 1428 addPass(createRegAllocPass(true)); 1429 1430 // Allow targets to change the register assignments before rewriting. 1431 addPreRewrite(); 1432 1433 // Finally rewrite virtual registers. 1434 addPass(&VirtRegRewriterID); 1435 1436 // Regalloc scoring for ML-driven eviction - noop except when learning a new 1437 // eviction policy. 1438 addPass(createRegAllocScoringPass()); 1439 return true; 1440 } 1441 1442 /// Return true if the default global register allocator is in use and 1443 /// has not be overriden on the command line with '-regalloc=...' 1444 bool TargetPassConfig::usingDefaultRegAlloc() const { 1445 return RegAlloc.getNumOccurrences() == 0; 1446 } 1447 1448 /// Add the minimum set of target-independent passes that are required for 1449 /// register allocation. No coalescing or scheduling. 1450 void TargetPassConfig::addFastRegAlloc() { 1451 addPass(&PHIEliminationID); 1452 addPass(&TwoAddressInstructionPassID); 1453 1454 addRegAssignAndRewriteFast(); 1455 } 1456 1457 /// Add standard target-independent passes that are tightly coupled with 1458 /// optimized register allocation, including coalescing, machine instruction 1459 /// scheduling, and register allocation itself. 1460 void TargetPassConfig::addOptimizedRegAlloc() { 1461 addPass(&DetectDeadLanesID); 1462 1463 addPass(&ProcessImplicitDefsID); 1464 1465 // LiveVariables currently requires pure SSA form. 1466 // 1467 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags, 1468 // LiveVariables can be removed completely, and LiveIntervals can be directly 1469 // computed. (We still either need to regenerate kill flags after regalloc, or 1470 // preferably fix the scavenger to not depend on them). 1471 // FIXME: UnreachableMachineBlockElim is a dependant pass of LiveVariables. 1472 // When LiveVariables is removed this has to be removed/moved either. 1473 // Explicit addition of UnreachableMachineBlockElim allows stopping before or 1474 // after it with -stop-before/-stop-after. 1475 addPass(&UnreachableMachineBlockElimID); 1476 addPass(&LiveVariablesID); 1477 1478 // Edge splitting is smarter with machine loop info. 1479 addPass(&MachineLoopInfoID); 1480 addPass(&PHIEliminationID); 1481 1482 // Eventually, we want to run LiveIntervals before PHI elimination. 1483 if (EarlyLiveIntervals) 1484 addPass(&LiveIntervalsID); 1485 1486 addPass(&TwoAddressInstructionPassID); 1487 addPass(&RegisterCoalescerID); 1488 1489 // The machine scheduler may accidentally create disconnected components 1490 // when moving subregister definitions around, avoid this by splitting them to 1491 // separate vregs before. Splitting can also improve reg. allocation quality. 1492 addPass(&RenameIndependentSubregsID); 1493 1494 // PreRA instruction scheduling. 1495 addPass(&MachineSchedulerID); 1496 1497 if (addRegAssignAndRewriteOptimized()) { 1498 // Perform stack slot coloring and post-ra machine LICM. 1499 addPass(&StackSlotColoringID); 1500 1501 // Allow targets to expand pseudo instructions depending on the choice of 1502 // registers before MachineCopyPropagation. 1503 addPostRewrite(); 1504 1505 // Copy propagate to forward register uses and try to eliminate COPYs that 1506 // were not coalesced. 1507 addPass(&MachineCopyPropagationID); 1508 1509 // Run post-ra machine LICM to hoist reloads / remats. 1510 // 1511 // FIXME: can this move into MachineLateOptimization? 1512 addPass(&MachineLICMID); 1513 } 1514 } 1515 1516 //===---------------------------------------------------------------------===// 1517 /// Post RegAlloc Pass Configuration 1518 //===---------------------------------------------------------------------===// 1519 1520 /// Add passes that optimize machine instructions after register allocation. 1521 void TargetPassConfig::addMachineLateOptimization() { 1522 // Branch folding must be run after regalloc and prolog/epilog insertion. 1523 addPass(&BranchFolderPassID); 1524 1525 // Tail duplication. 1526 // Note that duplicating tail just increases code size and degrades 1527 // performance for targets that require Structured Control Flow. 1528 // In addition it can also make CFG irreducible. Thus we disable it. 1529 if (!TM->requiresStructuredCFG()) 1530 addPass(&TailDuplicateID); 1531 1532 // Copy propagation. 1533 addPass(&MachineCopyPropagationID); 1534 } 1535 1536 /// Add standard GC passes. 1537 bool TargetPassConfig::addGCPasses() { 1538 addPass(&GCMachineCodeAnalysisID); 1539 return true; 1540 } 1541 1542 /// Add standard basic block placement passes. 1543 void TargetPassConfig::addBlockPlacement() { 1544 if (EnableFSDiscriminator) { 1545 addPass(createMIRAddFSDiscriminatorsPass( 1546 sampleprof::FSDiscriminatorPass::Pass2)); 1547 const std::string ProfileFile = getFSProfileFile(TM); 1548 if (!ProfileFile.empty() && !DisableLayoutFSProfileLoader) 1549 addPass( 1550 createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM), 1551 sampleprof::FSDiscriminatorPass::Pass2)); 1552 } 1553 if (addPass(&MachineBlockPlacementID)) { 1554 // Run a separate pass to collect block placement statistics. 1555 if (EnableBlockPlacementStats) 1556 addPass(&MachineBlockPlacementStatsID); 1557 } 1558 } 1559 1560 //===---------------------------------------------------------------------===// 1561 /// GlobalISel Configuration 1562 //===---------------------------------------------------------------------===// 1563 bool TargetPassConfig::isGlobalISelAbortEnabled() const { 1564 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::Enable; 1565 } 1566 1567 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const { 1568 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::DisableWithDiag; 1569 } 1570 1571 bool TargetPassConfig::isGISelCSEEnabled() const { 1572 return true; 1573 } 1574 1575 std::unique_ptr<CSEConfigBase> TargetPassConfig::getCSEConfig() const { 1576 return std::make_unique<CSEConfigBase>(); 1577 } 1578