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/FunctionInfo.h" 22 #include "llvm/IR/Verifier.h" 23 #include "llvm/IR/LegacyPassManager.h" 24 #include "llvm/Support/CommandLine.h" 25 #include "llvm/Support/ManagedStatic.h" 26 #include "llvm/Analysis/BasicAliasAnalysis.h" 27 #include "llvm/Analysis/CFLAliasAnalysis.h" 28 #include "llvm/Analysis/GlobalsModRef.h" 29 #include "llvm/Analysis/ScopedNoAliasAA.h" 30 #include "llvm/Analysis/TargetLibraryInfo.h" 31 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 32 #include "llvm/Target/TargetMachine.h" 33 #include "llvm/Transforms/IPO.h" 34 #include "llvm/Transforms/Scalar.h" 35 #include "llvm/Transforms/Vectorize.h" 36 37 using namespace llvm; 38 39 static cl::opt<bool> 40 RunLoopVectorization("vectorize-loops", cl::Hidden, 41 cl::desc("Run the Loop vectorization passes")); 42 43 static cl::opt<bool> 44 RunSLPVectorization("vectorize-slp", cl::Hidden, 45 cl::desc("Run the SLP vectorization passes")); 46 47 static cl::opt<bool> 48 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden, 49 cl::desc("Run the BB vectorization passes")); 50 51 static cl::opt<bool> 52 UseGVNAfterVectorization("use-gvn-after-vectorization", 53 cl::init(false), cl::Hidden, 54 cl::desc("Run GVN instead of Early CSE after vectorization passes")); 55 56 static cl::opt<bool> ExtraVectorizerPasses( 57 "extra-vectorizer-passes", cl::init(false), cl::Hidden, 58 cl::desc("Run cleanup optimization passes after vectorization.")); 59 60 static cl::opt<bool> UseNewSROA("use-new-sroa", 61 cl::init(true), cl::Hidden, 62 cl::desc("Enable the new, experimental SROA pass")); 63 64 static cl::opt<bool> 65 RunLoopRerolling("reroll-loops", cl::Hidden, 66 cl::desc("Run the loop rerolling pass")); 67 68 static cl::opt<bool> 69 RunFloat2Int("float-to-int", cl::Hidden, cl::init(true), 70 cl::desc("Run the float2int (float demotion) pass")); 71 72 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false), 73 cl::Hidden, 74 cl::desc("Run the load combining pass")); 75 76 static cl::opt<bool> 77 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization", 78 cl::init(true), cl::Hidden, 79 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop " 80 "vectorizer instead of before")); 81 82 static cl::opt<bool> UseCFLAA("use-cfl-aa", 83 cl::init(false), cl::Hidden, 84 cl::desc("Enable the new, experimental CFL alias analysis")); 85 86 static cl::opt<bool> 87 EnableMLSM("mlsm", cl::init(true), cl::Hidden, 88 cl::desc("Enable motion of merged load and store")); 89 90 static cl::opt<bool> EnableLoopInterchange( 91 "enable-loopinterchange", cl::init(false), cl::Hidden, 92 cl::desc("Enable the new, experimental LoopInterchange Pass")); 93 94 static cl::opt<bool> EnableLoopDistribute( 95 "enable-loop-distribute", cl::init(false), cl::Hidden, 96 cl::desc("Enable the new, experimental LoopDistribution Pass")); 97 98 static cl::opt<bool> EnableNonLTOGlobalsModRef( 99 "enable-non-lto-gmr", cl::init(true), cl::Hidden, 100 cl::desc( 101 "Enable the GlobalsModRef AliasAnalysis outside of the LTO pipeline.")); 102 103 static cl::opt<bool> EnableLoopLoadElim( 104 "enable-loop-load-elim", cl::init(false), cl::Hidden, 105 cl::desc("Enable the new, experimental LoopLoadElimination Pass")); 106 107 PassManagerBuilder::PassManagerBuilder() { 108 OptLevel = 2; 109 SizeLevel = 0; 110 LibraryInfo = nullptr; 111 Inliner = nullptr; 112 FunctionIndex = nullptr; 113 DisableUnitAtATime = false; 114 DisableUnrollLoops = false; 115 BBVectorize = RunBBVectorization; 116 SLPVectorize = RunSLPVectorization; 117 LoopVectorize = RunLoopVectorization; 118 RerollLoops = RunLoopRerolling; 119 LoadCombine = RunLoadCombine; 120 DisableGVNLoadPRE = false; 121 VerifyInput = false; 122 VerifyOutput = false; 123 MergeFunctions = false; 124 PrepareForLTO = false; 125 } 126 127 PassManagerBuilder::~PassManagerBuilder() { 128 delete LibraryInfo; 129 delete Inliner; 130 } 131 132 /// Set of global extensions, automatically added as part of the standard set. 133 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy, 134 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions; 135 136 void PassManagerBuilder::addGlobalExtension( 137 PassManagerBuilder::ExtensionPointTy Ty, 138 PassManagerBuilder::ExtensionFn Fn) { 139 GlobalExtensions->push_back(std::make_pair(Ty, Fn)); 140 } 141 142 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) { 143 Extensions.push_back(std::make_pair(Ty, Fn)); 144 } 145 146 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy, 147 legacy::PassManagerBase &PM) const { 148 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i) 149 if ((*GlobalExtensions)[i].first == ETy) 150 (*GlobalExtensions)[i].second(*this, PM); 151 for (unsigned i = 0, e = Extensions.size(); i != e; ++i) 152 if (Extensions[i].first == ETy) 153 Extensions[i].second(*this, PM); 154 } 155 156 void PassManagerBuilder::addInitialAliasAnalysisPasses( 157 legacy::PassManagerBase &PM) const { 158 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 159 // BasicAliasAnalysis wins if they disagree. This is intended to help 160 // support "obvious" type-punning idioms. 161 if (UseCFLAA) 162 PM.add(createCFLAAWrapperPass()); 163 PM.add(createTypeBasedAAWrapperPass()); 164 PM.add(createScopedNoAliasAAWrapperPass()); 165 } 166 167 void PassManagerBuilder::populateFunctionPassManager( 168 legacy::FunctionPassManager &FPM) { 169 addExtensionsToPM(EP_EarlyAsPossible, FPM); 170 171 // Add LibraryInfo if we have some. 172 if (LibraryInfo) 173 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 174 175 if (OptLevel == 0) return; 176 177 addInitialAliasAnalysisPasses(FPM); 178 179 FPM.add(createCFGSimplificationPass()); 180 if (UseNewSROA) 181 FPM.add(createSROAPass()); 182 else 183 FPM.add(createScalarReplAggregatesPass()); 184 FPM.add(createEarlyCSEPass()); 185 FPM.add(createLowerExpectIntrinsicPass()); 186 } 187 188 void PassManagerBuilder::populateModulePassManager( 189 legacy::PassManagerBase &MPM) { 190 // If all optimizations are disabled, just run the always-inline pass and, 191 // if enabled, the function merging pass. 192 if (OptLevel == 0) { 193 if (Inliner) { 194 MPM.add(Inliner); 195 Inliner = nullptr; 196 } 197 198 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly 199 // creates a CGSCC pass manager, but we don't want to add extensions into 200 // that pass manager. To prevent this we insert a no-op module pass to reset 201 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0 202 // builds. The function merging pass is 203 if (MergeFunctions) 204 MPM.add(createMergeFunctionsPass()); 205 else if (!GlobalExtensions->empty() || !Extensions.empty()) 206 MPM.add(createBarrierNoopPass()); 207 208 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM); 209 return; 210 } 211 212 // Add LibraryInfo if we have some. 213 if (LibraryInfo) 214 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 215 216 addInitialAliasAnalysisPasses(MPM); 217 218 if (!DisableUnitAtATime) { 219 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM); 220 221 MPM.add(createIPSCCPPass()); // IP SCCP 222 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars 223 224 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination 225 226 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE 227 addExtensionsToPM(EP_Peephole, MPM); 228 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE 229 } 230 231 if (EnableNonLTOGlobalsModRef) 232 // We add a module alias analysis pass here. In part due to bugs in the 233 // analysis infrastructure this "works" in that the analysis stays alive 234 // for the entire SCC pass run below. 235 MPM.add(createGlobalsAAWrapperPass()); 236 237 // Start of CallGraph SCC passes. 238 if (!DisableUnitAtATime) 239 MPM.add(createPruneEHPass()); // Remove dead EH info 240 if (Inliner) { 241 MPM.add(Inliner); 242 Inliner = nullptr; 243 } 244 if (!DisableUnitAtATime) 245 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs 246 if (OptLevel > 2) 247 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args 248 249 // Start of function pass. 250 // Break up aggregate allocas, using SSAUpdater. 251 if (UseNewSROA) 252 MPM.add(createSROAPass()); 253 else 254 MPM.add(createScalarReplAggregatesPass(-1, false)); 255 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 256 MPM.add(createJumpThreadingPass()); // Thread jumps. 257 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals 258 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 259 MPM.add(createInstructionCombiningPass()); // Combine silly seq's 260 addExtensionsToPM(EP_Peephole, MPM); 261 262 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls 263 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 264 MPM.add(createReassociatePass()); // Reassociate expressions 265 // Rotate Loop - disable header duplication at -Oz 266 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 267 MPM.add(createLICMPass()); // Hoist loop invariants 268 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3)); 269 MPM.add(createCFGSimplificationPass()); 270 MPM.add(createInstructionCombiningPass()); 271 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars 272 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset. 273 MPM.add(createLoopDeletionPass()); // Delete dead loops 274 if (EnableLoopInterchange) { 275 MPM.add(createLoopInterchangePass()); // Interchange loops 276 MPM.add(createCFGSimplificationPass()); 277 } 278 if (!DisableUnrollLoops) 279 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops 280 addExtensionsToPM(EP_LoopOptimizerEnd, MPM); 281 282 if (OptLevel > 1) { 283 if (EnableMLSM) 284 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds 285 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 286 } 287 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset 288 MPM.add(createSCCPPass()); // Constant prop with SCCP 289 290 // Delete dead bit computations (instcombine runs after to fold away the dead 291 // computations, and then ADCE will run later to exploit any new DCE 292 // opportunities that creates). 293 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations 294 295 // Run instcombine after redundancy elimination to exploit opportunities 296 // opened up by them. 297 MPM.add(createInstructionCombiningPass()); 298 addExtensionsToPM(EP_Peephole, MPM); 299 MPM.add(createJumpThreadingPass()); // Thread jumps 300 MPM.add(createCorrelatedValuePropagationPass()); 301 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores 302 MPM.add(createLICMPass()); 303 304 addExtensionsToPM(EP_ScalarOptimizerLate, MPM); 305 306 if (RerollLoops) 307 MPM.add(createLoopRerollPass()); 308 if (!RunSLPAfterLoopVectorization) { 309 if (SLPVectorize) 310 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 311 312 if (BBVectorize) { 313 MPM.add(createBBVectorizePass()); 314 MPM.add(createInstructionCombiningPass()); 315 addExtensionsToPM(EP_Peephole, MPM); 316 if (OptLevel > 1 && UseGVNAfterVectorization) 317 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 318 else 319 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 320 321 // BBVectorize may have significantly shortened a loop body; unroll again. 322 if (!DisableUnrollLoops) 323 MPM.add(createLoopUnrollPass()); 324 } 325 } 326 327 if (LoadCombine) 328 MPM.add(createLoadCombinePass()); 329 330 MPM.add(createAggressiveDCEPass()); // Delete dead instructions 331 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 332 MPM.add(createInstructionCombiningPass()); // Clean up after everything. 333 addExtensionsToPM(EP_Peephole, MPM); 334 335 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC 336 // pass manager that we are specifically trying to avoid. To prevent this 337 // we must insert a no-op module pass to reset the pass manager. 338 MPM.add(createBarrierNoopPass()); 339 340 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO) { 341 // Remove avail extern fns and globals definitions if we aren't 342 // compiling an object file for later LTO. For LTO we want to preserve 343 // these so they are eligible for inlining at link-time. Note if they 344 // are unreferenced they will be removed by GlobalDCE later, so 345 // this only impacts referenced available externally globals. 346 // Eventually they will be suppressed during codegen, but eliminating 347 // here enables more opportunity for GlobalDCE as it may make 348 // globals referenced by available external functions dead 349 // and saves running remaining passes on the eliminated functions. 350 MPM.add(createEliminateAvailableExternallyPass()); 351 } 352 353 if (EnableNonLTOGlobalsModRef) 354 // We add a fresh GlobalsModRef run at this point. This is particularly 355 // useful as the above will have inlined, DCE'ed, and function-attr 356 // propagated everything. We should at this point have a reasonably minimal 357 // and richly annotated call graph. By computing aliasing and mod/ref 358 // information for all local globals here, the late loop passes and notably 359 // the vectorizer will be able to use them to help recognize vectorizable 360 // memory operations. 361 // 362 // Note that this relies on a bug in the pass manager which preserves 363 // a module analysis into a function pass pipeline (and throughout it) so 364 // long as the first function pass doesn't invalidate the module analysis. 365 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for 366 // this to work. Fortunately, it is trivial to preserve AliasAnalysis 367 // (doing nothing preserves it as it is required to be conservatively 368 // correct in the face of IR changes). 369 MPM.add(createGlobalsAAWrapperPass()); 370 371 if (RunFloat2Int) 372 MPM.add(createFloat2IntPass()); 373 374 addExtensionsToPM(EP_VectorizerStart, MPM); 375 376 // Re-rotate loops in all our loop nests. These may have fallout out of 377 // rotated form due to GVN or other transformations, and the vectorizer relies 378 // on the rotated form. Disable header duplication at -Oz. 379 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 380 381 // Distribute loops to allow partial vectorization. I.e. isolate dependences 382 // into separate loop that would otherwise inhibit vectorization. 383 if (EnableLoopDistribute) 384 MPM.add(createLoopDistributePass()); 385 386 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize)); 387 388 // Eliminate loads by forwarding stores from the previous iteration to loads 389 // of the current iteration. 390 if (EnableLoopLoadElim) 391 MPM.add(createLoopLoadEliminationPass()); 392 393 // FIXME: Because of #pragma vectorize enable, the passes below are always 394 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when 395 // on -O1 and no #pragma is found). Would be good to have these two passes 396 // as function calls, so that we can only pass them when the vectorizer 397 // changed the code. 398 MPM.add(createInstructionCombiningPass()); 399 if (OptLevel > 1 && ExtraVectorizerPasses) { 400 // At higher optimization levels, try to clean up any runtime overlap and 401 // alignment checks inserted by the vectorizer. We want to track correllated 402 // runtime checks for two inner loops in the same outer loop, fold any 403 // common computations, hoist loop-invariant aspects out of any outer loop, 404 // and unswitch the runtime checks if possible. Once hoisted, we may have 405 // dead (or speculatable) control flows or more combining opportunities. 406 MPM.add(createEarlyCSEPass()); 407 MPM.add(createCorrelatedValuePropagationPass()); 408 MPM.add(createInstructionCombiningPass()); 409 MPM.add(createLICMPass()); 410 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3)); 411 MPM.add(createCFGSimplificationPass()); 412 MPM.add(createInstructionCombiningPass()); 413 } 414 415 if (RunSLPAfterLoopVectorization) { 416 if (SLPVectorize) { 417 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 418 if (OptLevel > 1 && ExtraVectorizerPasses) { 419 MPM.add(createEarlyCSEPass()); 420 } 421 } 422 423 if (BBVectorize) { 424 MPM.add(createBBVectorizePass()); 425 MPM.add(createInstructionCombiningPass()); 426 addExtensionsToPM(EP_Peephole, MPM); 427 if (OptLevel > 1 && UseGVNAfterVectorization) 428 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 429 else 430 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 431 432 // BBVectorize may have significantly shortened a loop body; unroll again. 433 if (!DisableUnrollLoops) 434 MPM.add(createLoopUnrollPass()); 435 } 436 } 437 438 addExtensionsToPM(EP_Peephole, MPM); 439 MPM.add(createCFGSimplificationPass()); 440 MPM.add(createInstructionCombiningPass()); 441 442 if (!DisableUnrollLoops) { 443 MPM.add(createLoopUnrollPass()); // Unroll small loops 444 445 // LoopUnroll may generate some redundency to cleanup. 446 MPM.add(createInstructionCombiningPass()); 447 448 // Runtime unrolling will introduce runtime check in loop prologue. If the 449 // unrolled loop is a inner loop, then the prologue will be inside the 450 // outer loop. LICM pass can help to promote the runtime check out if the 451 // checked value is loop invariant. 452 MPM.add(createLICMPass()); 453 } 454 455 // After vectorization and unrolling, assume intrinsics may tell us more 456 // about pointer alignments. 457 MPM.add(createAlignmentFromAssumptionsPass()); 458 459 if (!DisableUnitAtATime) { 460 // FIXME: We shouldn't bother with this anymore. 461 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes 462 463 // GlobalOpt already deletes dead functions and globals, at -O2 try a 464 // late pass of GlobalDCE. It is capable of deleting dead cycles. 465 if (OptLevel > 1) { 466 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals. 467 MPM.add(createConstantMergePass()); // Merge dup global constants 468 } 469 } 470 471 if (MergeFunctions) 472 MPM.add(createMergeFunctionsPass()); 473 474 addExtensionsToPM(EP_OptimizerLast, MPM); 475 } 476 477 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) { 478 // Provide AliasAnalysis services for optimizations. 479 addInitialAliasAnalysisPasses(PM); 480 481 if (FunctionIndex) 482 PM.add(createFunctionImportPass(FunctionIndex)); 483 484 // Propagate constants at call sites into the functions they call. This 485 // opens opportunities for globalopt (and inlining) by substituting function 486 // pointers passed as arguments to direct uses of functions. 487 PM.add(createIPSCCPPass()); 488 489 // Now that we internalized some globals, see if we can hack on them! 490 PM.add(createFunctionAttrsPass()); // Add norecurse if possible. 491 PM.add(createGlobalOptimizerPass()); 492 493 // Linking modules together can lead to duplicated global constants, only 494 // keep one copy of each constant. 495 PM.add(createConstantMergePass()); 496 497 // Remove unused arguments from functions. 498 PM.add(createDeadArgEliminationPass()); 499 500 // Reduce the code after globalopt and ipsccp. Both can open up significant 501 // simplification opportunities, and both can propagate functions through 502 // function pointers. When this happens, we often have to resolve varargs 503 // calls, etc, so let instcombine do this. 504 PM.add(createInstructionCombiningPass()); 505 addExtensionsToPM(EP_Peephole, PM); 506 507 // Inline small functions 508 bool RunInliner = Inliner; 509 if (RunInliner) { 510 PM.add(Inliner); 511 Inliner = nullptr; 512 } 513 514 PM.add(createPruneEHPass()); // Remove dead EH info. 515 516 // Optimize globals again if we ran the inliner. 517 if (RunInliner) 518 PM.add(createGlobalOptimizerPass()); 519 PM.add(createGlobalDCEPass()); // Remove dead functions. 520 521 // If we didn't decide to inline a function, check to see if we can 522 // transform it to pass arguments by value instead of by reference. 523 PM.add(createArgumentPromotionPass()); 524 525 // The IPO passes may leave cruft around. Clean up after them. 526 PM.add(createInstructionCombiningPass()); 527 addExtensionsToPM(EP_Peephole, PM); 528 PM.add(createJumpThreadingPass()); 529 530 // Break up allocas 531 if (UseNewSROA) 532 PM.add(createSROAPass()); 533 else 534 PM.add(createScalarReplAggregatesPass()); 535 536 // Run a few AA driven optimizations here and now, to cleanup the code. 537 PM.add(createFunctionAttrsPass()); // Add nocapture. 538 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis. 539 540 PM.add(createLICMPass()); // Hoist loop invariants. 541 if (EnableMLSM) 542 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds. 543 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies. 544 PM.add(createMemCpyOptPass()); // Remove dead memcpys. 545 546 // Nuke dead stores. 547 PM.add(createDeadStoreEliminationPass()); 548 549 // More loops are countable; try to optimize them. 550 PM.add(createIndVarSimplifyPass()); 551 PM.add(createLoopDeletionPass()); 552 if (EnableLoopInterchange) 553 PM.add(createLoopInterchangePass()); 554 555 PM.add(createLoopVectorizePass(true, LoopVectorize)); 556 557 // More scalar chains could be vectorized due to more alias information 558 if (RunSLPAfterLoopVectorization) 559 if (SLPVectorize) 560 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 561 562 // After vectorization, assume intrinsics may tell us more about pointer 563 // alignments. 564 PM.add(createAlignmentFromAssumptionsPass()); 565 566 if (LoadCombine) 567 PM.add(createLoadCombinePass()); 568 569 // Cleanup and simplify the code after the scalar optimizations. 570 PM.add(createInstructionCombiningPass()); 571 addExtensionsToPM(EP_Peephole, PM); 572 573 PM.add(createJumpThreadingPass()); 574 } 575 576 void PassManagerBuilder::addLateLTOOptimizationPasses( 577 legacy::PassManagerBase &PM) { 578 // Delete basic blocks, which optimization passes may have killed. 579 PM.add(createCFGSimplificationPass()); 580 581 // Drop bodies of available externally objects to improve GlobalDCE. 582 PM.add(createEliminateAvailableExternallyPass()); 583 584 // Now that we have optimized the program, discard unreachable functions. 585 PM.add(createGlobalDCEPass()); 586 587 // FIXME: this is profitable (for compiler time) to do at -O0 too, but 588 // currently it damages debug info. 589 if (MergeFunctions) 590 PM.add(createMergeFunctionsPass()); 591 } 592 593 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) { 594 if (LibraryInfo) 595 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 596 597 if (VerifyInput) 598 PM.add(createVerifierPass()); 599 600 if (OptLevel > 1) 601 addLTOOptimizationPasses(PM); 602 603 // Lower bit sets to globals. This pass supports Clang's control flow 604 // integrity mechanisms (-fsanitize=cfi*) and needs to run at link time if CFI 605 // is enabled. The pass does nothing if CFI is disabled. 606 PM.add(createLowerBitSetsPass()); 607 608 if (OptLevel != 0) 609 addLateLTOOptimizationPasses(PM); 610 611 if (VerifyOutput) 612 PM.add(createVerifierPass()); 613 } 614 615 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) { 616 return reinterpret_cast<PassManagerBuilder*>(P); 617 } 618 619 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) { 620 return reinterpret_cast<LLVMPassManagerBuilderRef>(P); 621 } 622 623 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() { 624 PassManagerBuilder *PMB = new PassManagerBuilder(); 625 return wrap(PMB); 626 } 627 628 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) { 629 PassManagerBuilder *Builder = unwrap(PMB); 630 delete Builder; 631 } 632 633 void 634 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB, 635 unsigned OptLevel) { 636 PassManagerBuilder *Builder = unwrap(PMB); 637 Builder->OptLevel = OptLevel; 638 } 639 640 void 641 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB, 642 unsigned SizeLevel) { 643 PassManagerBuilder *Builder = unwrap(PMB); 644 Builder->SizeLevel = SizeLevel; 645 } 646 647 void 648 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB, 649 LLVMBool Value) { 650 PassManagerBuilder *Builder = unwrap(PMB); 651 Builder->DisableUnitAtATime = Value; 652 } 653 654 void 655 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB, 656 LLVMBool Value) { 657 PassManagerBuilder *Builder = unwrap(PMB); 658 Builder->DisableUnrollLoops = Value; 659 } 660 661 void 662 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB, 663 LLVMBool Value) { 664 // NOTE: The simplify-libcalls pass has been removed. 665 } 666 667 void 668 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB, 669 unsigned Threshold) { 670 PassManagerBuilder *Builder = unwrap(PMB); 671 Builder->Inliner = createFunctionInliningPass(Threshold); 672 } 673 674 void 675 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB, 676 LLVMPassManagerRef PM) { 677 PassManagerBuilder *Builder = unwrap(PMB); 678 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM); 679 Builder->populateFunctionPassManager(*FPM); 680 } 681 682 void 683 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB, 684 LLVMPassManagerRef PM) { 685 PassManagerBuilder *Builder = unwrap(PMB); 686 legacy::PassManagerBase *MPM = unwrap(PM); 687 Builder->populateModulePassManager(*MPM); 688 } 689 690 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB, 691 LLVMPassManagerRef PM, 692 LLVMBool Internalize, 693 LLVMBool RunInliner) { 694 PassManagerBuilder *Builder = unwrap(PMB); 695 legacy::PassManagerBase *LPM = unwrap(PM); 696 697 // A small backwards compatibility hack. populateLTOPassManager used to take 698 // an RunInliner option. 699 if (RunInliner && !Builder->Inliner) 700 Builder->Inliner = createFunctionInliningPass(); 701 702 Builder->populateLTOPassManager(*LPM); 703 } 704