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/IR/LegacyPassManager.h" 23 #include "llvm/Support/CommandLine.h" 24 #include "llvm/Support/ManagedStatic.h" 25 #include "llvm/Analysis/BasicAliasAnalysis.h" 26 #include "llvm/Analysis/CFLAliasAnalysis.h" 27 #include "llvm/Analysis/GlobalsModRef.h" 28 #include "llvm/Analysis/ScopedNoAliasAA.h" 29 #include "llvm/Analysis/TargetLibraryInfo.h" 30 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 31 #include "llvm/Target/TargetMachine.h" 32 #include "llvm/Transforms/IPO.h" 33 #include "llvm/Transforms/Scalar.h" 34 #include "llvm/Transforms/Vectorize.h" 35 36 using namespace llvm; 37 38 static cl::opt<bool> 39 RunLoopVectorization("vectorize-loops", cl::Hidden, 40 cl::desc("Run the Loop vectorization passes")); 41 42 static cl::opt<bool> 43 RunSLPVectorization("vectorize-slp", cl::Hidden, 44 cl::desc("Run the SLP vectorization passes")); 45 46 static cl::opt<bool> 47 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden, 48 cl::desc("Run the BB vectorization passes")); 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 static 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> UseNewSROA("use-new-sroa", 60 cl::init(true), cl::Hidden, 61 cl::desc("Enable the new, experimental SROA pass")); 62 63 static cl::opt<bool> 64 RunLoopRerolling("reroll-loops", cl::Hidden, 65 cl::desc("Run the loop rerolling pass")); 66 67 static cl::opt<bool> 68 RunFloat2Int("float-to-int", cl::Hidden, cl::init(true), 69 cl::desc("Run the float2int (float demotion) pass")); 70 71 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false), 72 cl::Hidden, 73 cl::desc("Run the load combining pass")); 74 75 static cl::opt<bool> 76 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization", 77 cl::init(true), cl::Hidden, 78 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop " 79 "vectorizer instead of before")); 80 81 static cl::opt<bool> UseCFLAA("use-cfl-aa", 82 cl::init(false), cl::Hidden, 83 cl::desc("Enable the new, experimental CFL alias analysis")); 84 85 static cl::opt<bool> 86 EnableMLSM("mlsm", cl::init(true), cl::Hidden, 87 cl::desc("Enable motion of merged load and store")); 88 89 static cl::opt<bool> EnableLoopInterchange( 90 "enable-loopinterchange", cl::init(false), cl::Hidden, 91 cl::desc("Enable the new, experimental LoopInterchange Pass")); 92 93 static cl::opt<bool> EnableLoopDistribute( 94 "enable-loop-distribute", cl::init(false), cl::Hidden, 95 cl::desc("Enable the new, experimental LoopDistribution Pass")); 96 97 static cl::opt<bool> EnableNonLTOGlobalsModRef( 98 "enable-non-lto-gmr", cl::init(false), cl::Hidden, 99 cl::desc( 100 "Enable the GlobalsModRef AliasAnalysis outside of the LTO pipeline.")); 101 102 PassManagerBuilder::PassManagerBuilder() { 103 OptLevel = 2; 104 SizeLevel = 0; 105 LibraryInfo = nullptr; 106 Inliner = nullptr; 107 DisableUnitAtATime = false; 108 DisableUnrollLoops = false; 109 BBVectorize = RunBBVectorization; 110 SLPVectorize = RunSLPVectorization; 111 LoopVectorize = RunLoopVectorization; 112 RerollLoops = RunLoopRerolling; 113 LoadCombine = RunLoadCombine; 114 DisableGVNLoadPRE = false; 115 VerifyInput = false; 116 VerifyOutput = false; 117 MergeFunctions = false; 118 PrepareForLTO = false; 119 } 120 121 PassManagerBuilder::~PassManagerBuilder() { 122 delete LibraryInfo; 123 delete Inliner; 124 } 125 126 /// Set of global extensions, automatically added as part of the standard set. 127 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy, 128 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions; 129 130 void PassManagerBuilder::addGlobalExtension( 131 PassManagerBuilder::ExtensionPointTy Ty, 132 PassManagerBuilder::ExtensionFn Fn) { 133 GlobalExtensions->push_back(std::make_pair(Ty, Fn)); 134 } 135 136 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) { 137 Extensions.push_back(std::make_pair(Ty, Fn)); 138 } 139 140 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy, 141 legacy::PassManagerBase &PM) const { 142 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i) 143 if ((*GlobalExtensions)[i].first == ETy) 144 (*GlobalExtensions)[i].second(*this, PM); 145 for (unsigned i = 0, e = Extensions.size(); i != e; ++i) 146 if (Extensions[i].first == ETy) 147 Extensions[i].second(*this, PM); 148 } 149 150 void PassManagerBuilder::addInitialAliasAnalysisPasses( 151 legacy::PassManagerBase &PM) const { 152 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 153 // BasicAliasAnalysis wins if they disagree. This is intended to help 154 // support "obvious" type-punning idioms. 155 if (UseCFLAA) 156 PM.add(createCFLAAWrapperPass()); 157 PM.add(createTypeBasedAAWrapperPass()); 158 PM.add(createScopedNoAliasAAWrapperPass()); 159 } 160 161 void PassManagerBuilder::populateFunctionPassManager( 162 legacy::FunctionPassManager &FPM) { 163 addExtensionsToPM(EP_EarlyAsPossible, FPM); 164 165 // Add LibraryInfo if we have some. 166 if (LibraryInfo) 167 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 168 169 if (OptLevel == 0) return; 170 171 addInitialAliasAnalysisPasses(FPM); 172 173 FPM.add(createCFGSimplificationPass()); 174 if (UseNewSROA) 175 FPM.add(createSROAPass()); 176 else 177 FPM.add(createScalarReplAggregatesPass()); 178 FPM.add(createEarlyCSEPass()); 179 FPM.add(createLowerExpectIntrinsicPass()); 180 } 181 182 void PassManagerBuilder::populateModulePassManager( 183 legacy::PassManagerBase &MPM) { 184 // If all optimizations are disabled, just run the always-inline pass and, 185 // if enabled, the function merging pass. 186 if (OptLevel == 0) { 187 if (Inliner) { 188 MPM.add(Inliner); 189 Inliner = nullptr; 190 } 191 192 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly 193 // creates a CGSCC pass manager, but we don't want to add extensions into 194 // that pass manager. To prevent this we insert a no-op module pass to reset 195 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0 196 // builds. The function merging pass is 197 if (MergeFunctions) 198 MPM.add(createMergeFunctionsPass()); 199 else if (!GlobalExtensions->empty() || !Extensions.empty()) 200 MPM.add(createBarrierNoopPass()); 201 202 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM); 203 return; 204 } 205 206 // Add LibraryInfo if we have some. 207 if (LibraryInfo) 208 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 209 210 addInitialAliasAnalysisPasses(MPM); 211 212 if (!DisableUnitAtATime) { 213 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM); 214 215 MPM.add(createIPSCCPPass()); // IP SCCP 216 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars 217 218 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination 219 220 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE 221 addExtensionsToPM(EP_Peephole, MPM); 222 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE 223 } 224 225 if (EnableNonLTOGlobalsModRef) 226 // We add a module alias analysis pass here. In part due to bugs in the 227 // analysis infrastructure this "works" in that the analysis stays alive 228 // for the entire SCC pass run below. 229 MPM.add(createGlobalsAAWrapperPass()); 230 231 // Start of CallGraph SCC passes. 232 if (!DisableUnitAtATime) 233 MPM.add(createPruneEHPass()); // Remove dead EH info 234 if (Inliner) { 235 MPM.add(Inliner); 236 Inliner = nullptr; 237 } 238 if (!DisableUnitAtATime) 239 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs 240 if (OptLevel > 2) 241 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args 242 243 // Start of function pass. 244 // Break up aggregate allocas, using SSAUpdater. 245 if (UseNewSROA) 246 MPM.add(createSROAPass()); 247 else 248 MPM.add(createScalarReplAggregatesPass(-1, false)); 249 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 250 MPM.add(createJumpThreadingPass()); // Thread jumps. 251 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals 252 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 253 MPM.add(createInstructionCombiningPass()); // Combine silly seq's 254 addExtensionsToPM(EP_Peephole, MPM); 255 256 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls 257 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 258 MPM.add(createReassociatePass()); // Reassociate expressions 259 // Rotate Loop - disable header duplication at -Oz 260 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 261 MPM.add(createLICMPass()); // Hoist loop invariants 262 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3)); 263 MPM.add(createCFGSimplificationPass()); 264 MPM.add(createInstructionCombiningPass()); 265 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars 266 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset. 267 MPM.add(createLoopDeletionPass()); // Delete dead loops 268 if (EnableLoopInterchange) { 269 MPM.add(createLoopInterchangePass()); // Interchange loops 270 MPM.add(createCFGSimplificationPass()); 271 } 272 if (!DisableUnrollLoops) 273 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops 274 addExtensionsToPM(EP_LoopOptimizerEnd, MPM); 275 276 if (OptLevel > 1) { 277 if (EnableMLSM) 278 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds 279 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 280 } 281 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset 282 MPM.add(createSCCPPass()); // Constant prop with SCCP 283 284 // Delete dead bit computations (instcombine runs after to fold away the dead 285 // computations, and then ADCE will run later to exploit any new DCE 286 // opportunities that creates). 287 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations 288 289 // Run instcombine after redundancy elimination to exploit opportunities 290 // opened up by them. 291 MPM.add(createInstructionCombiningPass()); 292 addExtensionsToPM(EP_Peephole, MPM); 293 MPM.add(createJumpThreadingPass()); // Thread jumps 294 MPM.add(createCorrelatedValuePropagationPass()); 295 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores 296 MPM.add(createLICMPass()); 297 298 addExtensionsToPM(EP_ScalarOptimizerLate, MPM); 299 300 if (RerollLoops) 301 MPM.add(createLoopRerollPass()); 302 if (!RunSLPAfterLoopVectorization) { 303 if (SLPVectorize) 304 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 305 306 if (BBVectorize) { 307 MPM.add(createBBVectorizePass()); 308 MPM.add(createInstructionCombiningPass()); 309 addExtensionsToPM(EP_Peephole, MPM); 310 if (OptLevel > 1 && UseGVNAfterVectorization) 311 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 312 else 313 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 314 315 // BBVectorize may have significantly shortened a loop body; unroll again. 316 if (!DisableUnrollLoops) 317 MPM.add(createLoopUnrollPass()); 318 } 319 } 320 321 if (LoadCombine) 322 MPM.add(createLoadCombinePass()); 323 324 MPM.add(createAggressiveDCEPass()); // Delete dead instructions 325 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 326 MPM.add(createInstructionCombiningPass()); // Clean up after everything. 327 addExtensionsToPM(EP_Peephole, MPM); 328 329 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC 330 // pass manager that we are specifically trying to avoid. To prevent this 331 // we must insert a no-op module pass to reset the pass manager. 332 MPM.add(createBarrierNoopPass()); 333 334 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO) { 335 // Remove avail extern fns and globals definitions if we aren't 336 // compiling an object file for later LTO. For LTO we want to preserve 337 // these so they are eligible for inlining at link-time. Note if they 338 // are unreferenced they will be removed by GlobalDCE later, so 339 // this only impacts referenced available externally globals. 340 // Eventually they will be suppressed during codegen, but eliminating 341 // here enables more opportunity for GlobalDCE as it may make 342 // globals referenced by available external functions dead 343 // and saves running remaining passes on the eliminated functions. 344 MPM.add(createEliminateAvailableExternallyPass()); 345 } 346 347 if (EnableNonLTOGlobalsModRef) 348 // We add a fresh GlobalsModRef run at this point. This is particularly 349 // useful as the above will have inlined, DCE'ed, and function-attr 350 // propagated everything. We should at this point have a reasonably minimal 351 // and richly annotated call graph. By computing aliasing and mod/ref 352 // information for all local globals here, the late loop passes and notably 353 // the vectorizer will be able to use them to help recognize vectorizable 354 // memory operations. 355 // 356 // Note that this relies on a bug in the pass manager which preserves 357 // a module analysis into a function pass pipeline (and throughout it) so 358 // long as the first function pass doesn't invalidate the module analysis. 359 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for 360 // this to work. Fortunately, it is trivial to preserve AliasAnalysis 361 // (doing nothing preserves it as it is required to be conservatively 362 // correct in the face of IR changes). 363 MPM.add(createGlobalsAAWrapperPass()); 364 365 if (RunFloat2Int) 366 MPM.add(createFloat2IntPass()); 367 368 addExtensionsToPM(EP_VectorizerStart, MPM); 369 370 // Re-rotate loops in all our loop nests. These may have fallout out of 371 // rotated form due to GVN or other transformations, and the vectorizer relies 372 // on the rotated form. Disable header duplication at -Oz. 373 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 374 375 // Distribute loops to allow partial vectorization. I.e. isolate dependences 376 // into separate loop that would otherwise inhibit vectorization. 377 if (EnableLoopDistribute) 378 MPM.add(createLoopDistributePass()); 379 380 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize)); 381 // FIXME: Because of #pragma vectorize enable, the passes below are always 382 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when 383 // on -O1 and no #pragma is found). Would be good to have these two passes 384 // as function calls, so that we can only pass them when the vectorizer 385 // changed the code. 386 MPM.add(createInstructionCombiningPass()); 387 if (OptLevel > 1 && ExtraVectorizerPasses) { 388 // At higher optimization levels, try to clean up any runtime overlap and 389 // alignment checks inserted by the vectorizer. We want to track correllated 390 // runtime checks for two inner loops in the same outer loop, fold any 391 // common computations, hoist loop-invariant aspects out of any outer loop, 392 // and unswitch the runtime checks if possible. Once hoisted, we may have 393 // dead (or speculatable) control flows or more combining opportunities. 394 MPM.add(createEarlyCSEPass()); 395 MPM.add(createCorrelatedValuePropagationPass()); 396 MPM.add(createInstructionCombiningPass()); 397 MPM.add(createLICMPass()); 398 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3)); 399 MPM.add(createCFGSimplificationPass()); 400 MPM.add(createInstructionCombiningPass()); 401 } 402 403 if (RunSLPAfterLoopVectorization) { 404 if (SLPVectorize) { 405 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 406 if (OptLevel > 1 && ExtraVectorizerPasses) { 407 MPM.add(createEarlyCSEPass()); 408 } 409 } 410 411 if (BBVectorize) { 412 MPM.add(createBBVectorizePass()); 413 MPM.add(createInstructionCombiningPass()); 414 addExtensionsToPM(EP_Peephole, MPM); 415 if (OptLevel > 1 && UseGVNAfterVectorization) 416 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 417 else 418 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 419 420 // BBVectorize may have significantly shortened a loop body; unroll again. 421 if (!DisableUnrollLoops) 422 MPM.add(createLoopUnrollPass()); 423 } 424 } 425 426 addExtensionsToPM(EP_Peephole, MPM); 427 MPM.add(createCFGSimplificationPass()); 428 MPM.add(createInstructionCombiningPass()); 429 430 if (!DisableUnrollLoops) { 431 MPM.add(createLoopUnrollPass()); // Unroll small loops 432 433 // LoopUnroll may generate some redundency to cleanup. 434 MPM.add(createInstructionCombiningPass()); 435 436 // Runtime unrolling will introduce runtime check in loop prologue. If the 437 // unrolled loop is a inner loop, then the prologue will be inside the 438 // outer loop. LICM pass can help to promote the runtime check out if the 439 // checked value is loop invariant. 440 MPM.add(createLICMPass()); 441 } 442 443 // After vectorization and unrolling, assume intrinsics may tell us more 444 // about pointer alignments. 445 MPM.add(createAlignmentFromAssumptionsPass()); 446 447 if (!DisableUnitAtATime) { 448 // FIXME: We shouldn't bother with this anymore. 449 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes 450 451 // GlobalOpt already deletes dead functions and globals, at -O2 try a 452 // late pass of GlobalDCE. It is capable of deleting dead cycles. 453 if (OptLevel > 1) { 454 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals. 455 MPM.add(createConstantMergePass()); // Merge dup global constants 456 } 457 } 458 459 if (MergeFunctions) 460 MPM.add(createMergeFunctionsPass()); 461 462 addExtensionsToPM(EP_OptimizerLast, MPM); 463 } 464 465 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) { 466 // Provide AliasAnalysis services for optimizations. 467 addInitialAliasAnalysisPasses(PM); 468 469 // Propagate constants at call sites into the functions they call. This 470 // opens opportunities for globalopt (and inlining) by substituting function 471 // pointers passed as arguments to direct uses of functions. 472 PM.add(createIPSCCPPass()); 473 474 // Now that we internalized some globals, see if we can hack on them! 475 PM.add(createGlobalOptimizerPass()); 476 477 // Linking modules together can lead to duplicated global constants, only 478 // keep one copy of each constant. 479 PM.add(createConstantMergePass()); 480 481 // Remove unused arguments from functions. 482 PM.add(createDeadArgEliminationPass()); 483 484 // Reduce the code after globalopt and ipsccp. Both can open up significant 485 // simplification opportunities, and both can propagate functions through 486 // function pointers. When this happens, we often have to resolve varargs 487 // calls, etc, so let instcombine do this. 488 PM.add(createInstructionCombiningPass()); 489 addExtensionsToPM(EP_Peephole, PM); 490 491 // Inline small functions 492 bool RunInliner = Inliner; 493 if (RunInliner) { 494 PM.add(Inliner); 495 Inliner = nullptr; 496 } 497 498 PM.add(createPruneEHPass()); // Remove dead EH info. 499 500 // Optimize globals again if we ran the inliner. 501 if (RunInliner) 502 PM.add(createGlobalOptimizerPass()); 503 PM.add(createGlobalDCEPass()); // Remove dead functions. 504 505 // If we didn't decide to inline a function, check to see if we can 506 // transform it to pass arguments by value instead of by reference. 507 PM.add(createArgumentPromotionPass()); 508 509 // The IPO passes may leave cruft around. Clean up after them. 510 PM.add(createInstructionCombiningPass()); 511 addExtensionsToPM(EP_Peephole, PM); 512 PM.add(createJumpThreadingPass()); 513 514 // Break up allocas 515 if (UseNewSROA) 516 PM.add(createSROAPass()); 517 else 518 PM.add(createScalarReplAggregatesPass()); 519 520 // Run a few AA driven optimizations here and now, to cleanup the code. 521 PM.add(createFunctionAttrsPass()); // Add nocapture. 522 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis. 523 524 PM.add(createLICMPass()); // Hoist loop invariants. 525 if (EnableMLSM) 526 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds. 527 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies. 528 PM.add(createMemCpyOptPass()); // Remove dead memcpys. 529 530 // Nuke dead stores. 531 PM.add(createDeadStoreEliminationPass()); 532 533 // More loops are countable; try to optimize them. 534 PM.add(createIndVarSimplifyPass()); 535 PM.add(createLoopDeletionPass()); 536 if (EnableLoopInterchange) 537 PM.add(createLoopInterchangePass()); 538 539 PM.add(createLoopVectorizePass(true, LoopVectorize)); 540 541 // More scalar chains could be vectorized due to more alias information 542 if (RunSLPAfterLoopVectorization) 543 if (SLPVectorize) 544 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 545 546 // After vectorization, assume intrinsics may tell us more about pointer 547 // alignments. 548 PM.add(createAlignmentFromAssumptionsPass()); 549 550 if (LoadCombine) 551 PM.add(createLoadCombinePass()); 552 553 // Cleanup and simplify the code after the scalar optimizations. 554 PM.add(createInstructionCombiningPass()); 555 addExtensionsToPM(EP_Peephole, PM); 556 557 PM.add(createJumpThreadingPass()); 558 } 559 560 void PassManagerBuilder::addLateLTOOptimizationPasses( 561 legacy::PassManagerBase &PM) { 562 // Delete basic blocks, which optimization passes may have killed. 563 PM.add(createCFGSimplificationPass()); 564 565 // Drop bodies of available externally objects to improve GlobalDCE. 566 PM.add(createEliminateAvailableExternallyPass()); 567 568 // Now that we have optimized the program, discard unreachable functions. 569 PM.add(createGlobalDCEPass()); 570 571 // FIXME: this is profitable (for compiler time) to do at -O0 too, but 572 // currently it damages debug info. 573 if (MergeFunctions) 574 PM.add(createMergeFunctionsPass()); 575 } 576 577 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) { 578 if (LibraryInfo) 579 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 580 581 if (VerifyInput) 582 PM.add(createVerifierPass()); 583 584 if (OptLevel > 1) 585 addLTOOptimizationPasses(PM); 586 587 // Lower bit sets to globals. This pass supports Clang's control flow 588 // integrity mechanisms (-fsanitize=cfi*) and needs to run at link time if CFI 589 // is enabled. The pass does nothing if CFI is disabled. 590 PM.add(createLowerBitSetsPass()); 591 592 if (OptLevel != 0) 593 addLateLTOOptimizationPasses(PM); 594 595 if (VerifyOutput) 596 PM.add(createVerifierPass()); 597 } 598 599 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) { 600 return reinterpret_cast<PassManagerBuilder*>(P); 601 } 602 603 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) { 604 return reinterpret_cast<LLVMPassManagerBuilderRef>(P); 605 } 606 607 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() { 608 PassManagerBuilder *PMB = new PassManagerBuilder(); 609 return wrap(PMB); 610 } 611 612 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) { 613 PassManagerBuilder *Builder = unwrap(PMB); 614 delete Builder; 615 } 616 617 void 618 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB, 619 unsigned OptLevel) { 620 PassManagerBuilder *Builder = unwrap(PMB); 621 Builder->OptLevel = OptLevel; 622 } 623 624 void 625 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB, 626 unsigned SizeLevel) { 627 PassManagerBuilder *Builder = unwrap(PMB); 628 Builder->SizeLevel = SizeLevel; 629 } 630 631 void 632 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB, 633 LLVMBool Value) { 634 PassManagerBuilder *Builder = unwrap(PMB); 635 Builder->DisableUnitAtATime = Value; 636 } 637 638 void 639 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB, 640 LLVMBool Value) { 641 PassManagerBuilder *Builder = unwrap(PMB); 642 Builder->DisableUnrollLoops = Value; 643 } 644 645 void 646 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB, 647 LLVMBool Value) { 648 // NOTE: The simplify-libcalls pass has been removed. 649 } 650 651 void 652 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB, 653 unsigned Threshold) { 654 PassManagerBuilder *Builder = unwrap(PMB); 655 Builder->Inliner = createFunctionInliningPass(Threshold); 656 } 657 658 void 659 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB, 660 LLVMPassManagerRef PM) { 661 PassManagerBuilder *Builder = unwrap(PMB); 662 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM); 663 Builder->populateFunctionPassManager(*FPM); 664 } 665 666 void 667 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB, 668 LLVMPassManagerRef PM) { 669 PassManagerBuilder *Builder = unwrap(PMB); 670 legacy::PassManagerBase *MPM = unwrap(PM); 671 Builder->populateModulePassManager(*MPM); 672 } 673 674 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB, 675 LLVMPassManagerRef PM, 676 LLVMBool Internalize, 677 LLVMBool RunInliner) { 678 PassManagerBuilder *Builder = unwrap(PMB); 679 legacy::PassManagerBase *LPM = unwrap(PM); 680 681 // A small backwards compatibility hack. populateLTOPassManager used to take 682 // an RunInliner option. 683 if (RunInliner && !Builder->Inliner) 684 Builder->Inliner = createFunctionInliningPass(); 685 686 Builder->populateLTOPassManager(*LPM); 687 } 688