1 //===- opt.cpp - The LLVM Modular Optimizer -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Optimizations may be specified an arbitrary number of times on the command
11 // line, They are run in the order specified.
13 //===----------------------------------------------------------------------===//
15 #include "BreakpointPrinter.h"
16 #include "NewPMDriver.h"
17 #include "PassPrinters.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/Analysis/CallGraph.h"
20 #include "llvm/Analysis/CallGraphSCCPass.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/RegionPass.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/Bitcode/BitcodeWriterPass.h"
26 #include "llvm/CodeGen/CommandFlags.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DebugInfo.h"
29 #include "llvm/IR/IRPrintingPasses.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/LegacyPassManager.h"
32 #include "llvm/IR/LegacyPassNameParser.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Verifier.h"
35 #include "llvm/IRReader/IRReader.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/LinkAllIR.h"
38 #include "llvm/LinkAllPasses.h"
39 #include "llvm/MC/SubtargetFeature.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/FileSystem.h"
42 #include "llvm/Support/Host.h"
43 #include "llvm/Support/ManagedStatic.h"
44 #include "llvm/Support/PluginLoader.h"
45 #include "llvm/Support/PrettyStackTrace.h"
46 #include "llvm/Support/Signals.h"
47 #include "llvm/Support/SourceMgr.h"
48 #include "llvm/Support/SystemUtils.h"
49 #include "llvm/Support/TargetRegistry.h"
50 #include "llvm/Support/TargetSelect.h"
51 #include "llvm/Support/ToolOutputFile.h"
52 #include "llvm/Support/YAMLTraits.h"
53 #include "llvm/Target/TargetMachine.h"
54 #include "llvm/Transforms/Coroutines.h"
55 #include "llvm/Transforms/IPO/AlwaysInliner.h"
56 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
57 #include "llvm/Transforms/Utils/Cloning.h"
61 using namespace opt_tool;
63 // The OptimizationList is automatically populated with registered Passes by the
66 static cl::list<const PassInfo*, bool, PassNameParser>
67 PassList(cl::desc("Optimizations available:"));
69 // This flag specifies a textual description of the optimization pass pipeline
70 // to run over the module. This flag switches opt to use the new pass manager
71 // infrastructure, completely disabling all of the flags specific to the old
73 static cl::opt<std::string> PassPipeline(
75 cl::desc("A textual description of the pass pipeline for optimizing"),
78 // Other command line options...
80 static cl::opt<std::string>
81 InputFilename(cl::Positional, cl::desc("<input bitcode file>"),
82 cl::init("-"), cl::value_desc("filename"));
84 static cl::opt<std::string>
85 OutputFilename("o", cl::desc("Override output filename"),
86 cl::value_desc("filename"));
89 Force("f", cl::desc("Enable binary output on terminals"));
92 PrintEachXForm("p", cl::desc("Print module after each transformation"));
95 NoOutput("disable-output",
96 cl::desc("Do not write result bitcode file"), cl::Hidden);
99 OutputAssembly("S", cl::desc("Write output as LLVM assembly"));
102 OutputThinLTOBC("thinlto-bc",
103 cl::desc("Write output as ThinLTO-ready bitcode"));
106 NoVerify("disable-verify", cl::desc("Do not run the verifier"), cl::Hidden);
109 VerifyEach("verify-each", cl::desc("Verify after each transform"));
112 DisableDITypeMap("disable-debug-info-type-map",
113 cl::desc("Don't use a uniquing type map for debug info"));
116 StripDebug("strip-debug",
117 cl::desc("Strip debugger symbol info from translation unit"));
120 DisableInline("disable-inlining", cl::desc("Do not run the inliner pass"));
123 DisableOptimizations("disable-opt",
124 cl::desc("Do not run any optimization passes"));
127 StandardLinkOpts("std-link-opts",
128 cl::desc("Include the standard link time optimizations"));
132 cl::desc("Optimization level 0. Similar to clang -O0"));
136 cl::desc("Optimization level 1. Similar to clang -O1"));
140 cl::desc("Optimization level 2. Similar to clang -O2"));
144 cl::desc("Like -O2 with extra optimizations for size. Similar to clang -Os"));
148 cl::desc("Like -Os but reduces code size further. Similar to clang -Oz"));
152 cl::desc("Optimization level 3. Similar to clang -O3"));
154 static cl::opt<unsigned>
155 CodeGenOptLevel("codegen-opt-level",
156 cl::desc("Override optimization level for codegen hooks"));
158 static cl::opt<std::string>
159 TargetTriple("mtriple", cl::desc("Override target triple for module"));
162 UnitAtATime("funit-at-a-time",
163 cl::desc("Enable IPO. This corresponds to gcc's -funit-at-a-time"),
167 DisableLoopUnrolling("disable-loop-unrolling",
168 cl::desc("Disable loop unrolling in all relevant passes"),
171 DisableLoopVectorization("disable-loop-vectorization",
172 cl::desc("Disable the loop vectorization pass"),
176 DisableSLPVectorization("disable-slp-vectorization",
177 cl::desc("Disable the slp vectorization pass"),
180 static cl::opt<bool> EmitSummaryIndex("module-summary",
181 cl::desc("Emit module summary index"),
184 static cl::opt<bool> EmitModuleHash("module-hash", cl::desc("Emit module hash"),
188 DisableSimplifyLibCalls("disable-simplify-libcalls",
189 cl::desc("Disable simplify-libcalls"));
192 Quiet("q", cl::desc("Obsolete option"), cl::Hidden);
195 QuietA("quiet", cl::desc("Alias for -q"), cl::aliasopt(Quiet));
198 AnalyzeOnly("analyze", cl::desc("Only perform analysis, no optimization"));
201 PrintBreakpoints("print-breakpoints-for-testing",
202 cl::desc("Print select breakpoints location for testing"));
204 static cl::opt<std::string>
205 DefaultDataLayout("default-data-layout",
206 cl::desc("data layout string to use if not specified by module"),
207 cl::value_desc("layout-string"), cl::init(""));
209 static cl::opt<bool> PreserveBitcodeUseListOrder(
210 "preserve-bc-uselistorder",
211 cl::desc("Preserve use-list order when writing LLVM bitcode."),
212 cl::init(true), cl::Hidden);
214 static cl::opt<bool> PreserveAssemblyUseListOrder(
215 "preserve-ll-uselistorder",
216 cl::desc("Preserve use-list order when writing LLVM assembly."),
217 cl::init(false), cl::Hidden);
220 RunTwice("run-twice",
221 cl::desc("Run all passes twice, re-using the same pass manager."),
222 cl::init(false), cl::Hidden);
224 static cl::opt<bool> DiscardValueNames(
225 "discard-value-names",
226 cl::desc("Discard names from Value (other than GlobalValue)."),
227 cl::init(false), cl::Hidden);
229 static cl::opt<bool> Coroutines(
231 cl::desc("Enable coroutine passes."),
232 cl::init(false), cl::Hidden);
234 static cl::opt<bool> PassRemarksWithHotness(
235 "pass-remarks-with-hotness",
236 cl::desc("With PGO, include profile count in optimization remarks"),
239 static cl::opt<std::string>
240 RemarksFilename("pass-remarks-output",
241 cl::desc("YAML output filename for pass remarks"),
242 cl::value_desc("filename"));
244 static inline void addPass(legacy::PassManagerBase &PM, Pass *P) {
245 // Add the pass to the pass manager...
248 // If we are verifying all of the intermediate steps, add the verifier...
250 PM.add(createVerifierPass());
253 /// This routine adds optimization passes based on selected optimization level,
256 /// OptLevel - Optimization Level
257 static void AddOptimizationPasses(legacy::PassManagerBase &MPM,
258 legacy::FunctionPassManager &FPM,
259 TargetMachine *TM, unsigned OptLevel,
260 unsigned SizeLevel) {
261 if (!NoVerify || VerifyEach)
262 FPM.add(createVerifierPass()); // Verify that input is correct
264 PassManagerBuilder Builder;
265 Builder.OptLevel = OptLevel;
266 Builder.SizeLevel = SizeLevel;
270 } else if (OptLevel > 1) {
271 Builder.Inliner = createFunctionInliningPass(OptLevel, SizeLevel);
273 Builder.Inliner = createAlwaysInlinerLegacyPass();
275 Builder.DisableUnitAtATime = !UnitAtATime;
276 Builder.DisableUnrollLoops = (DisableLoopUnrolling.getNumOccurrences() > 0) ?
277 DisableLoopUnrolling : OptLevel == 0;
279 // This is final, unless there is a #pragma vectorize enable
280 if (DisableLoopVectorization)
281 Builder.LoopVectorize = false;
282 // If option wasn't forced via cmd line (-vectorize-loops, -loop-vectorize)
283 else if (!Builder.LoopVectorize)
284 Builder.LoopVectorize = OptLevel > 1 && SizeLevel < 2;
286 // When #pragma vectorize is on for SLP, do the same as above
287 Builder.SLPVectorize =
288 DisableSLPVectorization ? false : OptLevel > 1 && SizeLevel < 2;
290 // Add target-specific passes that need to run as early as possible.
292 Builder.addExtension(
293 PassManagerBuilder::EP_EarlyAsPossible,
294 [&](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
295 TM->addEarlyAsPossiblePasses(PM);
299 addCoroutinePassesToExtensionPoints(Builder);
301 Builder.populateFunctionPassManager(FPM);
302 Builder.populateModulePassManager(MPM);
305 static void AddStandardLinkPasses(legacy::PassManagerBase &PM) {
306 PassManagerBuilder Builder;
307 Builder.VerifyInput = true;
308 if (DisableOptimizations)
309 Builder.OptLevel = 0;
312 Builder.Inliner = createFunctionInliningPass();
313 Builder.populateLTOPassManager(PM);
316 //===----------------------------------------------------------------------===//
317 // CodeGen-related helper functions.
320 static CodeGenOpt::Level GetCodeGenOptLevel() {
321 if (CodeGenOptLevel.getNumOccurrences())
322 return static_cast<CodeGenOpt::Level>(unsigned(CodeGenOptLevel));
324 return CodeGenOpt::Less;
326 return CodeGenOpt::Default;
328 return CodeGenOpt::Aggressive;
329 return CodeGenOpt::None;
332 // Returns the TargetMachine instance or zero if no triple is provided.
333 static TargetMachine* GetTargetMachine(Triple TheTriple, StringRef CPUStr,
334 StringRef FeaturesStr,
335 const TargetOptions &Options) {
337 const Target *TheTarget = TargetRegistry::lookupTarget(MArch, TheTriple,
339 // Some modules don't specify a triple, and this is okay.
344 return TheTarget->createTargetMachine(TheTriple.getTriple(), CPUStr,
345 FeaturesStr, Options, getRelocModel(),
346 CMModel, GetCodeGenOptLevel());
349 #ifdef LINK_POLLY_INTO_TOOLS
351 void initializePollyPasses(llvm::PassRegistry &Registry);
355 //===----------------------------------------------------------------------===//
358 int main(int argc, char **argv) {
359 sys::PrintStackTraceOnErrorSignal(argv[0]);
360 llvm::PrettyStackTraceProgram X(argc, argv);
362 // Enable debug stream buffering.
363 EnableDebugBuffering = true;
365 llvm_shutdown_obj Y; // Call llvm_shutdown() on exit.
368 InitializeAllTargets();
369 InitializeAllTargetMCs();
370 InitializeAllAsmPrinters();
371 InitializeAllAsmParsers();
374 PassRegistry &Registry = *PassRegistry::getPassRegistry();
375 initializeCore(Registry);
376 initializeCoroutines(Registry);
377 initializeScalarOpts(Registry);
378 initializeObjCARCOpts(Registry);
379 initializeVectorization(Registry);
380 initializeIPO(Registry);
381 initializeAnalysis(Registry);
382 initializeTransformUtils(Registry);
383 initializeInstCombine(Registry);
384 initializeInstrumentation(Registry);
385 initializeTarget(Registry);
386 // For codegen passes, only passes that do IR to IR transformation are
388 initializeCodeGenPreparePass(Registry);
389 initializeAtomicExpandPass(Registry);
390 initializeRewriteSymbolsLegacyPassPass(Registry);
391 initializeWinEHPreparePass(Registry);
392 initializeDwarfEHPreparePass(Registry);
393 initializeSafeStackPass(Registry);
394 initializeSjLjEHPreparePass(Registry);
395 initializePreISelIntrinsicLoweringLegacyPassPass(Registry);
396 initializeGlobalMergePass(Registry);
397 initializeInterleavedAccessPass(Registry);
398 initializeCountingFunctionInserterPass(Registry);
399 initializeUnreachableBlockElimLegacyPassPass(Registry);
401 #ifdef LINK_POLLY_INTO_TOOLS
402 polly::initializePollyPasses(Registry);
405 cl::ParseCommandLineOptions(argc, argv,
406 "llvm .bc -> .bc modular optimizer and analysis printer\n");
408 if (AnalyzeOnly && NoOutput) {
409 errs() << argv[0] << ": analyze mode conflicts with no-output mode.\n";
415 Context.setDiscardValueNames(DiscardValueNames);
416 if (!DisableDITypeMap)
417 Context.enableDebugTypeODRUniquing();
419 if (PassRemarksWithHotness)
420 Context.setDiagnosticHotnessRequested(true);
422 std::unique_ptr<tool_output_file> YamlFile;
423 if (RemarksFilename != "") {
425 YamlFile = llvm::make_unique<tool_output_file>(RemarksFilename, EC,
428 errs() << EC.message() << '\n';
431 Context.setDiagnosticsOutputFile(
432 llvm::make_unique<yaml::Output>(YamlFile->os()));
435 // Load the input module...
436 std::unique_ptr<Module> M = parseIRFile(InputFilename, Err, Context);
439 Err.print(argv[0], errs());
443 // Strip debug info before running the verifier.
447 // Immediately run the verifier to catch any problems before starting up the
448 // pass pipelines. Otherwise we can crash on broken code during
449 // doInitialization().
450 if (!NoVerify && verifyModule(*M, &errs())) {
451 errs() << argv[0] << ": " << InputFilename
452 << ": error: input module is broken!\n";
456 // If we are supposed to override the target triple, do so now.
457 if (!TargetTriple.empty())
458 M->setTargetTriple(Triple::normalize(TargetTriple));
460 // Figure out what stream we are supposed to write to...
461 std::unique_ptr<tool_output_file> Out;
463 if (!OutputFilename.empty())
464 errs() << "WARNING: The -o (output filename) option is ignored when\n"
465 "the --disable-output option is used.\n";
467 // Default to standard output.
468 if (OutputFilename.empty())
469 OutputFilename = "-";
472 Out.reset(new tool_output_file(OutputFilename, EC, sys::fs::F_None));
474 errs() << EC.message() << '\n';
479 Triple ModuleTriple(M->getTargetTriple());
480 std::string CPUStr, FeaturesStr;
481 TargetMachine *Machine = nullptr;
482 const TargetOptions Options = InitTargetOptionsFromCodeGenFlags();
484 if (ModuleTriple.getArch()) {
485 CPUStr = getCPUStr();
486 FeaturesStr = getFeaturesStr();
487 Machine = GetTargetMachine(ModuleTriple, CPUStr, FeaturesStr, Options);
490 std::unique_ptr<TargetMachine> TM(Machine);
492 // Override function attributes based on CPUStr, FeaturesStr, and command line
494 setFunctionAttributes(CPUStr, FeaturesStr, *M);
496 // If the output is set to be emitted to standard out, and standard out is a
497 // console, print out a warning message and refuse to do it. We don't
498 // impress anyone by spewing tons of binary goo to a terminal.
499 if (!Force && !NoOutput && !AnalyzeOnly && !OutputAssembly)
500 if (CheckBitcodeOutputToConsole(Out->os(), !Quiet))
503 if (PassPipeline.getNumOccurrences() > 0) {
504 OutputKind OK = OK_NoOutput;
506 OK = OutputAssembly ? OK_OutputAssembly : OK_OutputBitcode;
508 VerifierKind VK = VK_VerifyInAndOut;
512 VK = VK_VerifyEachPass;
514 // The user has asked to use the new pass manager and provided a pipeline
515 // string. Hand off the rest of the functionality to the new code for that
517 return runPassPipeline(argv[0], *M, TM.get(), Out.get(),
518 PassPipeline, OK, VK, PreserveAssemblyUseListOrder,
519 PreserveBitcodeUseListOrder, EmitSummaryIndex,
525 // Create a PassManager to hold and optimize the collection of passes we are
528 legacy::PassManager Passes;
530 // Add an appropriate TargetLibraryInfo pass for the module's triple.
531 TargetLibraryInfoImpl TLII(ModuleTriple);
533 // The -disable-simplify-libcalls flag actually disables all builtin optzns.
534 if (DisableSimplifyLibCalls)
535 TLII.disableAllFunctions();
536 Passes.add(new TargetLibraryInfoWrapperPass(TLII));
538 // Add an appropriate DataLayout instance for this module.
539 const DataLayout &DL = M->getDataLayout();
540 if (DL.isDefault() && !DefaultDataLayout.empty()) {
541 M->setDataLayout(DefaultDataLayout);
544 // Add internal analysis passes from the target machine.
545 Passes.add(createTargetTransformInfoWrapperPass(TM ? TM->getTargetIRAnalysis()
546 : TargetIRAnalysis()));
548 std::unique_ptr<legacy::FunctionPassManager> FPasses;
549 if (OptLevelO0 || OptLevelO1 || OptLevelO2 || OptLevelOs || OptLevelOz ||
551 FPasses.reset(new legacy::FunctionPassManager(M.get()));
552 FPasses->add(createTargetTransformInfoWrapperPass(
553 TM ? TM->getTargetIRAnalysis() : TargetIRAnalysis()));
556 if (PrintBreakpoints) {
557 // Default to standard output.
559 if (OutputFilename.empty())
560 OutputFilename = "-";
563 Out = llvm::make_unique<tool_output_file>(OutputFilename, EC,
566 errs() << EC.message() << '\n';
570 Passes.add(createBreakpointPrinter(Out->os()));
574 // Create a new optimization pass for each one specified on the command line
575 for (unsigned i = 0; i < PassList.size(); ++i) {
576 if (StandardLinkOpts &&
577 StandardLinkOpts.getPosition() < PassList.getPosition(i)) {
578 AddStandardLinkPasses(Passes);
579 StandardLinkOpts = false;
582 if (OptLevelO0 && OptLevelO0.getPosition() < PassList.getPosition(i)) {
583 AddOptimizationPasses(Passes, *FPasses, TM.get(), 0, 0);
587 if (OptLevelO1 && OptLevelO1.getPosition() < PassList.getPosition(i)) {
588 AddOptimizationPasses(Passes, *FPasses, TM.get(), 1, 0);
592 if (OptLevelO2 && OptLevelO2.getPosition() < PassList.getPosition(i)) {
593 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 0);
597 if (OptLevelOs && OptLevelOs.getPosition() < PassList.getPosition(i)) {
598 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 1);
602 if (OptLevelOz && OptLevelOz.getPosition() < PassList.getPosition(i)) {
603 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 2);
607 if (OptLevelO3 && OptLevelO3.getPosition() < PassList.getPosition(i)) {
608 AddOptimizationPasses(Passes, *FPasses, TM.get(), 3, 0);
612 const PassInfo *PassInf = PassList[i];
614 if (PassInf->getTargetMachineCtor())
615 P = PassInf->getTargetMachineCtor()(TM.get());
616 else if (PassInf->getNormalCtor())
617 P = PassInf->getNormalCtor()();
619 errs() << argv[0] << ": cannot create pass: "
620 << PassInf->getPassName() << "\n";
622 PassKind Kind = P->getPassKind();
628 Passes.add(createBasicBlockPassPrinter(PassInf, Out->os(), Quiet));
631 Passes.add(createRegionPassPrinter(PassInf, Out->os(), Quiet));
634 Passes.add(createLoopPassPrinter(PassInf, Out->os(), Quiet));
637 Passes.add(createFunctionPassPrinter(PassInf, Out->os(), Quiet));
639 case PT_CallGraphSCC:
640 Passes.add(createCallGraphPassPrinter(PassInf, Out->os(), Quiet));
643 Passes.add(createModulePassPrinter(PassInf, Out->os(), Quiet));
651 createPrintModulePass(errs(), "", PreserveAssemblyUseListOrder));
654 if (StandardLinkOpts) {
655 AddStandardLinkPasses(Passes);
656 StandardLinkOpts = false;
660 AddOptimizationPasses(Passes, *FPasses, TM.get(), 0, 0);
663 AddOptimizationPasses(Passes, *FPasses, TM.get(), 1, 0);
666 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 0);
669 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 1);
672 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 2);
675 AddOptimizationPasses(Passes, *FPasses, TM.get(), 3, 0);
678 FPasses->doInitialization();
679 for (Function &F : *M)
681 FPasses->doFinalization();
684 // Check that the module is well formed on completion of optimization
685 if (!NoVerify && !VerifyEach)
686 Passes.add(createVerifierPass());
688 // In run twice mode, we want to make sure the output is bit-by-bit
689 // equivalent if we run the pass manager again, so setup two buffers and
690 // a stream to write to them. Note that llc does something similar and it
691 // may be worth to abstract this out in the future.
692 SmallVector<char, 0> Buffer;
693 SmallVector<char, 0> CompileTwiceBuffer;
694 std::unique_ptr<raw_svector_ostream> BOS;
695 raw_ostream *OS = nullptr;
697 // Write bitcode or assembly to the output as the last step...
698 if (!NoOutput && !AnalyzeOnly) {
702 BOS = make_unique<raw_svector_ostream>(Buffer);
705 if (OutputAssembly) {
706 if (EmitSummaryIndex)
707 report_fatal_error("Text output is incompatible with -module-summary");
709 report_fatal_error("Text output is incompatible with -module-hash");
710 Passes.add(createPrintModulePass(*OS, "", PreserveAssemblyUseListOrder));
711 } else if (OutputThinLTOBC)
712 Passes.add(createWriteThinLTOBitcodePass(*OS));
714 Passes.add(createBitcodeWriterPass(*OS, PreserveBitcodeUseListOrder,
715 EmitSummaryIndex, EmitModuleHash));
718 // Before executing passes, print the final values of the LLVM options.
719 cl::PrintOptionValues();
721 // If requested, run all passes again with the same pass manager to catch
722 // bugs caused by persistent state in the passes
724 std::unique_ptr<Module> M2(CloneModule(M.get()));
726 CompileTwiceBuffer = Buffer;
730 // Now that we have all of the passes ready, run them.
733 // Compare the two outputs and make sure they're the same
736 if (Buffer.size() != CompileTwiceBuffer.size() ||
737 (memcmp(Buffer.data(), CompileTwiceBuffer.data(), Buffer.size()) !=
739 errs() << "Running the pass manager twice changed the output.\n"
740 "Writing the result of the second run to the specified output.\n"
741 "To generate the one-run comparison binary, just run without\n"
742 "the compile-twice option\n";
743 Out->os() << BOS->str();
749 Out->os() << BOS->str();
753 if (!NoOutput || PrintBreakpoints)