1 //===- opt.cpp - The LLVM Modular Optimizer -------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Optimizations may be specified an arbitrary number of times on the command
10 // line, They are run in the order specified.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "BreakpointPrinter.h"
15 #include "NewPMDriver.h"
16 #include "PassPrinters.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/Analysis/CallGraph.h"
19 #include "llvm/Analysis/CallGraphSCCPass.h"
20 #include "llvm/Analysis/LoopPass.h"
21 #include "llvm/Analysis/RegionPass.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/AsmParser/Parser.h"
25 #include "llvm/CodeGen/CommandFlags.h"
26 #include "llvm/CodeGen/TargetPassConfig.h"
27 #include "llvm/Config/llvm-config.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DebugInfo.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/LLVMRemarkStreamer.h"
32 #include "llvm/IR/LegacyPassManager.h"
33 #include "llvm/IR/LegacyPassNameParser.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/Verifier.h"
36 #include "llvm/IRReader/IRReader.h"
37 #include "llvm/InitializePasses.h"
38 #include "llvm/LinkAllIR.h"
39 #include "llvm/LinkAllPasses.h"
40 #include "llvm/MC/SubtargetFeature.h"
41 #include "llvm/Remarks/HotnessThresholdParser.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/FileSystem.h"
44 #include "llvm/Support/Host.h"
45 #include "llvm/Support/InitLLVM.h"
46 #include "llvm/Support/PluginLoader.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/IPO/WholeProgramDevirt.h"
58 #include "llvm/Transforms/Utils/Cloning.h"
59 #include "llvm/Transforms/Utils/Debugify.h"
60 #include <algorithm>
61 #include <memory>
62 using namespace llvm;
63 using namespace opt_tool;
64
65 static codegen::RegisterCodeGenFlags CFG;
66
67 // The OptimizationList is automatically populated with registered Passes by the
68 // PassNameParser.
69 static cl::list<const PassInfo *, bool, PassNameParser> PassList(cl::desc(
70 "Optimizations available (use '-passes=' for the new pass manager)"));
71
72 static cl::opt<bool> EnableNewPassManager(
73 "enable-new-pm",
74 cl::desc("Enable the new pass manager, translating "
75 "'opt -foo' to 'opt -passes=foo'. This is strictly for the new PM "
76 "migration, use '-passes=' when possible."),
77 cl::init(LLVM_ENABLE_NEW_PASS_MANAGER));
78
79 // This flag specifies a textual description of the optimization pass pipeline
80 // to run over the module. This flag switches opt to use the new pass manager
81 // infrastructure, completely disabling all of the flags specific to the old
82 // pass management.
83 static cl::opt<std::string> PassPipeline(
84 "passes",
85 cl::desc(
86 "A textual description of the pass pipeline. To have analysis passes "
87 "available before a certain pass, add 'require<foo-analysis>'."));
88
89 static cl::opt<bool> PrintPasses("print-passes",
90 cl::desc("Print available passes that can be "
91 "specified in -passes=foo and exit"));
92
93 static cl::opt<std::string>
94 InputFilename(cl::Positional, cl::desc("<input bitcode file>"),
95 cl::init("-"), cl::value_desc("filename"));
96
97 static cl::opt<std::string>
98 OutputFilename("o", cl::desc("Override output filename"),
99 cl::value_desc("filename"));
100
101 static cl::opt<bool>
102 Force("f", cl::desc("Enable binary output on terminals"));
103
104 static cl::opt<bool>
105 PrintEachXForm("p", cl::desc("Print module after each transformation"));
106
107 static cl::opt<bool>
108 NoOutput("disable-output",
109 cl::desc("Do not write result bitcode file"), cl::Hidden);
110
111 static cl::opt<bool>
112 OutputAssembly("S", cl::desc("Write output as LLVM assembly"));
113
114 static cl::opt<bool>
115 OutputThinLTOBC("thinlto-bc",
116 cl::desc("Write output as ThinLTO-ready bitcode"));
117
118 static cl::opt<bool>
119 SplitLTOUnit("thinlto-split-lto-unit",
120 cl::desc("Enable splitting of a ThinLTO LTOUnit"));
121
122 static cl::opt<std::string> ThinLinkBitcodeFile(
123 "thin-link-bitcode-file", cl::value_desc("filename"),
124 cl::desc(
125 "A file in which to write minimized bitcode for the thin link only"));
126
127 static cl::opt<bool>
128 NoVerify("disable-verify", cl::desc("Do not run the verifier"), cl::Hidden);
129
130 static cl::opt<bool> NoUpgradeDebugInfo("disable-upgrade-debug-info",
131 cl::desc("Generate invalid output"),
132 cl::ReallyHidden);
133
134 static cl::opt<bool> VerifyEach("verify-each",
135 cl::desc("Verify after each transform"));
136
137 static cl::opt<bool>
138 DisableDITypeMap("disable-debug-info-type-map",
139 cl::desc("Don't use a uniquing type map for debug info"));
140
141 static cl::opt<bool>
142 StripDebug("strip-debug",
143 cl::desc("Strip debugger symbol info from translation unit"));
144
145 static cl::opt<bool>
146 StripNamedMetadata("strip-named-metadata",
147 cl::desc("Strip module-level named metadata"));
148
149 static cl::opt<bool>
150 DisableInline("disable-inlining",
151 cl::desc("Do not run the inliner pass (legacy PM only)"));
152
153 static cl::opt<bool>
154 DisableOptimizations("disable-opt",
155 cl::desc("Do not run any optimization passes"));
156
157 static cl::opt<bool> StandardLinkOpts(
158 "std-link-opts",
159 cl::desc("Include the standard link time optimizations (legacy PM only)"));
160
161 static cl::opt<bool>
162 OptLevelO0("O0", cl::desc("Optimization level 0. Similar to clang -O0. "
163 "Use -passes='default<O0>' for the new PM"));
164
165 static cl::opt<bool>
166 OptLevelO1("O1", cl::desc("Optimization level 1. Similar to clang -O1. "
167 "Use -passes='default<O1>' for the new PM"));
168
169 static cl::opt<bool>
170 OptLevelO2("O2", cl::desc("Optimization level 2. Similar to clang -O2. "
171 "Use -passes='default<O2>' for the new PM"));
172
173 static cl::opt<bool>
174 OptLevelOs("Os", cl::desc("Like -O2 but size-conscious. Similar to clang "
175 "-Os. Use -passes='default<Os>' for the new PM"));
176
177 static cl::opt<bool> OptLevelOz(
178 "Oz",
179 cl::desc("Like -O2 but optimize for code size above all else. Similar to "
180 "clang -Oz. Use -passes='default<Oz>' for the new PM"));
181
182 static cl::opt<bool>
183 OptLevelO3("O3", cl::desc("Optimization level 3. Similar to clang -O3. "
184 "Use -passes='default<O3>' for the new PM"));
185
186 static cl::opt<unsigned> CodeGenOptLevel(
187 "codegen-opt-level",
188 cl::desc("Override optimization level for codegen hooks, legacy PM only"));
189
190 static cl::opt<std::string>
191 TargetTriple("mtriple", cl::desc("Override target triple for module"));
192
193 cl::opt<bool> DisableLoopUnrolling(
194 "disable-loop-unrolling",
195 cl::desc("Disable loop unrolling in all relevant passes"), cl::init(false));
196
197 static cl::opt<bool> EmitSummaryIndex("module-summary",
198 cl::desc("Emit module summary index"),
199 cl::init(false));
200
201 static cl::opt<bool> EmitModuleHash("module-hash", cl::desc("Emit module hash"),
202 cl::init(false));
203
204 static cl::opt<bool>
205 DisableSimplifyLibCalls("disable-simplify-libcalls",
206 cl::desc("Disable simplify-libcalls"));
207
208 static cl::list<std::string>
209 DisableBuiltins("disable-builtin",
210 cl::desc("Disable specific target library builtin function"),
211 cl::ZeroOrMore);
212
213 static cl::opt<bool>
214 AnalyzeOnly("analyze", cl::desc("Only perform analysis, no optimization. "
215 "Legacy pass manager only."));
216
217 static cl::opt<bool> EnableDebugify(
218 "enable-debugify",
219 cl::desc(
220 "Start the pipeline with debugify and end it with check-debugify"));
221
222 static cl::opt<bool> VerifyDebugInfoPreserve(
223 "verify-debuginfo-preserve",
224 cl::desc("Start the pipeline with collecting and end it with checking of "
225 "debug info preservation."));
226
227 static cl::opt<bool> VerifyEachDebugInfoPreserve(
228 "verify-each-debuginfo-preserve",
229 cl::desc("Start each pass with collecting and end it with checking of "
230 "debug info preservation."));
231
232 static cl::opt<std::string>
233 VerifyDIPreserveExport("verify-di-preserve-export",
234 cl::desc("Export debug info preservation failures into "
235 "specified (JSON) file (should be abs path as we use"
236 " append mode to insert new JSON objects)"),
237 cl::value_desc("filename"), cl::init(""));
238
239 static cl::opt<bool>
240 PrintBreakpoints("print-breakpoints-for-testing",
241 cl::desc("Print select breakpoints location for testing"));
242
243 static cl::opt<std::string> ClDataLayout("data-layout",
244 cl::desc("data layout string to use"),
245 cl::value_desc("layout-string"),
246 cl::init(""));
247
248 static cl::opt<bool> PreserveBitcodeUseListOrder(
249 "preserve-bc-uselistorder",
250 cl::desc("Preserve use-list order when writing LLVM bitcode."),
251 cl::init(true), cl::Hidden);
252
253 static cl::opt<bool> PreserveAssemblyUseListOrder(
254 "preserve-ll-uselistorder",
255 cl::desc("Preserve use-list order when writing LLVM assembly."),
256 cl::init(false), cl::Hidden);
257
258 static cl::opt<bool> RunTwice("run-twice",
259 cl::desc("Run all passes twice, re-using the "
260 "same pass manager (legacy PM only)."),
261 cl::init(false), cl::Hidden);
262
263 static cl::opt<bool> DiscardValueNames(
264 "discard-value-names",
265 cl::desc("Discard names from Value (other than GlobalValue)."),
266 cl::init(false), cl::Hidden);
267
268 static cl::opt<bool> Coroutines(
269 "enable-coroutines",
270 cl::desc("Enable coroutine passes."),
271 cl::init(false), cl::Hidden);
272
273 static cl::opt<bool> TimeTrace(
274 "time-trace",
275 cl::desc("Record time trace"));
276
277 static cl::opt<unsigned> TimeTraceGranularity(
278 "time-trace-granularity",
279 cl::desc("Minimum time granularity (in microseconds) traced by time profiler"),
280 cl::init(500), cl::Hidden);
281
282 static cl::opt<std::string>
283 TimeTraceFile("time-trace-file",
284 cl::desc("Specify time trace file destination"),
285 cl::value_desc("filename"));
286
287 static cl::opt<bool> RemarksWithHotness(
288 "pass-remarks-with-hotness",
289 cl::desc("With PGO, include profile count in optimization remarks"),
290 cl::Hidden);
291
292 static cl::opt<Optional<uint64_t>, false, remarks::HotnessThresholdParser>
293 RemarksHotnessThreshold(
294 "pass-remarks-hotness-threshold",
295 cl::desc("Minimum profile count required for "
296 "an optimization remark to be output. "
297 "Use 'auto' to apply the threshold from profile summary."),
298 cl::value_desc("N or 'auto'"), cl::init(0), cl::Hidden);
299
300 static cl::opt<std::string>
301 RemarksFilename("pass-remarks-output",
302 cl::desc("Output filename for pass remarks"),
303 cl::value_desc("filename"));
304
305 static cl::opt<std::string>
306 RemarksPasses("pass-remarks-filter",
307 cl::desc("Only record optimization remarks from passes whose "
308 "names match the given regular expression"),
309 cl::value_desc("regex"));
310
311 static cl::opt<std::string> RemarksFormat(
312 "pass-remarks-format",
313 cl::desc("The format used for serializing remarks (default: YAML)"),
314 cl::value_desc("format"), cl::init("yaml"));
315
316 namespace llvm {
317 cl::opt<PGOKind>
318 PGOKindFlag("pgo-kind", cl::init(NoPGO), cl::Hidden,
319 cl::desc("The kind of profile guided optimization"),
320 cl::values(clEnumValN(NoPGO, "nopgo", "Do not use PGO."),
321 clEnumValN(InstrGen, "pgo-instr-gen-pipeline",
322 "Instrument the IR to generate profile."),
323 clEnumValN(InstrUse, "pgo-instr-use-pipeline",
324 "Use instrumented profile to guide PGO."),
325 clEnumValN(SampleUse, "pgo-sample-use-pipeline",
326 "Use sampled profile to guide PGO.")));
327 cl::opt<std::string> ProfileFile("profile-file",
328 cl::desc("Path to the profile."), cl::Hidden);
329
330 cl::opt<CSPGOKind> CSPGOKindFlag(
331 "cspgo-kind", cl::init(NoCSPGO), cl::Hidden,
332 cl::desc("The kind of context sensitive profile guided optimization"),
333 cl::values(
334 clEnumValN(NoCSPGO, "nocspgo", "Do not use CSPGO."),
335 clEnumValN(
336 CSInstrGen, "cspgo-instr-gen-pipeline",
337 "Instrument (context sensitive) the IR to generate profile."),
338 clEnumValN(
339 CSInstrUse, "cspgo-instr-use-pipeline",
340 "Use instrumented (context sensitive) profile to guide PGO.")));
341 cl::opt<std::string> CSProfileGenFile(
342 "cs-profilegen-file",
343 cl::desc("Path to the instrumented context sensitive profile."),
344 cl::Hidden);
345 } // namespace llvm
346
addPass(legacy::PassManagerBase & PM,Pass * P)347 static inline void addPass(legacy::PassManagerBase &PM, Pass *P) {
348 // Add the pass to the pass manager...
349 PM.add(P);
350
351 // If we are verifying all of the intermediate steps, add the verifier...
352 if (VerifyEach)
353 PM.add(createVerifierPass());
354 }
355
356 /// This routine adds optimization passes based on selected optimization level,
357 /// OptLevel.
358 ///
359 /// OptLevel - Optimization Level
AddOptimizationPasses(legacy::PassManagerBase & MPM,legacy::FunctionPassManager & FPM,TargetMachine * TM,unsigned OptLevel,unsigned SizeLevel)360 static void AddOptimizationPasses(legacy::PassManagerBase &MPM,
361 legacy::FunctionPassManager &FPM,
362 TargetMachine *TM, unsigned OptLevel,
363 unsigned SizeLevel) {
364 if (!NoVerify || VerifyEach)
365 FPM.add(createVerifierPass()); // Verify that input is correct
366
367 PassManagerBuilder Builder;
368 Builder.OptLevel = OptLevel;
369 Builder.SizeLevel = SizeLevel;
370
371 if (DisableInline) {
372 // No inlining pass
373 } else if (OptLevel > 1) {
374 Builder.Inliner = createFunctionInliningPass(OptLevel, SizeLevel, false);
375 } else {
376 Builder.Inliner = createAlwaysInlinerLegacyPass();
377 }
378 Builder.DisableUnrollLoops = (DisableLoopUnrolling.getNumOccurrences() > 0) ?
379 DisableLoopUnrolling : OptLevel == 0;
380
381 Builder.LoopVectorize = OptLevel > 1 && SizeLevel < 2;
382
383 Builder.SLPVectorize = OptLevel > 1 && SizeLevel < 2;
384
385 if (TM)
386 TM->adjustPassManager(Builder);
387
388 if (Coroutines)
389 addCoroutinePassesToExtensionPoints(Builder);
390
391 switch (PGOKindFlag) {
392 case InstrGen:
393 Builder.EnablePGOInstrGen = true;
394 Builder.PGOInstrGen = ProfileFile;
395 break;
396 case InstrUse:
397 Builder.PGOInstrUse = ProfileFile;
398 break;
399 case SampleUse:
400 Builder.PGOSampleUse = ProfileFile;
401 break;
402 default:
403 break;
404 }
405
406 switch (CSPGOKindFlag) {
407 case CSInstrGen:
408 Builder.EnablePGOCSInstrGen = true;
409 break;
410 case CSInstrUse:
411 Builder.EnablePGOCSInstrUse = true;
412 break;
413 default:
414 break;
415 }
416
417 Builder.populateFunctionPassManager(FPM);
418 Builder.populateModulePassManager(MPM);
419 }
420
AddStandardLinkPasses(legacy::PassManagerBase & PM)421 static void AddStandardLinkPasses(legacy::PassManagerBase &PM) {
422 PassManagerBuilder Builder;
423 Builder.VerifyInput = true;
424 if (DisableOptimizations)
425 Builder.OptLevel = 0;
426
427 if (!DisableInline)
428 Builder.Inliner = createFunctionInliningPass();
429 Builder.populateLTOPassManager(PM);
430 }
431
432 //===----------------------------------------------------------------------===//
433 // CodeGen-related helper functions.
434 //
435
GetCodeGenOptLevel()436 static CodeGenOpt::Level GetCodeGenOptLevel() {
437 if (CodeGenOptLevel.getNumOccurrences())
438 return static_cast<CodeGenOpt::Level>(unsigned(CodeGenOptLevel));
439 if (OptLevelO1)
440 return CodeGenOpt::Less;
441 if (OptLevelO2)
442 return CodeGenOpt::Default;
443 if (OptLevelO3)
444 return CodeGenOpt::Aggressive;
445 return CodeGenOpt::None;
446 }
447
448 // Returns the TargetMachine instance or zero if no triple is provided.
GetTargetMachine(Triple TheTriple,StringRef CPUStr,StringRef FeaturesStr,const TargetOptions & Options)449 static TargetMachine* GetTargetMachine(Triple TheTriple, StringRef CPUStr,
450 StringRef FeaturesStr,
451 const TargetOptions &Options) {
452 std::string Error;
453 const Target *TheTarget =
454 TargetRegistry::lookupTarget(codegen::getMArch(), TheTriple, Error);
455 // Some modules don't specify a triple, and this is okay.
456 if (!TheTarget) {
457 return nullptr;
458 }
459
460 return TheTarget->createTargetMachine(
461 TheTriple.getTriple(), codegen::getCPUStr(), codegen::getFeaturesStr(),
462 Options, codegen::getExplicitRelocModel(),
463 codegen::getExplicitCodeModel(), GetCodeGenOptLevel());
464 }
465
466 #ifdef BUILD_EXAMPLES
467 void initializeExampleIRTransforms(llvm::PassRegistry &Registry);
468 #endif
469
470 struct TimeTracerRAII {
TimeTracerRAIITimeTracerRAII471 TimeTracerRAII(StringRef ProgramName) {
472 if (TimeTrace)
473 timeTraceProfilerInitialize(TimeTraceGranularity, ProgramName);
474 }
~TimeTracerRAIITimeTracerRAII475 ~TimeTracerRAII() {
476 if (TimeTrace) {
477 if (auto E = timeTraceProfilerWrite(TimeTraceFile, OutputFilename)) {
478 handleAllErrors(std::move(E), [&](const StringError &SE) {
479 errs() << SE.getMessage() << "\n";
480 });
481 return;
482 }
483 timeTraceProfilerCleanup();
484 }
485 }
486 };
487
488 // For use in NPM transition. Currently this contains most codegen-specific
489 // passes. Remove passes from here when porting to the NPM.
490 // TODO: use a codegen version of PassRegistry.def/PassBuilder::is*Pass() once
491 // it exists.
shouldPinPassToLegacyPM(StringRef Pass)492 static bool shouldPinPassToLegacyPM(StringRef Pass) {
493 std::vector<StringRef> PassNameExactToIgnore = {
494 "nvvm-reflect",
495 "nvvm-intr-range",
496 "amdgpu-simplifylib",
497 "amdgpu-usenative",
498 "amdgpu-promote-alloca",
499 "amdgpu-promote-alloca-to-vector",
500 "amdgpu-lower-kernel-attributes",
501 "amdgpu-propagate-attributes-early",
502 "amdgpu-propagate-attributes-late",
503 "amdgpu-unify-metadata",
504 "amdgpu-printf-runtime-binding",
505 "amdgpu-always-inline"};
506 if (llvm::is_contained(PassNameExactToIgnore, Pass))
507 return false;
508
509 std::vector<StringRef> PassNamePrefix = {
510 "x86-", "xcore-", "wasm-", "systemz-", "ppc-", "nvvm-", "nvptx-",
511 "mips-", "lanai-", "hexagon-", "bpf-", "avr-", "thumb2-", "arm-",
512 "si-", "gcn-", "amdgpu-", "aarch64-", "amdgcn-", "polly-"};
513 std::vector<StringRef> PassNameContain = {"ehprepare"};
514 std::vector<StringRef> PassNameExact = {
515 "safe-stack", "cost-model",
516 "codegenprepare", "interleaved-load-combine",
517 "unreachableblockelim", "verify-safepoint-ir",
518 "atomic-expand", "expandvp",
519 "hardware-loops", "type-promotion",
520 "mve-tail-predication", "interleaved-access",
521 "global-merge", "pre-isel-intrinsic-lowering",
522 "expand-reductions", "indirectbr-expand",
523 "generic-to-nvvm", "expandmemcmp",
524 "loop-reduce", "lower-amx-type",
525 "pre-amx-config", "lower-amx-intrinsics",
526 "polyhedral-info", "replace-with-veclib"};
527 for (const auto &P : PassNamePrefix)
528 if (Pass.startswith(P))
529 return true;
530 for (const auto &P : PassNameContain)
531 if (Pass.contains(P))
532 return true;
533 return llvm::is_contained(PassNameExact, Pass);
534 }
535
536 // For use in NPM transition.
shouldForceLegacyPM()537 static bool shouldForceLegacyPM() {
538 for (const auto &P : PassList) {
539 StringRef Arg = P->getPassArgument();
540 if (shouldPinPassToLegacyPM(Arg))
541 return true;
542 }
543 return false;
544 }
545
546 //===----------------------------------------------------------------------===//
547 // main for opt
548 //
main(int argc,char ** argv)549 int main(int argc, char **argv) {
550 InitLLVM X(argc, argv);
551
552 // Enable debug stream buffering.
553 EnableDebugBuffering = true;
554
555 InitializeAllTargets();
556 InitializeAllTargetMCs();
557 InitializeAllAsmPrinters();
558 InitializeAllAsmParsers();
559
560 // Initialize passes
561 PassRegistry &Registry = *PassRegistry::getPassRegistry();
562 initializeCore(Registry);
563 initializeCoroutines(Registry);
564 initializeScalarOpts(Registry);
565 initializeObjCARCOpts(Registry);
566 initializeVectorization(Registry);
567 initializeIPO(Registry);
568 initializeAnalysis(Registry);
569 initializeTransformUtils(Registry);
570 initializeInstCombine(Registry);
571 initializeAggressiveInstCombine(Registry);
572 initializeInstrumentation(Registry);
573 initializeTarget(Registry);
574 // For codegen passes, only passes that do IR to IR transformation are
575 // supported.
576 initializeExpandMemCmpPassPass(Registry);
577 initializeScalarizeMaskedMemIntrinLegacyPassPass(Registry);
578 initializeCodeGenPreparePass(Registry);
579 initializeAtomicExpandPass(Registry);
580 initializeRewriteSymbolsLegacyPassPass(Registry);
581 initializeWinEHPreparePass(Registry);
582 initializeDwarfEHPrepareLegacyPassPass(Registry);
583 initializeSafeStackLegacyPassPass(Registry);
584 initializeSjLjEHPreparePass(Registry);
585 initializePreISelIntrinsicLoweringLegacyPassPass(Registry);
586 initializeGlobalMergePass(Registry);
587 initializeIndirectBrExpandPassPass(Registry);
588 initializeInterleavedLoadCombinePass(Registry);
589 initializeInterleavedAccessPass(Registry);
590 initializeEntryExitInstrumenterPass(Registry);
591 initializePostInlineEntryExitInstrumenterPass(Registry);
592 initializeUnreachableBlockElimLegacyPassPass(Registry);
593 initializeExpandReductionsPass(Registry);
594 initializeExpandVectorPredicationPass(Registry);
595 initializeWasmEHPreparePass(Registry);
596 initializeWriteBitcodePassPass(Registry);
597 initializeHardwareLoopsPass(Registry);
598 initializeTypePromotionPass(Registry);
599 initializeReplaceWithVeclibLegacyPass(Registry);
600
601 #ifdef BUILD_EXAMPLES
602 initializeExampleIRTransforms(Registry);
603 #endif
604
605 cl::ParseCommandLineOptions(argc, argv,
606 "llvm .bc -> .bc modular optimizer and analysis printer\n");
607
608 LLVMContext Context;
609
610 if (AnalyzeOnly && NoOutput) {
611 errs() << argv[0] << ": analyze mode conflicts with no-output mode.\n";
612 return 1;
613 }
614
615 // FIXME: once the legacy PM code is deleted, move runPassPipeline() here and
616 // construct the PassBuilder before parsing IR so we can reuse the same
617 // PassBuilder for print passes.
618 if (PrintPasses) {
619 printPasses(outs());
620 return 0;
621 }
622
623 TimeTracerRAII TimeTracer(argv[0]);
624
625 SMDiagnostic Err;
626
627 Context.setDiscardValueNames(DiscardValueNames);
628 if (!DisableDITypeMap)
629 Context.enableDebugTypeODRUniquing();
630
631 Expected<std::unique_ptr<ToolOutputFile>> RemarksFileOrErr =
632 setupLLVMOptimizationRemarks(Context, RemarksFilename, RemarksPasses,
633 RemarksFormat, RemarksWithHotness,
634 RemarksHotnessThreshold);
635 if (Error E = RemarksFileOrErr.takeError()) {
636 errs() << toString(std::move(E)) << '\n';
637 return 1;
638 }
639 std::unique_ptr<ToolOutputFile> RemarksFile = std::move(*RemarksFileOrErr);
640
641 // Load the input module...
642 auto SetDataLayout = [](StringRef) -> Optional<std::string> {
643 if (ClDataLayout.empty())
644 return None;
645 return ClDataLayout;
646 };
647 std::unique_ptr<Module> M;
648 if (NoUpgradeDebugInfo)
649 M = parseAssemblyFileWithIndexNoUpgradeDebugInfo(
650 InputFilename, Err, Context, nullptr, SetDataLayout)
651 .Mod;
652 else
653 M = parseIRFile(InputFilename, Err, Context, SetDataLayout);
654
655 if (!M) {
656 Err.print(argv[0], errs());
657 return 1;
658 }
659
660 // Strip debug info before running the verifier.
661 if (StripDebug)
662 StripDebugInfo(*M);
663
664 // Erase module-level named metadata, if requested.
665 if (StripNamedMetadata) {
666 while (!M->named_metadata_empty()) {
667 NamedMDNode *NMD = &*M->named_metadata_begin();
668 M->eraseNamedMetadata(NMD);
669 }
670 }
671
672 // If we are supposed to override the target triple or data layout, do so now.
673 if (!TargetTriple.empty())
674 M->setTargetTriple(Triple::normalize(TargetTriple));
675
676 // Immediately run the verifier to catch any problems before starting up the
677 // pass pipelines. Otherwise we can crash on broken code during
678 // doInitialization().
679 if (!NoVerify && verifyModule(*M, &errs())) {
680 errs() << argv[0] << ": " << InputFilename
681 << ": error: input module is broken!\n";
682 return 1;
683 }
684
685 // Enable testing of whole program devirtualization on this module by invoking
686 // the facility for updating public visibility to linkage unit visibility when
687 // specified by an internal option. This is normally done during LTO which is
688 // not performed via opt.
689 updateVCallVisibilityInModule(*M,
690 /* WholeProgramVisibilityEnabledInLTO */ false,
691 /* DynamicExportSymbols */ {});
692
693 // Figure out what stream we are supposed to write to...
694 std::unique_ptr<ToolOutputFile> Out;
695 std::unique_ptr<ToolOutputFile> ThinLinkOut;
696 if (NoOutput) {
697 if (!OutputFilename.empty())
698 errs() << "WARNING: The -o (output filename) option is ignored when\n"
699 "the --disable-output option is used.\n";
700 } else {
701 // Default to standard output.
702 if (OutputFilename.empty())
703 OutputFilename = "-";
704
705 std::error_code EC;
706 sys::fs::OpenFlags Flags =
707 OutputAssembly ? sys::fs::OF_TextWithCRLF : sys::fs::OF_None;
708 Out.reset(new ToolOutputFile(OutputFilename, EC, Flags));
709 if (EC) {
710 errs() << EC.message() << '\n';
711 return 1;
712 }
713
714 if (!ThinLinkBitcodeFile.empty()) {
715 ThinLinkOut.reset(
716 new ToolOutputFile(ThinLinkBitcodeFile, EC, sys::fs::OF_None));
717 if (EC) {
718 errs() << EC.message() << '\n';
719 return 1;
720 }
721 }
722 }
723
724 Triple ModuleTriple(M->getTargetTriple());
725 std::string CPUStr, FeaturesStr;
726 TargetMachine *Machine = nullptr;
727 const TargetOptions Options =
728 codegen::InitTargetOptionsFromCodeGenFlags(ModuleTriple);
729
730 if (ModuleTriple.getArch()) {
731 CPUStr = codegen::getCPUStr();
732 FeaturesStr = codegen::getFeaturesStr();
733 Machine = GetTargetMachine(ModuleTriple, CPUStr, FeaturesStr, Options);
734 } else if (ModuleTriple.getArchName() != "unknown" &&
735 ModuleTriple.getArchName() != "") {
736 errs() << argv[0] << ": unrecognized architecture '"
737 << ModuleTriple.getArchName() << "' provided.\n";
738 return 1;
739 }
740
741 std::unique_ptr<TargetMachine> TM(Machine);
742
743 // Override function attributes based on CPUStr, FeaturesStr, and command line
744 // flags.
745 codegen::setFunctionAttributes(CPUStr, FeaturesStr, *M);
746
747 // If the output is set to be emitted to standard out, and standard out is a
748 // console, print out a warning message and refuse to do it. We don't
749 // impress anyone by spewing tons of binary goo to a terminal.
750 if (!Force && !NoOutput && !AnalyzeOnly && !OutputAssembly)
751 if (CheckBitcodeOutputToConsole(Out->os()))
752 NoOutput = true;
753
754 if (OutputThinLTOBC)
755 M->addModuleFlag(Module::Error, "EnableSplitLTOUnit", SplitLTOUnit);
756
757 // Add an appropriate TargetLibraryInfo pass for the module's triple.
758 TargetLibraryInfoImpl TLII(ModuleTriple);
759
760 // The -disable-simplify-libcalls flag actually disables all builtin optzns.
761 if (DisableSimplifyLibCalls)
762 TLII.disableAllFunctions();
763 else {
764 // Disable individual builtin functions in TargetLibraryInfo.
765 LibFunc F;
766 for (auto &FuncName : DisableBuiltins)
767 if (TLII.getLibFunc(FuncName, F))
768 TLII.setUnavailable(F);
769 else {
770 errs() << argv[0] << ": cannot disable nonexistent builtin function "
771 << FuncName << '\n';
772 return 1;
773 }
774 }
775
776 // If `-passes=` is specified, use NPM.
777 // If `-enable-new-pm` is specified and there are no codegen passes, use NPM.
778 // e.g. `-enable-new-pm -sroa` will use NPM.
779 // but `-enable-new-pm -codegenprepare` will still revert to legacy PM.
780 if ((EnableNewPassManager && !shouldForceLegacyPM()) ||
781 PassPipeline.getNumOccurrences() > 0) {
782 if (AnalyzeOnly) {
783 errs() << "Cannot specify -analyze under new pass manager, either "
784 "specify '-enable-new-pm=0', or use the corresponding new pass "
785 "manager pass, e.g. '-passes=print<scalar-evolution>'. For a "
786 "full list of passes, see the '--print-passes' flag.\n";
787 return 1;
788 }
789 if (legacy::debugPassSpecified()) {
790 errs()
791 << "-debug-pass does not work with the new PM, either use "
792 "-debug-pass-manager, or use the legacy PM (-enable-new-pm=0)\n";
793 return 1;
794 }
795 if (PassPipeline.getNumOccurrences() > 0 && PassList.size() > 0) {
796 errs()
797 << "Cannot specify passes via both -foo-pass and --passes=foo-pass\n";
798 return 1;
799 }
800 SmallVector<StringRef, 4> Passes;
801 if (OptLevelO0)
802 Passes.push_back("default<O0>");
803 if (OptLevelO1)
804 Passes.push_back("default<O1>");
805 if (OptLevelO2)
806 Passes.push_back("default<O2>");
807 if (OptLevelO3)
808 Passes.push_back("default<O3>");
809 if (OptLevelOs)
810 Passes.push_back("default<Os>");
811 if (OptLevelOz)
812 Passes.push_back("default<Oz>");
813 for (const auto &P : PassList)
814 Passes.push_back(P->getPassArgument());
815 OutputKind OK = OK_NoOutput;
816 if (!NoOutput)
817 OK = OutputAssembly
818 ? OK_OutputAssembly
819 : (OutputThinLTOBC ? OK_OutputThinLTOBitcode : OK_OutputBitcode);
820
821 VerifierKind VK = VK_VerifyInAndOut;
822 if (NoVerify)
823 VK = VK_NoVerifier;
824 else if (VerifyEach)
825 VK = VK_VerifyEachPass;
826
827 // The user has asked to use the new pass manager and provided a pipeline
828 // string. Hand off the rest of the functionality to the new code for that
829 // layer.
830 return runPassPipeline(argv[0], *M, TM.get(), &TLII, Out.get(),
831 ThinLinkOut.get(), RemarksFile.get(), PassPipeline,
832 Passes, OK, VK, PreserveAssemblyUseListOrder,
833 PreserveBitcodeUseListOrder, EmitSummaryIndex,
834 EmitModuleHash, EnableDebugify)
835 ? 0
836 : 1;
837 }
838
839 // Create a PassManager to hold and optimize the collection of passes we are
840 // about to build. If the -debugify-each option is set, wrap each pass with
841 // the (-check)-debugify passes.
842 DebugifyCustomPassManager Passes;
843 DebugifyStatsMap DIStatsMap;
844 DebugInfoPerPassMap DIPreservationMap;
845 if (DebugifyEach) {
846 Passes.setDebugifyMode(DebugifyMode::SyntheticDebugInfo);
847 Passes.setDIStatsMap(DIStatsMap);
848 } else if (VerifyEachDebugInfoPreserve) {
849 Passes.setDebugifyMode(DebugifyMode::OriginalDebugInfo);
850 Passes.setDIPreservationMap(DIPreservationMap);
851 if (!VerifyDIPreserveExport.empty())
852 Passes.setOrigDIVerifyBugsReportFilePath(VerifyDIPreserveExport);
853 }
854
855 bool AddOneTimeDebugifyPasses =
856 (EnableDebugify && !DebugifyEach) ||
857 (VerifyDebugInfoPreserve && !VerifyEachDebugInfoPreserve);
858
859 Passes.add(new TargetLibraryInfoWrapperPass(TLII));
860
861 // Add internal analysis passes from the target machine.
862 Passes.add(createTargetTransformInfoWrapperPass(TM ? TM->getTargetIRAnalysis()
863 : TargetIRAnalysis()));
864
865 if (AddOneTimeDebugifyPasses) {
866 if (EnableDebugify) {
867 Passes.setDIStatsMap(DIStatsMap);
868 Passes.add(createDebugifyModulePass());
869 } else if (VerifyDebugInfoPreserve) {
870 Passes.setDIPreservationMap(DIPreservationMap);
871 Passes.add(createDebugifyModulePass(
872 DebugifyMode::OriginalDebugInfo, "",
873 &(Passes.getDebugInfoPerPassMap())));
874 }
875 }
876
877 std::unique_ptr<legacy::FunctionPassManager> FPasses;
878 if (OptLevelO0 || OptLevelO1 || OptLevelO2 || OptLevelOs || OptLevelOz ||
879 OptLevelO3) {
880 FPasses.reset(new legacy::FunctionPassManager(M.get()));
881 FPasses->add(createTargetTransformInfoWrapperPass(
882 TM ? TM->getTargetIRAnalysis() : TargetIRAnalysis()));
883 }
884
885 if (PrintBreakpoints) {
886 // Default to standard output.
887 if (!Out) {
888 if (OutputFilename.empty())
889 OutputFilename = "-";
890
891 std::error_code EC;
892 Out = std::make_unique<ToolOutputFile>(OutputFilename, EC,
893 sys::fs::OF_None);
894 if (EC) {
895 errs() << EC.message() << '\n';
896 return 1;
897 }
898 }
899 Passes.add(createBreakpointPrinter(Out->os()));
900 NoOutput = true;
901 }
902
903 if (TM) {
904 // FIXME: We should dyn_cast this when supported.
905 auto <M = static_cast<LLVMTargetMachine &>(*TM);
906 Pass *TPC = LTM.createPassConfig(Passes);
907 Passes.add(TPC);
908 }
909
910 // Create a new optimization pass for each one specified on the command line
911 for (unsigned i = 0; i < PassList.size(); ++i) {
912 if (StandardLinkOpts &&
913 StandardLinkOpts.getPosition() < PassList.getPosition(i)) {
914 AddStandardLinkPasses(Passes);
915 StandardLinkOpts = false;
916 }
917
918 if (OptLevelO0 && OptLevelO0.getPosition() < PassList.getPosition(i)) {
919 AddOptimizationPasses(Passes, *FPasses, TM.get(), 0, 0);
920 OptLevelO0 = false;
921 }
922
923 if (OptLevelO1 && OptLevelO1.getPosition() < PassList.getPosition(i)) {
924 AddOptimizationPasses(Passes, *FPasses, TM.get(), 1, 0);
925 OptLevelO1 = false;
926 }
927
928 if (OptLevelO2 && OptLevelO2.getPosition() < PassList.getPosition(i)) {
929 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 0);
930 OptLevelO2 = false;
931 }
932
933 if (OptLevelOs && OptLevelOs.getPosition() < PassList.getPosition(i)) {
934 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 1);
935 OptLevelOs = false;
936 }
937
938 if (OptLevelOz && OptLevelOz.getPosition() < PassList.getPosition(i)) {
939 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 2);
940 OptLevelOz = false;
941 }
942
943 if (OptLevelO3 && OptLevelO3.getPosition() < PassList.getPosition(i)) {
944 AddOptimizationPasses(Passes, *FPasses, TM.get(), 3, 0);
945 OptLevelO3 = false;
946 }
947
948 const PassInfo *PassInf = PassList[i];
949 Pass *P = nullptr;
950 if (PassInf->getNormalCtor())
951 P = PassInf->getNormalCtor()();
952 else
953 errs() << argv[0] << ": cannot create pass: "
954 << PassInf->getPassName() << "\n";
955 if (P) {
956 PassKind Kind = P->getPassKind();
957 addPass(Passes, P);
958
959 if (AnalyzeOnly) {
960 switch (Kind) {
961 case PT_Region:
962 Passes.add(createRegionPassPrinter(PassInf, Out->os()));
963 break;
964 case PT_Loop:
965 Passes.add(createLoopPassPrinter(PassInf, Out->os()));
966 break;
967 case PT_Function:
968 Passes.add(createFunctionPassPrinter(PassInf, Out->os()));
969 break;
970 case PT_CallGraphSCC:
971 Passes.add(createCallGraphPassPrinter(PassInf, Out->os()));
972 break;
973 default:
974 Passes.add(createModulePassPrinter(PassInf, Out->os()));
975 break;
976 }
977 }
978 }
979
980 if (PrintEachXForm)
981 Passes.add(
982 createPrintModulePass(errs(), "", PreserveAssemblyUseListOrder));
983 }
984
985 if (StandardLinkOpts) {
986 AddStandardLinkPasses(Passes);
987 StandardLinkOpts = false;
988 }
989
990 if (OptLevelO0)
991 AddOptimizationPasses(Passes, *FPasses, TM.get(), 0, 0);
992
993 if (OptLevelO1)
994 AddOptimizationPasses(Passes, *FPasses, TM.get(), 1, 0);
995
996 if (OptLevelO2)
997 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 0);
998
999 if (OptLevelOs)
1000 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 1);
1001
1002 if (OptLevelOz)
1003 AddOptimizationPasses(Passes, *FPasses, TM.get(), 2, 2);
1004
1005 if (OptLevelO3)
1006 AddOptimizationPasses(Passes, *FPasses, TM.get(), 3, 0);
1007
1008 if (FPasses) {
1009 FPasses->doInitialization();
1010 for (Function &F : *M)
1011 FPasses->run(F);
1012 FPasses->doFinalization();
1013 }
1014
1015 // Check that the module is well formed on completion of optimization
1016 if (!NoVerify && !VerifyEach)
1017 Passes.add(createVerifierPass());
1018
1019 if (AddOneTimeDebugifyPasses) {
1020 if (EnableDebugify)
1021 Passes.add(createCheckDebugifyModulePass(false));
1022 else if (VerifyDebugInfoPreserve) {
1023 if (!VerifyDIPreserveExport.empty())
1024 Passes.setOrigDIVerifyBugsReportFilePath(VerifyDIPreserveExport);
1025 Passes.add(createCheckDebugifyModulePass(
1026 false, "", nullptr, DebugifyMode::OriginalDebugInfo,
1027 &(Passes.getDebugInfoPerPassMap()), VerifyDIPreserveExport));
1028 }
1029 }
1030
1031 // In run twice mode, we want to make sure the output is bit-by-bit
1032 // equivalent if we run the pass manager again, so setup two buffers and
1033 // a stream to write to them. Note that llc does something similar and it
1034 // may be worth to abstract this out in the future.
1035 SmallVector<char, 0> Buffer;
1036 SmallVector<char, 0> FirstRunBuffer;
1037 std::unique_ptr<raw_svector_ostream> BOS;
1038 raw_ostream *OS = nullptr;
1039
1040 const bool ShouldEmitOutput = !NoOutput && !AnalyzeOnly;
1041
1042 // Write bitcode or assembly to the output as the last step...
1043 if (ShouldEmitOutput || RunTwice) {
1044 assert(Out);
1045 OS = &Out->os();
1046 if (RunTwice) {
1047 BOS = std::make_unique<raw_svector_ostream>(Buffer);
1048 OS = BOS.get();
1049 }
1050 if (OutputAssembly) {
1051 if (EmitSummaryIndex)
1052 report_fatal_error("Text output is incompatible with -module-summary");
1053 if (EmitModuleHash)
1054 report_fatal_error("Text output is incompatible with -module-hash");
1055 Passes.add(createPrintModulePass(*OS, "", PreserveAssemblyUseListOrder));
1056 } else if (OutputThinLTOBC)
1057 Passes.add(createWriteThinLTOBitcodePass(
1058 *OS, ThinLinkOut ? &ThinLinkOut->os() : nullptr));
1059 else
1060 Passes.add(createBitcodeWriterPass(*OS, PreserveBitcodeUseListOrder,
1061 EmitSummaryIndex, EmitModuleHash));
1062 }
1063
1064 // Before executing passes, print the final values of the LLVM options.
1065 cl::PrintOptionValues();
1066
1067 if (!RunTwice) {
1068 // Now that we have all of the passes ready, run them.
1069 Passes.run(*M);
1070 } else {
1071 // If requested, run all passes twice with the same pass manager to catch
1072 // bugs caused by persistent state in the passes.
1073 std::unique_ptr<Module> M2(CloneModule(*M));
1074 // Run all passes on the original module first, so the second run processes
1075 // the clone to catch CloneModule bugs.
1076 Passes.run(*M);
1077 FirstRunBuffer = Buffer;
1078 Buffer.clear();
1079
1080 Passes.run(*M2);
1081
1082 // Compare the two outputs and make sure they're the same
1083 assert(Out);
1084 if (Buffer.size() != FirstRunBuffer.size() ||
1085 (memcmp(Buffer.data(), FirstRunBuffer.data(), Buffer.size()) != 0)) {
1086 errs()
1087 << "Running the pass manager twice changed the output.\n"
1088 "Writing the result of the second run to the specified output.\n"
1089 "To generate the one-run comparison binary, just run without\n"
1090 "the compile-twice option\n";
1091 if (ShouldEmitOutput) {
1092 Out->os() << BOS->str();
1093 Out->keep();
1094 }
1095 if (RemarksFile)
1096 RemarksFile->keep();
1097 return 1;
1098 }
1099 if (ShouldEmitOutput)
1100 Out->os() << BOS->str();
1101 }
1102
1103 if (DebugifyEach && !DebugifyExport.empty())
1104 exportDebugifyStats(DebugifyExport, Passes.getDebugifyStatsMap());
1105
1106 // Declare success.
1107 if (!NoOutput || PrintBreakpoints)
1108 Out->keep();
1109
1110 if (RemarksFile)
1111 RemarksFile->keep();
1112
1113 if (ThinLinkOut)
1114 ThinLinkOut->keep();
1115
1116 return 0;
1117 }
1118