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(true), 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(createInstructionCombiningPass()); 264 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars 265 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset. 266 MPM.add(createLoopDeletionPass()); // Delete dead loops 267 if (EnableLoopInterchange) { 268 MPM.add(createLoopInterchangePass()); // Interchange loops 269 MPM.add(createCFGSimplificationPass()); 270 } 271 if (!DisableUnrollLoops) 272 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops 273 addExtensionsToPM(EP_LoopOptimizerEnd, MPM); 274 275 if (OptLevel > 1) { 276 if (EnableMLSM) 277 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds 278 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 279 } 280 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset 281 MPM.add(createSCCPPass()); // Constant prop with SCCP 282 283 // Delete dead bit computations (instcombine runs after to fold away the dead 284 // computations, and then ADCE will run later to exploit any new DCE 285 // opportunities that creates). 286 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations 287 288 // Run instcombine after redundancy elimination to exploit opportunities 289 // opened up by them. 290 MPM.add(createInstructionCombiningPass()); 291 addExtensionsToPM(EP_Peephole, MPM); 292 MPM.add(createJumpThreadingPass()); // Thread jumps 293 MPM.add(createCorrelatedValuePropagationPass()); 294 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores 295 MPM.add(createLICMPass()); 296 297 addExtensionsToPM(EP_ScalarOptimizerLate, MPM); 298 299 if (RerollLoops) 300 MPM.add(createLoopRerollPass()); 301 if (!RunSLPAfterLoopVectorization) { 302 if (SLPVectorize) 303 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 304 305 if (BBVectorize) { 306 MPM.add(createBBVectorizePass()); 307 MPM.add(createInstructionCombiningPass()); 308 addExtensionsToPM(EP_Peephole, MPM); 309 if (OptLevel > 1 && UseGVNAfterVectorization) 310 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 311 else 312 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 313 314 // BBVectorize may have significantly shortened a loop body; unroll again. 315 if (!DisableUnrollLoops) 316 MPM.add(createLoopUnrollPass()); 317 } 318 } 319 320 if (LoadCombine) 321 MPM.add(createLoadCombinePass()); 322 323 MPM.add(createAggressiveDCEPass()); // Delete dead instructions 324 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 325 MPM.add(createInstructionCombiningPass()); // Clean up after everything. 326 addExtensionsToPM(EP_Peephole, MPM); 327 328 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC 329 // pass manager that we are specifically trying to avoid. To prevent this 330 // we must insert a no-op module pass to reset the pass manager. 331 MPM.add(createBarrierNoopPass()); 332 333 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO) { 334 // Remove avail extern fns and globals definitions if we aren't 335 // compiling an object file for later LTO. For LTO we want to preserve 336 // these so they are eligible for inlining at link-time. Note if they 337 // are unreferenced they will be removed by GlobalDCE later, so 338 // this only impacts referenced available externally globals. 339 // Eventually they will be suppressed during codegen, but eliminating 340 // here enables more opportunity for GlobalDCE as it may make 341 // globals referenced by available external functions dead 342 // and saves running remaining passes on the eliminated functions. 343 MPM.add(createEliminateAvailableExternallyPass()); 344 } 345 346 if (EnableNonLTOGlobalsModRef) 347 // We add a fresh GlobalsModRef run at this point. This is particularly 348 // useful as the above will have inlined, DCE'ed, and function-attr 349 // propagated everything. We should at this point have a reasonably minimal 350 // and richly annotated call graph. By computing aliasing and mod/ref 351 // information for all local globals here, the late loop passes and notably 352 // the vectorizer will be able to use them to help recognize vectorizable 353 // memory operations. 354 // 355 // Note that this relies on a bug in the pass manager which preserves 356 // a module analysis into a function pass pipeline (and throughout it) so 357 // long as the first function pass doesn't invalidate the module analysis. 358 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for 359 // this to work. Fortunately, it is trivial to preserve AliasAnalysis 360 // (doing nothing preserves it as it is required to be conservatively 361 // correct in the face of IR changes). 362 MPM.add(createGlobalsAAWrapperPass()); 363 364 if (RunFloat2Int) 365 MPM.add(createFloat2IntPass()); 366 367 addExtensionsToPM(EP_VectorizerStart, MPM); 368 369 // Re-rotate loops in all our loop nests. These may have fallout out of 370 // rotated form due to GVN or other transformations, and the vectorizer relies 371 // on the rotated form. Disable header duplication at -Oz. 372 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 373 374 // Distribute loops to allow partial vectorization. I.e. isolate dependences 375 // into separate loop that would otherwise inhibit vectorization. 376 if (EnableLoopDistribute) 377 MPM.add(createLoopDistributePass()); 378 379 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize)); 380 // FIXME: Because of #pragma vectorize enable, the passes below are always 381 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when 382 // on -O1 and no #pragma is found). Would be good to have these two passes 383 // as function calls, so that we can only pass them when the vectorizer 384 // changed the code. 385 MPM.add(createInstructionCombiningPass()); 386 if (OptLevel > 1 && ExtraVectorizerPasses) { 387 // At higher optimization levels, try to clean up any runtime overlap and 388 // alignment checks inserted by the vectorizer. We want to track correllated 389 // runtime checks for two inner loops in the same outer loop, fold any 390 // common computations, hoist loop-invariant aspects out of any outer loop, 391 // and unswitch the runtime checks if possible. Once hoisted, we may have 392 // dead (or speculatable) control flows or more combining opportunities. 393 MPM.add(createEarlyCSEPass()); 394 MPM.add(createCorrelatedValuePropagationPass()); 395 MPM.add(createInstructionCombiningPass()); 396 MPM.add(createLICMPass()); 397 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3)); 398 MPM.add(createCFGSimplificationPass()); 399 MPM.add(createInstructionCombiningPass()); 400 } 401 402 if (RunSLPAfterLoopVectorization) { 403 if (SLPVectorize) { 404 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 405 if (OptLevel > 1 && ExtraVectorizerPasses) { 406 MPM.add(createEarlyCSEPass()); 407 } 408 } 409 410 if (BBVectorize) { 411 MPM.add(createBBVectorizePass()); 412 MPM.add(createInstructionCombiningPass()); 413 addExtensionsToPM(EP_Peephole, MPM); 414 if (OptLevel > 1 && UseGVNAfterVectorization) 415 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 416 else 417 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 418 419 // BBVectorize may have significantly shortened a loop body; unroll again. 420 if (!DisableUnrollLoops) 421 MPM.add(createLoopUnrollPass()); 422 } 423 } 424 425 addExtensionsToPM(EP_Peephole, MPM); 426 MPM.add(createCFGSimplificationPass()); 427 MPM.add(createInstructionCombiningPass()); 428 429 if (!DisableUnrollLoops) { 430 MPM.add(createLoopUnrollPass()); // Unroll small loops 431 432 // LoopUnroll may generate some redundency to cleanup. 433 MPM.add(createInstructionCombiningPass()); 434 435 // Runtime unrolling will introduce runtime check in loop prologue. If the 436 // unrolled loop is a inner loop, then the prologue will be inside the 437 // outer loop. LICM pass can help to promote the runtime check out if the 438 // checked value is loop invariant. 439 MPM.add(createLICMPass()); 440 } 441 442 // After vectorization and unrolling, assume intrinsics may tell us more 443 // about pointer alignments. 444 MPM.add(createAlignmentFromAssumptionsPass()); 445 446 if (!DisableUnitAtATime) { 447 // FIXME: We shouldn't bother with this anymore. 448 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes 449 450 // GlobalOpt already deletes dead functions and globals, at -O2 try a 451 // late pass of GlobalDCE. It is capable of deleting dead cycles. 452 if (OptLevel > 1) { 453 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals. 454 MPM.add(createConstantMergePass()); // Merge dup global constants 455 } 456 } 457 458 if (MergeFunctions) 459 MPM.add(createMergeFunctionsPass()); 460 461 addExtensionsToPM(EP_OptimizerLast, MPM); 462 } 463 464 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) { 465 // Provide AliasAnalysis services for optimizations. 466 addInitialAliasAnalysisPasses(PM); 467 468 // Propagate constants at call sites into the functions they call. This 469 // opens opportunities for globalopt (and inlining) by substituting function 470 // pointers passed as arguments to direct uses of functions. 471 PM.add(createIPSCCPPass()); 472 473 // Now that we internalized some globals, see if we can hack on them! 474 PM.add(createGlobalOptimizerPass()); 475 476 // Linking modules together can lead to duplicated global constants, only 477 // keep one copy of each constant. 478 PM.add(createConstantMergePass()); 479 480 // Remove unused arguments from functions. 481 PM.add(createDeadArgEliminationPass()); 482 483 // Reduce the code after globalopt and ipsccp. Both can open up significant 484 // simplification opportunities, and both can propagate functions through 485 // function pointers. When this happens, we often have to resolve varargs 486 // calls, etc, so let instcombine do this. 487 PM.add(createInstructionCombiningPass()); 488 addExtensionsToPM(EP_Peephole, PM); 489 490 // Inline small functions 491 bool RunInliner = Inliner; 492 if (RunInliner) { 493 PM.add(Inliner); 494 Inliner = nullptr; 495 } 496 497 PM.add(createPruneEHPass()); // Remove dead EH info. 498 499 // Optimize globals again if we ran the inliner. 500 if (RunInliner) 501 PM.add(createGlobalOptimizerPass()); 502 PM.add(createGlobalDCEPass()); // Remove dead functions. 503 504 // If we didn't decide to inline a function, check to see if we can 505 // transform it to pass arguments by value instead of by reference. 506 PM.add(createArgumentPromotionPass()); 507 508 // The IPO passes may leave cruft around. Clean up after them. 509 PM.add(createInstructionCombiningPass()); 510 addExtensionsToPM(EP_Peephole, PM); 511 PM.add(createJumpThreadingPass()); 512 513 // Break up allocas 514 if (UseNewSROA) 515 PM.add(createSROAPass()); 516 else 517 PM.add(createScalarReplAggregatesPass()); 518 519 // Run a few AA driven optimizations here and now, to cleanup the code. 520 PM.add(createFunctionAttrsPass()); // Add nocapture. 521 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis. 522 523 PM.add(createLICMPass()); // Hoist loop invariants. 524 if (EnableMLSM) 525 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds. 526 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies. 527 PM.add(createMemCpyOptPass()); // Remove dead memcpys. 528 529 // Nuke dead stores. 530 PM.add(createDeadStoreEliminationPass()); 531 532 // More loops are countable; try to optimize them. 533 PM.add(createIndVarSimplifyPass()); 534 PM.add(createLoopDeletionPass()); 535 if (EnableLoopInterchange) 536 PM.add(createLoopInterchangePass()); 537 538 PM.add(createLoopVectorizePass(true, LoopVectorize)); 539 540 // More scalar chains could be vectorized due to more alias information 541 if (RunSLPAfterLoopVectorization) 542 if (SLPVectorize) 543 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 544 545 // After vectorization, assume intrinsics may tell us more about pointer 546 // alignments. 547 PM.add(createAlignmentFromAssumptionsPass()); 548 549 if (LoadCombine) 550 PM.add(createLoadCombinePass()); 551 552 // Cleanup and simplify the code after the scalar optimizations. 553 PM.add(createInstructionCombiningPass()); 554 addExtensionsToPM(EP_Peephole, PM); 555 556 PM.add(createJumpThreadingPass()); 557 } 558 559 void PassManagerBuilder::addLateLTOOptimizationPasses( 560 legacy::PassManagerBase &PM) { 561 // Delete basic blocks, which optimization passes may have killed. 562 PM.add(createCFGSimplificationPass()); 563 564 // Drop bodies of available externally objects to improve GlobalDCE. 565 PM.add(createEliminateAvailableExternallyPass()); 566 567 // Now that we have optimized the program, discard unreachable functions. 568 PM.add(createGlobalDCEPass()); 569 570 // FIXME: this is profitable (for compiler time) to do at -O0 too, but 571 // currently it damages debug info. 572 if (MergeFunctions) 573 PM.add(createMergeFunctionsPass()); 574 } 575 576 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) { 577 if (LibraryInfo) 578 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 579 580 if (VerifyInput) 581 PM.add(createVerifierPass()); 582 583 if (OptLevel > 1) 584 addLTOOptimizationPasses(PM); 585 586 // Lower bit sets to globals. This pass supports Clang's control flow 587 // integrity mechanisms (-fsanitize=cfi*) and needs to run at link time if CFI 588 // is enabled. The pass does nothing if CFI is disabled. 589 PM.add(createLowerBitSetsPass()); 590 591 if (OptLevel != 0) 592 addLateLTOOptimizationPasses(PM); 593 594 if (VerifyOutput) 595 PM.add(createVerifierPass()); 596 } 597 598 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) { 599 return reinterpret_cast<PassManagerBuilder*>(P); 600 } 601 602 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) { 603 return reinterpret_cast<LLVMPassManagerBuilderRef>(P); 604 } 605 606 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() { 607 PassManagerBuilder *PMB = new PassManagerBuilder(); 608 return wrap(PMB); 609 } 610 611 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) { 612 PassManagerBuilder *Builder = unwrap(PMB); 613 delete Builder; 614 } 615 616 void 617 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB, 618 unsigned OptLevel) { 619 PassManagerBuilder *Builder = unwrap(PMB); 620 Builder->OptLevel = OptLevel; 621 } 622 623 void 624 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB, 625 unsigned SizeLevel) { 626 PassManagerBuilder *Builder = unwrap(PMB); 627 Builder->SizeLevel = SizeLevel; 628 } 629 630 void 631 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB, 632 LLVMBool Value) { 633 PassManagerBuilder *Builder = unwrap(PMB); 634 Builder->DisableUnitAtATime = Value; 635 } 636 637 void 638 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB, 639 LLVMBool Value) { 640 PassManagerBuilder *Builder = unwrap(PMB); 641 Builder->DisableUnrollLoops = Value; 642 } 643 644 void 645 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB, 646 LLVMBool Value) { 647 // NOTE: The simplify-libcalls pass has been removed. 648 } 649 650 void 651 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB, 652 unsigned Threshold) { 653 PassManagerBuilder *Builder = unwrap(PMB); 654 Builder->Inliner = createFunctionInliningPass(Threshold); 655 } 656 657 void 658 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB, 659 LLVMPassManagerRef PM) { 660 PassManagerBuilder *Builder = unwrap(PMB); 661 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM); 662 Builder->populateFunctionPassManager(*FPM); 663 } 664 665 void 666 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB, 667 LLVMPassManagerRef PM) { 668 PassManagerBuilder *Builder = unwrap(PMB); 669 legacy::PassManagerBase *MPM = unwrap(PM); 670 Builder->populateModulePassManager(*MPM); 671 } 672 673 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB, 674 LLVMPassManagerRef PM, 675 LLVMBool Internalize, 676 LLVMBool RunInliner) { 677 PassManagerBuilder *Builder = unwrap(PMB); 678 legacy::PassManagerBase *LPM = unwrap(PM); 679 680 // A small backwards compatibility hack. populateLTOPassManager used to take 681 // an RunInliner option. 682 if (RunInliner && !Builder->Inliner) 683 Builder->Inliner = createFunctionInliningPass(); 684 685 Builder->populateLTOPassManager(*LPM); 686 } 687