lib/Transforms/IPO/PassManagerBuilder.cpp

   1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file defines the PassManagerBuilder class, which is used to set up a
  11 // "standard" optimization sequence suitable for languages like C and C++.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15
  16 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
  17 #include "llvm-c/Transforms/PassManagerBuilder.h"
  18 #include "llvm/ADT/SmallVector.h"
  19 #include "llvm/Analysis/Passes.h"
  20 #include "llvm/IR/DataLayout.h"
  21 #include "llvm/IR/Verifier.h"
  22 #include "llvm/PassManager.h"
  23 #include "llvm/Support/CommandLine.h"
  24 #include "llvm/Support/ManagedStatic.h"
  25 #include "llvm/Target/TargetLibraryInfo.h"
  26 #include "llvm/Target/TargetMachine.h"
  27 #include "llvm/Target/TargetSubtargetInfo.h"
  28 #include "llvm/Transforms/IPO.h"
  29 #include "llvm/Transforms/Scalar.h"
  30 #include "llvm/Transforms/Vectorize.h"
  31
  32 using namespace llvm;
  33
  34 static cl::opt<bool>
  35 RunLoopVectorization("vectorize-loops", cl::Hidden,
  36                      cl::desc("Run the Loop vectorization passes"));
  37
  38 static cl::opt<bool>
  39 RunSLPVectorization("vectorize-slp", cl::Hidden,
  40                     cl::desc("Run the SLP vectorization passes"));
  41
  42 static cl::opt<bool>
  43 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden,
  44                     cl::desc("Run the BB vectorization passes"));
  45
  46 static cl::opt<bool>
  47 UseGVNAfterVectorization("use-gvn-after-vectorization",
  48   cl::init(false), cl::Hidden,
  49   cl::desc("Run GVN instead of Early CSE after vectorization passes"));
  50
  51 static cl::opt<bool> UseNewSROA("use-new-sroa",
  52   cl::init(true), cl::Hidden,
  53   cl::desc("Enable the new, experimental SROA pass"));
  54
  55 static cl::opt<bool>
  56 RunLoopRerolling("reroll-loops", cl::Hidden,
  57                  cl::desc("Run the loop rerolling pass"));
  58
  59 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
  60                                     cl::Hidden,
  61                                     cl::desc("Run the load combining pass"));
  62
  63 static cl::opt<bool>
  64 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization",
  65   cl::init(true), cl::Hidden,
  66   cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop "
  67            "vectorizer instead of before"));
  68
  69 static cl::opt<bool> UseCFLAA("use-cfl-aa",
  70   cl::init(false), cl::Hidden,
  71   cl::desc("Enable the new, experimental CFL alias analysis"));
  72
  73 PassManagerBuilder::PassManagerBuilder() {
  74     OptLevel = 2;
  75     SizeLevel = 0;
  76     LibraryInfo = nullptr;
  77     Inliner = nullptr;
  78     DisableTailCalls = false;
  79     DisableUnitAtATime = false;
  80     DisableUnrollLoops = false;
  81     BBVectorize = RunBBVectorization;
  82     SLPVectorize = RunSLPVectorization;
  83     LoopVectorize = RunLoopVectorization;
  84     RerollLoops = RunLoopRerolling;
  85     LoadCombine = RunLoadCombine;
  86     DisableGVNLoadPRE = false;
  87     VerifyInput = false;
  88     VerifyOutput = false;
  89     StripDebug = false;
  90 }
  91
  92 PassManagerBuilder::~PassManagerBuilder() {
  93   delete LibraryInfo;
  94   delete Inliner;
  95 }
  96
  97 /// Set of global extensions, automatically added as part of the standard set.
  98 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
  99    PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
 100
 101 void PassManagerBuilder::addGlobalExtension(
 102     PassManagerBuilder::ExtensionPointTy Ty,
 103     PassManagerBuilder::ExtensionFn Fn) {
 104   GlobalExtensions->push_back(std::make_pair(Ty, Fn));
 105 }
 106
 107 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
 108   Extensions.push_back(std::make_pair(Ty, Fn));
 109 }
 110
 111 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
 112                                            PassManagerBase &PM) const {
 113   for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
 114     if ((*GlobalExtensions)[i].first == ETy)
 115       (*GlobalExtensions)[i].second(*this, PM);
 116   for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
 117     if (Extensions[i].first == ETy)
 118       Extensions[i].second(*this, PM);
 119 }
 120
 121 void
 122 PassManagerBuilder::addInitialAliasAnalysisPasses(PassManagerBase &PM) const {
 123   // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
 124   // BasicAliasAnalysis wins if they disagree. This is intended to help
 125   // support "obvious" type-punning idioms.
 126   if (UseCFLAA)
 127     PM.add(createCFLAliasAnalysisPass());
 128   PM.add(createTypeBasedAliasAnalysisPass());
 129   PM.add(createScopedNoAliasAAPass());
 130   PM.add(createBasicAliasAnalysisPass());
 131 }
 132
 133 void PassManagerBuilder::populateFunctionPassManager(FunctionPassManager &FPM) {
 134   addExtensionsToPM(EP_EarlyAsPossible, FPM);
 135
 136   // Add LibraryInfo if we have some.
 137   if (LibraryInfo) FPM.add(new TargetLibraryInfo(*LibraryInfo));
 138
 139   if (OptLevel == 0) return;
 140
 141   addInitialAliasAnalysisPasses(FPM);
 142
 143   FPM.add(createCFGSimplificationPass());
 144   if (UseNewSROA)
 145     FPM.add(createSROAPass());
 146   else
 147     FPM.add(createScalarReplAggregatesPass());
 148   FPM.add(createEarlyCSEPass());
 149   FPM.add(createLowerExpectIntrinsicPass());
 150 }
 151
 152 void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
 153   // If all optimizations are disabled, just run the always-inline pass.
 154   if (OptLevel == 0) {
 155     if (Inliner) {
 156       MPM.add(Inliner);
 157       Inliner = nullptr;
 158     }
 159
 160     // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
 161     // pass manager, but we don't want to add extensions into that pass manager.
 162     // To prevent this we must insert a no-op module pass to reset the pass
 163     // manager to get the same behavior as EP_OptimizerLast in non-O0 builds.
 164     if (!GlobalExtensions->empty() || !Extensions.empty())
 165       MPM.add(createBarrierNoopPass());
 166
 167     addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
 168     return;
 169   }
 170
 171   // Add LibraryInfo if we have some.
 172   if (LibraryInfo) MPM.add(new TargetLibraryInfo(*LibraryInfo));
 173
 174   addInitialAliasAnalysisPasses(MPM);
 175
 176   if (!DisableUnitAtATime) {
 177     addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
 178
 179     MPM.add(createIPSCCPPass());              // IP SCCP
 180     MPM.add(createGlobalOptimizerPass());     // Optimize out global vars
 181
 182     MPM.add(createDeadArgEliminationPass());  // Dead argument elimination
 183
 184     MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
 185     addExtensionsToPM(EP_Peephole, MPM);
 186     MPM.add(createCFGSimplificationPass());   // Clean up after IPCP & DAE
 187   }
 188
 189   // Start of CallGraph SCC passes.
 190   if (!DisableUnitAtATime)
 191     MPM.add(createPruneEHPass());             // Remove dead EH info
 192   if (Inliner) {
 193     MPM.add(Inliner);
 194     Inliner = nullptr;
 195   }
 196   if (!DisableUnitAtATime)
 197     MPM.add(createFunctionAttrsPass());       // Set readonly/readnone attrs
 198   if (OptLevel > 2)
 199     MPM.add(createArgumentPromotionPass());   // Scalarize uninlined fn args
 200
 201   // Start of function pass.
 202   // Break up aggregate allocas, using SSAUpdater.
 203   if (UseNewSROA)
 204     MPM.add(createSROAPass(/*RequiresDomTree*/ false));
 205   else
 206     MPM.add(createScalarReplAggregatesPass(-1, false));
 207   MPM.add(createEarlyCSEPass());              // Catch trivial redundancies
 208   MPM.add(createJumpThreadingPass());         // Thread jumps.
 209   MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
 210   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
 211   MPM.add(createInstructionCombiningPass());  // Combine silly seq's
 212   addExtensionsToPM(EP_Peephole, MPM);
 213
 214   if (!DisableTailCalls)
 215     MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
 216   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
 217   MPM.add(createReassociatePass());           // Reassociate expressions
 218   MPM.add(createLoopRotatePass());            // Rotate Loop
 219   MPM.add(createLICMPass());                  // Hoist loop invariants
 220   MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
 221   MPM.add(createInstructionCombiningPass());
 222   MPM.add(createIndVarSimplifyPass());        // Canonicalize indvars
 223   MPM.add(createLoopIdiomPass());             // Recognize idioms like memset.
 224   MPM.add(createLoopDeletionPass());          // Delete dead loops
 225
 226   if (!DisableUnrollLoops)
 227     MPM.add(createSimpleLoopUnrollPass());    // Unroll small loops
 228   addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
 229
 230   if (OptLevel > 1) {
 231     MPM.add(createMergedLoadStoreMotionPass()); // Merge load/stores in diamond
 232     MPM.add(createGVNPass(DisableGVNLoadPRE));  // Remove redundancies
 233   }
 234   MPM.add(createMemCpyOptPass());             // Remove memcpy / form memset
 235   MPM.add(createSCCPPass());                  // Constant prop with SCCP
 236
 237   // Run instcombine after redundancy elimination to exploit opportunities
 238   // opened up by them.
 239   MPM.add(createInstructionCombiningPass());
 240   addExtensionsToPM(EP_Peephole, MPM);
 241   MPM.add(createJumpThreadingPass());         // Thread jumps
 242   MPM.add(createCorrelatedValuePropagationPass());
 243   MPM.add(createDeadStoreEliminationPass());  // Delete dead stores
 244
 245   addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
 246
 247   if (RerollLoops)
 248     MPM.add(createLoopRerollPass());
 249   if (!RunSLPAfterLoopVectorization) {
 250     if (SLPVectorize)
 251       MPM.add(createSLPVectorizerPass());   // Vectorize parallel scalar chains.
 252
 253     if (BBVectorize) {
 254       MPM.add(createBBVectorizePass());
 255       MPM.add(createInstructionCombiningPass());
 256       addExtensionsToPM(EP_Peephole, MPM);
 257       if (OptLevel > 1 && UseGVNAfterVectorization)
 258         MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
 259       else
 260         MPM.add(createEarlyCSEPass());      // Catch trivial redundancies
 261
 262       // BBVectorize may have significantly shortened a loop body; unroll again.
 263       if (!DisableUnrollLoops)
 264         MPM.add(createLoopUnrollPass());
 265     }
 266   }
 267
 268   if (LoadCombine)
 269     MPM.add(createLoadCombinePass());
 270
 271   MPM.add(createAggressiveDCEPass());         // Delete dead instructions
 272   MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
 273   MPM.add(createInstructionCombiningPass());  // Clean up after everything.
 274   addExtensionsToPM(EP_Peephole, MPM);
 275
 276   // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
 277   // pass manager that we are specifically trying to avoid. To prevent this
 278   // we must insert a no-op module pass to reset the pass manager.
 279   MPM.add(createBarrierNoopPass());
 280   MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
 281   // FIXME: Because of #pragma vectorize enable, the passes below are always
 282   // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
 283   // on -O1 and no #pragma is found). Would be good to have these two passes
 284   // as function calls, so that we can only pass them when the vectorizer
 285   // changed the code.
 286   MPM.add(createInstructionCombiningPass());
 287
 288   if (RunSLPAfterLoopVectorization) {
 289     if (SLPVectorize)
 290       MPM.add(createSLPVectorizerPass());   // Vectorize parallel scalar chains.
 291
 292     if (BBVectorize) {
 293       MPM.add(createBBVectorizePass());
 294       MPM.add(createInstructionCombiningPass());
 295       addExtensionsToPM(EP_Peephole, MPM);
 296       if (OptLevel > 1 && UseGVNAfterVectorization)
 297         MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
 298       else
 299         MPM.add(createEarlyCSEPass());      // Catch trivial redundancies
 300
 301       // BBVectorize may have significantly shortened a loop body; unroll again.
 302       if (!DisableUnrollLoops)
 303         MPM.add(createLoopUnrollPass());
 304     }
 305   }
 306
 307   addExtensionsToPM(EP_Peephole, MPM);
 308   MPM.add(createCFGSimplificationPass());
 309
 310   if (!DisableUnrollLoops)
 311     MPM.add(createLoopUnrollPass());    // Unroll small loops
 312
 313   if (!DisableUnitAtATime) {
 314     // FIXME: We shouldn't bother with this anymore.
 315     MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
 316
 317     // GlobalOpt already deletes dead functions and globals, at -O2 try a
 318     // late pass of GlobalDCE.  It is capable of deleting dead cycles.
 319     if (OptLevel > 1) {
 320       MPM.add(createGlobalDCEPass());         // Remove dead fns and globals.
 321       MPM.add(createConstantMergePass());     // Merge dup global constants
 322     }
 323   }
 324   addExtensionsToPM(EP_OptimizerLast, MPM);
 325 }
 326
 327 void PassManagerBuilder::addLTOOptimizationPasses(PassManagerBase &PM) {
 328   // Provide AliasAnalysis services for optimizations.
 329   addInitialAliasAnalysisPasses(PM);
 330
 331   // Propagate constants at call sites into the functions they call.  This
 332   // opens opportunities for globalopt (and inlining) by substituting function
 333   // pointers passed as arguments to direct uses of functions.
 334   PM.add(createIPSCCPPass());
 335
 336   // Now that we internalized some globals, see if we can hack on them!
 337   PM.add(createGlobalOptimizerPass());
 338
 339   // Linking modules together can lead to duplicated global constants, only
 340   // keep one copy of each constant.
 341   PM.add(createConstantMergePass());
 342
 343   // Remove unused arguments from functions.
 344   PM.add(createDeadArgEliminationPass());
 345
 346   // Reduce the code after globalopt and ipsccp.  Both can open up significant
 347   // simplification opportunities, and both can propagate functions through
 348   // function pointers.  When this happens, we often have to resolve varargs
 349   // calls, etc, so let instcombine do this.
 350   PM.add(createInstructionCombiningPass());
 351   addExtensionsToPM(EP_Peephole, PM);
 352
 353   // Inline small functions
 354   bool RunInliner = Inliner;
 355   if (RunInliner) {
 356     PM.add(Inliner);
 357     Inliner = nullptr;
 358   }
 359
 360   PM.add(createPruneEHPass());   // Remove dead EH info.
 361
 362   // Optimize globals again if we ran the inliner.
 363   if (RunInliner)
 364     PM.add(createGlobalOptimizerPass());
 365   PM.add(createGlobalDCEPass()); // Remove dead functions.
 366
 367   // If we didn't decide to inline a function, check to see if we can
 368   // transform it to pass arguments by value instead of by reference.
 369   PM.add(createArgumentPromotionPass());
 370
 371   // The IPO passes may leave cruft around.  Clean up after them.
 372   PM.add(createInstructionCombiningPass());
 373   addExtensionsToPM(EP_Peephole, PM);
 374   PM.add(createJumpThreadingPass());
 375
 376   // Break up allocas
 377   if (UseNewSROA)
 378     PM.add(createSROAPass());
 379   else
 380     PM.add(createScalarReplAggregatesPass());
 381
 382   // Run a few AA driven optimizations here and now, to cleanup the code.
 383   PM.add(createFunctionAttrsPass()); // Add nocapture.
 384   PM.add(createGlobalsModRefPass()); // IP alias analysis.
 385
 386   PM.add(createLICMPass());                 // Hoist loop invariants.
 387   PM.add(createMergedLoadStoreMotionPass()); // Merge load/stores in diamonds
 388   PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
 389   PM.add(createMemCpyOptPass());            // Remove dead memcpys.
 390
 391   // Nuke dead stores.
 392   PM.add(createDeadStoreEliminationPass());
 393
 394   // More loops are countable; try to optimize them.
 395   PM.add(createIndVarSimplifyPass());
 396   PM.add(createLoopDeletionPass());
 397   PM.add(createLoopVectorizePass(true, true));
 398
 399   // More scalar chains could be vectorized due to more alias information
 400   PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
 401
 402   if (LoadCombine)
 403     PM.add(createLoadCombinePass());
 404
 405   // Cleanup and simplify the code after the scalar optimizations.
 406   PM.add(createInstructionCombiningPass());
 407   addExtensionsToPM(EP_Peephole, PM);
 408
 409   PM.add(createJumpThreadingPass());
 410
 411   // Delete basic blocks, which optimization passes may have killed.
 412   PM.add(createCFGSimplificationPass());
 413
 414   // Now that we have optimized the program, discard unreachable functions.
 415   PM.add(createGlobalDCEPass());
 416 }
 417
 418 void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM,
 419                                                 TargetMachine *TM) {
 420   if (TM) {
 421     const DataLayout *DL = TM->getSubtargetImpl()->getDataLayout();
 422     PM.add(new DataLayoutPass(*DL));
 423     TM->addAnalysisPasses(PM);
 424   }
 425
 426   if (LibraryInfo)
 427     PM.add(new TargetLibraryInfo(*LibraryInfo));
 428
 429   if (VerifyInput)
 430     PM.add(createVerifierPass());
 431
 432   if (StripDebug)
 433     PM.add(createStripSymbolsPass(true));
 434
 435   if (VerifyInput)
 436     PM.add(createDebugInfoVerifierPass());
 437
 438   if (OptLevel != 0)
 439     addLTOOptimizationPasses(PM);
 440
 441   if (VerifyOutput) {
 442     PM.add(createVerifierPass());
 443     PM.add(createDebugInfoVerifierPass());
 444   }
 445 }
 446
 447 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
 448     return reinterpret_cast<PassManagerBuilder*>(P);
 449 }
 450
 451 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
 452   return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
 453 }
 454
 455 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
 456   PassManagerBuilder *PMB = new PassManagerBuilder();
 457   return wrap(PMB);
 458 }
 459
 460 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
 461   PassManagerBuilder *Builder = unwrap(PMB);
 462   delete Builder;
 463 }
 464
 465 void
 466 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
 467                                   unsigned OptLevel) {
 468   PassManagerBuilder *Builder = unwrap(PMB);
 469   Builder->OptLevel = OptLevel;
 470 }
 471
 472 void
 473 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
 474                                    unsigned SizeLevel) {
 475   PassManagerBuilder *Builder = unwrap(PMB);
 476   Builder->SizeLevel = SizeLevel;
 477 }
 478
 479 void
 480 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
 481                                             LLVMBool Value) {
 482   PassManagerBuilder *Builder = unwrap(PMB);
 483   Builder->DisableUnitAtATime = Value;
 484 }
 485
 486 void
 487 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
 488                                             LLVMBool Value) {
 489   PassManagerBuilder *Builder = unwrap(PMB);
 490   Builder->DisableUnrollLoops = Value;
 491 }
 492
 493 void
 494 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
 495                                                  LLVMBool Value) {
 496   // NOTE: The simplify-libcalls pass has been removed.
 497 }
 498
 499 void
 500 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
 501                                               unsigned Threshold) {
 502   PassManagerBuilder *Builder = unwrap(PMB);
 503   Builder->Inliner = createFunctionInliningPass(Threshold);
 504 }
 505
 506 void
 507 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
 508                                                   LLVMPassManagerRef PM) {
 509   PassManagerBuilder *Builder = unwrap(PMB);
 510   FunctionPassManager *FPM = unwrap<FunctionPassManager>(PM);
 511   Builder->populateFunctionPassManager(*FPM);
 512 }
 513
 514 void
 515 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
 516                                                 LLVMPassManagerRef PM) {
 517   PassManagerBuilder *Builder = unwrap(PMB);
 518   PassManagerBase *MPM = unwrap(PM);
 519   Builder->populateModulePassManager(*MPM);
 520 }
 521
 522 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
 523                                                   LLVMPassManagerRef PM,
 524                                                   LLVMBool Internalize,
 525                                                   LLVMBool RunInliner) {
 526   PassManagerBuilder *Builder = unwrap(PMB);
 527   PassManagerBase *LPM = unwrap(PM);
 528
 529   // A small backwards compatibility hack. populateLTOPassManager used to take
 530   // an RunInliner option.
 531   if (RunInliner && !Builder->Inliner)
 532     Builder->Inliner = createFunctionInliningPass();
 533
 534   Builder->populateLTOPassManager(*LPM);
 535 }