1 //===-- Passes.cpp - Target independent code generation passes ------------===//
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 // This file defines interfaces to access the target independent code
11 // generation passes provided by the LLVM backend.
13 //===---------------------------------------------------------------------===//
15 #include "llvm/Analysis/Passes.h"
16 #include "llvm/Analysis/Verifier.h"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/CodeGen/GCStrategy.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/Passes.h"
22 #include "llvm/CodeGen/RegAllocRegistry.h"
23 #include "llvm/Target/TargetLowering.h"
24 #include "llvm/Target/TargetOptions.h"
25 #include "llvm/Assembly/PrintModulePass.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
32 static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden,
33 cl::desc("Disable Post Regalloc"));
34 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
35 cl::desc("Disable branch folding"));
36 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
37 cl::desc("Disable tail duplication"));
38 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
39 cl::desc("Disable pre-register allocation tail duplication"));
40 static cl::opt<bool> EnableBlockPlacement("enable-block-placement",
41 cl::Hidden, cl::desc("Enable probability-driven block placement"));
42 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats",
43 cl::Hidden, cl::desc("Collect probability-driven block placement stats"));
44 static cl::opt<bool> DisableCodePlace("disable-code-place", cl::Hidden,
45 cl::desc("Disable code placement"));
46 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
47 cl::desc("Disable Stack Slot Coloring"));
48 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden,
49 cl::desc("Disable Machine Dead Code Elimination"));
50 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
51 cl::desc("Disable Machine LICM"));
52 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden,
53 cl::desc("Disable Machine Common Subexpression Elimination"));
54 static cl::opt<cl::boolOrDefault>
55 OptimizeRegAlloc("optimize-regalloc", cl::Hidden,
56 cl::desc("Enable optimized register allocation compilation path."));
57 static cl::opt<cl::boolOrDefault>
58 EnableMachineSched("enable-misched", cl::Hidden,
59 cl::desc("Enable the machine instruction scheduling pass."));
60 static cl::opt<bool> EnableStrongPHIElim("strong-phi-elim", cl::Hidden,
61 cl::desc("Use strong PHI elimination."));
62 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
64 cl::desc("Disable Machine LICM"));
65 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
66 cl::desc("Disable Machine Sinking"));
67 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
68 cl::desc("Disable Loop Strength Reduction Pass"));
69 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
70 cl::desc("Disable Codegen Prepare"));
71 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden,
72 cl::desc("Disable Copy Propagation pass"));
73 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
74 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
75 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
76 cl::desc("Print LLVM IR input to isel pass"));
77 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
78 cl::desc("Dump garbage collector data"));
79 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
80 cl::desc("Verify generated machine code"),
81 cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
83 // Allow Pass selection to be overriden by command line options.
85 // DefaultID is the default pass to run which may be NoPassID, or may be
86 // overriden by the target.
88 // OptionalID is a pass that may be forcibly enabled by the user when the
89 // default is NoPassID.
90 char &enablePass(char &DefaultID, cl::boolOrDefault Override,
91 char *OptionalIDPtr = &NoPassID) {
96 if (&DefaultID != &NoPassID)
98 if (OptionalIDPtr == &NoPassID)
99 report_fatal_error("Target cannot enable pass");
100 return *OptionalIDPtr;
104 llvm_unreachable("Invalid command line option state");
107 //===---------------------------------------------------------------------===//
109 //===---------------------------------------------------------------------===//
111 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig",
112 "Target Pass Configuration", false, false)
113 char TargetPassConfig::ID = 0;
115 static char NoPassIDAnchor = 0;
116 char &llvm::NoPassID = NoPassIDAnchor;
118 // Out of line virtual method.
119 TargetPassConfig::~TargetPassConfig() {}
121 // Out of line constructor provides default values for pass options and
122 // registers all common codegen passes.
123 TargetPassConfig::TargetPassConfig(TargetMachine *tm, PassManagerBase &pm)
124 : ImmutablePass(ID), TM(tm), PM(pm), Initialized(false),
125 DisableVerify(false),
126 EnableTailMerge(true) {
128 // Register all target independent codegen passes to activate their PassIDs,
129 // including this pass itself.
130 initializeCodeGen(*PassRegistry::getPassRegistry());
133 /// createPassConfig - Create a pass configuration object to be used by
134 /// addPassToEmitX methods for generating a pipeline of CodeGen passes.
136 /// Targets may override this to extend TargetPassConfig.
137 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) {
138 return new TargetPassConfig(this, PM);
141 TargetPassConfig::TargetPassConfig()
142 : ImmutablePass(ID), PM(*(PassManagerBase*)0) {
143 llvm_unreachable("TargetPassConfig should not be constructed on-the-fly");
146 // Helper to verify the analysis is really immutable.
147 void TargetPassConfig::setOpt(bool &Opt, bool Val) {
148 assert(!Initialized && "PassConfig is immutable");
152 void TargetPassConfig::addPass(char &ID) {
153 if (&ID == &NoPassID)
156 // FIXME: check user overrides
157 Pass *P = Pass::createPass(ID);
159 llvm_unreachable("Pass ID not registered");
163 void TargetPassConfig::printNoVerify(const char *Banner) const {
164 if (TM->shouldPrintMachineCode())
165 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
168 void TargetPassConfig::printAndVerify(const char *Banner) const {
169 if (TM->shouldPrintMachineCode())
170 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
172 if (VerifyMachineCode)
173 PM.add(createMachineVerifierPass(Banner));
176 /// Add common target configurable passes that perform LLVM IR to IR transforms
177 /// following machine independent optimization.
178 void TargetPassConfig::addIRPasses() {
179 // Basic AliasAnalysis support.
180 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
181 // BasicAliasAnalysis wins if they disagree. This is intended to help
182 // support "obvious" type-punning idioms.
183 PM.add(createTypeBasedAliasAnalysisPass());
184 PM.add(createBasicAliasAnalysisPass());
186 // Before running any passes, run the verifier to determine if the input
187 // coming from the front-end and/or optimizer is valid.
189 PM.add(createVerifierPass());
191 // Run loop strength reduction before anything else.
192 if (getOptLevel() != CodeGenOpt::None && !DisableLSR) {
193 PM.add(createLoopStrengthReducePass(getTargetLowering()));
195 PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &dbgs()));
198 PM.add(createGCLoweringPass());
200 // Make sure that no unreachable blocks are instruction selected.
201 PM.add(createUnreachableBlockEliminationPass());
204 /// Add common passes that perform LLVM IR to IR transforms in preparation for
205 /// instruction selection.
206 void TargetPassConfig::addISelPrepare() {
207 if (getOptLevel() != CodeGenOpt::None && !DisableCGP)
208 PM.add(createCodeGenPreparePass(getTargetLowering()));
210 PM.add(createStackProtectorPass(getTargetLowering()));
215 PM.add(createPrintFunctionPass("\n\n"
216 "*** Final LLVM Code input to ISel ***\n",
219 // All passes which modify the LLVM IR are now complete; run the verifier
220 // to ensure that the IR is valid.
222 PM.add(createVerifierPass());
225 /// Add the complete set of target-independent postISel code generator passes.
227 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
228 /// with nontrivial configuration or multiple passes are broken out below in
229 /// add%Stage routines.
231 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
232 /// addPre/Post methods with empty header implementations allow injecting
233 /// target-specific fixups just before or after major stages. Additionally,
234 /// targets have the flexibility to change pass order within a stage by
235 /// overriding default implementation of add%Stage routines below. Each
236 /// technique has maintainability tradeoffs because alternate pass orders are
237 /// not well supported. addPre/Post works better if the target pass is easily
238 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
239 /// the target should override the stage instead.
241 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection
242 /// before/after any target-independent pass. But it's currently overkill.
243 void TargetPassConfig::addMachinePasses() {
244 // Print the instruction selected machine code...
245 printAndVerify("After Instruction Selection");
247 // Expand pseudo-instructions emitted by ISel.
248 addPass(ExpandISelPseudosID);
250 // Add passes that optimize machine instructions in SSA form.
251 if (getOptLevel() != CodeGenOpt::None) {
252 addMachineSSAOptimization();
255 // If the target requests it, assign local variables to stack slots relative
256 // to one another and simplify frame index references where possible.
257 addPass(LocalStackSlotAllocationID);
260 // Run pre-ra passes.
261 if (addPreRegAlloc())
262 printAndVerify("After PreRegAlloc passes");
264 // Run register allocation and passes that are tightly coupled with it,
265 // including phi elimination and scheduling.
266 if (getOptimizeRegAlloc())
267 addOptimizedRegAlloc(createRegAllocPass(true));
269 addFastRegAlloc(createRegAllocPass(false));
271 // Run post-ra passes.
272 if (addPostRegAlloc())
273 printAndVerify("After PostRegAlloc passes");
275 // Insert prolog/epilog code. Eliminate abstract frame index references...
276 addPass(PrologEpilogCodeInserterID);
277 printAndVerify("After PrologEpilogCodeInserter");
279 /// Add passes that optimize machine instructions after register allocation.
280 if (getOptLevel() != CodeGenOpt::None)
281 addMachineLateOptimization();
283 // Expand pseudo instructions before second scheduling pass.
284 addPass(ExpandPostRAPseudosID);
285 printNoVerify("After ExpandPostRAPseudos");
287 // Run pre-sched2 passes.
289 printNoVerify("After PreSched2 passes");
291 // Second pass scheduler.
292 if (getOptLevel() != CodeGenOpt::None && !DisablePostRA) {
293 addPass(PostRASchedulerID);
294 printNoVerify("After PostRAScheduler");
298 addPass(GCMachineCodeAnalysisID);
300 PM.add(createGCInfoPrinter(dbgs()));
302 // Basic block placement.
303 if (getOptLevel() != CodeGenOpt::None && !DisableCodePlace)
306 if (addPreEmitPass())
307 printNoVerify("After PreEmit passes");
310 /// Add passes that optimize machine instructions in SSA form.
311 void TargetPassConfig::addMachineSSAOptimization() {
312 // Pre-ra tail duplication.
313 if (!DisableEarlyTailDup) {
314 addPass(TailDuplicateID);
315 printAndVerify("After Pre-RegAlloc TailDuplicate");
318 // Optimize PHIs before DCE: removing dead PHI cycles may make more
319 // instructions dead.
320 addPass(OptimizePHIsID);
322 // If the target requests it, assign local variables to stack slots relative
323 // to one another and simplify frame index references where possible.
324 addPass(LocalStackSlotAllocationID);
326 // With optimization, dead code should already be eliminated. However
327 // there is one known exception: lowered code for arguments that are only
328 // used by tail calls, where the tail calls reuse the incoming stack
329 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
330 if (!DisableMachineDCE)
331 addPass(DeadMachineInstructionElimID);
332 printAndVerify("After codegen DCE pass");
334 if (!DisableMachineLICM)
335 addPass(MachineLICMID);
336 if (!DisableMachineCSE)
337 addPass(MachineCSEID);
338 if (!DisableMachineSink)
339 addPass(MachineSinkingID);
340 printAndVerify("After Machine LICM, CSE and Sinking passes");
342 addPass(PeepholeOptimizerID);
343 printAndVerify("After codegen peephole optimization pass");
346 //===---------------------------------------------------------------------===//
347 /// Register Allocation Pass Configuration
348 //===---------------------------------------------------------------------===//
350 bool TargetPassConfig::getOptimizeRegAlloc() const {
351 switch (OptimizeRegAlloc) {
352 case cl::BOU_UNSET: return getOptLevel() != CodeGenOpt::None;
353 case cl::BOU_TRUE: return true;
354 case cl::BOU_FALSE: return false;
356 llvm_unreachable("Invalid optimize-regalloc state");
359 /// RegisterRegAlloc's global Registry tracks allocator registration.
360 MachinePassRegistry RegisterRegAlloc::Registry;
362 /// A dummy default pass factory indicates whether the register allocator is
363 /// overridden on the command line.
364 static FunctionPass *useDefaultRegisterAllocator() { return 0; }
365 static RegisterRegAlloc
366 defaultRegAlloc("default",
367 "pick register allocator based on -O option",
368 useDefaultRegisterAllocator);
370 /// -regalloc=... command line option.
371 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false,
372 RegisterPassParser<RegisterRegAlloc> >
374 cl::init(&useDefaultRegisterAllocator),
375 cl::desc("Register allocator to use"));
378 /// Instantiate the default register allocator pass for this target for either
379 /// the optimized or unoptimized allocation path. This will be added to the pass
380 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc
381 /// in the optimized case.
383 /// A target that uses the standard regalloc pass order for fast or optimized
384 /// allocation may still override this for per-target regalloc
385 /// selection. But -regalloc=... always takes precedence.
386 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) {
388 return createGreedyRegisterAllocator();
390 return createFastRegisterAllocator();
393 /// Find and instantiate the register allocation pass requested by this target
394 /// at the current optimization level. Different register allocators are
395 /// defined as separate passes because they may require different analysis.
397 /// This helper ensures that the regalloc= option is always available,
398 /// even for targets that override the default allocator.
400 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs,
401 /// this can be folded into addPass.
402 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) {
403 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault();
405 // Initialize the global default.
408 RegisterRegAlloc::setDefault(RegAlloc);
410 if (Ctor != useDefaultRegisterAllocator)
413 // With no -regalloc= override, ask the target for a regalloc pass.
414 return createTargetRegisterAllocator(Optimized);
417 /// Add the minimum set of target-independent passes that are required for
418 /// register allocation. No coalescing or scheduling.
419 void TargetPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
420 addPass(PHIEliminationID);
421 addPass(TwoAddressInstructionPassID);
423 PM.add(RegAllocPass);
424 printAndVerify("After Register Allocation");
427 /// Add standard target-independent passes that are tightly coupled with
428 /// optimized register allocation, including coalescing, machine instruction
429 /// scheduling, and register allocation itself.
430 void TargetPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
431 // LiveVariables currently requires pure SSA form.
433 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags,
434 // LiveVariables can be removed completely, and LiveIntervals can be directly
435 // computed. (We still either need to regenerate kill flags after regalloc, or
436 // preferably fix the scavenger to not depend on them).
437 addPass(LiveVariablesID);
439 // Add passes that move from transformed SSA into conventional SSA. This is a
440 // "copy coalescing" problem.
442 if (!EnableStrongPHIElim) {
443 // Edge splitting is smarter with machine loop info.
444 addPass(MachineLoopInfoID);
445 addPass(PHIEliminationID);
447 addPass(TwoAddressInstructionPassID);
449 // FIXME: Either remove this pass completely, or fix it so that it works on
450 // SSA form. We could modify LiveIntervals to be independent of this pass, But
451 // it would be even better to simply eliminate *all* IMPLICIT_DEFs before
453 addPass(ProcessImplicitDefsID);
455 if (EnableStrongPHIElim)
456 addPass(StrongPHIEliminationID);
458 addPass(RegisterCoalescerID);
460 // PreRA instruction scheduling.
461 addPass(enablePass(getSchedPass(), EnableMachineSched, &MachineSchedulerID));
463 // Add the selected register allocation pass.
464 PM.add(RegAllocPass);
465 printAndVerify("After Register Allocation");
467 // FinalizeRegAlloc is convenient until MachineInstrBundles is more mature,
468 // but eventually, all users of it should probably be moved to addPostRA and
469 // it can go away. Currently, it's the intended place for targets to run
470 // FinalizeMachineBundles, because passes other than MachineScheduling an
471 // RegAlloc itself may not be aware of bundles.
472 if (addFinalizeRegAlloc())
473 printAndVerify("After RegAlloc finalization");
475 // Perform stack slot coloring and post-ra machine LICM.
477 // FIXME: Re-enable coloring with register when it's capable of adding
480 addPass(StackSlotColoringID);
482 // Run post-ra machine LICM to hoist reloads / remats.
484 // FIXME: can this move into MachineLateOptimization?
485 if (!DisablePostRAMachineLICM)
486 addPass(MachineLICMID);
488 printAndVerify("After StackSlotColoring and postra Machine LICM");
491 //===---------------------------------------------------------------------===//
492 /// Post RegAlloc Pass Configuration
493 //===---------------------------------------------------------------------===//
495 /// Add passes that optimize machine instructions after register allocation.
496 void TargetPassConfig::addMachineLateOptimization() {
497 // Branch folding must be run after regalloc and prolog/epilog insertion.
498 if (!DisableBranchFold) {
499 addPass(BranchFolderPassID);
500 printNoVerify("After BranchFolding");
504 if (!DisableTailDuplicate) {
505 addPass(TailDuplicateID);
506 printNoVerify("After TailDuplicate");
510 if (!DisableCopyProp) {
511 addPass(MachineCopyPropagationID);
512 printNoVerify("After copy propagation pass");
516 /// Add standard basic block placement passes.
517 void TargetPassConfig::addBlockPlacement() {
518 if (EnableBlockPlacement) {
519 // MachineBlockPlacement is an experimental pass which is disabled by
520 // default currently. Eventually it should subsume CodePlacementOpt, so
521 // when enabled, the other is disabled.
522 addPass(MachineBlockPlacementID);
523 printNoVerify("After MachineBlockPlacement");
525 addPass(CodePlacementOptID);
526 printNoVerify("After CodePlacementOpt");
529 // Run a separate pass to collect block placement statistics.
530 if (EnableBlockPlacementStats) {
531 addPass(MachineBlockPlacementStatsID);
532 printNoVerify("After MachineBlockPlacementStats");