1 //===-- LLVMTargetMachine.cpp - Implement the LLVMTargetMachine class -----===//
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 implements the LLVMTargetMachine class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetMachine.h"
15 #include "llvm/PassManager.h"
16 #include "llvm/Analysis/Passes.h"
17 #include "llvm/Analysis/Verifier.h"
18 #include "llvm/Assembly/PrintModulePass.h"
19 #include "llvm/CodeGen/AsmPrinter.h"
20 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
21 #include "llvm/CodeGen/MachineModuleInfo.h"
22 #include "llvm/CodeGen/GCStrategy.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/Target/TargetLowering.h"
25 #include "llvm/Target/TargetOptions.h"
26 #include "llvm/MC/MCAsmInfo.h"
27 #include "llvm/MC/MCInstrInfo.h"
28 #include "llvm/MC/MCStreamer.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetLoweringObjectFile.h"
34 #include "llvm/Target/TargetRegisterInfo.h"
35 #include "llvm/Target/TargetSubtargetInfo.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/ADT/OwningPtr.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/FormattedStream.h"
41 #include "llvm/Support/TargetRegistry.h"
48 static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden,
49 cl::desc("Disable Post Regalloc"));
50 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
51 cl::desc("Disable branch folding"));
52 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
53 cl::desc("Disable tail duplication"));
54 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
55 cl::desc("Disable pre-register allocation tail duplication"));
56 static cl::opt<bool> EnableBlockPlacement("enable-block-placement",
57 cl::Hidden, cl::desc("Enable probability-driven block placement"));
58 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats",
59 cl::Hidden, cl::desc("Collect probability-driven block placement stats"));
60 static cl::opt<bool> DisableCodePlace("disable-code-place", cl::Hidden,
61 cl::desc("Disable code placement"));
62 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
63 cl::desc("Disable Stack Slot Coloring"));
64 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden,
65 cl::desc("Disable Machine Dead Code Elimination"));
66 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
67 cl::desc("Disable Machine LICM"));
68 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden,
69 cl::desc("Disable Machine Common Subexpression Elimination"));
70 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
72 cl::desc("Disable Machine LICM"));
73 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
74 cl::desc("Disable Machine Sinking"));
75 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
76 cl::desc("Disable Loop Strength Reduction Pass"));
77 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
78 cl::desc("Disable Codegen Prepare"));
79 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
80 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
81 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
82 cl::desc("Print LLVM IR input to isel pass"));
83 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
84 cl::desc("Dump garbage collector data"));
85 static cl::opt<bool> ShowMCEncoding("show-mc-encoding", cl::Hidden,
86 cl::desc("Show encoding in .s output"));
87 static cl::opt<bool> ShowMCInst("show-mc-inst", cl::Hidden,
88 cl::desc("Show instruction structure in .s output"));
89 static cl::opt<bool> EnableMCLogging("enable-mc-api-logging", cl::Hidden,
90 cl::desc("Enable MC API logging"));
91 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
92 cl::desc("Verify generated machine code"),
93 cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
95 static cl::opt<cl::boolOrDefault>
96 AsmVerbose("asm-verbose", cl::desc("Add comments to directives."),
97 cl::init(cl::BOU_UNSET));
99 static bool getVerboseAsm() {
100 switch (AsmVerbose) {
102 case cl::BOU_UNSET: return TargetMachine::getAsmVerbosityDefault();
103 case cl::BOU_TRUE: return true;
104 case cl::BOU_FALSE: return false;
108 // Enable or disable FastISel. Both options are needed, because
109 // FastISel is enabled by default with -fast, and we wish to be
110 // able to enable or disable fast-isel independently from -O0.
111 static cl::opt<cl::boolOrDefault>
112 EnableFastISelOption("fast-isel", cl::Hidden,
113 cl::desc("Enable the \"fast\" instruction selector"));
115 LLVMTargetMachine::LLVMTargetMachine(const Target &T, StringRef Triple,
116 StringRef CPU, StringRef FS,
117 Reloc::Model RM, CodeModel::Model CM,
118 CodeGenOpt::Level OL)
119 : TargetMachine(T, Triple, CPU, FS) {
120 CodeGenInfo = T.createMCCodeGenInfo(Triple, RM, CM, OL);
121 AsmInfo = T.createMCAsmInfo(Triple);
122 // TargetSelect.h moved to a different directory between LLVM 2.9 and 3.0,
123 // and if the old one gets included then MCAsmInfo will be NULL and
124 // we'll crash later.
125 // Provide the user with a useful error message about what's wrong.
126 assert(AsmInfo && "MCAsmInfo not initialized."
127 "Make sure you include the correct TargetSelect.h"
128 "and that InitializeAllTargetMCs() is being invoked!");
131 bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
132 formatted_raw_ostream &Out,
133 CodeGenFileType FileType,
134 bool DisableVerify) {
135 // Add common CodeGen passes.
136 MCContext *Context = 0;
137 if (addCommonCodeGenPasses(PM, DisableVerify, Context))
139 assert(Context != 0 && "Failed to get MCContext");
141 if (hasMCSaveTempLabels())
142 Context->setAllowTemporaryLabels(false);
144 const MCAsmInfo &MAI = *getMCAsmInfo();
145 const MCSubtargetInfo &STI = getSubtarget<MCSubtargetInfo>();
146 OwningPtr<MCStreamer> AsmStreamer;
149 default: return true;
150 case CGFT_AssemblyFile: {
151 MCInstPrinter *InstPrinter =
152 getTarget().createMCInstPrinter(MAI.getAssemblerDialect(), MAI, STI);
154 // Create a code emitter if asked to show the encoding.
155 MCCodeEmitter *MCE = 0;
156 MCAsmBackend *MAB = 0;
157 if (ShowMCEncoding) {
158 const MCSubtargetInfo &STI = getSubtarget<MCSubtargetInfo>();
159 MCE = getTarget().createMCCodeEmitter(*getInstrInfo(), STI, *Context);
160 MAB = getTarget().createMCAsmBackend(getTargetTriple());
163 MCStreamer *S = getTarget().createAsmStreamer(*Context, Out,
167 hasMCUseDwarfDirectory(),
171 AsmStreamer.reset(S);
174 case CGFT_ObjectFile: {
175 // Create the code emitter for the target if it exists. If not, .o file
177 MCCodeEmitter *MCE = getTarget().createMCCodeEmitter(*getInstrInfo(), STI,
179 MCAsmBackend *MAB = getTarget().createMCAsmBackend(getTargetTriple());
180 if (MCE == 0 || MAB == 0)
183 AsmStreamer.reset(getTarget().createMCObjectStreamer(getTargetTriple(),
185 MCE, hasMCRelaxAll(),
186 hasMCNoExecStack()));
187 AsmStreamer.get()->InitSections();
191 // The Null output is intended for use for performance analysis and testing,
193 AsmStreamer.reset(createNullStreamer(*Context));
198 AsmStreamer.reset(createLoggingStreamer(AsmStreamer.take(), errs()));
200 // Create the AsmPrinter, which takes ownership of AsmStreamer if successful.
201 FunctionPass *Printer = getTarget().createAsmPrinter(*this, *AsmStreamer);
205 // If successful, createAsmPrinter took ownership of AsmStreamer.
210 PM.add(createGCInfoDeleter());
214 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to
215 /// get machine code emitted. This uses a JITCodeEmitter object to handle
216 /// actually outputting the machine code and resolving things like the address
217 /// of functions. This method should returns true if machine code emission is
220 bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
222 bool DisableVerify) {
223 // Add common CodeGen passes.
225 if (addCommonCodeGenPasses(PM, DisableVerify, Ctx))
228 addCodeEmitter(PM, JCE);
229 PM.add(createGCInfoDeleter());
231 return false; // success!
234 /// addPassesToEmitMC - Add passes to the specified pass manager to get
235 /// machine code emitted with the MCJIT. This method returns true if machine
236 /// code is not supported. It fills the MCContext Ctx pointer which can be
237 /// used to build custom MCStreamer.
239 bool LLVMTargetMachine::addPassesToEmitMC(PassManagerBase &PM,
242 bool DisableVerify) {
243 // Add common CodeGen passes.
244 if (addCommonCodeGenPasses(PM, DisableVerify, Ctx))
247 if (hasMCSaveTempLabels())
248 Ctx->setAllowTemporaryLabels(false);
250 // Create the code emitter for the target if it exists. If not, .o file
252 const MCSubtargetInfo &STI = getSubtarget<MCSubtargetInfo>();
253 MCCodeEmitter *MCE = getTarget().createMCCodeEmitter(*getInstrInfo(),STI, *Ctx);
254 MCAsmBackend *MAB = getTarget().createMCAsmBackend(getTargetTriple());
255 if (MCE == 0 || MAB == 0)
258 OwningPtr<MCStreamer> AsmStreamer;
259 AsmStreamer.reset(getTarget().createMCObjectStreamer(getTargetTriple(), *Ctx,
262 hasMCNoExecStack()));
263 AsmStreamer.get()->InitSections();
265 // Create the AsmPrinter, which takes ownership of AsmStreamer if successful.
266 FunctionPass *Printer = getTarget().createAsmPrinter(*this, *AsmStreamer);
270 // If successful, createAsmPrinter took ownership of AsmStreamer.
275 return false; // success!
278 static void printNoVerify(PassManagerBase &PM, const char *Banner) {
279 if (PrintMachineCode)
280 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
283 static void printAndVerify(PassManagerBase &PM,
284 const char *Banner) {
285 if (PrintMachineCode)
286 PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
288 if (VerifyMachineCode)
289 PM.add(createMachineVerifierPass(Banner));
292 /// addCommonCodeGenPasses - Add standard LLVM codegen passes used for both
293 /// emitting to assembly files or machine code output.
295 bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
297 MCContext *&OutContext) {
298 // Standard LLVM-Level Passes.
300 // Basic AliasAnalysis support.
301 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
302 // BasicAliasAnalysis wins if they disagree. This is intended to help
303 // support "obvious" type-punning idioms.
304 PM.add(createTypeBasedAliasAnalysisPass());
305 PM.add(createBasicAliasAnalysisPass());
307 // Before running any passes, run the verifier to determine if the input
308 // coming from the front-end and/or optimizer is valid.
310 PM.add(createVerifierPass());
312 // Run loop strength reduction before anything else.
313 if (getOptLevel() != CodeGenOpt::None && !DisableLSR) {
314 PM.add(createLoopStrengthReducePass(getTargetLowering()));
316 PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &dbgs()));
319 PM.add(createGCLoweringPass());
321 // Make sure that no unreachable blocks are instruction selected.
322 PM.add(createUnreachableBlockEliminationPass());
324 // Turn exception handling constructs into something the code generators can
326 switch (getMCAsmInfo()->getExceptionHandlingType()) {
327 case ExceptionHandling::SjLj:
328 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
329 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
330 // catch info can get misplaced when a selector ends up more than one block
331 // removed from the parent invoke(s). This could happen when a landing
332 // pad is shared by multiple invokes and is also a target of a normal
333 // edge from elsewhere.
334 PM.add(createSjLjEHPass(getTargetLowering()));
336 case ExceptionHandling::DwarfCFI:
337 case ExceptionHandling::ARM:
338 case ExceptionHandling::Win64:
339 PM.add(createDwarfEHPass(this));
341 case ExceptionHandling::None:
342 PM.add(createLowerInvokePass(getTargetLowering()));
344 // The lower invoke pass may create unreachable code. Remove it.
345 PM.add(createUnreachableBlockEliminationPass());
349 if (getOptLevel() != CodeGenOpt::None && !DisableCGP)
350 PM.add(createCodeGenPreparePass(getTargetLowering()));
352 PM.add(createStackProtectorPass(getTargetLowering()));
357 PM.add(createPrintFunctionPass("\n\n"
358 "*** Final LLVM Code input to ISel ***\n",
361 // All passes which modify the LLVM IR are now complete; run the verifier
362 // to ensure that the IR is valid.
364 PM.add(createVerifierPass());
366 // Standard Lower-Level Passes.
368 // Install a MachineModuleInfo class, which is an immutable pass that holds
369 // all the per-module stuff we're generating, including MCContext.
370 MachineModuleInfo *MMI = new MachineModuleInfo(*getMCAsmInfo(),
372 &getTargetLowering()->getObjFileLowering());
374 OutContext = &MMI->getContext(); // Return the MCContext specifically by-ref.
376 // Set up a MachineFunction for the rest of CodeGen to work on.
377 PM.add(new MachineFunctionAnalysis(*this));
379 // Enable FastISel with -fast, but allow that to be overridden.
380 if (EnableFastISelOption == cl::BOU_TRUE ||
381 (getOptLevel() == CodeGenOpt::None &&
382 EnableFastISelOption != cl::BOU_FALSE))
383 EnableFastISel = true;
385 // Ask the target for an isel.
386 if (addInstSelector(PM))
389 // Print the instruction selected machine code...
390 printAndVerify(PM, "After Instruction Selection");
392 // Expand pseudo-instructions emitted by ISel.
393 PM.add(createExpandISelPseudosPass());
395 // Pre-ra tail duplication.
396 if (getOptLevel() != CodeGenOpt::None && !DisableEarlyTailDup) {
397 PM.add(createTailDuplicatePass(true));
398 printAndVerify(PM, "After Pre-RegAlloc TailDuplicate");
401 // Optimize PHIs before DCE: removing dead PHI cycles may make more
402 // instructions dead.
403 if (getOptLevel() != CodeGenOpt::None)
404 PM.add(createOptimizePHIsPass());
406 // If the target requests it, assign local variables to stack slots relative
407 // to one another and simplify frame index references where possible.
408 PM.add(createLocalStackSlotAllocationPass());
410 if (getOptLevel() != CodeGenOpt::None) {
411 // With optimization, dead code should already be eliminated. However
412 // there is one known exception: lowered code for arguments that are only
413 // used by tail calls, where the tail calls reuse the incoming stack
414 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
415 if (!DisableMachineDCE)
416 PM.add(createDeadMachineInstructionElimPass());
417 printAndVerify(PM, "After codegen DCE pass");
419 if (!DisableMachineLICM)
420 PM.add(createMachineLICMPass());
421 if (!DisableMachineCSE)
422 PM.add(createMachineCSEPass());
423 if (!DisableMachineSink)
424 PM.add(createMachineSinkingPass());
425 printAndVerify(PM, "After Machine LICM, CSE and Sinking passes");
427 PM.add(createPeepholeOptimizerPass());
428 printAndVerify(PM, "After codegen peephole optimization pass");
431 // Run pre-ra passes.
432 if (addPreRegAlloc(PM))
433 printAndVerify(PM, "After PreRegAlloc passes");
435 // Perform register allocation.
436 PM.add(createRegisterAllocator(getOptLevel()));
437 printAndVerify(PM, "After Register Allocation");
439 // Perform stack slot coloring and post-ra machine LICM.
440 if (getOptLevel() != CodeGenOpt::None) {
441 // FIXME: Re-enable coloring with register when it's capable of adding
444 PM.add(createStackSlotColoringPass(false));
446 // Run post-ra machine LICM to hoist reloads / remats.
447 if (!DisablePostRAMachineLICM)
448 PM.add(createMachineLICMPass(false));
450 printAndVerify(PM, "After StackSlotColoring and postra Machine LICM");
453 // Run post-ra passes.
454 if (addPostRegAlloc(PM))
455 printAndVerify(PM, "After PostRegAlloc passes");
457 PM.add(createExpandPostRAPseudosPass());
458 printAndVerify(PM, "After ExpandPostRAPseudos");
460 // Insert prolog/epilog code. Eliminate abstract frame index references...
461 PM.add(createPrologEpilogCodeInserter());
462 printAndVerify(PM, "After PrologEpilogCodeInserter");
464 // Run pre-sched2 passes.
465 if (addPreSched2(PM))
466 printAndVerify(PM, "After PreSched2 passes");
468 // Second pass scheduler.
469 if (getOptLevel() != CodeGenOpt::None && !DisablePostRA) {
470 PM.add(createPostRAScheduler(getOptLevel()));
471 printAndVerify(PM, "After PostRAScheduler");
474 // Branch folding must be run after regalloc and prolog/epilog insertion.
475 if (getOptLevel() != CodeGenOpt::None && !DisableBranchFold) {
476 PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
477 printNoVerify(PM, "After BranchFolding");
481 if (getOptLevel() != CodeGenOpt::None && !DisableTailDuplicate) {
482 PM.add(createTailDuplicatePass(false));
483 printNoVerify(PM, "After TailDuplicate");
486 PM.add(createGCMachineCodeAnalysisPass());
489 PM.add(createGCInfoPrinter(dbgs()));
491 if (getOptLevel() != CodeGenOpt::None && !DisableCodePlace) {
492 if (EnableBlockPlacement) {
493 // MachineBlockPlacement is an experimental pass which is disabled by
494 // default currently. Eventually it should subsume CodePlacementOpt, so
495 // when enabled, the other is disabled.
496 PM.add(createMachineBlockPlacementPass());
497 printNoVerify(PM, "After MachineBlockPlacement");
499 PM.add(createCodePlacementOptPass());
500 printNoVerify(PM, "After CodePlacementOpt");
503 // Run a separate pass to collect block placement statistics.
504 if (EnableBlockPlacementStats) {
505 PM.add(createMachineBlockPlacementStatsPass());
506 printNoVerify(PM, "After MachineBlockPlacementStats");
510 if (addPreEmitPass(PM))
511 printNoVerify(PM, "After PreEmit passes");