llvm/lib/Transforms/IPO/PassManagerBuilder.cpp

   1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===//
   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 the PassManagerBuilder class, which is used to set up a
  10 // "standard" optimization sequence suitable for languages like C and C++.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
  15 #include "llvm-c/Transforms/PassManagerBuilder.h"
  16 #include "llvm/ADT/STLExtras.h"
  17 #include "llvm/ADT/SmallVector.h"
  18 #include "llvm/Analysis/CFLAndersAliasAnalysis.h"
  19 #include "llvm/Analysis/CFLSteensAliasAnalysis.h"
  20 #include "llvm/Analysis/GlobalsModRef.h"
  21 #include "llvm/Analysis/ScopedNoAliasAA.h"
  22 #include "llvm/Analysis/TargetLibraryInfo.h"
  23 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
  24 #include "llvm/IR/LegacyPassManager.h"
  25 #include "llvm/Support/CommandLine.h"
  26 #include "llvm/Support/ManagedStatic.h"
  27 #include "llvm/Target/CGPassBuilderOption.h"
  28 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"
  29 #include "llvm/Transforms/IPO.h"
  30 #include "llvm/Transforms/IPO/Attributor.h"
  31 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h"
  32 #include "llvm/Transforms/IPO/FunctionAttrs.h"
  33 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
  34 #include "llvm/Transforms/InstCombine/InstCombine.h"
  35 #include "llvm/Transforms/Instrumentation.h"
  36 #include "llvm/Transforms/Scalar.h"
  37 #include "llvm/Transforms/Scalar/GVN.h"
  38 #include "llvm/Transforms/Scalar/LICM.h"
  39 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
  40 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
  41 #include "llvm/Transforms/Utils.h"
  42 #include "llvm/Transforms/Vectorize.h"
  43
  44 using namespace llvm;
  45
  46 namespace llvm {
  47 cl::opt<bool> RunPartialInlining("enable-partial-inlining", cl::Hidden,
  48                                  cl::desc("Run Partial inlinining pass"));
  49
  50 static cl::opt<bool>
  51 UseGVNAfterVectorization("use-gvn-after-vectorization",
  52   cl::init(false), cl::Hidden,
  53   cl::desc("Run GVN instead of Early CSE after vectorization passes"));
  54
  55 cl::opt<bool> ExtraVectorizerPasses(
  56     "extra-vectorizer-passes", cl::init(false), cl::Hidden,
  57     cl::desc("Run cleanup optimization passes after vectorization."));
  58
  59 static cl::opt<bool>
  60 RunLoopRerolling("reroll-loops", cl::Hidden,
  61                  cl::desc("Run the loop rerolling pass"));
  62
  63 cl::opt<bool> RunNewGVN("enable-newgvn", cl::init(false), cl::Hidden,
  64                         cl::desc("Run the NewGVN pass"));
  65
  66 // Experimental option to use CFL-AA
  67 static cl::opt<::CFLAAType>
  68     UseCFLAA("use-cfl-aa", cl::init(::CFLAAType::None), cl::Hidden,
  69              cl::desc("Enable the new, experimental CFL alias analysis"),
  70              cl::values(clEnumValN(::CFLAAType::None, "none", "Disable CFL-AA"),
  71                         clEnumValN(::CFLAAType::Steensgaard, "steens",
  72                                    "Enable unification-based CFL-AA"),
  73                         clEnumValN(::CFLAAType::Andersen, "anders",
  74                                    "Enable inclusion-based CFL-AA"),
  75                         clEnumValN(::CFLAAType::Both, "both",
  76                                    "Enable both variants of CFL-AA")));
  77
  78 cl::opt<bool> EnableLoopInterchange(
  79     "enable-loopinterchange", cl::init(false), cl::Hidden,
  80     cl::desc("Enable the experimental LoopInterchange Pass"));
  81
  82 cl::opt<bool> EnableUnrollAndJam("enable-unroll-and-jam", cl::init(false),
  83                                  cl::Hidden,
  84                                  cl::desc("Enable Unroll And Jam Pass"));
  85
  86 cl::opt<bool> EnableLoopFlatten("enable-loop-flatten", cl::init(false),
  87                                 cl::Hidden,
  88                                 cl::desc("Enable the LoopFlatten Pass"));
  89
  90 cl::opt<bool> EnableDFAJumpThreading("enable-dfa-jump-thread",
  91                                      cl::desc("Enable DFA jump threading."),
  92                                      cl::init(false), cl::Hidden);
  93
  94 cl::opt<bool> EnableHotColdSplit("hot-cold-split",
  95                                  cl::desc("Enable hot-cold splitting pass"));
  96
  97 cl::opt<bool> EnableIROutliner("ir-outliner", cl::init(false), cl::Hidden,
  98     cl::desc("Enable ir outliner pass"));
  99
 100 static cl::opt<bool> UseLoopVersioningLICM(
 101     "enable-loop-versioning-licm", cl::init(false), cl::Hidden,
 102     cl::desc("Enable the experimental Loop Versioning LICM pass"));
 103
 104 cl::opt<bool>
 105     DisablePreInliner("disable-preinline", cl::init(false), cl::Hidden,
 106                       cl::desc("Disable pre-instrumentation inliner"));
 107
 108 cl::opt<int> PreInlineThreshold(
 109     "preinline-threshold", cl::Hidden, cl::init(75),
 110     cl::desc("Control the amount of inlining in pre-instrumentation inliner "
 111              "(default = 75)"));
 112
 113 cl::opt<bool>
 114     EnableGVNHoist("enable-gvn-hoist",
 115                    cl::desc("Enable the GVN hoisting pass (default = off)"));
 116
 117 static cl::opt<bool>
 118     DisableLibCallsShrinkWrap("disable-libcalls-shrinkwrap", cl::init(false),
 119                               cl::Hidden,
 120                               cl::desc("Disable shrink-wrap library calls"));
 121
 122 cl::opt<bool>
 123     EnableGVNSink("enable-gvn-sink",
 124                   cl::desc("Enable the GVN sinking pass (default = off)"));
 125
 126 // This option is used in simplifying testing SampleFDO optimizations for
 127 // profile loading.
 128 cl::opt<bool>
 129     EnableCHR("enable-chr", cl::init(true), cl::Hidden,
 130               cl::desc("Enable control height reduction optimization (CHR)"));
 131
 132 cl::opt<bool> FlattenedProfileUsed(
 133     "flattened-profile-used", cl::init(false), cl::Hidden,
 134     cl::desc("Indicate the sample profile being used is flattened, i.e., "
 135              "no inline hierachy exists in the profile. "));
 136
 137 cl::opt<bool> EnableOrderFileInstrumentation(
 138     "enable-order-file-instrumentation", cl::init(false), cl::Hidden,
 139     cl::desc("Enable order file instrumentation (default = off)"));
 140
 141 cl::opt<bool> EnableMatrix(
 142     "enable-matrix", cl::init(false), cl::Hidden,
 143     cl::desc("Enable lowering of the matrix intrinsics"));
 144
 145 cl::opt<bool> EnableConstraintElimination(
 146     "enable-constraint-elimination", cl::init(false), cl::Hidden,
 147     cl::desc(
 148         "Enable pass to eliminate conditions based on linear constraints."));
 149
 150 cl::opt<bool> EnableFunctionSpecialization(
 151     "enable-function-specialization", cl::init(false), cl::Hidden,
 152     cl::desc("Enable Function Specialization pass"));
 153
 154 cl::opt<AttributorRunOption> AttributorRun(
 155     "attributor-enable", cl::Hidden, cl::init(AttributorRunOption::NONE),
 156     cl::desc("Enable the attributor inter-procedural deduction pass."),
 157     cl::values(clEnumValN(AttributorRunOption::ALL, "all",
 158                           "enable all attributor runs"),
 159                clEnumValN(AttributorRunOption::MODULE, "module",
 160                           "enable module-wide attributor runs"),
 161                clEnumValN(AttributorRunOption::CGSCC, "cgscc",
 162                           "enable call graph SCC attributor runs"),
 163                clEnumValN(AttributorRunOption::NONE, "none",
 164                           "disable attributor runs")));
 165
 166 extern cl::opt<bool> EnableKnowledgeRetention;
 167 } // namespace llvm
 168
 169 PassManagerBuilder::PassManagerBuilder() {
 170     OptLevel = 2;
 171     SizeLevel = 0;
 172     LibraryInfo = nullptr;
 173     Inliner = nullptr;
 174     DisableUnrollLoops = false;
 175     SLPVectorize = false;
 176     LoopVectorize = true;
 177     LoopsInterleaved = true;
 178     RerollLoops = RunLoopRerolling;
 179     NewGVN = RunNewGVN;
 180     LicmMssaOptCap = SetLicmMssaOptCap;
 181     LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
 182     DisableGVNLoadPRE = false;
 183     ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
 184     VerifyInput = false;
 185     VerifyOutput = false;
 186     MergeFunctions = false;
 187     DivergentTarget = false;
 188     CallGraphProfile = true;
 189 }
 190
 191 PassManagerBuilder::~PassManagerBuilder() {
 192   delete LibraryInfo;
 193   delete Inliner;
 194 }
 195
 196 /// Set of global extensions, automatically added as part of the standard set.
 197 static ManagedStatic<
 198     SmallVector<std::tuple<PassManagerBuilder::ExtensionPointTy,
 199                            PassManagerBuilder::ExtensionFn,
 200                            PassManagerBuilder::GlobalExtensionID>,
 201                 8>>
 202     GlobalExtensions;
 203 static PassManagerBuilder::GlobalExtensionID GlobalExtensionsCounter;
 204
 205 /// Check if GlobalExtensions is constructed and not empty.
 206 /// Since GlobalExtensions is a managed static, calling 'empty()' will trigger
 207 /// the construction of the object.
 208 static bool GlobalExtensionsNotEmpty() {
 209   return GlobalExtensions.isConstructed() && !GlobalExtensions->empty();
 210 }
 211
 212 PassManagerBuilder::GlobalExtensionID
 213 PassManagerBuilder::addGlobalExtension(PassManagerBuilder::ExtensionPointTy Ty,
 214                                        PassManagerBuilder::ExtensionFn Fn) {
 215   auto ExtensionID = GlobalExtensionsCounter++;
 216   GlobalExtensions->push_back(std::make_tuple(Ty, std::move(Fn), ExtensionID));
 217   return ExtensionID;
 218 }
 219
 220 void PassManagerBuilder::removeGlobalExtension(
 221     PassManagerBuilder::GlobalExtensionID ExtensionID) {
 222   // RegisterStandardPasses may try to call this function after GlobalExtensions
 223   // has already been destroyed; doing so should not generate an error.
 224   if (!GlobalExtensions.isConstructed())
 225     return;
 226
 227   auto GlobalExtension =
 228       llvm::find_if(*GlobalExtensions, [ExtensionID](const auto &elem) {
 229         return std::get<2>(elem) == ExtensionID;
 230       });
 231   assert(GlobalExtension != GlobalExtensions->end() &&
 232          "The extension ID to be removed should always be valid.");
 233
 234   GlobalExtensions->erase(GlobalExtension);
 235 }
 236
 237 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
 238   Extensions.push_back(std::make_pair(Ty, std::move(Fn)));
 239 }
 240
 241 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
 242                                            legacy::PassManagerBase &PM) const {
 243   if (GlobalExtensionsNotEmpty()) {
 244     for (auto &Ext : *GlobalExtensions) {
 245       if (std::get<0>(Ext) == ETy)
 246         std::get<1>(Ext)(*this, PM);
 247     }
 248   }
 249   for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
 250     if (Extensions[i].first == ETy)
 251       Extensions[i].second(*this, PM);
 252 }
 253
 254 void PassManagerBuilder::addInitialAliasAnalysisPasses(
 255     legacy::PassManagerBase &PM) const {
 256   switch (UseCFLAA) {
 257   case ::CFLAAType::Steensgaard:
 258     PM.add(createCFLSteensAAWrapperPass());
 259     break;
 260   case ::CFLAAType::Andersen:
 261     PM.add(createCFLAndersAAWrapperPass());
 262     break;
 263   case ::CFLAAType::Both:
 264     PM.add(createCFLSteensAAWrapperPass());
 265     PM.add(createCFLAndersAAWrapperPass());
 266     break;
 267   default:
 268     break;
 269   }
 270
 271   // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
 272   // BasicAliasAnalysis wins if they disagree. This is intended to help
 273   // support "obvious" type-punning idioms.
 274   PM.add(createTypeBasedAAWrapperPass());
 275   PM.add(createScopedNoAliasAAWrapperPass());
 276 }
 277
 278 void PassManagerBuilder::populateFunctionPassManager(
 279     legacy::FunctionPassManager &FPM) {
 280   addExtensionsToPM(EP_EarlyAsPossible, FPM);
 281
 282   // Add LibraryInfo if we have some.
 283   if (LibraryInfo)
 284     FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
 285
 286   // The backends do not handle matrix intrinsics currently.
 287   // Make sure they are also lowered in O0.
 288   // FIXME: A lightweight version of the pass should run in the backend
 289   //        pipeline on demand.
 290   if (EnableMatrix && OptLevel == 0)
 291     FPM.add(createLowerMatrixIntrinsicsMinimalPass());
 292
 293   if (OptLevel == 0) return;
 294
 295   addInitialAliasAnalysisPasses(FPM);
 296
 297   // Lower llvm.expect to metadata before attempting transforms.
 298   // Compare/branch metadata may alter the behavior of passes like SimplifyCFG.
 299   FPM.add(createLowerExpectIntrinsicPass());
 300   FPM.add(createCFGSimplificationPass());
 301   FPM.add(createSROAPass());
 302   FPM.add(createEarlyCSEPass());
 303 }
 304
 305 void PassManagerBuilder::addFunctionSimplificationPasses(
 306     legacy::PassManagerBase &MPM) {
 307   // Start of function pass.
 308   // Break up aggregate allocas, using SSAUpdater.
 309   assert(OptLevel >= 1 && "Calling function optimizer with no optimization level!");
 310   MPM.add(createSROAPass());
 311   MPM.add(createEarlyCSEPass(true /* Enable mem-ssa. */)); // Catch trivial redundancies
 312   if (EnableKnowledgeRetention)
 313     MPM.add(createAssumeSimplifyPass());
 314
 315   if (OptLevel > 1) {
 316     if (EnableGVNHoist)
 317       MPM.add(createGVNHoistPass());
 318     if (EnableGVNSink) {
 319       MPM.add(createGVNSinkPass());
 320       MPM.add(createCFGSimplificationPass(
 321           SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
 322     }
 323   }
 324
 325   if (EnableConstraintElimination)
 326     MPM.add(createConstraintEliminationPass());
 327
 328   if (OptLevel > 1) {
 329     // Speculative execution if the target has divergent branches; otherwise nop.
 330     MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass());
 331
 332     MPM.add(createJumpThreadingPass());         // Thread jumps.
 333     MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
 334   }
 335   MPM.add(
 336       createCFGSimplificationPass(SimplifyCFGOptions().convertSwitchRangeToICmp(
 337           true))); // Merge & remove BBs
 338   // Combine silly seq's
 339   if (OptLevel > 2)
 340     MPM.add(createAggressiveInstCombinerPass());
 341   MPM.add(createInstructionCombiningPass());
 342   if (SizeLevel == 0 && !DisableLibCallsShrinkWrap)
 343     MPM.add(createLibCallsShrinkWrapPass());
 344   addExtensionsToPM(EP_Peephole, MPM);
 345
 346   // TODO: Investigate the cost/benefit of tail call elimination on debugging.
 347   if (OptLevel > 1)
 348     MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
 349   MPM.add(
 350       createCFGSimplificationPass(SimplifyCFGOptions().convertSwitchRangeToICmp(
 351           true)));                            // Merge & remove BBs
 352   MPM.add(createReassociatePass());           // Reassociate expressions
 353
 354   // The matrix extension can introduce large vector operations early, which can
 355   // benefit from running vector-combine early on.
 356   if (EnableMatrix)
 357     MPM.add(createVectorCombinePass());
 358
 359   // Begin the loop pass pipeline.
 360
 361   // The simple loop unswitch pass relies on separate cleanup passes. Schedule
 362   // them first so when we re-process a loop they run before other loop
 363   // passes.
 364   MPM.add(createLoopInstSimplifyPass());
 365   MPM.add(createLoopSimplifyCFGPass());
 366
 367   // Try to remove as much code from the loop header as possible,
 368   // to reduce amount of IR that will have to be duplicated. However,
 369   // do not perform speculative hoisting the first time as LICM
 370   // will destroy metadata that may not need to be destroyed if run
 371   // after loop rotation.
 372   // TODO: Investigate promotion cap for O1.
 373   MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 374                          /*AllowSpeculation=*/false));
 375   // Rotate Loop - disable header duplication at -Oz
 376   MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1, false));
 377   // TODO: Investigate promotion cap for O1.
 378   MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 379                          /*AllowSpeculation=*/true));
 380   MPM.add(createSimpleLoopUnswitchLegacyPass(OptLevel == 3));
 381   // FIXME: We break the loop pass pipeline here in order to do full
 382   // simplifycfg. Eventually loop-simplifycfg should be enhanced to replace the
 383   // need for this.
 384   MPM.add(createCFGSimplificationPass(
 385       SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
 386   MPM.add(createInstructionCombiningPass());
 387   // We resume loop passes creating a second loop pipeline here.
 388   if (EnableLoopFlatten) {
 389     MPM.add(createLoopFlattenPass()); // Flatten loops
 390     MPM.add(createLoopSimplifyCFGPass());
 391   }
 392   MPM.add(createLoopIdiomPass());             // Recognize idioms like memset.
 393   MPM.add(createIndVarSimplifyPass());        // Canonicalize indvars
 394   addExtensionsToPM(EP_LateLoopOptimizations, MPM);
 395   MPM.add(createLoopDeletionPass());          // Delete dead loops
 396
 397   if (EnableLoopInterchange)
 398     MPM.add(createLoopInterchangePass()); // Interchange loops
 399
 400   // Unroll small loops and perform peeling.
 401   MPM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops,
 402                                      ForgetAllSCEVInLoopUnroll));
 403   addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
 404   // This ends the loop pass pipelines.
 405
 406   // Break up allocas that may now be splittable after loop unrolling.
 407   MPM.add(createSROAPass());
 408
 409   if (OptLevel > 1) {
 410     MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
 411     MPM.add(NewGVN ? createNewGVNPass()
 412                    : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
 413   }
 414   MPM.add(createSCCPPass());                  // Constant prop with SCCP
 415
 416   if (EnableConstraintElimination)
 417     MPM.add(createConstraintEliminationPass());
 418
 419   // Delete dead bit computations (instcombine runs after to fold away the dead
 420   // computations, and then ADCE will run later to exploit any new DCE
 421   // opportunities that creates).
 422   MPM.add(createBitTrackingDCEPass());        // Delete dead bit computations
 423
 424   // Run instcombine after redundancy elimination to exploit opportunities
 425   // opened up by them.
 426   MPM.add(createInstructionCombiningPass());
 427   addExtensionsToPM(EP_Peephole, MPM);
 428   if (OptLevel > 1) {
 429     if (EnableDFAJumpThreading && SizeLevel == 0)
 430       MPM.add(createDFAJumpThreadingPass());
 431
 432     MPM.add(createJumpThreadingPass());         // Thread jumps
 433     MPM.add(createCorrelatedValuePropagationPass());
 434   }
 435   MPM.add(createAggressiveDCEPass()); // Delete dead instructions
 436
 437   MPM.add(createMemCpyOptPass());               // Remove memcpy / form memset
 438   // TODO: Investigate if this is too expensive at O1.
 439   if (OptLevel > 1) {
 440     MPM.add(createDeadStoreEliminationPass());  // Delete dead stores
 441     MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 442                            /*AllowSpeculation=*/true));
 443   }
 444
 445   addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
 446
 447   if (RerollLoops)
 448     MPM.add(createLoopRerollPass());
 449
 450   // Merge & remove BBs and sink & hoist common instructions.
 451   MPM.add(createCFGSimplificationPass(
 452       SimplifyCFGOptions().hoistCommonInsts(true).sinkCommonInsts(true)));
 453   // Clean up after everything.
 454   MPM.add(createInstructionCombiningPass());
 455   addExtensionsToPM(EP_Peephole, MPM);
 456 }
 457
 458 /// FIXME: Should LTO cause any differences to this set of passes?
 459 void PassManagerBuilder::addVectorPasses(legacy::PassManagerBase &PM,
 460                                          bool IsFullLTO) {
 461   PM.add(createLoopVectorizePass(!LoopsInterleaved, !LoopVectorize));
 462
 463   if (IsFullLTO) {
 464     // The vectorizer may have significantly shortened a loop body; unroll
 465     // again. Unroll small loops to hide loop backedge latency and saturate any
 466     // parallel execution resources of an out-of-order processor. We also then
 467     // need to clean up redundancies and loop invariant code.
 468     // FIXME: It would be really good to use a loop-integrated instruction
 469     // combiner for cleanup here so that the unrolling and LICM can be pipelined
 470     // across the loop nests.
 471     // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
 472     if (EnableUnrollAndJam && !DisableUnrollLoops)
 473       PM.add(createLoopUnrollAndJamPass(OptLevel));
 474     PM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops,
 475                                 ForgetAllSCEVInLoopUnroll));
 476     PM.add(createWarnMissedTransformationsPass());
 477   }
 478
 479   if (!IsFullLTO) {
 480     // Eliminate loads by forwarding stores from the previous iteration to loads
 481     // of the current iteration.
 482     PM.add(createLoopLoadEliminationPass());
 483   }
 484   // Cleanup after the loop optimization passes.
 485   PM.add(createInstructionCombiningPass());
 486
 487   if (OptLevel > 1 && ExtraVectorizerPasses) {
 488     // At higher optimization levels, try to clean up any runtime overlap and
 489     // alignment checks inserted by the vectorizer. We want to track correlated
 490     // runtime checks for two inner loops in the same outer loop, fold any
 491     // common computations, hoist loop-invariant aspects out of any outer loop,
 492     // and unswitch the runtime checks if possible. Once hoisted, we may have
 493     // dead (or speculatable) control flows or more combining opportunities.
 494     PM.add(createEarlyCSEPass());
 495     PM.add(createCorrelatedValuePropagationPass());
 496     PM.add(createInstructionCombiningPass());
 497     PM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 498                           /*AllowSpeculation=*/true));
 499     PM.add(createSimpleLoopUnswitchLegacyPass());
 500     PM.add(createCFGSimplificationPass(
 501         SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
 502     PM.add(createInstructionCombiningPass());
 503   }
 504
 505   // Now that we've formed fast to execute loop structures, we do further
 506   // optimizations. These are run afterward as they might block doing complex
 507   // analyses and transforms such as what are needed for loop vectorization.
 508
 509   // Cleanup after loop vectorization, etc. Simplification passes like CVP and
 510   // GVN, loop transforms, and others have already run, so it's now better to
 511   // convert to more optimized IR using more aggressive simplify CFG options.
 512   // The extra sinking transform can create larger basic blocks, so do this
 513   // before SLP vectorization.
 514   PM.add(createCFGSimplificationPass(SimplifyCFGOptions()
 515                                          .forwardSwitchCondToPhi(true)
 516                                          .convertSwitchRangeToICmp(true)
 517                                          .convertSwitchToLookupTable(true)
 518                                          .needCanonicalLoops(false)
 519                                          .hoistCommonInsts(true)
 520                                          .sinkCommonInsts(true)));
 521
 522   if (IsFullLTO) {
 523     PM.add(createSCCPPass());                 // Propagate exposed constants
 524     PM.add(createInstructionCombiningPass()); // Clean up again
 525     PM.add(createBitTrackingDCEPass());
 526   }
 527
 528   // Optimize parallel scalar instruction chains into SIMD instructions.
 529   if (SLPVectorize) {
 530     PM.add(createSLPVectorizerPass());
 531     if (OptLevel > 1 && ExtraVectorizerPasses)
 532       PM.add(createEarlyCSEPass());
 533   }
 534
 535   // Enhance/cleanup vector code.
 536   PM.add(createVectorCombinePass());
 537
 538   if (!IsFullLTO) {
 539     addExtensionsToPM(EP_Peephole, PM);
 540     PM.add(createInstructionCombiningPass());
 541
 542     if (EnableUnrollAndJam && !DisableUnrollLoops) {
 543       // Unroll and Jam. We do this before unroll but need to be in a separate
 544       // loop pass manager in order for the outer loop to be processed by
 545       // unroll and jam before the inner loop is unrolled.
 546       PM.add(createLoopUnrollAndJamPass(OptLevel));
 547     }
 548
 549     // Unroll small loops
 550     PM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops,
 551                                 ForgetAllSCEVInLoopUnroll));
 552
 553     if (!DisableUnrollLoops) {
 554       // LoopUnroll may generate some redundency to cleanup.
 555       PM.add(createInstructionCombiningPass());
 556
 557       // Runtime unrolling will introduce runtime check in loop prologue. If the
 558       // unrolled loop is a inner loop, then the prologue will be inside the
 559       // outer loop. LICM pass can help to promote the runtime check out if the
 560       // checked value is loop invariant.
 561       PM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 562                             /*AllowSpeculation=*/true));
 563     }
 564
 565     PM.add(createWarnMissedTransformationsPass());
 566   }
 567
 568   // After vectorization and unrolling, assume intrinsics may tell us more
 569   // about pointer alignments.
 570   PM.add(createAlignmentFromAssumptionsPass());
 571
 572   if (IsFullLTO)
 573     PM.add(createInstructionCombiningPass());
 574 }
 575
 576 void PassManagerBuilder::populateModulePassManager(
 577     legacy::PassManagerBase &MPM) {
 578   MPM.add(createAnnotation2MetadataLegacyPass());
 579
 580   // Allow forcing function attributes as a debugging and tuning aid.
 581   MPM.add(createForceFunctionAttrsLegacyPass());
 582
 583   // If all optimizations are disabled, just run the always-inline pass and,
 584   // if enabled, the function merging pass.
 585   if (OptLevel == 0) {
 586     if (Inliner) {
 587       MPM.add(Inliner);
 588       Inliner = nullptr;
 589     }
 590
 591     // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
 592     // creates a CGSCC pass manager, but we don't want to add extensions into
 593     // that pass manager. To prevent this we insert a no-op module pass to reset
 594     // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
 595     // builds. The function merging pass is
 596     if (MergeFunctions)
 597       MPM.add(createMergeFunctionsPass());
 598     else if (GlobalExtensionsNotEmpty() || !Extensions.empty())
 599       MPM.add(createBarrierNoopPass());
 600
 601     addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
 602
 603     MPM.add(createAnnotationRemarksLegacyPass());
 604     return;
 605   }
 606
 607   // Add LibraryInfo if we have some.
 608   if (LibraryInfo)
 609     MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
 610
 611   addInitialAliasAnalysisPasses(MPM);
 612
 613   // Infer attributes about declarations if possible.
 614   MPM.add(createInferFunctionAttrsLegacyPass());
 615
 616   // Infer attributes on declarations, call sites, arguments, etc.
 617   if (AttributorRun & AttributorRunOption::MODULE)
 618     MPM.add(createAttributorLegacyPass());
 619
 620   addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
 621
 622   if (OptLevel > 2)
 623     MPM.add(createCallSiteSplittingPass());
 624
 625   // Propage constant function arguments by specializing the functions.
 626   if (OptLevel > 2 && EnableFunctionSpecialization)
 627     MPM.add(createFunctionSpecializationPass());
 628
 629   MPM.add(createIPSCCPPass());          // IP SCCP
 630   MPM.add(createCalledValuePropagationPass());
 631
 632   MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
 633   // Promote any localized global vars.
 634   MPM.add(createPromoteMemoryToRegisterPass());
 635
 636   MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
 637
 638   MPM.add(createInstructionCombiningPass()); // Clean up after IPCP & DAE
 639   addExtensionsToPM(EP_Peephole, MPM);
 640   MPM.add(
 641       createCFGSimplificationPass(SimplifyCFGOptions().convertSwitchRangeToICmp(
 642           true))); // Clean up after IPCP & DAE
 643
 644   // We add a module alias analysis pass here. In part due to bugs in the
 645   // analysis infrastructure this "works" in that the analysis stays alive
 646   // for the entire SCC pass run below.
 647   MPM.add(createGlobalsAAWrapperPass());
 648
 649   // Start of CallGraph SCC passes.
 650   MPM.add(createPruneEHPass()); // Remove dead EH info
 651   bool RunInliner = false;
 652   if (Inliner) {
 653     MPM.add(Inliner);
 654     Inliner = nullptr;
 655     RunInliner = true;
 656   }
 657
 658   // Infer attributes on declarations, call sites, arguments, etc. for an SCC.
 659   if (AttributorRun & AttributorRunOption::CGSCC)
 660     MPM.add(createAttributorCGSCCLegacyPass());
 661
 662   // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if
 663   // there are no OpenMP runtime calls present in the module.
 664   if (OptLevel > 1)
 665     MPM.add(createOpenMPOptCGSCCLegacyPass());
 666
 667   MPM.add(createPostOrderFunctionAttrsLegacyPass());
 668
 669   addExtensionsToPM(EP_CGSCCOptimizerLate, MPM);
 670   addFunctionSimplificationPasses(MPM);
 671
 672   // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
 673   // pass manager that we are specifically trying to avoid. To prevent this
 674   // we must insert a no-op module pass to reset the pass manager.
 675   MPM.add(createBarrierNoopPass());
 676
 677   if (RunPartialInlining)
 678     MPM.add(createPartialInliningPass());
 679
 680   if (OptLevel > 1)
 681     // Remove avail extern fns and globals definitions if we aren't
 682     // compiling an object file for later LTO. For LTO we want to preserve
 683     // these so they are eligible for inlining at link-time. Note if they
 684     // are unreferenced they will be removed by GlobalDCE later, so
 685     // this only impacts referenced available externally globals.
 686     // Eventually they will be suppressed during codegen, but eliminating
 687     // here enables more opportunity for GlobalDCE as it may make
 688     // globals referenced by available external functions dead
 689     // and saves running remaining passes on the eliminated functions.
 690     MPM.add(createEliminateAvailableExternallyPass());
 691
 692   MPM.add(createReversePostOrderFunctionAttrsPass());
 693
 694   // The inliner performs some kind of dead code elimination as it goes,
 695   // but there are cases that are not really caught by it. We might
 696   // at some point consider teaching the inliner about them, but it
 697   // is OK for now to run GlobalOpt + GlobalDCE in tandem as their
 698   // benefits generally outweight the cost, making the whole pipeline
 699   // faster.
 700   if (RunInliner) {
 701     MPM.add(createGlobalOptimizerPass());
 702     MPM.add(createGlobalDCEPass());
 703   }
 704
 705   // Scheduling LoopVersioningLICM when inlining is over, because after that
 706   // we may see more accurate aliasing. Reason to run this late is that too
 707   // early versioning may prevent further inlining due to increase of code
 708   // size. By placing it just after inlining other optimizations which runs
 709   // later might get benefit of no-alias assumption in clone loop.
 710   if (UseLoopVersioningLICM) {
 711     MPM.add(createLoopVersioningLICMPass());    // Do LoopVersioningLICM
 712     MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
 713                            /*AllowSpeculation=*/true));
 714   }
 715
 716   // We add a fresh GlobalsModRef run at this point. This is particularly
 717   // useful as the above will have inlined, DCE'ed, and function-attr
 718   // propagated everything. We should at this point have a reasonably minimal
 719   // and richly annotated call graph. By computing aliasing and mod/ref
 720   // information for all local globals here, the late loop passes and notably
 721   // the vectorizer will be able to use them to help recognize vectorizable
 722   // memory operations.
 723   //
 724   // Note that this relies on a bug in the pass manager which preserves
 725   // a module analysis into a function pass pipeline (and throughout it) so
 726   // long as the first function pass doesn't invalidate the module analysis.
 727   // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for
 728   // this to work. Fortunately, it is trivial to preserve AliasAnalysis
 729   // (doing nothing preserves it as it is required to be conservatively
 730   // correct in the face of IR changes).
 731   MPM.add(createGlobalsAAWrapperPass());
 732
 733   MPM.add(createFloat2IntPass());
 734   MPM.add(createLowerConstantIntrinsicsPass());
 735
 736   if (EnableMatrix) {
 737     MPM.add(createLowerMatrixIntrinsicsPass());
 738     // CSE the pointer arithmetic of the column vectors.  This allows alias
 739     // analysis to establish no-aliasing between loads and stores of different
 740     // columns of the same matrix.
 741     MPM.add(createEarlyCSEPass(false));
 742   }
 743
 744   addExtensionsToPM(EP_VectorizerStart, MPM);
 745
 746   // Re-rotate loops in all our loop nests. These may have fallout out of
 747   // rotated form due to GVN or other transformations, and the vectorizer relies
 748   // on the rotated form. Disable header duplication at -Oz.
 749   MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1, false));
 750
 751   // Distribute loops to allow partial vectorization.  I.e. isolate dependences
 752   // into separate loop that would otherwise inhibit vectorization.  This is
 753   // currently only performed for loops marked with the metadata
 754   // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
 755   MPM.add(createLoopDistributePass());
 756
 757   addVectorPasses(MPM, /* IsFullLTO */ false);
 758
 759   // FIXME: We shouldn't bother with this anymore.
 760   MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
 761
 762   // GlobalOpt already deletes dead functions and globals, at -O2 try a
 763   // late pass of GlobalDCE.  It is capable of deleting dead cycles.
 764   if (OptLevel > 1) {
 765     MPM.add(createGlobalDCEPass());         // Remove dead fns and globals.
 766     MPM.add(createConstantMergePass());     // Merge dup global constants
 767   }
 768
 769   // See comment in the new PM for justification of scheduling splitting at
 770   // this stage (\ref buildModuleSimplificationPipeline).
 771   if (EnableHotColdSplit)
 772     MPM.add(createHotColdSplittingPass());
 773
 774   if (EnableIROutliner)
 775     MPM.add(createIROutlinerPass());
 776
 777   if (MergeFunctions)
 778     MPM.add(createMergeFunctionsPass());
 779
 780   // LoopSink pass sinks instructions hoisted by LICM, which serves as a
 781   // canonicalization pass that enables other optimizations. As a result,
 782   // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
 783   // result too early.
 784   MPM.add(createLoopSinkPass());
 785   // Get rid of LCSSA nodes.
 786   MPM.add(createInstSimplifyLegacyPass());
 787
 788   // This hoists/decomposes div/rem ops. It should run after other sink/hoist
 789   // passes to avoid re-sinking, but before SimplifyCFG because it can allow
 790   // flattening of blocks.
 791   MPM.add(createDivRemPairsPass());
 792
 793   // LoopSink (and other loop passes since the last simplifyCFG) might have
 794   // resulted in single-entry-single-exit or empty blocks. Clean up the CFG.
 795   MPM.add(createCFGSimplificationPass(
 796       SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
 797
 798   addExtensionsToPM(EP_OptimizerLast, MPM);
 799
 800   MPM.add(createAnnotationRemarksLegacyPass());
 801 }
 802
 803 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
 804   PassManagerBuilder *PMB = new PassManagerBuilder();
 805   return wrap(PMB);
 806 }
 807
 808 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
 809   PassManagerBuilder *Builder = unwrap(PMB);
 810   delete Builder;
 811 }
 812
 813 void
 814 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
 815                                   unsigned OptLevel) {
 816   PassManagerBuilder *Builder = unwrap(PMB);
 817   Builder->OptLevel = OptLevel;
 818 }
 819
 820 void
 821 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
 822                                    unsigned SizeLevel) {
 823   PassManagerBuilder *Builder = unwrap(PMB);
 824   Builder->SizeLevel = SizeLevel;
 825 }
 826
 827 void
 828 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
 829                                             LLVMBool Value) {
 830   // NOTE: The DisableUnitAtATime switch has been removed.
 831 }
 832
 833 void
 834 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
 835                                             LLVMBool Value) {
 836   PassManagerBuilder *Builder = unwrap(PMB);
 837   Builder->DisableUnrollLoops = Value;
 838 }
 839
 840 void
 841 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
 842                                                  LLVMBool Value) {
 843   // NOTE: The simplify-libcalls pass has been removed.
 844 }
 845
 846 void
 847 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
 848                                               unsigned Threshold) {
 849   PassManagerBuilder *Builder = unwrap(PMB);
 850   Builder->Inliner = createFunctionInliningPass(Threshold);
 851 }
 852
 853 void
 854 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
 855                                                   LLVMPassManagerRef PM) {
 856   PassManagerBuilder *Builder = unwrap(PMB);
 857   legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
 858   Builder->populateFunctionPassManager(*FPM);
 859 }
 860
 861 void
 862 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
 863                                                 LLVMPassManagerRef PM) {
 864   PassManagerBuilder *Builder = unwrap(PMB);
 865   legacy::PassManagerBase *MPM = unwrap(PM);
 866   Builder->populateModulePassManager(*MPM);
 867 }