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