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