Enable the loop vectorizer only on O2 and above. (Still disabled by default)
[oota-llvm.git] / lib / Transforms / IPO / PassManagerBuilder.cpp
1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the PassManagerBuilder class, which is used to set up a
11 // "standard" optimization sequence suitable for languages like C and C++.
12 //
13 //===----------------------------------------------------------------------===//
14
15
16 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
17 #include "llvm-c/Transforms/PassManagerBuilder.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/Analysis/Passes.h"
20 #include "llvm/Analysis/Verifier.h"
21 #include "llvm/DefaultPasses.h"
22 #include "llvm/PassManager.h"
23 #include "llvm/PassManager.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/ManagedStatic.h"
26 #include "llvm/Target/TargetLibraryInfo.h"
27 #include "llvm/Transforms/IPO.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/Transforms/Vectorize.h"
30
31 using namespace llvm;
32
33 static cl::opt<bool>
34 RunLoopVectorization("vectorize-loops",
35                      cl::desc("Run the Loop vectorization passes"));
36
37 static cl::opt<bool>
38 RunBBVectorization("vectorize", cl::desc("Run the BB vectorization passes"));
39
40 static cl::opt<bool>
41 UseGVNAfterVectorization("use-gvn-after-vectorization",
42   cl::init(false), cl::Hidden,
43   cl::desc("Run GVN instead of Early CSE after vectorization passes"));
44
45 static cl::opt<bool> UseNewSROA("use-new-sroa",
46   cl::init(true), cl::Hidden,
47   cl::desc("Enable the new, experimental SROA pass"));
48
49 PassManagerBuilder::PassManagerBuilder() {
50     OptLevel = 2;
51     SizeLevel = 0;
52     LibraryInfo = 0;
53     Inliner = 0;
54     DisableSimplifyLibCalls = false;
55     DisableUnitAtATime = false;
56     DisableUnrollLoops = false;
57     Vectorize = RunBBVectorization;
58     LoopVectorize = RunLoopVectorization;
59 }
60
61 PassManagerBuilder::~PassManagerBuilder() {
62   delete LibraryInfo;
63   delete Inliner;
64 }
65
66 /// Set of global extensions, automatically added as part of the standard set.
67 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
68    PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
69
70 void PassManagerBuilder::addGlobalExtension(
71     PassManagerBuilder::ExtensionPointTy Ty,
72     PassManagerBuilder::ExtensionFn Fn) {
73   GlobalExtensions->push_back(std::make_pair(Ty, Fn));
74 }
75
76 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
77   Extensions.push_back(std::make_pair(Ty, Fn));
78 }
79
80 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
81                                            PassManagerBase &PM) const {
82   for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
83     if ((*GlobalExtensions)[i].first == ETy)
84       (*GlobalExtensions)[i].second(*this, PM);
85   for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
86     if (Extensions[i].first == ETy)
87       Extensions[i].second(*this, PM);
88 }
89
90 void
91 PassManagerBuilder::addInitialAliasAnalysisPasses(PassManagerBase &PM) const {
92   // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
93   // BasicAliasAnalysis wins if they disagree. This is intended to help
94   // support "obvious" type-punning idioms.
95   PM.add(createTypeBasedAliasAnalysisPass());
96   PM.add(createBasicAliasAnalysisPass());
97 }
98
99 void PassManagerBuilder::populateFunctionPassManager(FunctionPassManager &FPM) {
100   addExtensionsToPM(EP_EarlyAsPossible, FPM);
101
102   // Add LibraryInfo if we have some.
103   if (LibraryInfo) FPM.add(new TargetLibraryInfo(*LibraryInfo));
104
105   if (OptLevel == 0) return;
106
107   addInitialAliasAnalysisPasses(FPM);
108
109   FPM.add(createCFGSimplificationPass());
110   if (UseNewSROA)
111     FPM.add(createSROAPass());
112   else
113     FPM.add(createScalarReplAggregatesPass());
114   FPM.add(createEarlyCSEPass());
115   FPM.add(createLowerExpectIntrinsicPass());
116 }
117
118 void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
119   // If all optimizations are disabled, just run the always-inline pass.
120   if (OptLevel == 0) {
121     if (Inliner) {
122       MPM.add(Inliner);
123       Inliner = 0;
124     }
125
126     // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
127     // pass manager, but we don't want to add extensions into that pass manager.
128     // To prevent this we must insert a no-op module pass to reset the pass
129     // manager to get the same behavior as EP_OptimizerLast in non-O0 builds.
130     if (!GlobalExtensions->empty() || !Extensions.empty())
131       MPM.add(createBarrierNoopPass());
132
133     addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
134     return;
135   }
136
137   // Add LibraryInfo if we have some.
138   if (LibraryInfo) MPM.add(new TargetLibraryInfo(*LibraryInfo));
139
140   addInitialAliasAnalysisPasses(MPM);
141
142   if (!DisableUnitAtATime) {
143     addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
144
145     MPM.add(createGlobalOptimizerPass());     // Optimize out global vars
146
147     MPM.add(createIPSCCPPass());              // IP SCCP
148     MPM.add(createDeadArgEliminationPass());  // Dead argument elimination
149
150     MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
151     MPM.add(createCFGSimplificationPass());   // Clean up after IPCP & DAE
152   }
153
154   // Start of CallGraph SCC passes.
155   if (!DisableUnitAtATime)
156     MPM.add(createPruneEHPass());             // Remove dead EH info
157   if (Inliner) {
158     MPM.add(Inliner);
159     Inliner = 0;
160   }
161   if (!DisableUnitAtATime)
162     MPM.add(createFunctionAttrsPass());       // Set readonly/readnone attrs
163   if (OptLevel > 2)
164     MPM.add(createArgumentPromotionPass());   // Scalarize uninlined fn args
165
166   // Start of function pass.
167   // Break up aggregate allocas, using SSAUpdater.
168   if (UseNewSROA)
169     MPM.add(createSROAPass(/*RequiresDomTree*/ false));
170   else
171     MPM.add(createScalarReplAggregatesPass(-1, false));
172   MPM.add(createEarlyCSEPass());              // Catch trivial redundancies
173   if (!DisableSimplifyLibCalls)
174     MPM.add(createSimplifyLibCallsPass());    // Library Call Optimizations
175   MPM.add(createJumpThreadingPass());         // Thread jumps.
176   MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
177   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
178   MPM.add(createInstructionCombiningPass());  // Combine silly seq's
179
180   MPM.add(createTailCallEliminationPass());   // Eliminate tail calls
181   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
182   MPM.add(createReassociatePass());           // Reassociate expressions
183   MPM.add(createLoopRotatePass());            // Rotate Loop
184   MPM.add(createLICMPass());                  // Hoist loop invariants
185   MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
186   MPM.add(createInstructionCombiningPass());
187   MPM.add(createIndVarSimplifyPass());        // Canonicalize indvars
188   MPM.add(createLoopIdiomPass());             // Recognize idioms like memset.
189   MPM.add(createLoopDeletionPass());          // Delete dead loops
190
191   if (LoopVectorize && OptLevel > 1)
192     MPM.add(createLoopVectorizePass());
193
194   if (!DisableUnrollLoops)
195     MPM.add(createLoopUnrollPass());          // Unroll small loops
196   addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
197
198   if (OptLevel > 1)
199     MPM.add(createGVNPass());                 // Remove redundancies
200   MPM.add(createMemCpyOptPass());             // Remove memcpy / form memset
201   MPM.add(createSCCPPass());                  // Constant prop with SCCP
202
203   // Run instcombine after redundancy elimination to exploit opportunities
204   // opened up by them.
205   MPM.add(createInstructionCombiningPass());
206   MPM.add(createJumpThreadingPass());         // Thread jumps
207   MPM.add(createCorrelatedValuePropagationPass());
208   MPM.add(createDeadStoreEliminationPass());  // Delete dead stores
209
210   addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
211
212   if (Vectorize) {
213     MPM.add(createBBVectorizePass());
214     MPM.add(createInstructionCombiningPass());
215     if (OptLevel > 1 && UseGVNAfterVectorization)
216       MPM.add(createGVNPass());                   // Remove redundancies
217     else
218       MPM.add(createEarlyCSEPass());              // Catch trivial redundancies
219   }
220
221   MPM.add(createAggressiveDCEPass());         // Delete dead instructions
222   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
223   MPM.add(createInstructionCombiningPass());  // Clean up after everything.
224
225   if (!DisableUnitAtATime) {
226     // FIXME: We shouldn't bother with this anymore.
227     MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
228
229     // GlobalOpt already deletes dead functions and globals, at -O2 try a
230     // late pass of GlobalDCE.  It is capable of deleting dead cycles.
231     if (OptLevel > 1) {
232       MPM.add(createGlobalDCEPass());         // Remove dead fns and globals.
233       MPM.add(createConstantMergePass());     // Merge dup global constants
234     }
235   }
236   addExtensionsToPM(EP_OptimizerLast, MPM);
237 }
238
239 void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM,
240                                                 bool Internalize,
241                                                 bool RunInliner,
242                                                 bool DisableGVNLoadPRE) {
243   // Provide AliasAnalysis services for optimizations.
244   addInitialAliasAnalysisPasses(PM);
245
246   // Now that composite has been compiled, scan through the module, looking
247   // for a main function.  If main is defined, mark all other functions
248   // internal.
249   if (Internalize) {
250     std::vector<const char*> E;
251     E.push_back("main");
252     PM.add(createInternalizePass(E));
253   }
254
255   // Propagate constants at call sites into the functions they call.  This
256   // opens opportunities for globalopt (and inlining) by substituting function
257   // pointers passed as arguments to direct uses of functions.
258   PM.add(createIPSCCPPass());
259
260   // Now that we internalized some globals, see if we can hack on them!
261   PM.add(createGlobalOptimizerPass());
262
263   // Linking modules together can lead to duplicated global constants, only
264   // keep one copy of each constant.
265   PM.add(createConstantMergePass());
266
267   // Remove unused arguments from functions.
268   PM.add(createDeadArgEliminationPass());
269
270   // Reduce the code after globalopt and ipsccp.  Both can open up significant
271   // simplification opportunities, and both can propagate functions through
272   // function pointers.  When this happens, we often have to resolve varargs
273   // calls, etc, so let instcombine do this.
274   PM.add(createInstructionCombiningPass());
275
276   // Inline small functions
277   if (RunInliner)
278     PM.add(createFunctionInliningPass());
279
280   PM.add(createPruneEHPass());   // Remove dead EH info.
281
282   // Optimize globals again if we ran the inliner.
283   if (RunInliner)
284     PM.add(createGlobalOptimizerPass());
285   PM.add(createGlobalDCEPass()); // Remove dead functions.
286
287   // If we didn't decide to inline a function, check to see if we can
288   // transform it to pass arguments by value instead of by reference.
289   PM.add(createArgumentPromotionPass());
290
291   // The IPO passes may leave cruft around.  Clean up after them.
292   PM.add(createInstructionCombiningPass());
293   PM.add(createJumpThreadingPass());
294   // Break up allocas
295   if (UseNewSROA)
296     PM.add(createSROAPass());
297   else
298     PM.add(createScalarReplAggregatesPass());
299
300   // Run a few AA driven optimizations here and now, to cleanup the code.
301   PM.add(createFunctionAttrsPass()); // Add nocapture.
302   PM.add(createGlobalsModRefPass()); // IP alias analysis.
303
304   PM.add(createLICMPass());                 // Hoist loop invariants.
305   PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
306   PM.add(createMemCpyOptPass());            // Remove dead memcpys.
307   // Nuke dead stores.
308   PM.add(createDeadStoreEliminationPass());
309
310   // Cleanup and simplify the code after the scalar optimizations.
311   PM.add(createInstructionCombiningPass());
312
313   PM.add(createJumpThreadingPass());
314
315   // Delete basic blocks, which optimization passes may have killed.
316   PM.add(createCFGSimplificationPass());
317
318   // Now that we have optimized the program, discard unreachable functions.
319   PM.add(createGlobalDCEPass());
320 }
321
322 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
323   PassManagerBuilder *PMB = new PassManagerBuilder();
324   return wrap(PMB);
325 }
326
327 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
328   PassManagerBuilder *Builder = unwrap(PMB);
329   delete Builder;
330 }
331
332 void
333 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
334                                   unsigned OptLevel) {
335   PassManagerBuilder *Builder = unwrap(PMB);
336   Builder->OptLevel = OptLevel;
337 }
338
339 void
340 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
341                                    unsigned SizeLevel) {
342   PassManagerBuilder *Builder = unwrap(PMB);
343   Builder->SizeLevel = SizeLevel;
344 }
345
346 void
347 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
348                                             LLVMBool Value) {
349   PassManagerBuilder *Builder = unwrap(PMB);
350   Builder->DisableUnitAtATime = Value;
351 }
352
353 void
354 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
355                                             LLVMBool Value) {
356   PassManagerBuilder *Builder = unwrap(PMB);
357   Builder->DisableUnrollLoops = Value;
358 }
359
360 void
361 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
362                                                  LLVMBool Value) {
363   PassManagerBuilder *Builder = unwrap(PMB);
364   Builder->DisableSimplifyLibCalls = Value;
365 }
366
367 void
368 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
369                                               unsigned Threshold) {
370   PassManagerBuilder *Builder = unwrap(PMB);
371   Builder->Inliner = createFunctionInliningPass(Threshold);
372 }
373
374 void
375 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
376                                                   LLVMPassManagerRef PM) {
377   PassManagerBuilder *Builder = unwrap(PMB);
378   FunctionPassManager *FPM = unwrap<FunctionPassManager>(PM);
379   Builder->populateFunctionPassManager(*FPM);
380 }
381
382 void
383 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
384                                                 LLVMPassManagerRef PM) {
385   PassManagerBuilder *Builder = unwrap(PMB);
386   PassManagerBase *MPM = unwrap(PM);
387   Builder->populateModulePassManager(*MPM);
388 }
389
390 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
391                                                   LLVMPassManagerRef PM,
392                                                   bool Internalize,
393                                                   bool RunInliner) {
394   PassManagerBuilder *Builder = unwrap(PMB);
395   PassManagerBase *LPM = unwrap(PM);
396   Builder->populateLTOPassManager(*LPM, Internalize, RunInliner);
397 }