Bug Summary

File:lib/CodeGen/CodeGenPrepare.cpp
Warning:line 3153, column 41
Called C++ object pointer is null

Annotated Source Code

1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
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 pass munges the code in the input function to better prepare it for
11// SelectionDAG-based code generation. This works around limitations in it's
12// basic-block-at-a-time approach. It should eventually be removed.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/ADT/APInt.h"
17#include "llvm/ADT/ArrayRef.h"
18#include "llvm/ADT/DenseMap.h"
19#include "llvm/ADT/PointerIntPair.h"
20#include "llvm/ADT/STLExtras.h"
21#include "llvm/ADT/SetVector.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/Statistic.h"
25#include "llvm/Analysis/BlockFrequencyInfo.h"
26#include "llvm/Analysis/BranchProbabilityInfo.h"
27#include "llvm/Analysis/ConstantFolding.h"
28#include "llvm/Analysis/InstructionSimplify.h"
29#include "llvm/Analysis/LoopInfo.h"
30#include "llvm/Analysis/MemoryBuiltins.h"
31#include "llvm/Analysis/ProfileSummaryInfo.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/TargetTransformInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/CodeGen/Analysis.h"
36#include "llvm/CodeGen/ISDOpcodes.h"
37#include "llvm/CodeGen/MachineValueType.h"
38#include "llvm/CodeGen/SelectionDAGNodes.h"
39#include "llvm/CodeGen/TargetLowering.h"
40#include "llvm/CodeGen/TargetPassConfig.h"
41#include "llvm/CodeGen/TargetSubtargetInfo.h"
42#include "llvm/CodeGen/ValueTypes.h"
43#include "llvm/IR/Argument.h"
44#include "llvm/IR/Attributes.h"
45#include "llvm/IR/BasicBlock.h"
46#include "llvm/IR/CallSite.h"
47#include "llvm/IR/Constant.h"
48#include "llvm/IR/Constants.h"
49#include "llvm/IR/DataLayout.h"
50#include "llvm/IR/DerivedTypes.h"
51#include "llvm/IR/Dominators.h"
52#include "llvm/IR/Function.h"
53#include "llvm/IR/GetElementPtrTypeIterator.h"
54#include "llvm/IR/GlobalValue.h"
55#include "llvm/IR/GlobalVariable.h"
56#include "llvm/IR/IRBuilder.h"
57#include "llvm/IR/InlineAsm.h"
58#include "llvm/IR/InstrTypes.h"
59#include "llvm/IR/Instruction.h"
60#include "llvm/IR/Instructions.h"
61#include "llvm/IR/IntrinsicInst.h"
62#include "llvm/IR/Intrinsics.h"
63#include "llvm/IR/LLVMContext.h"
64#include "llvm/IR/MDBuilder.h"
65#include "llvm/IR/Module.h"
66#include "llvm/IR/Operator.h"
67#include "llvm/IR/PatternMatch.h"
68#include "llvm/IR/Statepoint.h"
69#include "llvm/IR/Type.h"
70#include "llvm/IR/Use.h"
71#include "llvm/IR/User.h"
72#include "llvm/IR/Value.h"
73#include "llvm/IR/ValueHandle.h"
74#include "llvm/IR/ValueMap.h"
75#include "llvm/Pass.h"
76#include "llvm/Support/BlockFrequency.h"
77#include "llvm/Support/BranchProbability.h"
78#include "llvm/Support/Casting.h"
79#include "llvm/Support/CommandLine.h"
80#include "llvm/Support/Compiler.h"
81#include "llvm/Support/Debug.h"
82#include "llvm/Support/ErrorHandling.h"
83#include "llvm/Support/MathExtras.h"
84#include "llvm/Support/raw_ostream.h"
85#include "llvm/Target/TargetMachine.h"
86#include "llvm/Target/TargetOptions.h"
87#include "llvm/Transforms/Utils/BasicBlockUtils.h"
88#include "llvm/Transforms/Utils/BypassSlowDivision.h"
89#include "llvm/Transforms/Utils/Cloning.h"
90#include "llvm/Transforms/Utils/Local.h"
91#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
92#include "llvm/Transforms/Utils/ValueMapper.h"
93#include <algorithm>
94#include <cassert>
95#include <cstdint>
96#include <iterator>
97#include <limits>
98#include <memory>
99#include <utility>
100#include <vector>
101
102using namespace llvm;
103using namespace llvm::PatternMatch;
104
105#define DEBUG_TYPE"codegenprepare" "codegenprepare"
106
107STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim"
, "Number of blocks eliminated", {0}, false}
;
108STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim"
, "Number of trivial PHIs eliminated", {0}, false}
;
109STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim"
, "Number of GEPs converted to casts", {0}, false}
;
110STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses"
, "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps"
, {0}, false}
111 "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses"
, "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps"
, {0}, false}
;
112STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses"
, "Number of uses of Cast expressions replaced with uses " "of sunken Casts"
, {0}, false}
113 "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses"
, "Number of uses of Cast expressions replaced with uses " "of sunken Casts"
, {0}, false}
;
114STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
, {0}, false}
115 "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
, {0}, false}
;
116STATISTIC(NumMemoryInstsPhiCreated,static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions", {0}, false}
117 "Number of phis created when address "static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions", {0}, false}
118 "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions", {0}, false}
;
119STATISTIC(NumMemoryInstsSelectCreated,static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions", {0}, false}
120 "Number of select created when address "static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions", {0}, false}
121 "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions", {0}, false}
;
122STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads")static llvm::Statistic NumExtsMoved = {"codegenprepare", "NumExtsMoved"
, "Number of [s|z]ext instructions combined with loads", {0},
false}
;
123STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized")static llvm::Statistic NumExtUses = {"codegenprepare", "NumExtUses"
, "Number of uses of [s|z]ext instructions optimized", {0}, false
}
;
124STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded"
, "Number of and mask instructions added to form ext loads", {
0}, false}
125 "Number of and mask instructions added to form ext loads")static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded"
, "Number of and mask instructions added to form ext loads", {
0}, false}
;
126STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized")static llvm::Statistic NumAndUses = {"codegenprepare", "NumAndUses"
, "Number of uses of and mask instructions optimized", {0}, false
}
;
127STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup"
, "Number of return instructions duplicated", {0}, false}
;
128STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved"
, "Number of debug value instructions moved", {0}, false}
;
129STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare"
, "NumSelectsExpanded", "Number of selects turned into branches"
, {0}, false}
;
130STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare"
, "NumStoreExtractExposed", "Number of store(extractelement) exposed"
, {0}, false}
;
131
132static cl::opt<bool> DisableBranchOpts(
133 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
134 cl::desc("Disable branch optimizations in CodeGenPrepare"));
135
136static cl::opt<bool>
137 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
138 cl::desc("Disable GC optimizations in CodeGenPrepare"));
139
140static cl::opt<bool> DisableSelectToBranch(
141 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
142 cl::desc("Disable select to branch conversion."));
143
144static cl::opt<bool> AddrSinkUsingGEPs(
145 "addr-sink-using-gep", cl::Hidden, cl::init(true),
146 cl::desc("Address sinking in CGP using GEPs."));
147
148static cl::opt<bool> EnableAndCmpSinking(
149 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
150 cl::desc("Enable sinkinig and/cmp into branches."));
151
152static cl::opt<bool> DisableStoreExtract(
153 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
154 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
155
156static cl::opt<bool> StressStoreExtract(
157 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
158 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
159
160static cl::opt<bool> DisableExtLdPromotion(
161 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
162 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
163 "CodeGenPrepare"));
164
165static cl::opt<bool> StressExtLdPromotion(
166 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
167 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
168 "optimization in CodeGenPrepare"));
169
170static cl::opt<bool> DisablePreheaderProtect(
171 "disable-preheader-prot", cl::Hidden, cl::init(false),
172 cl::desc("Disable protection against removing loop preheaders"));
173
174static cl::opt<bool> ProfileGuidedSectionPrefix(
175 "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
176 cl::desc("Use profile info to add section prefix for hot/cold functions"));
177
178static cl::opt<unsigned> FreqRatioToSkipMerge(
179 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
180 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
181 "(frequency of destination block) is greater than this ratio"));
182
183static cl::opt<bool> ForceSplitStore(
184 "force-split-store", cl::Hidden, cl::init(false),
185 cl::desc("Force store splitting no matter what the target query says."));
186
187static cl::opt<bool>
188EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
189 cl::desc("Enable merging of redundant sexts when one is dominating"
190 " the other."), cl::init(true));
191
192static cl::opt<bool> DisableComplexAddrModes(
193 "disable-complex-addr-modes", cl::Hidden, cl::init(true),
194 cl::desc("Disables combining addressing modes with different parts "
195 "in optimizeMemoryInst."));
196
197static cl::opt<bool>
198AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
199 cl::desc("Allow creation of Phis in Address sinking."));
200
201static cl::opt<bool>
202AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(false),
203 cl::desc("Allow creation of selects in Address sinking."));
204
205namespace {
206
207using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
208using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
209using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
210using SExts = SmallVector<Instruction *, 16>;
211using ValueToSExts = DenseMap<Value *, SExts>;
212
213class TypePromotionTransaction;
214
215 class CodeGenPrepare : public FunctionPass {
216 const TargetMachine *TM = nullptr;
217 const TargetSubtargetInfo *SubtargetInfo;
218 const TargetLowering *TLI = nullptr;
219 const TargetRegisterInfo *TRI;
220 const TargetTransformInfo *TTI = nullptr;
221 const TargetLibraryInfo *TLInfo;
222 const LoopInfo *LI;
223 std::unique_ptr<BlockFrequencyInfo> BFI;
224 std::unique_ptr<BranchProbabilityInfo> BPI;
225
226 /// As we scan instructions optimizing them, this is the next instruction
227 /// to optimize. Transforms that can invalidate this should update it.
228 BasicBlock::iterator CurInstIterator;
229
230 /// Keeps track of non-local addresses that have been sunk into a block.
231 /// This allows us to avoid inserting duplicate code for blocks with
232 /// multiple load/stores of the same address.
233 ValueMap<Value*, Value*> SunkAddrs;
234
235 /// Keeps track of all instructions inserted for the current function.
236 SetOfInstrs InsertedInsts;
237
238 /// Keeps track of the type of the related instruction before their
239 /// promotion for the current function.
240 InstrToOrigTy PromotedInsts;
241
242 /// Keep track of instructions removed during promotion.
243 SetOfInstrs RemovedInsts;
244
245 /// Keep track of sext chains based on their initial value.
246 DenseMap<Value *, Instruction *> SeenChainsForSExt;
247
248 /// Keep track of SExt promoted.
249 ValueToSExts ValToSExtendedUses;
250
251 /// True if CFG is modified in any way.
252 bool ModifiedDT;
253
254 /// True if optimizing for size.
255 bool OptSize;
256
257 /// DataLayout for the Function being processed.
258 const DataLayout *DL = nullptr;
259
260 public:
261 static char ID; // Pass identification, replacement for typeid
262
263 CodeGenPrepare() : FunctionPass(ID) {
264 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
265 }
266
267 bool runOnFunction(Function &F) override;
268
269 StringRef getPassName() const override { return "CodeGen Prepare"; }
270
271 void getAnalysisUsage(AnalysisUsage &AU) const override {
272 // FIXME: When we can selectively preserve passes, preserve the domtree.
273 AU.addRequired<ProfileSummaryInfoWrapperPass>();
274 AU.addRequired<TargetLibraryInfoWrapperPass>();
275 AU.addRequired<TargetTransformInfoWrapperPass>();
276 AU.addRequired<LoopInfoWrapperPass>();
277 }
278
279 private:
280 bool eliminateFallThrough(Function &F);
281 bool eliminateMostlyEmptyBlocks(Function &F);
282 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
283 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
284 void eliminateMostlyEmptyBlock(BasicBlock *BB);
285 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
286 bool isPreheader);
287 bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
288 bool optimizeInst(Instruction *I, bool &ModifiedDT);
289 bool optimizeMemoryInst(Instruction *I, Value *Addr,
290 Type *AccessTy, unsigned AS);
291 bool optimizeInlineAsmInst(CallInst *CS);
292 bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
293 bool optimizeExt(Instruction *&I);
294 bool optimizeExtUses(Instruction *I);
295 bool optimizeLoadExt(LoadInst *I);
296 bool optimizeSelectInst(SelectInst *SI);
297 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
298 bool optimizeSwitchInst(SwitchInst *CI);
299 bool optimizeExtractElementInst(Instruction *Inst);
300 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
301 bool placeDbgValues(Function &F);
302 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
303 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
304 bool tryToPromoteExts(TypePromotionTransaction &TPT,
305 const SmallVectorImpl<Instruction *> &Exts,
306 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
307 unsigned CreatedInstsCost = 0);
308 bool mergeSExts(Function &F);
309 bool performAddressTypePromotion(
310 Instruction *&Inst,
311 bool AllowPromotionWithoutCommonHeader,
312 bool HasPromoted, TypePromotionTransaction &TPT,
313 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
314 bool splitBranchCondition(Function &F);
315 bool simplifyOffsetableRelocate(Instruction &I);
316 bool splitIndirectCriticalEdges(Function &F);
317 };
318
319} // end anonymous namespace
320
321char CodeGenPrepare::ID = 0;
322
323INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
324 "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
325INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry);
326INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare"
, &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor
<CodeGenPrepare>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag
; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry
) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce
, std::ref(Registry)); }
327 "Optimize for code generation", false, false)PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare"
, &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor
<CodeGenPrepare>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag
; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry
) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce
, std::ref(Registry)); }
328
329FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
330
331bool CodeGenPrepare::runOnFunction(Function &F) {
332 if (skipFunction(F))
333 return false;
334
335 DL = &F.getParent()->getDataLayout();
336
337 bool EverMadeChange = false;
338 // Clear per function information.
339 InsertedInsts.clear();
340 PromotedInsts.clear();
341 BFI.reset();
342 BPI.reset();
343
344 ModifiedDT = false;
345 if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
346 TM = &TPC->getTM<TargetMachine>();
347 SubtargetInfo = TM->getSubtargetImpl(F);
348 TLI = SubtargetInfo->getTargetLowering();
349 TRI = SubtargetInfo->getRegisterInfo();
350 }
351 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
352 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
353 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
354 OptSize = F.optForSize();
355
356 ProfileSummaryInfo *PSI =
357 getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
358 if (ProfileGuidedSectionPrefix) {
359 if (PSI->isFunctionHotInCallGraph(&F))
360 F.setSectionPrefix(".hot");
361 else if (PSI->isFunctionColdInCallGraph(&F))
362 F.setSectionPrefix(".unlikely");
363 }
364
365 /// This optimization identifies DIV instructions that can be
366 /// profitably bypassed and carried out with a shorter, faster divide.
367 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
368 TLI->isSlowDivBypassed()) {
369 const DenseMap<unsigned int, unsigned int> &BypassWidths =
370 TLI->getBypassSlowDivWidths();
371 BasicBlock* BB = &*F.begin();
372 while (BB != nullptr) {
373 // bypassSlowDivision may create new BBs, but we don't want to reapply the
374 // optimization to those blocks.
375 BasicBlock* Next = BB->getNextNode();
376 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
377 BB = Next;
378 }
379 }
380
381 // Eliminate blocks that contain only PHI nodes and an
382 // unconditional branch.
383 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
384
385 // llvm.dbg.value is far away from the value then iSel may not be able
386 // handle it properly. iSel will drop llvm.dbg.value if it can not
387 // find a node corresponding to the value.
388 EverMadeChange |= placeDbgValues(F);
389
390 if (!DisableBranchOpts)
391 EverMadeChange |= splitBranchCondition(F);
392
393 // Split some critical edges where one of the sources is an indirect branch,
394 // to help generate sane code for PHIs involving such edges.
395 EverMadeChange |= splitIndirectCriticalEdges(F);
396
397 bool MadeChange = true;
398 while (MadeChange) {
399 MadeChange = false;
400 SeenChainsForSExt.clear();
401 ValToSExtendedUses.clear();
402 RemovedInsts.clear();
403 for (Function::iterator I = F.begin(); I != F.end(); ) {
404 BasicBlock *BB = &*I++;
405 bool ModifiedDTOnIteration = false;
406 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
407
408 // Restart BB iteration if the dominator tree of the Function was changed
409 if (ModifiedDTOnIteration)
410 break;
411 }
412 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
413 MadeChange |= mergeSExts(F);
414
415 // Really free removed instructions during promotion.
416 for (Instruction *I : RemovedInsts)
417 I->deleteValue();
418
419 EverMadeChange |= MadeChange;
420 }
421
422 SunkAddrs.clear();
423
424 if (!DisableBranchOpts) {
425 MadeChange = false;
426 SmallPtrSet<BasicBlock*, 8> WorkList;
427 for (BasicBlock &BB : F) {
428 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
429 MadeChange |= ConstantFoldTerminator(&BB, true);
430 if (!MadeChange) continue;
431
432 for (SmallVectorImpl<BasicBlock*>::iterator
433 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
434 if (pred_begin(*II) == pred_end(*II))
435 WorkList.insert(*II);
436 }
437
438 // Delete the dead blocks and any of their dead successors.
439 MadeChange |= !WorkList.empty();
440 while (!WorkList.empty()) {
441 BasicBlock *BB = *WorkList.begin();
442 WorkList.erase(BB);
443 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
444
445 DeleteDeadBlock(BB);
446
447 for (SmallVectorImpl<BasicBlock*>::iterator
448 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
449 if (pred_begin(*II) == pred_end(*II))
450 WorkList.insert(*II);
451 }
452
453 // Merge pairs of basic blocks with unconditional branches, connected by
454 // a single edge.
455 if (EverMadeChange || MadeChange)
456 MadeChange |= eliminateFallThrough(F);
457
458 EverMadeChange |= MadeChange;
459 }
460
461 if (!DisableGCOpts) {
462 SmallVector<Instruction *, 2> Statepoints;
463 for (BasicBlock &BB : F)
464 for (Instruction &I : BB)
465 if (isStatepoint(I))
466 Statepoints.push_back(&I);
467 for (auto &I : Statepoints)
468 EverMadeChange |= simplifyOffsetableRelocate(*I);
469 }
470
471 return EverMadeChange;
472}
473
474/// Merge basic blocks which are connected by a single edge, where one of the
475/// basic blocks has a single successor pointing to the other basic block,
476/// which has a single predecessor.
477bool CodeGenPrepare::eliminateFallThrough(Function &F) {
478 bool Changed = false;
479 // Scan all of the blocks in the function, except for the entry block.
480 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
481 BasicBlock *BB = &*I++;
482 // If the destination block has a single pred, then this is a trivial
483 // edge, just collapse it.
484 BasicBlock *SinglePred = BB->getSinglePredecessor();
485
486 // Don't merge if BB's address is taken.
487 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
488
489 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
490 if (Term && !Term->isConditional()) {
491 Changed = true;
492 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "To merge:\n"<< *
SinglePred << "\n\n\n"; } } while (false)
;
493 // Remember if SinglePred was the entry block of the function.
494 // If so, we will need to move BB back to the entry position.
495 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
496 MergeBasicBlockIntoOnlyPred(BB, nullptr);
497
498 if (isEntry && BB != &BB->getParent()->getEntryBlock())
499 BB->moveBefore(&BB->getParent()->getEntryBlock());
500
501 // We have erased a block. Update the iterator.
502 I = BB->getIterator();
503 }
504 }
505 return Changed;
506}
507
508/// Find a destination block from BB if BB is mergeable empty block.
509BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
510 // If this block doesn't end with an uncond branch, ignore it.
511 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
512 if (!BI || !BI->isUnconditional())
513 return nullptr;
514
515 // If the instruction before the branch (skipping debug info) isn't a phi
516 // node, then other stuff is happening here.
517 BasicBlock::iterator BBI = BI->getIterator();
518 if (BBI != BB->begin()) {
519 --BBI;
520 while (isa<DbgInfoIntrinsic>(BBI)) {
521 if (BBI == BB->begin())
522 break;
523 --BBI;
524 }
525 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
526 return nullptr;
527 }
528
529 // Do not break infinite loops.
530 BasicBlock *DestBB = BI->getSuccessor(0);
531 if (DestBB == BB)
532 return nullptr;
533
534 if (!canMergeBlocks(BB, DestBB))
535 DestBB = nullptr;
536
537 return DestBB;
538}
539
540// Return the unique indirectbr predecessor of a block. This may return null
541// even if such a predecessor exists, if it's not useful for splitting.
542// If a predecessor is found, OtherPreds will contain all other (non-indirectbr)
543// predecessors of BB.
544static BasicBlock *
545findIBRPredecessor(BasicBlock *BB, SmallVectorImpl<BasicBlock *> &OtherPreds) {
546 // If the block doesn't have any PHIs, we don't care about it, since there's
547 // no point in splitting it.
548 PHINode *PN = dyn_cast<PHINode>(BB->begin());
549 if (!PN)
550 return nullptr;
551
552 // Verify we have exactly one IBR predecessor.
553 // Conservatively bail out if one of the other predecessors is not a "regular"
554 // terminator (that is, not a switch or a br).
555 BasicBlock *IBB = nullptr;
556 for (unsigned Pred = 0, E = PN->getNumIncomingValues(); Pred != E; ++Pred) {
557 BasicBlock *PredBB = PN->getIncomingBlock(Pred);
558 TerminatorInst *PredTerm = PredBB->getTerminator();
559 switch (PredTerm->getOpcode()) {
560 case Instruction::IndirectBr:
561 if (IBB)
562 return nullptr;
563 IBB = PredBB;
564 break;
565 case Instruction::Br:
566 case Instruction::Switch:
567 OtherPreds.push_back(PredBB);
568 continue;
569 default:
570 return nullptr;
571 }
572 }
573
574 return IBB;
575}
576
577// Split critical edges where the source of the edge is an indirectbr
578// instruction. This isn't always possible, but we can handle some easy cases.
579// This is useful because MI is unable to split such critical edges,
580// which means it will not be able to sink instructions along those edges.
581// This is especially painful for indirect branches with many successors, where
582// we end up having to prepare all outgoing values in the origin block.
583//
584// Our normal algorithm for splitting critical edges requires us to update
585// the outgoing edges of the edge origin block, but for an indirectbr this
586// is hard, since it would require finding and updating the block addresses
587// the indirect branch uses. But if a block only has a single indirectbr
588// predecessor, with the others being regular branches, we can do it in a
589// different way.
590// Say we have A -> D, B -> D, I -> D where only I -> D is an indirectbr.
591// We can split D into D0 and D1, where D0 contains only the PHIs from D,
592// and D1 is the D block body. We can then duplicate D0 as D0A and D0B, and
593// create the following structure:
594// A -> D0A, B -> D0A, I -> D0B, D0A -> D1, D0B -> D1
595bool CodeGenPrepare::splitIndirectCriticalEdges(Function &F) {
596 // Check whether the function has any indirectbrs, and collect which blocks
597 // they may jump to. Since most functions don't have indirect branches,
598 // this lowers the common case's overhead to O(Blocks) instead of O(Edges).
599 SmallSetVector<BasicBlock *, 16> Targets;
600 for (auto &BB : F) {
601 auto *IBI = dyn_cast<IndirectBrInst>(BB.getTerminator());
602 if (!IBI)
603 continue;
604
605 for (unsigned Succ = 0, E = IBI->getNumSuccessors(); Succ != E; ++Succ)
606 Targets.insert(IBI->getSuccessor(Succ));
607 }
608
609 if (Targets.empty())
610 return false;
611
612 bool Changed = false;
613 for (BasicBlock *Target : Targets) {
614 SmallVector<BasicBlock *, 16> OtherPreds;
615 BasicBlock *IBRPred = findIBRPredecessor(Target, OtherPreds);
616 // If we did not found an indirectbr, or the indirectbr is the only
617 // incoming edge, this isn't the kind of edge we're looking for.
618 if (!IBRPred || OtherPreds.empty())
619 continue;
620
621 // Don't even think about ehpads/landingpads.
622 Instruction *FirstNonPHI = Target->getFirstNonPHI();
623 if (FirstNonPHI->isEHPad() || Target->isLandingPad())
624 continue;
625
626 BasicBlock *BodyBlock = Target->splitBasicBlock(FirstNonPHI, ".split");
627 // It's possible Target was its own successor through an indirectbr.
628 // In this case, the indirectbr now comes from BodyBlock.
629 if (IBRPred == Target)
630 IBRPred = BodyBlock;
631
632 // At this point Target only has PHIs, and BodyBlock has the rest of the
633 // block's body. Create a copy of Target that will be used by the "direct"
634 // preds.
635 ValueToValueMapTy VMap;
636 BasicBlock *DirectSucc = CloneBasicBlock(Target, VMap, ".clone", &F);
637
638 for (BasicBlock *Pred : OtherPreds) {
639 // If the target is a loop to itself, then the terminator of the split
640 // block needs to be updated.
641 if (Pred == Target)
642 BodyBlock->getTerminator()->replaceUsesOfWith(Target, DirectSucc);
643 else
644 Pred->getTerminator()->replaceUsesOfWith(Target, DirectSucc);
645 }
646
647 // Ok, now fix up the PHIs. We know the two blocks only have PHIs, and that
648 // they are clones, so the number of PHIs are the same.
649 // (a) Remove the edge coming from IBRPred from the "Direct" PHI
650 // (b) Leave that as the only edge in the "Indirect" PHI.
651 // (c) Merge the two in the body block.
652 BasicBlock::iterator Indirect = Target->begin(),
653 End = Target->getFirstNonPHI()->getIterator();
654 BasicBlock::iterator Direct = DirectSucc->begin();
655 BasicBlock::iterator MergeInsert = BodyBlock->getFirstInsertionPt();
656
657 assert(&*End == Target->getTerminator() &&(static_cast <bool> (&*End == Target->getTerminator
() && "Block was expected to only contain PHIs") ? void
(0) : __assert_fail ("&*End == Target->getTerminator() && \"Block was expected to only contain PHIs\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 658, __extension__ __PRETTY_FUNCTION__))
658 "Block was expected to only contain PHIs")(static_cast <bool> (&*End == Target->getTerminator
() && "Block was expected to only contain PHIs") ? void
(0) : __assert_fail ("&*End == Target->getTerminator() && \"Block was expected to only contain PHIs\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 658, __extension__ __PRETTY_FUNCTION__))
;
659
660 while (Indirect != End) {
661 PHINode *DirPHI = cast<PHINode>(Direct);
662 PHINode *IndPHI = cast<PHINode>(Indirect);
663
664 // Now, clean up - the direct block shouldn't get the indirect value,
665 // and vice versa.
666 DirPHI->removeIncomingValue(IBRPred);
667 Direct++;
668
669 // Advance the pointer here, to avoid invalidation issues when the old
670 // PHI is erased.
671 Indirect++;
672
673 PHINode *NewIndPHI = PHINode::Create(IndPHI->getType(), 1, "ind", IndPHI);
674 NewIndPHI->addIncoming(IndPHI->getIncomingValueForBlock(IBRPred),
675 IBRPred);
676
677 // Create a PHI in the body block, to merge the direct and indirect
678 // predecessors.
679 PHINode *MergePHI =
680 PHINode::Create(IndPHI->getType(), 2, "merge", &*MergeInsert);
681 MergePHI->addIncoming(NewIndPHI, Target);
682 MergePHI->addIncoming(DirPHI, DirectSucc);
683
684 IndPHI->replaceAllUsesWith(MergePHI);
685 IndPHI->eraseFromParent();
686 }
687
688 Changed = true;
689 }
690
691 return Changed;
692}
693
694/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
695/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
696/// edges in ways that are non-optimal for isel. Start by eliminating these
697/// blocks so we can split them the way we want them.
698bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
699 SmallPtrSet<BasicBlock *, 16> Preheaders;
700 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
701 while (!LoopList.empty()) {
702 Loop *L = LoopList.pop_back_val();
703 LoopList.insert(LoopList.end(), L->begin(), L->end());
704 if (BasicBlock *Preheader = L->getLoopPreheader())
705 Preheaders.insert(Preheader);
706 }
707
708 bool MadeChange = false;
709 // Note that this intentionally skips the entry block.
710 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
711 BasicBlock *BB = &*I++;
712 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
713 if (!DestBB ||
714 !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
715 continue;
716
717 eliminateMostlyEmptyBlock(BB);
718 MadeChange = true;
719 }
720 return MadeChange;
721}
722
723bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
724 BasicBlock *DestBB,
725 bool isPreheader) {
726 // Do not delete loop preheaders if doing so would create a critical edge.
727 // Loop preheaders can be good locations to spill registers. If the
728 // preheader is deleted and we create a critical edge, registers may be
729 // spilled in the loop body instead.
730 if (!DisablePreheaderProtect && isPreheader &&
731 !(BB->getSinglePredecessor() &&
732 BB->getSinglePredecessor()->getSingleSuccessor()))
733 return false;
734
735 // Try to skip merging if the unique predecessor of BB is terminated by a
736 // switch or indirect branch instruction, and BB is used as an incoming block
737 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
738 // add COPY instructions in the predecessor of BB instead of BB (if it is not
739 // merged). Note that the critical edge created by merging such blocks wont be
740 // split in MachineSink because the jump table is not analyzable. By keeping
741 // such empty block (BB), ISel will place COPY instructions in BB, not in the
742 // predecessor of BB.
743 BasicBlock *Pred = BB->getUniquePredecessor();
744 if (!Pred ||
745 !(isa<SwitchInst>(Pred->getTerminator()) ||
746 isa<IndirectBrInst>(Pred->getTerminator())))
747 return true;
748
749 if (BB->getTerminator() != BB->getFirstNonPHI())
750 return true;
751
752 // We use a simple cost heuristic which determine skipping merging is
753 // profitable if the cost of skipping merging is less than the cost of
754 // merging : Cost(skipping merging) < Cost(merging BB), where the
755 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
756 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
757 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
758 // Freq(Pred) / Freq(BB) > 2.
759 // Note that if there are multiple empty blocks sharing the same incoming
760 // value for the PHIs in the DestBB, we consider them together. In such
761 // case, Cost(merging BB) will be the sum of their frequencies.
762
763 if (!isa<PHINode>(DestBB->begin()))
764 return true;
765
766 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
767
768 // Find all other incoming blocks from which incoming values of all PHIs in
769 // DestBB are the same as the ones from BB.
770 for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
771 ++PI) {
772 BasicBlock *DestBBPred = *PI;
773 if (DestBBPred == BB)
774 continue;
775
776 bool HasAllSameValue = true;
777 BasicBlock::const_iterator DestBBI = DestBB->begin();
778 while (const PHINode *DestPN = dyn_cast<PHINode>(DestBBI++)) {
779 if (DestPN->getIncomingValueForBlock(BB) !=
780 DestPN->getIncomingValueForBlock(DestBBPred)) {
781 HasAllSameValue = false;
782 break;
783 }
784 }
785 if (HasAllSameValue)
786 SameIncomingValueBBs.insert(DestBBPred);
787 }
788
789 // See if all BB's incoming values are same as the value from Pred. In this
790 // case, no reason to skip merging because COPYs are expected to be place in
791 // Pred already.
792 if (SameIncomingValueBBs.count(Pred))
793 return true;
794
795 if (!BFI) {
796 Function &F = *BB->getParent();
797 LoopInfo LI{DominatorTree(F)};
798 BPI.reset(new BranchProbabilityInfo(F, LI));
799 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
800 }
801
802 BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
803 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
804
805 for (auto SameValueBB : SameIncomingValueBBs)
806 if (SameValueBB->getUniquePredecessor() == Pred &&
807 DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
808 BBFreq += BFI->getBlockFreq(SameValueBB);
809
810 return PredFreq.getFrequency() <=
811 BBFreq.getFrequency() * FreqRatioToSkipMerge;
812}
813
814/// Return true if we can merge BB into DestBB if there is a single
815/// unconditional branch between them, and BB contains no other non-phi
816/// instructions.
817bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
818 const BasicBlock *DestBB) const {
819 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
820 // the successor. If there are more complex condition (e.g. preheaders),
821 // don't mess around with them.
822 BasicBlock::const_iterator BBI = BB->begin();
823 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
824 for (const User *U : PN->users()) {
825 const Instruction *UI = cast<Instruction>(U);
826 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
827 return false;
828 // If User is inside DestBB block and it is a PHINode then check
829 // incoming value. If incoming value is not from BB then this is
830 // a complex condition (e.g. preheaders) we want to avoid here.
831 if (UI->getParent() == DestBB) {
832 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
833 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
834 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
835 if (Insn && Insn->getParent() == BB &&
836 Insn->getParent() != UPN->getIncomingBlock(I))
837 return false;
838 }
839 }
840 }
841 }
842
843 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
844 // and DestBB may have conflicting incoming values for the block. If so, we
845 // can't merge the block.
846 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
847 if (!DestBBPN) return true; // no conflict.
848
849 // Collect the preds of BB.
850 SmallPtrSet<const BasicBlock*, 16> BBPreds;
851 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
852 // It is faster to get preds from a PHI than with pred_iterator.
853 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
854 BBPreds.insert(BBPN->getIncomingBlock(i));
855 } else {
856 BBPreds.insert(pred_begin(BB), pred_end(BB));
857 }
858
859 // Walk the preds of DestBB.
860 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
861 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
862 if (BBPreds.count(Pred)) { // Common predecessor?
863 BBI = DestBB->begin();
864 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
865 const Value *V1 = PN->getIncomingValueForBlock(Pred);
866 const Value *V2 = PN->getIncomingValueForBlock(BB);
867
868 // If V2 is a phi node in BB, look up what the mapped value will be.
869 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
870 if (V2PN->getParent() == BB)
871 V2 = V2PN->getIncomingValueForBlock(Pred);
872
873 // If there is a conflict, bail out.
874 if (V1 != V2) return false;
875 }
876 }
877 }
878
879 return true;
880}
881
882/// Eliminate a basic block that has only phi's and an unconditional branch in
883/// it.
884void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
885 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
886 BasicBlock *DestBB = BI->getSuccessor(0);
887
888 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
<< *BB << *DestBB; } } while (false)
;
889
890 // If the destination block has a single pred, then this is a trivial edge,
891 // just collapse it.
892 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
893 if (SinglePred != DestBB) {
894 // Remember if SinglePred was the entry block of the function. If so, we
895 // will need to move BB back to the entry position.
896 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
897 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
898
899 if (isEntry && BB != &BB->getParent()->getEntryBlock())
900 BB->moveBefore(&BB->getParent()->getEntryBlock());
901
902 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB
<< "\n\n\n"; } } while (false)
;
903 return;
904 }
905 }
906
907 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
908 // to handle the new incoming edges it is about to have.
909 PHINode *PN;
910 for (BasicBlock::iterator BBI = DestBB->begin();
911 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
912 // Remove the incoming value for BB, and remember it.
913 Value *InVal = PN->removeIncomingValue(BB, false);
914
915 // Two options: either the InVal is a phi node defined in BB or it is some
916 // value that dominates BB.
917 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
918 if (InValPhi && InValPhi->getParent() == BB) {
919 // Add all of the input values of the input PHI as inputs of this phi.
920 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
921 PN->addIncoming(InValPhi->getIncomingValue(i),
922 InValPhi->getIncomingBlock(i));
923 } else {
924 // Otherwise, add one instance of the dominating value for each edge that
925 // we will be adding.
926 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
927 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
928 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
929 } else {
930 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
931 PN->addIncoming(InVal, *PI);
932 }
933 }
934 }
935
936 // The PHIs are now updated, change everything that refers to BB to use
937 // DestBB and remove BB.
938 BB->replaceAllUsesWith(DestBB);
939 BB->eraseFromParent();
940 ++NumBlocksElim;
941
942 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB
<< "\n\n\n"; } } while (false)
;
943}
944
945// Computes a map of base pointer relocation instructions to corresponding
946// derived pointer relocation instructions given a vector of all relocate calls
947static void computeBaseDerivedRelocateMap(
948 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
949 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
950 &RelocateInstMap) {
951 // Collect information in two maps: one primarily for locating the base object
952 // while filling the second map; the second map is the final structure holding
953 // a mapping between Base and corresponding Derived relocate calls
954 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
955 for (auto *ThisRelocate : AllRelocateCalls) {
956 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
957 ThisRelocate->getDerivedPtrIndex());
958 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
959 }
960 for (auto &Item : RelocateIdxMap) {
961 std::pair<unsigned, unsigned> Key = Item.first;
962 if (Key.first == Key.second)
963 // Base relocation: nothing to insert
964 continue;
965
966 GCRelocateInst *I = Item.second;
967 auto BaseKey = std::make_pair(Key.first, Key.first);
968
969 // We're iterating over RelocateIdxMap so we cannot modify it.
970 auto MaybeBase = RelocateIdxMap.find(BaseKey);
971 if (MaybeBase == RelocateIdxMap.end())
972 // TODO: We might want to insert a new base object relocate and gep off
973 // that, if there are enough derived object relocates.
974 continue;
975
976 RelocateInstMap[MaybeBase->second].push_back(I);
977 }
978}
979
980// Accepts a GEP and extracts the operands into a vector provided they're all
981// small integer constants
982static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
983 SmallVectorImpl<Value *> &OffsetV) {
984 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
985 // Only accept small constant integer operands
986 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
987 if (!Op || Op->getZExtValue() > 20)
988 return false;
989 }
990
991 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
992 OffsetV.push_back(GEP->getOperand(i));
993 return true;
994}
995
996// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
997// replace, computes a replacement, and affects it.
998static bool
999simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1000 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1001 bool MadeChange = false;
1002 // We must ensure the relocation of derived pointer is defined after
1003 // relocation of base pointer. If we find a relocation corresponding to base
1004 // defined earlier than relocation of base then we move relocation of base
1005 // right before found relocation. We consider only relocation in the same
1006 // basic block as relocation of base. Relocations from other basic block will
1007 // be skipped by optimization and we do not care about them.
1008 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1009 &*R != RelocatedBase; ++R)
1010 if (auto RI = dyn_cast<GCRelocateInst>(R))
1011 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1012 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1013 RelocatedBase->moveBefore(RI);
1014 break;
1015 }
1016
1017 for (GCRelocateInst *ToReplace : Targets) {
1018 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&(static_cast <bool> (ToReplace->getBasePtrIndex() ==
RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object"
) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1019, __extension__ __PRETTY_FUNCTION__))
1019 "Not relocating a derived object of the original base object")(static_cast <bool> (ToReplace->getBasePtrIndex() ==
RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object"
) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1019, __extension__ __PRETTY_FUNCTION__))
;
1020 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1021 // A duplicate relocate call. TODO: coalesce duplicates.
1022 continue;
1023 }
1024
1025 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1026 // Base and derived relocates are in different basic blocks.
1027 // In this case transform is only valid when base dominates derived
1028 // relocate. However it would be too expensive to check dominance
1029 // for each such relocate, so we skip the whole transformation.
1030 continue;
1031 }
1032
1033 Value *Base = ToReplace->getBasePtr();
1034 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1035 if (!Derived || Derived->getPointerOperand() != Base)
1036 continue;
1037
1038 SmallVector<Value *, 2> OffsetV;
1039 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1040 continue;
1041
1042 // Create a Builder and replace the target callsite with a gep
1043 assert(RelocatedBase->getNextNode() &&(static_cast <bool> (RelocatedBase->getNextNode() &&
"Should always have one since it's not a terminator") ? void
(0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1044, __extension__ __PRETTY_FUNCTION__))
1044 "Should always have one since it's not a terminator")(static_cast <bool> (RelocatedBase->getNextNode() &&
"Should always have one since it's not a terminator") ? void
(0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1044, __extension__ __PRETTY_FUNCTION__))
;
1045
1046 // Insert after RelocatedBase
1047 IRBuilder<> Builder(RelocatedBase->getNextNode());
1048 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1049
1050 // If gc_relocate does not match the actual type, cast it to the right type.
1051 // In theory, there must be a bitcast after gc_relocate if the type does not
1052 // match, and we should reuse it to get the derived pointer. But it could be
1053 // cases like this:
1054 // bb1:
1055 // ...
1056 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1057 // br label %merge
1058 //
1059 // bb2:
1060 // ...
1061 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1062 // br label %merge
1063 //
1064 // merge:
1065 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1066 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1067 //
1068 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1069 // no matter there is already one or not. In this way, we can handle all cases, and
1070 // the extra bitcast should be optimized away in later passes.
1071 Value *ActualRelocatedBase = RelocatedBase;
1072 if (RelocatedBase->getType() != Base->getType()) {
1073 ActualRelocatedBase =
1074 Builder.CreateBitCast(RelocatedBase, Base->getType());
1075 }
1076 Value *Replacement = Builder.CreateGEP(
1077 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1078 Replacement->takeName(ToReplace);
1079 // If the newly generated derived pointer's type does not match the original derived
1080 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1081 Value *ActualReplacement = Replacement;
1082 if (Replacement->getType() != ToReplace->getType()) {
1083 ActualReplacement =
1084 Builder.CreateBitCast(Replacement, ToReplace->getType());
1085 }
1086 ToReplace->replaceAllUsesWith(ActualReplacement);
1087 ToReplace->eraseFromParent();
1088
1089 MadeChange = true;
1090 }
1091 return MadeChange;
1092}
1093
1094// Turns this:
1095//
1096// %base = ...
1097// %ptr = gep %base + 15
1098// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1099// %base' = relocate(%tok, i32 4, i32 4)
1100// %ptr' = relocate(%tok, i32 4, i32 5)
1101// %val = load %ptr'
1102//
1103// into this:
1104//
1105// %base = ...
1106// %ptr = gep %base + 15
1107// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1108// %base' = gc.relocate(%tok, i32 4, i32 4)
1109// %ptr' = gep %base' + 15
1110// %val = load %ptr'
1111bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1112 bool MadeChange = false;
1113 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1114
1115 for (auto *U : I.users())
1116 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1117 // Collect all the relocate calls associated with a statepoint
1118 AllRelocateCalls.push_back(Relocate);
1119
1120 // We need atleast one base pointer relocation + one derived pointer
1121 // relocation to mangle
1122 if (AllRelocateCalls.size() < 2)
1123 return false;
1124
1125 // RelocateInstMap is a mapping from the base relocate instruction to the
1126 // corresponding derived relocate instructions
1127 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1128 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1129 if (RelocateInstMap.empty())
1130 return false;
1131
1132 for (auto &Item : RelocateInstMap)
1133 // Item.first is the RelocatedBase to offset against
1134 // Item.second is the vector of Targets to replace
1135 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1136 return MadeChange;
1137}
1138
1139/// SinkCast - Sink the specified cast instruction into its user blocks
1140static bool SinkCast(CastInst *CI) {
1141 BasicBlock *DefBB = CI->getParent();
1142
1143 /// InsertedCasts - Only insert a cast in each block once.
1144 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1145
1146 bool MadeChange = false;
1147 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1148 UI != E; ) {
1149 Use &TheUse = UI.getUse();
1150 Instruction *User = cast<Instruction>(*UI);
1151
1152 // Figure out which BB this cast is used in. For PHI's this is the
1153 // appropriate predecessor block.
1154 BasicBlock *UserBB = User->getParent();
1155 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1156 UserBB = PN->getIncomingBlock(TheUse);
1157 }
1158
1159 // Preincrement use iterator so we don't invalidate it.
1160 ++UI;
1161
1162 // The first insertion point of a block containing an EH pad is after the
1163 // pad. If the pad is the user, we cannot sink the cast past the pad.
1164 if (User->isEHPad())
1165 continue;
1166
1167 // If the block selected to receive the cast is an EH pad that does not
1168 // allow non-PHI instructions before the terminator, we can't sink the
1169 // cast.
1170 if (UserBB->getTerminator()->isEHPad())
1171 continue;
1172
1173 // If this user is in the same block as the cast, don't change the cast.
1174 if (UserBB == DefBB) continue;
1175
1176 // If we have already inserted a cast into this block, use it.
1177 CastInst *&InsertedCast = InsertedCasts[UserBB];
1178
1179 if (!InsertedCast) {
1180 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1181 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1181, __extension__ __PRETTY_FUNCTION__))
;
1182 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1183 CI->getType(), "", &*InsertPt);
1184 }
1185
1186 // Replace a use of the cast with a use of the new cast.
1187 TheUse = InsertedCast;
1188 MadeChange = true;
1189 ++NumCastUses;
1190 }
1191
1192 // If we removed all uses, nuke the cast.
1193 if (CI->use_empty()) {
1194 salvageDebugInfo(*CI);
1195 CI->eraseFromParent();
1196 MadeChange = true;
1197 }
1198
1199 return MadeChange;
1200}
1201
1202/// If the specified cast instruction is a noop copy (e.g. it's casting from
1203/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1204/// reduce the number of virtual registers that must be created and coalesced.
1205///
1206/// Return true if any changes are made.
1207static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1208 const DataLayout &DL) {
1209 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1210 // than sinking only nop casts, but is helpful on some platforms.
1211 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1212 if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1213 ASC->getDestAddressSpace()))
1214 return false;
1215 }
1216
1217 // If this is a noop copy,
1218 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1219 EVT DstVT = TLI.getValueType(DL, CI->getType());
1220
1221 // This is an fp<->int conversion?
1222 if (SrcVT.isInteger() != DstVT.isInteger())
1223 return false;
1224
1225 // If this is an extension, it will be a zero or sign extension, which
1226 // isn't a noop.
1227 if (SrcVT.bitsLT(DstVT)) return false;
1228
1229 // If these values will be promoted, find out what they will be promoted
1230 // to. This helps us consider truncates on PPC as noop copies when they
1231 // are.
1232 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1233 TargetLowering::TypePromoteInteger)
1234 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1235 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1236 TargetLowering::TypePromoteInteger)
1237 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1238
1239 // If, after promotion, these are the same types, this is a noop copy.
1240 if (SrcVT != DstVT)
1241 return false;
1242
1243 return SinkCast(CI);
1244}
1245
1246/// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1247/// possible.
1248///
1249/// Return true if any changes were made.
1250static bool CombineUAddWithOverflow(CmpInst *CI) {
1251 Value *A, *B;
1252 Instruction *AddI;
1253 if (!match(CI,
1254 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1255 return false;
1256
1257 Type *Ty = AddI->getType();
1258 if (!isa<IntegerType>(Ty))
1259 return false;
1260
1261 // We don't want to move around uses of condition values this late, so we we
1262 // check if it is legal to create the call to the intrinsic in the basic
1263 // block containing the icmp:
1264
1265 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1266 return false;
1267
1268#ifndef NDEBUG
1269 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1270 // for now:
1271 if (AddI->hasOneUse())
1272 assert(*AddI->user_begin() == CI && "expected!")(static_cast <bool> (*AddI->user_begin() == CI &&
"expected!") ? void (0) : __assert_fail ("*AddI->user_begin() == CI && \"expected!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1272, __extension__ __PRETTY_FUNCTION__))
;
1273#endif
1274
1275 Module *M = CI->getModule();
1276 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1277
1278 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1279
1280 auto *UAddWithOverflow =
1281 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1282 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1283 auto *Overflow =
1284 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1285
1286 CI->replaceAllUsesWith(Overflow);
1287 AddI->replaceAllUsesWith(UAdd);
1288 CI->eraseFromParent();
1289 AddI->eraseFromParent();
1290 return true;
1291}
1292
1293/// Sink the given CmpInst into user blocks to reduce the number of virtual
1294/// registers that must be created and coalesced. This is a clear win except on
1295/// targets with multiple condition code registers (PowerPC), where it might
1296/// lose; some adjustment may be wanted there.
1297///
1298/// Return true if any changes are made.
1299static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1300 BasicBlock *DefBB = CI->getParent();
1301
1302 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1303 if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1304 return false;
1305
1306 // Only insert a cmp in each block once.
1307 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1308
1309 bool MadeChange = false;
1310 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1311 UI != E; ) {
1312 Use &TheUse = UI.getUse();
1313 Instruction *User = cast<Instruction>(*UI);
1314
1315 // Preincrement use iterator so we don't invalidate it.
1316 ++UI;
1317
1318 // Don't bother for PHI nodes.
1319 if (isa<PHINode>(User))
1320 continue;
1321
1322 // Figure out which BB this cmp is used in.
1323 BasicBlock *UserBB = User->getParent();
1324
1325 // If this user is in the same block as the cmp, don't change the cmp.
1326 if (UserBB == DefBB) continue;
1327
1328 // If we have already inserted a cmp into this block, use it.
1329 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1330
1331 if (!InsertedCmp) {
1332 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1333 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1333, __extension__ __PRETTY_FUNCTION__))
;
1334 InsertedCmp =
1335 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1336 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1337 // Propagate the debug info.
1338 InsertedCmp->setDebugLoc(CI->getDebugLoc());
1339 }
1340
1341 // Replace a use of the cmp with a use of the new cmp.
1342 TheUse = InsertedCmp;
1343 MadeChange = true;
1344 ++NumCmpUses;
1345 }
1346
1347 // If we removed all uses, nuke the cmp.
1348 if (CI->use_empty()) {
1349 CI->eraseFromParent();
1350 MadeChange = true;
1351 }
1352
1353 return MadeChange;
1354}
1355
1356static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1357 if (SinkCmpExpression(CI, TLI))
1358 return true;
1359
1360 if (CombineUAddWithOverflow(CI))
1361 return true;
1362
1363 return false;
1364}
1365
1366/// Duplicate and sink the given 'and' instruction into user blocks where it is
1367/// used in a compare to allow isel to generate better code for targets where
1368/// this operation can be combined.
1369///
1370/// Return true if any changes are made.
1371static bool sinkAndCmp0Expression(Instruction *AndI,
1372 const TargetLowering &TLI,
1373 SetOfInstrs &InsertedInsts) {
1374 // Double-check that we're not trying to optimize an instruction that was
1375 // already optimized by some other part of this pass.
1376 assert(!InsertedInsts.count(AndI) &&(static_cast <bool> (!InsertedInsts.count(AndI) &&
"Attempting to optimize already optimized and instruction") ?
void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1377, __extension__ __PRETTY_FUNCTION__))
1377 "Attempting to optimize already optimized and instruction")(static_cast <bool> (!InsertedInsts.count(AndI) &&
"Attempting to optimize already optimized and instruction") ?
void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1377, __extension__ __PRETTY_FUNCTION__))
;
1378 (void) InsertedInsts;
1379
1380 // Nothing to do for single use in same basic block.
1381 if (AndI->hasOneUse() &&
1382 AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1383 return false;
1384
1385 // Try to avoid cases where sinking/duplicating is likely to increase register
1386 // pressure.
1387 if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1388 !isa<ConstantInt>(AndI->getOperand(1)) &&
1389 AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1390 return false;
1391
1392 for (auto *U : AndI->users()) {
1393 Instruction *User = cast<Instruction>(U);
1394
1395 // Only sink for and mask feeding icmp with 0.
1396 if (!isa<ICmpInst>(User))
1397 return false;
1398
1399 auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1400 if (!CmpC || !CmpC->isZero())
1401 return false;
1402 }
1403
1404 if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1405 return false;
1406
1407 DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "found 'and' feeding only icmp 0;\n"
; } } while (false)
;
1408 DEBUG(AndI->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { AndI->getParent()->dump(); } } while
(false)
;
1409
1410 // Push the 'and' into the same block as the icmp 0. There should only be
1411 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1412 // others, so we don't need to keep track of which BBs we insert into.
1413 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1414 UI != E; ) {
1415 Use &TheUse = UI.getUse();
1416 Instruction *User = cast<Instruction>(*UI);
1417
1418 // Preincrement use iterator so we don't invalidate it.
1419 ++UI;
1420
1421 DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "sinking 'and' use: " <<
*User << "\n"; } } while (false)
;
1422
1423 // Keep the 'and' in the same place if the use is already in the same block.
1424 Instruction *InsertPt =
1425 User->getParent() == AndI->getParent() ? AndI : User;
1426 Instruction *InsertedAnd =
1427 BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1428 AndI->getOperand(1), "", InsertPt);
1429 // Propagate the debug info.
1430 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1431
1432 // Replace a use of the 'and' with a use of the new 'and'.
1433 TheUse = InsertedAnd;
1434 ++NumAndUses;
1435 DEBUG(User->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { User->getParent()->dump(); } } while
(false)
;
1436 }
1437
1438 // We removed all uses, nuke the and.
1439 AndI->eraseFromParent();
1440 return true;
1441}
1442
1443/// Check if the candidates could be combined with a shift instruction, which
1444/// includes:
1445/// 1. Truncate instruction
1446/// 2. And instruction and the imm is a mask of the low bits:
1447/// imm & (imm+1) == 0
1448static bool isExtractBitsCandidateUse(Instruction *User) {
1449 if (!isa<TruncInst>(User)) {
1450 if (User->getOpcode() != Instruction::And ||
1451 !isa<ConstantInt>(User->getOperand(1)))
1452 return false;
1453
1454 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1455
1456 if ((Cimm & (Cimm + 1)).getBoolValue())
1457 return false;
1458 }
1459 return true;
1460}
1461
1462/// Sink both shift and truncate instruction to the use of truncate's BB.
1463static bool
1464SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1465 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1466 const TargetLowering &TLI, const DataLayout &DL) {
1467 BasicBlock *UserBB = User->getParent();
1468 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1469 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1470 bool MadeChange = false;
1471
1472 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1473 TruncE = TruncI->user_end();
1474 TruncUI != TruncE;) {
1475
1476 Use &TruncTheUse = TruncUI.getUse();
1477 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1478 // Preincrement use iterator so we don't invalidate it.
1479
1480 ++TruncUI;
1481
1482 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1483 if (!ISDOpcode)
1484 continue;
1485
1486 // If the use is actually a legal node, there will not be an
1487 // implicit truncate.
1488 // FIXME: always querying the result type is just an
1489 // approximation; some nodes' legality is determined by the
1490 // operand or other means. There's no good way to find out though.
1491 if (TLI.isOperationLegalOrCustom(
1492 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1493 continue;
1494
1495 // Don't bother for PHI nodes.
1496 if (isa<PHINode>(TruncUser))
1497 continue;
1498
1499 BasicBlock *TruncUserBB = TruncUser->getParent();
1500
1501 if (UserBB == TruncUserBB)
1502 continue;
1503
1504 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1505 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1506
1507 if (!InsertedShift && !InsertedTrunc) {
1508 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1509 assert(InsertPt != TruncUserBB->end())(static_cast <bool> (InsertPt != TruncUserBB->end())
? void (0) : __assert_fail ("InsertPt != TruncUserBB->end()"
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1509, __extension__ __PRETTY_FUNCTION__))
;
1510 // Sink the shift
1511 if (ShiftI->getOpcode() == Instruction::AShr)
1512 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1513 "", &*InsertPt);
1514 else
1515 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1516 "", &*InsertPt);
1517
1518 // Sink the trunc
1519 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1520 TruncInsertPt++;
1521 assert(TruncInsertPt != TruncUserBB->end())(static_cast <bool> (TruncInsertPt != TruncUserBB->end
()) ? void (0) : __assert_fail ("TruncInsertPt != TruncUserBB->end()"
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1521, __extension__ __PRETTY_FUNCTION__))
;
1522
1523 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1524 TruncI->getType(), "", &*TruncInsertPt);
1525
1526 MadeChange = true;
1527
1528 TruncTheUse = InsertedTrunc;
1529 }
1530 }
1531 return MadeChange;
1532}
1533
1534/// Sink the shift *right* instruction into user blocks if the uses could
1535/// potentially be combined with this shift instruction and generate BitExtract
1536/// instruction. It will only be applied if the architecture supports BitExtract
1537/// instruction. Here is an example:
1538/// BB1:
1539/// %x.extract.shift = lshr i64 %arg1, 32
1540/// BB2:
1541/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1542/// ==>
1543///
1544/// BB2:
1545/// %x.extract.shift.1 = lshr i64 %arg1, 32
1546/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1547///
1548/// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1549/// instruction.
1550/// Return true if any changes are made.
1551static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1552 const TargetLowering &TLI,
1553 const DataLayout &DL) {
1554 BasicBlock *DefBB = ShiftI->getParent();
1555
1556 /// Only insert instructions in each block once.
1557 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1558
1559 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1560
1561 bool MadeChange = false;
1562 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1563 UI != E;) {
1564 Use &TheUse = UI.getUse();
1565 Instruction *User = cast<Instruction>(*UI);
1566 // Preincrement use iterator so we don't invalidate it.
1567 ++UI;
1568
1569 // Don't bother for PHI nodes.
1570 if (isa<PHINode>(User))
1571 continue;
1572
1573 if (!isExtractBitsCandidateUse(User))
1574 continue;
1575
1576 BasicBlock *UserBB = User->getParent();
1577
1578 if (UserBB == DefBB) {
1579 // If the shift and truncate instruction are in the same BB. The use of
1580 // the truncate(TruncUse) may still introduce another truncate if not
1581 // legal. In this case, we would like to sink both shift and truncate
1582 // instruction to the BB of TruncUse.
1583 // for example:
1584 // BB1:
1585 // i64 shift.result = lshr i64 opnd, imm
1586 // trunc.result = trunc shift.result to i16
1587 //
1588 // BB2:
1589 // ----> We will have an implicit truncate here if the architecture does
1590 // not have i16 compare.
1591 // cmp i16 trunc.result, opnd2
1592 //
1593 if (isa<TruncInst>(User) && shiftIsLegal
1594 // If the type of the truncate is legal, no trucate will be
1595 // introduced in other basic blocks.
1596 &&
1597 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1598 MadeChange =
1599 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1600
1601 continue;
1602 }
1603 // If we have already inserted a shift into this block, use it.
1604 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1605
1606 if (!InsertedShift) {
1607 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1608 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 1608, __extension__ __PRETTY_FUNCTION__))
;
1609
1610 if (ShiftI->getOpcode() == Instruction::AShr)
1611 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1612 "", &*InsertPt);
1613 else
1614 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1615 "", &*InsertPt);
1616
1617 MadeChange = true;
1618 }
1619
1620 // Replace a use of the shift with a use of the new shift.
1621 TheUse = InsertedShift;
1622 }
1623
1624 // If we removed all uses, nuke the shift.
1625 if (ShiftI->use_empty())
1626 ShiftI->eraseFromParent();
1627
1628 return MadeChange;
1629}
1630
1631/// If counting leading or trailing zeros is an expensive operation and a zero
1632/// input is defined, add a check for zero to avoid calling the intrinsic.
1633///
1634/// We want to transform:
1635/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1636///
1637/// into:
1638/// entry:
1639/// %cmpz = icmp eq i64 %A, 0
1640/// br i1 %cmpz, label %cond.end, label %cond.false
1641/// cond.false:
1642/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1643/// br label %cond.end
1644/// cond.end:
1645/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1646///
1647/// If the transform is performed, return true and set ModifiedDT to true.
1648static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1649 const TargetLowering *TLI,
1650 const DataLayout *DL,
1651 bool &ModifiedDT) {
1652 if (!TLI || !DL)
1653 return false;
1654
1655 // If a zero input is undefined, it doesn't make sense to despeculate that.
1656 if (match(CountZeros->getOperand(1), m_One()))
1657 return false;
1658
1659 // If it's cheap to speculate, there's nothing to do.
1660 auto IntrinsicID = CountZeros->getIntrinsicID();
1661 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1662 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1663 return false;
1664
1665 // Only handle legal scalar cases. Anything else requires too much work.
1666 Type *Ty = CountZeros->getType();
1667 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1668 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1669 return false;
1670
1671 // The intrinsic will be sunk behind a compare against zero and branch.
1672 BasicBlock *StartBlock = CountZeros->getParent();
1673 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1674
1675 // Create another block after the count zero intrinsic. A PHI will be added
1676 // in this block to select the result of the intrinsic or the bit-width
1677 // constant if the input to the intrinsic is zero.
1678 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1679 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1680
1681 // Set up a builder to create a compare, conditional branch, and PHI.
1682 IRBuilder<> Builder(CountZeros->getContext());
1683 Builder.SetInsertPoint(StartBlock->getTerminator());
1684 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1685
1686 // Replace the unconditional branch that was created by the first split with
1687 // a compare against zero and a conditional branch.
1688 Value *Zero = Constant::getNullValue(Ty);
1689 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1690 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1691 StartBlock->getTerminator()->eraseFromParent();
1692
1693 // Create a PHI in the end block to select either the output of the intrinsic
1694 // or the bit width of the operand.
1695 Builder.SetInsertPoint(&EndBlock->front());
1696 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1697 CountZeros->replaceAllUsesWith(PN);
1698 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1699 PN->addIncoming(BitWidth, StartBlock);
1700 PN->addIncoming(CountZeros, CallBlock);
1701
1702 // We are explicitly handling the zero case, so we can set the intrinsic's
1703 // undefined zero argument to 'true'. This will also prevent reprocessing the
1704 // intrinsic; we only despeculate when a zero input is defined.
1705 CountZeros->setArgOperand(1, Builder.getTrue());
1706 ModifiedDT = true;
1707 return true;
1708}
1709
1710bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1711 BasicBlock *BB = CI->getParent();
1712
1713 // Lower inline assembly if we can.
1714 // If we found an inline asm expession, and if the target knows how to
1715 // lower it to normal LLVM code, do so now.
1716 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1717 if (TLI->ExpandInlineAsm(CI)) {
1718 // Avoid invalidating the iterator.
1719 CurInstIterator = BB->begin();
1720 // Avoid processing instructions out of order, which could cause
1721 // reuse before a value is defined.
1722 SunkAddrs.clear();
1723 return true;
1724 }
1725 // Sink address computing for memory operands into the block.
1726 if (optimizeInlineAsmInst(CI))
1727 return true;
1728 }
1729
1730 // Align the pointer arguments to this call if the target thinks it's a good
1731 // idea
1732 unsigned MinSize, PrefAlign;
1733 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1734 for (auto &Arg : CI->arg_operands()) {
1735 // We want to align both objects whose address is used directly and
1736 // objects whose address is used in casts and GEPs, though it only makes
1737 // sense for GEPs if the offset is a multiple of the desired alignment and
1738 // if size - offset meets the size threshold.
1739 if (!Arg->getType()->isPointerTy())
1740 continue;
1741 APInt Offset(DL->getPointerSizeInBits(
1742 cast<PointerType>(Arg->getType())->getAddressSpace()),
1743 0);
1744 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1745 uint64_t Offset2 = Offset.getLimitedValue();
1746 if ((Offset2 & (PrefAlign-1)) != 0)
1747 continue;
1748 AllocaInst *AI;
1749 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1750 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1751 AI->setAlignment(PrefAlign);
1752 // Global variables can only be aligned if they are defined in this
1753 // object (i.e. they are uniquely initialized in this object), and
1754 // over-aligning global variables that have an explicit section is
1755 // forbidden.
1756 GlobalVariable *GV;
1757 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1758 GV->getPointerAlignment(*DL) < PrefAlign &&
1759 DL->getTypeAllocSize(GV->getValueType()) >=
1760 MinSize + Offset2)
1761 GV->setAlignment(PrefAlign);
1762 }
1763 // If this is a memcpy (or similar) then we may be able to improve the
1764 // alignment
1765 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1766 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1767 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1768 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1769 if (Align > MI->getAlignment())
1770 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1771 }
1772 }
1773
1774 // If we have a cold call site, try to sink addressing computation into the
1775 // cold block. This interacts with our handling for loads and stores to
1776 // ensure that we can fold all uses of a potential addressing computation
1777 // into their uses. TODO: generalize this to work over profiling data
1778 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1779 for (auto &Arg : CI->arg_operands()) {
1780 if (!Arg->getType()->isPointerTy())
1781 continue;
1782 unsigned AS = Arg->getType()->getPointerAddressSpace();
1783 return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1784 }
1785
1786 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1787 if (II) {
1788 switch (II->getIntrinsicID()) {
1789 default: break;
1790 case Intrinsic::objectsize: {
1791 // Lower all uses of llvm.objectsize.*
1792 ConstantInt *RetVal =
1793 lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1794 // Substituting this can cause recursive simplifications, which can
1795 // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1796 // this
1797 // happens.
1798 Value *CurValue = &*CurInstIterator;
1799 WeakTrackingVH IterHandle(CurValue);
1800
1801 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1802
1803 // If the iterator instruction was recursively deleted, start over at the
1804 // start of the block.
1805 if (IterHandle != CurValue) {
1806 CurInstIterator = BB->begin();
1807 SunkAddrs.clear();
1808 }
1809 return true;
1810 }
1811 case Intrinsic::aarch64_stlxr:
1812 case Intrinsic::aarch64_stxr: {
1813 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1814 if (!ExtVal || !ExtVal->hasOneUse() ||
1815 ExtVal->getParent() == CI->getParent())
1816 return false;
1817 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1818 ExtVal->moveBefore(CI);
1819 // Mark this instruction as "inserted by CGP", so that other
1820 // optimizations don't touch it.
1821 InsertedInsts.insert(ExtVal);
1822 return true;
1823 }
1824 case Intrinsic::invariant_group_barrier:
1825 II->replaceAllUsesWith(II->getArgOperand(0));
1826 II->eraseFromParent();
1827 return true;
1828
1829 case Intrinsic::cttz:
1830 case Intrinsic::ctlz:
1831 // If counting zeros is expensive, try to avoid it.
1832 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1833 }
1834
1835 if (TLI) {
1836 SmallVector<Value*, 2> PtrOps;
1837 Type *AccessTy;
1838 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1839 while (!PtrOps.empty()) {
1840 Value *PtrVal = PtrOps.pop_back_val();
1841 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1842 if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1843 return true;
1844 }
1845 }
1846 }
1847
1848 // From here on out we're working with named functions.
1849 if (!CI->getCalledFunction()) return false;
1850
1851 // Lower all default uses of _chk calls. This is very similar
1852 // to what InstCombineCalls does, but here we are only lowering calls
1853 // to fortified library functions (e.g. __memcpy_chk) that have the default
1854 // "don't know" as the objectsize. Anything else should be left alone.
1855 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1856 if (Value *V = Simplifier.optimizeCall(CI)) {
1857 CI->replaceAllUsesWith(V);
1858 CI->eraseFromParent();
1859 return true;
1860 }
1861
1862 return false;
1863}
1864
1865/// Look for opportunities to duplicate return instructions to the predecessor
1866/// to enable tail call optimizations. The case it is currently looking for is:
1867/// @code
1868/// bb0:
1869/// %tmp0 = tail call i32 @f0()
1870/// br label %return
1871/// bb1:
1872/// %tmp1 = tail call i32 @f1()
1873/// br label %return
1874/// bb2:
1875/// %tmp2 = tail call i32 @f2()
1876/// br label %return
1877/// return:
1878/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1879/// ret i32 %retval
1880/// @endcode
1881///
1882/// =>
1883///
1884/// @code
1885/// bb0:
1886/// %tmp0 = tail call i32 @f0()
1887/// ret i32 %tmp0
1888/// bb1:
1889/// %tmp1 = tail call i32 @f1()
1890/// ret i32 %tmp1
1891/// bb2:
1892/// %tmp2 = tail call i32 @f2()
1893/// ret i32 %tmp2
1894/// @endcode
1895bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1896 if (!TLI)
1897 return false;
1898
1899 ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1900 if (!RetI)
1901 return false;
1902
1903 PHINode *PN = nullptr;
1904 BitCastInst *BCI = nullptr;
1905 Value *V = RetI->getReturnValue();
1906 if (V) {
1907 BCI = dyn_cast<BitCastInst>(V);
1908 if (BCI)
1909 V = BCI->getOperand(0);
1910
1911 PN = dyn_cast<PHINode>(V);
1912 if (!PN)
1913 return false;
1914 }
1915
1916 if (PN && PN->getParent() != BB)
1917 return false;
1918
1919 // Make sure there are no instructions between the PHI and return, or that the
1920 // return is the first instruction in the block.
1921 if (PN) {
1922 BasicBlock::iterator BI = BB->begin();
1923 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1924 if (&*BI == BCI)
1925 // Also skip over the bitcast.
1926 ++BI;
1927 if (&*BI != RetI)
1928 return false;
1929 } else {
1930 BasicBlock::iterator BI = BB->begin();
1931 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1932 if (&*BI != RetI)
1933 return false;
1934 }
1935
1936 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1937 /// call.
1938 const Function *F = BB->getParent();
1939 SmallVector<CallInst*, 4> TailCalls;
1940 if (PN) {
1941 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1942 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1943 // Make sure the phi value is indeed produced by the tail call.
1944 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1945 TLI->mayBeEmittedAsTailCall(CI) &&
1946 attributesPermitTailCall(F, CI, RetI, *TLI))
1947 TailCalls.push_back(CI);
1948 }
1949 } else {
1950 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1951 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1952 if (!VisitedBBs.insert(*PI).second)
1953 continue;
1954
1955 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1956 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1957 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1958 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1959 if (RI == RE)
1960 continue;
1961
1962 CallInst *CI = dyn_cast<CallInst>(&*RI);
1963 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1964 attributesPermitTailCall(F, CI, RetI, *TLI))
1965 TailCalls.push_back(CI);
1966 }
1967 }
1968
1969 bool Changed = false;
1970 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1971 CallInst *CI = TailCalls[i];
1972 CallSite CS(CI);
1973
1974 // Conservatively require the attributes of the call to match those of the
1975 // return. Ignore noalias because it doesn't affect the call sequence.
1976 AttributeList CalleeAttrs = CS.getAttributes();
1977 if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1978 .removeAttribute(Attribute::NoAlias) !=
1979 AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1980 .removeAttribute(Attribute::NoAlias))
1981 continue;
1982
1983 // Make sure the call instruction is followed by an unconditional branch to
1984 // the return block.
1985 BasicBlock *CallBB = CI->getParent();
1986 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1987 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1988 continue;
1989
1990 // Duplicate the return into CallBB.
1991 (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1992 ModifiedDT = Changed = true;
1993 ++NumRetsDup;
1994 }
1995
1996 // If we eliminated all predecessors of the block, delete the block now.
1997 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1998 BB->eraseFromParent();
1999
2000 return Changed;
2001}
2002
2003//===----------------------------------------------------------------------===//
2004// Memory Optimization
2005//===----------------------------------------------------------------------===//
2006
2007namespace {
2008
2009/// This is an extended version of TargetLowering::AddrMode
2010/// which holds actual Value*'s for register values.
2011struct ExtAddrMode : public TargetLowering::AddrMode {
2012 Value *BaseReg = nullptr;
2013 Value *ScaledReg = nullptr;
2014 Value *OriginalValue = nullptr;
2015
2016 enum FieldName {
2017 NoField = 0x00,
2018 BaseRegField = 0x01,
2019 BaseGVField = 0x02,
2020 BaseOffsField = 0x04,
2021 ScaledRegField = 0x08,
2022 ScaleField = 0x10,
2023 MultipleFields = 0xff
2024 };
2025
2026 ExtAddrMode() = default;
2027
2028 void print(raw_ostream &OS) const;
2029 void dump() const;
2030
2031 FieldName compare(const ExtAddrMode &other) {
2032 // First check that the types are the same on each field, as differing types
2033 // is something we can't cope with later on.
2034 if (BaseReg && other.BaseReg &&
2035 BaseReg->getType() != other.BaseReg->getType())
2036 return MultipleFields;
2037 if (BaseGV && other.BaseGV &&
2038 BaseGV->getType() != other.BaseGV->getType())
2039 return MultipleFields;
2040 if (ScaledReg && other.ScaledReg &&
2041 ScaledReg->getType() != other.ScaledReg->getType())
2042 return MultipleFields;
2043
2044 // Check each field to see if it differs.
2045 unsigned Result = NoField;
2046 if (BaseReg != other.BaseReg)
2047 Result |= BaseRegField;
2048 if (BaseGV != other.BaseGV)
2049 Result |= BaseGVField;
2050 if (BaseOffs != other.BaseOffs)
2051 Result |= BaseOffsField;
2052 if (ScaledReg != other.ScaledReg)
2053 Result |= ScaledRegField;
2054 // Don't count 0 as being a different scale, because that actually means
2055 // unscaled (which will already be counted by having no ScaledReg).
2056 if (Scale && other.Scale && Scale != other.Scale)
2057 Result |= ScaleField;
2058
2059 if (countPopulation(Result) > 1)
2060 return MultipleFields;
2061 else
2062 return static_cast<FieldName>(Result);
2063 }
2064
2065 // AddrModes with a baseReg or gv where the reg/gv is
2066 // the only populated field are trivial.
2067 bool isTrivial() {
2068 if (BaseGV && !BaseOffs && !Scale && !BaseReg)
2069 return true;
2070
2071 if (!BaseGV && !BaseOffs && !Scale && BaseReg)
2072 return true;
2073
2074 return false;
2075 }
2076};
2077
2078} // end anonymous namespace
2079
2080#ifndef NDEBUG
2081static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2082 AM.print(OS);
2083 return OS;
2084}
2085#endif
2086
2087#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2088void ExtAddrMode::print(raw_ostream &OS) const {
2089 bool NeedPlus = false;
2090 OS << "[";
2091 if (BaseGV) {
2092 OS << (NeedPlus ? " + " : "")
2093 << "GV:";
2094 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2095 NeedPlus = true;
2096 }
2097
2098 if (BaseOffs) {
2099 OS << (NeedPlus ? " + " : "")
2100 << BaseOffs;
2101 NeedPlus = true;
2102 }
2103
2104 if (BaseReg) {
2105 OS << (NeedPlus ? " + " : "")
2106 << "Base:";
2107 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2108 NeedPlus = true;
2109 }
2110 if (Scale) {
2111 OS << (NeedPlus ? " + " : "")
2112 << Scale << "*";
2113 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2114 }
2115
2116 OS << ']';
2117}
2118
2119LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void ExtAddrMode::dump() const {
2120 print(dbgs());
2121 dbgs() << '\n';
2122}
2123#endif
2124
2125namespace {
2126
2127/// \brief This class provides transaction based operation on the IR.
2128/// Every change made through this class is recorded in the internal state and
2129/// can be undone (rollback) until commit is called.
2130class TypePromotionTransaction {
2131 /// \brief This represents the common interface of the individual transaction.
2132 /// Each class implements the logic for doing one specific modification on
2133 /// the IR via the TypePromotionTransaction.
2134 class TypePromotionAction {
2135 protected:
2136 /// The Instruction modified.
2137 Instruction *Inst;
2138
2139 public:
2140 /// \brief Constructor of the action.
2141 /// The constructor performs the related action on the IR.
2142 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2143
2144 virtual ~TypePromotionAction() = default;
2145
2146 /// \brief Undo the modification done by this action.
2147 /// When this method is called, the IR must be in the same state as it was
2148 /// before this action was applied.
2149 /// \pre Undoing the action works if and only if the IR is in the exact same
2150 /// state as it was directly after this action was applied.
2151 virtual void undo() = 0;
2152
2153 /// \brief Advocate every change made by this action.
2154 /// When the results on the IR of the action are to be kept, it is important
2155 /// to call this function, otherwise hidden information may be kept forever.
2156 virtual void commit() {
2157 // Nothing to be done, this action is not doing anything.
2158 }
2159 };
2160
2161 /// \brief Utility to remember the position of an instruction.
2162 class InsertionHandler {
2163 /// Position of an instruction.
2164 /// Either an instruction:
2165 /// - Is the first in a basic block: BB is used.
2166 /// - Has a previous instructon: PrevInst is used.
2167 union {
2168 Instruction *PrevInst;
2169 BasicBlock *BB;
2170 } Point;
2171
2172 /// Remember whether or not the instruction had a previous instruction.
2173 bool HasPrevInstruction;
2174
2175 public:
2176 /// \brief Record the position of \p Inst.
2177 InsertionHandler(Instruction *Inst) {
2178 BasicBlock::iterator It = Inst->getIterator();
2179 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2180 if (HasPrevInstruction)
2181 Point.PrevInst = &*--It;
2182 else
2183 Point.BB = Inst->getParent();
2184 }
2185
2186 /// \brief Insert \p Inst at the recorded position.
2187 void insert(Instruction *Inst) {
2188 if (HasPrevInstruction) {
2189 if (Inst->getParent())
2190 Inst->removeFromParent();
2191 Inst->insertAfter(Point.PrevInst);
2192 } else {
2193 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2194 if (Inst->getParent())
2195 Inst->moveBefore(Position);
2196 else
2197 Inst->insertBefore(Position);
2198 }
2199 }
2200 };
2201
2202 /// \brief Move an instruction before another.
2203 class InstructionMoveBefore : public TypePromotionAction {
2204 /// Original position of the instruction.
2205 InsertionHandler Position;
2206
2207 public:
2208 /// \brief Move \p Inst before \p Before.
2209 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2210 : TypePromotionAction(Inst), Position(Inst) {
2211 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: move: " << *
Inst << "\nbefore: " << *Before << "\n"; } }
while (false)
;
2212 Inst->moveBefore(Before);
2213 }
2214
2215 /// \brief Move the instruction back to its original position.
2216 void undo() override {
2217 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: moveBefore: " <<
*Inst << "\n"; } } while (false)
;
2218 Position.insert(Inst);
2219 }
2220 };
2221
2222 /// \brief Set the operand of an instruction with a new value.
2223 class OperandSetter : public TypePromotionAction {
2224 /// Original operand of the instruction.
2225 Value *Origin;
2226
2227 /// Index of the modified instruction.
2228 unsigned Idx;
2229
2230 public:
2231 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2232 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2233 : TypePromotionAction(Inst), Idx(Idx) {
2234 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
2235 << "for:" << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
2236 << "with:" << *NewVal << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
;
2237 Origin = Inst->getOperand(Idx);
2238 Inst->setOperand(Idx, NewVal);
2239 }
2240
2241 /// \brief Restore the original value of the instruction.
2242 void undo() override {
2243 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
2244 << "for: " << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
2245 << "with: " << *Origin << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
;
2246 Inst->setOperand(Idx, Origin);
2247 }
2248 };
2249
2250 /// \brief Hide the operands of an instruction.
2251 /// Do as if this instruction was not using any of its operands.
2252 class OperandsHider : public TypePromotionAction {
2253 /// The list of original operands.
2254 SmallVector<Value *, 4> OriginalValues;
2255
2256 public:
2257 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2258 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2259 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: OperandsHider: " <<
*Inst << "\n"; } } while (false)
;
2260 unsigned NumOpnds = Inst->getNumOperands();
2261 OriginalValues.reserve(NumOpnds);
2262 for (unsigned It = 0; It < NumOpnds; ++It) {
2263 // Save the current operand.
2264 Value *Val = Inst->getOperand(It);
2265 OriginalValues.push_back(Val);
2266 // Set a dummy one.
2267 // We could use OperandSetter here, but that would imply an overhead
2268 // that we are not willing to pay.
2269 Inst->setOperand(It, UndefValue::get(Val->getType()));
2270 }
2271 }
2272
2273 /// \brief Restore the original list of uses.
2274 void undo() override {
2275 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: OperandsHider: "
<< *Inst << "\n"; } } while (false)
;
2276 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2277 Inst->setOperand(It, OriginalValues[It]);
2278 }
2279 };
2280
2281 /// \brief Build a truncate instruction.
2282 class TruncBuilder : public TypePromotionAction {
2283 Value *Val;
2284
2285 public:
2286 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2287 /// result.
2288 /// trunc Opnd to Ty.
2289 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2290 IRBuilder<> Builder(Opnd);
2291 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2292 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2293 }
2294
2295 /// \brief Get the built value.
2296 Value *getBuiltValue() { return Val; }
2297
2298 /// \brief Remove the built instruction.
2299 void undo() override {
2300 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2301 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2302 IVal->eraseFromParent();
2303 }
2304 };
2305
2306 /// \brief Build a sign extension instruction.
2307 class SExtBuilder : public TypePromotionAction {
2308 Value *Val;
2309
2310 public:
2311 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2312 /// result.
2313 /// sext Opnd to Ty.
2314 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2315 : TypePromotionAction(InsertPt) {
2316 IRBuilder<> Builder(InsertPt);
2317 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2318 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2319 }
2320
2321 /// \brief Get the built value.
2322 Value *getBuiltValue() { return Val; }
2323
2324 /// \brief Remove the built instruction.
2325 void undo() override {
2326 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2327 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2328 IVal->eraseFromParent();
2329 }
2330 };
2331
2332 /// \brief Build a zero extension instruction.
2333 class ZExtBuilder : public TypePromotionAction {
2334 Value *Val;
2335
2336 public:
2337 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2338 /// result.
2339 /// zext Opnd to Ty.
2340 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2341 : TypePromotionAction(InsertPt) {
2342 IRBuilder<> Builder(InsertPt);
2343 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2344 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2345 }
2346
2347 /// \brief Get the built value.
2348 Value *getBuiltValue() { return Val; }
2349
2350 /// \brief Remove the built instruction.
2351 void undo() override {
2352 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2353 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2354 IVal->eraseFromParent();
2355 }
2356 };
2357
2358 /// \brief Mutate an instruction to another type.
2359 class TypeMutator : public TypePromotionAction {
2360 /// Record the original type.
2361 Type *OrigTy;
2362
2363 public:
2364 /// \brief Mutate the type of \p Inst into \p NewTy.
2365 TypeMutator(Instruction *Inst, Type *NewTy)
2366 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2367 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
2368 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
;
2369 Inst->mutateType(NewTy);
2370 }
2371
2372 /// \brief Mutate the instruction back to its original type.
2373 void undo() override {
2374 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
2375 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
;
2376 Inst->mutateType(OrigTy);
2377 }
2378 };
2379
2380 /// \brief Replace the uses of an instruction by another instruction.
2381 class UsesReplacer : public TypePromotionAction {
2382 /// Helper structure to keep track of the replaced uses.
2383 struct InstructionAndIdx {
2384 /// The instruction using the instruction.
2385 Instruction *Inst;
2386
2387 /// The index where this instruction is used for Inst.
2388 unsigned Idx;
2389
2390 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2391 : Inst(Inst), Idx(Idx) {}
2392 };
2393
2394 /// Keep track of the original uses (pair Instruction, Index).
2395 SmallVector<InstructionAndIdx, 4> OriginalUses;
2396
2397 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2398
2399 public:
2400 /// \brief Replace all the use of \p Inst by \p New.
2401 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2402 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
2403 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
;
2404 // Record the original uses.
2405 for (Use &U : Inst->uses()) {
2406 Instruction *UserI = cast<Instruction>(U.getUser());
2407 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2408 }
2409 // Now, we can replace the uses.
2410 Inst->replaceAllUsesWith(New);
2411 }
2412
2413 /// \brief Reassign the original uses of Inst to Inst.
2414 void undo() override {
2415 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: UsersReplacer: "
<< *Inst << "\n"; } } while (false)
;
2416 for (use_iterator UseIt = OriginalUses.begin(),
2417 EndIt = OriginalUses.end();
2418 UseIt != EndIt; ++UseIt) {
2419 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2420 }
2421 }
2422 };
2423
2424 /// \brief Remove an instruction from the IR.
2425 class InstructionRemover : public TypePromotionAction {
2426 /// Original position of the instruction.
2427 InsertionHandler Inserter;
2428
2429 /// Helper structure to hide all the link to the instruction. In other
2430 /// words, this helps to do as if the instruction was removed.
2431 OperandsHider Hider;
2432
2433 /// Keep track of the uses replaced, if any.
2434 UsesReplacer *Replacer = nullptr;
2435
2436 /// Keep track of instructions removed.
2437 SetOfInstrs &RemovedInsts;
2438
2439 public:
2440 /// \brief Remove all reference of \p Inst and optinally replace all its
2441 /// uses with New.
2442 /// \p RemovedInsts Keep track of the instructions removed by this Action.
2443 /// \pre If !Inst->use_empty(), then New != nullptr
2444 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2445 Value *New = nullptr)
2446 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2447 RemovedInsts(RemovedInsts) {
2448 if (New)
2449 Replacer = new UsesReplacer(Inst, New);
2450 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2451 RemovedInsts.insert(Inst);
2452 /// The instructions removed here will be freed after completing
2453 /// optimizeBlock() for all blocks as we need to keep track of the
2454 /// removed instructions during promotion.
2455 Inst->removeFromParent();
2456 }
2457
2458 ~InstructionRemover() override { delete Replacer; }
2459
2460 /// \brief Resurrect the instruction and reassign it to the proper uses if
2461 /// new value was provided when build this action.
2462 void undo() override {
2463 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2464 Inserter.insert(Inst);
2465 if (Replacer)
2466 Replacer->undo();
2467 Hider.undo();
2468 RemovedInsts.erase(Inst);
2469 }
2470 };
2471
2472public:
2473 /// Restoration point.
2474 /// The restoration point is a pointer to an action instead of an iterator
2475 /// because the iterator may be invalidated but not the pointer.
2476 using ConstRestorationPt = const TypePromotionAction *;
2477
2478 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2479 : RemovedInsts(RemovedInsts) {}
2480
2481 /// Advocate every changes made in that transaction.
2482 void commit();
2483
2484 /// Undo all the changes made after the given point.
2485 void rollback(ConstRestorationPt Point);
2486
2487 /// Get the current restoration point.
2488 ConstRestorationPt getRestorationPoint() const;
2489
2490 /// \name API for IR modification with state keeping to support rollback.
2491 /// @{
2492 /// Same as Instruction::setOperand.
2493 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2494
2495 /// Same as Instruction::eraseFromParent.
2496 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2497
2498 /// Same as Value::replaceAllUsesWith.
2499 void replaceAllUsesWith(Instruction *Inst, Value *New);
2500
2501 /// Same as Value::mutateType.
2502 void mutateType(Instruction *Inst, Type *NewTy);
2503
2504 /// Same as IRBuilder::createTrunc.
2505 Value *createTrunc(Instruction *Opnd, Type *Ty);
2506
2507 /// Same as IRBuilder::createSExt.
2508 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2509
2510 /// Same as IRBuilder::createZExt.
2511 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2512
2513 /// Same as Instruction::moveBefore.
2514 void moveBefore(Instruction *Inst, Instruction *Before);
2515 /// @}
2516
2517private:
2518 /// The ordered list of actions made so far.
2519 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2520
2521 using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2522
2523 SetOfInstrs &RemovedInsts;
2524};
2525
2526} // end anonymous namespace
2527
2528void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2529 Value *NewVal) {
2530 Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2531 Inst, Idx, NewVal));
2532}
2533
2534void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2535 Value *NewVal) {
2536 Actions.push_back(
2537 llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2538 Inst, RemovedInsts, NewVal));
2539}
2540
2541void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2542 Value *New) {
2543 Actions.push_back(
2544 llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2545}
2546
2547void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2548 Actions.push_back(
2549 llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2550}
2551
2552Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2553 Type *Ty) {
2554 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2555 Value *Val = Ptr->getBuiltValue();
2556 Actions.push_back(std::move(Ptr));
2557 return Val;
2558}
2559
2560Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2561 Value *Opnd, Type *Ty) {
2562 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2563 Value *Val = Ptr->getBuiltValue();
2564 Actions.push_back(std::move(Ptr));
2565 return Val;
2566}
2567
2568Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2569 Value *Opnd, Type *Ty) {
2570 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2571 Value *Val = Ptr->getBuiltValue();
2572 Actions.push_back(std::move(Ptr));
2573 return Val;
2574}
2575
2576void TypePromotionTransaction::moveBefore(Instruction *Inst,
2577 Instruction *Before) {
2578 Actions.push_back(
2579 llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2580 Inst, Before));
2581}
2582
2583TypePromotionTransaction::ConstRestorationPt
2584TypePromotionTransaction::getRestorationPoint() const {
2585 return !Actions.empty() ? Actions.back().get() : nullptr;
2586}
2587
2588void TypePromotionTransaction::commit() {
2589 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2590 ++It)
2591 (*It)->commit();
2592 Actions.clear();
2593}
2594
2595void TypePromotionTransaction::rollback(
2596 TypePromotionTransaction::ConstRestorationPt Point) {
2597 while (!Actions.empty() && Point != Actions.back().get()) {
2598 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2599 Curr->undo();
2600 }
2601}
2602
2603namespace {
2604
2605/// \brief A helper class for matching addressing modes.
2606///
2607/// This encapsulates the logic for matching the target-legal addressing modes.
2608class AddressingModeMatcher {
2609 SmallVectorImpl<Instruction*> &AddrModeInsts;
2610 const TargetLowering &TLI;
2611 const TargetRegisterInfo &TRI;
2612 const DataLayout &DL;
2613
2614 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2615 /// the memory instruction that we're computing this address for.
2616 Type *AccessTy;
2617 unsigned AddrSpace;
2618 Instruction *MemoryInst;
2619
2620 /// This is the addressing mode that we're building up. This is
2621 /// part of the return value of this addressing mode matching stuff.
2622 ExtAddrMode &AddrMode;
2623
2624 /// The instructions inserted by other CodeGenPrepare optimizations.
2625 const SetOfInstrs &InsertedInsts;
2626
2627 /// A map from the instructions to their type before promotion.
2628 InstrToOrigTy &PromotedInsts;
2629
2630 /// The ongoing transaction where every action should be registered.
2631 TypePromotionTransaction &TPT;
2632
2633 /// This is set to true when we should not do profitability checks.
2634 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2635 bool IgnoreProfitability;
2636
2637 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2638 const TargetLowering &TLI,
2639 const TargetRegisterInfo &TRI,
2640 Type *AT, unsigned AS,
2641 Instruction *MI, ExtAddrMode &AM,
2642 const SetOfInstrs &InsertedInsts,
2643 InstrToOrigTy &PromotedInsts,
2644 TypePromotionTransaction &TPT)
2645 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2646 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2647 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2648 PromotedInsts(PromotedInsts), TPT(TPT) {
2649 IgnoreProfitability = false;
2650 }
2651
2652public:
2653 /// Find the maximal addressing mode that a load/store of V can fold,
2654 /// give an access type of AccessTy. This returns a list of involved
2655 /// instructions in AddrModeInsts.
2656 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2657 /// optimizations.
2658 /// \p PromotedInsts maps the instructions to their type before promotion.
2659 /// \p The ongoing transaction where every action should be registered.
2660 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2661 Instruction *MemoryInst,
2662 SmallVectorImpl<Instruction*> &AddrModeInsts,
2663 const TargetLowering &TLI,
2664 const TargetRegisterInfo &TRI,
2665 const SetOfInstrs &InsertedInsts,
2666 InstrToOrigTy &PromotedInsts,
2667 TypePromotionTransaction &TPT) {
2668 ExtAddrMode Result;
2669
2670 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
2671 AccessTy, AS,
2672 MemoryInst, Result, InsertedInsts,
2673 PromotedInsts, TPT).matchAddr(V, 0);
2674 (void)Success; assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?"
) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 2674, __extension__ __PRETTY_FUNCTION__))
;
2675 return Result;
2676 }
2677
2678private:
2679 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2680 bool matchAddr(Value *V, unsigned Depth);
2681 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2682 bool *MovedAway = nullptr);
2683 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2684 ExtAddrMode &AMBefore,
2685 ExtAddrMode &AMAfter);
2686 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2687 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2688 Value *PromotedOperand) const;
2689};
2690
2691/// \brief Keep track of simplification of Phi nodes.
2692/// Accept the set of all phi nodes and erase phi node from this set
2693/// if it is simplified.
2694class SimplificationTracker {
2695 DenseMap<Value *, Value *> Storage;
2696 const SimplifyQuery &SQ;
2697 SmallPtrSetImpl<PHINode *> &AllPhiNodes;
2698 SmallPtrSetImpl<SelectInst *> &AllSelectNodes;
2699
2700public:
2701 SimplificationTracker(const SimplifyQuery &sq,
2702 SmallPtrSetImpl<PHINode *> &APN,
2703 SmallPtrSetImpl<SelectInst *> &ASN)
2704 : SQ(sq), AllPhiNodes(APN), AllSelectNodes(ASN) {}
2705
2706 Value *Get(Value *V) {
2707 do {
2708 auto SV = Storage.find(V);
2709 if (SV == Storage.end())
2710 return V;
2711 V = SV->second;
2712 } while (true);
2713 }
2714
2715 Value *Simplify(Value *Val) {
2716 SmallVector<Value *, 32> WorkList;
2717 SmallPtrSet<Value *, 32> Visited;
2718 WorkList.push_back(Val);
2719 while (!WorkList.empty()) {
2720 auto P = WorkList.pop_back_val();
2721 if (!Visited.insert(P).second)
2722 continue;
2723 if (auto *PI = dyn_cast<Instruction>(P))
2724 if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2725 for (auto *U : PI->users())
2726 WorkList.push_back(cast<Value>(U));
2727 Put(PI, V);
2728 PI->replaceAllUsesWith(V);
2729 if (auto *PHI = dyn_cast<PHINode>(PI))
2730 AllPhiNodes.erase(PHI);
2731 if (auto *Select = dyn_cast<SelectInst>(PI))
2732 AllSelectNodes.erase(Select);
2733 PI->eraseFromParent();
2734 }
2735 }
2736 return Get(Val);
2737 }
2738
2739 void Put(Value *From, Value *To) {
2740 Storage.insert({ From, To });
2741 }
2742};
2743
2744/// \brief A helper class for combining addressing modes.
2745class AddressingModeCombiner {
2746 typedef std::pair<Value *, BasicBlock *> ValueInBB;
2747 typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2748 typedef std::pair<PHINode *, PHINode *> PHIPair;
2749
2750private:
2751 /// The addressing modes we've collected.
2752 SmallVector<ExtAddrMode, 16> AddrModes;
2753
2754 /// The field in which the AddrModes differ, when we have more than one.
2755 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2756
2757 /// Are the AddrModes that we have all just equal to their original values?
2758 bool AllAddrModesTrivial = true;
2759
2760 /// Common Type for all different fields in addressing modes.
2761 Type *CommonType;
2762
2763 /// SimplifyQuery for simplifyInstruction utility.
2764 const SimplifyQuery &SQ;
2765
2766 /// Original Address.
2767 ValueInBB Original;
2768
2769public:
2770 AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2771 : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2772
2773 /// \brief Get the combined AddrMode
2774 const ExtAddrMode &getAddrMode() const {
2775 return AddrModes[0];
2776 }
2777
2778 /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
2779 /// have.
2780 /// \return True iff we succeeded in doing so.
2781 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2782 // Take note of if we have any non-trivial AddrModes, as we need to detect
2783 // when all AddrModes are trivial as then we would introduce a phi or select
2784 // which just duplicates what's already there.
2785 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2786
2787 // If this is the first addrmode then everything is fine.
2788 if (AddrModes.empty()) {
2789 AddrModes.emplace_back(NewAddrMode);
2790 return true;
2791 }
2792
2793 // Figure out how different this is from the other address modes, which we
2794 // can do just by comparing against the first one given that we only care
2795 // about the cumulative difference.
2796 ExtAddrMode::FieldName ThisDifferentField =
2797 AddrModes[0].compare(NewAddrMode);
2798 if (DifferentField == ExtAddrMode::NoField)
2799 DifferentField = ThisDifferentField;
2800 else if (DifferentField != ThisDifferentField)
2801 DifferentField = ExtAddrMode::MultipleFields;
2802
2803 // If this AddrMode is the same as all the others then everything is fine
2804 // (which should only happen when there is actually only one AddrMode).
2805 if (DifferentField == ExtAddrMode::NoField) {
2806 assert(AddrModes.size() == 1)(static_cast <bool> (AddrModes.size() == 1) ? void (0) :
__assert_fail ("AddrModes.size() == 1", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 2806, __extension__ __PRETTY_FUNCTION__))
;
2807 return true;
2808 }
2809
2810 // If NewAddrMode differs in only one dimension then we can handle it by
2811 // inserting a phi/select later on.
2812 if (DifferentField != ExtAddrMode::MultipleFields) {
2813 AddrModes.emplace_back(NewAddrMode);
2814 return true;
2815 }
2816
2817 // We couldn't combine NewAddrMode with the rest, so return failure.
2818 AddrModes.clear();
2819 return false;
2820 }
2821
2822 /// \brief Combine the addressing modes we've collected into a single
2823 /// addressing mode.
2824 /// \return True iff we successfully combined them or we only had one so
2825 /// didn't need to combine them anyway.
2826 bool combineAddrModes() {
2827 // If we have no AddrModes then they can't be combined.
2828 if (AddrModes.size() == 0)
27
Assuming the condition is false
28
Taking false branch
2829 return false;
2830
2831 // A single AddrMode can trivially be combined.
2832 if (AddrModes.size() == 1)
29
Assuming the condition is false
30
Taking false branch
2833 return true;
2834
2835 // If the AddrModes we collected are all just equal to the value they are
2836 // derived from then combining them wouldn't do anything useful.
2837 if (AllAddrModesTrivial)
31
Assuming the condition is false
32
Taking false branch
2838 return false;
2839
2840 if (DisableComplexAddrModes)
33
Assuming the condition is false
34
Taking false branch
2841 return false;
2842
2843 // For now we support only different base registers.
2844 // TODO: enable others.
2845 if (DifferentField != ExtAddrMode::BaseRegField)
35
Assuming the condition is false
36
Taking false branch
2846 return false;
2847
2848 // Build a map between <original value, basic block where we saw it> to
2849 // value of base register.
2850 FoldAddrToValueMapping Map;
2851 initializeMap(Map);
2852
2853 Value *CommonValue = findCommon(Map);
37
Calling 'AddressingModeCombiner::findCommon'
2854 if (CommonValue)
2855 AddrModes[0].BaseReg = CommonValue;
2856 return CommonValue != nullptr;
2857 }
2858
2859private:
2860 /// \brief Initialize Map with anchor values. For address seen in some BB
2861 /// we set the value of different field saw in this address.
2862 /// If address is not an instruction than basic block is set to null.
2863 /// At the same time we find a common type for different field we will
2864 /// use to create new Phi/Select nodes. Keep it in CommonType field.
2865 void initializeMap(FoldAddrToValueMapping &Map) {
2866 // Keep track of keys where the value is null. We will need to replace it
2867 // with constant null when we know the common type.
2868 SmallVector<ValueInBB, 2> NullValue;
2869 for (auto &AM : AddrModes) {
2870 BasicBlock *BB = nullptr;
2871 if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2872 BB = I->getParent();
2873
2874 // For now we support only base register as different field.
2875 // TODO: Enable others.
2876 Value *DV = AM.BaseReg;
2877 if (DV) {
2878 if (CommonType)
2879 assert(CommonType == DV->getType() && "Different types detected!")(static_cast <bool> (CommonType == DV->getType() &&
"Different types detected!") ? void (0) : __assert_fail ("CommonType == DV->getType() && \"Different types detected!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 2879, __extension__ __PRETTY_FUNCTION__))
;
2880 else
2881 CommonType = DV->getType();
2882 Map[{ AM.OriginalValue, BB }] = DV;
2883 } else {
2884 NullValue.push_back({ AM.OriginalValue, BB });
2885 }
2886 }
2887 assert(CommonType && "At least one non-null value must be!")(static_cast <bool> (CommonType && "At least one non-null value must be!"
) ? void (0) : __assert_fail ("CommonType && \"At least one non-null value must be!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 2887, __extension__ __PRETTY_FUNCTION__))
;
2888 for (auto VIBB : NullValue)
2889 Map[VIBB] = Constant::getNullValue(CommonType);
2890 }
2891
2892 /// \brief We have mapping between value A and basic block where value A
2893 /// seen to other value B where B was a field in addressing mode represented
2894 /// by A. Also we have an original value C representin an address in some
2895 /// basic block. Traversing from C through phi and selects we ended up with
2896 /// A's in a map. This utility function tries to find a value V which is a
2897 /// field in addressing mode C and traversing through phi nodes and selects
2898 /// we will end up in corresponded values B in a map.
2899 /// The utility will create a new Phi/Selects if needed.
2900 // The simple example looks as follows:
2901 // BB1:
2902 // p1 = b1 + 40
2903 // br cond BB2, BB3
2904 // BB2:
2905 // p2 = b2 + 40
2906 // br BB3
2907 // BB3:
2908 // p = phi [p1, BB1], [p2, BB2]
2909 // v = load p
2910 // Map is
2911 // <p1, BB1> -> b1
2912 // <p2, BB2> -> b2
2913 // Request is
2914 // <p, BB3> -> ?
2915 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2916 Value *findCommon(FoldAddrToValueMapping &Map) {
2917 // Tracks of new created Phi nodes.
2918 SmallPtrSet<PHINode *, 32> NewPhiNodes;
2919 // Tracks of new created Select nodes.
2920 SmallPtrSet<SelectInst *, 32> NewSelectNodes;
2921 // Tracks the simplification of new created phi nodes. The reason we use
2922 // this mapping is because we will add new created Phi nodes in AddrToBase.
2923 // Simplification of Phi nodes is recursive, so some Phi node may
2924 // be simplified after we added it to AddrToBase.
2925 // Using this mapping we can find the current value in AddrToBase.
2926 SimplificationTracker ST(SQ, NewPhiNodes, NewSelectNodes);
2927
2928 // First step, DFS to create PHI nodes for all intermediate blocks.
2929 // Also fill traverse order for the second step.
2930 SmallVector<ValueInBB, 32> TraverseOrder;
2931 InsertPlaceholders(Map, TraverseOrder, NewPhiNodes, NewSelectNodes);
38
Calling 'AddressingModeCombiner::InsertPlaceholders'
2932
2933 // Second Step, fill new nodes by merged values and simplify if possible.
2934 FillPlaceholders(Map, TraverseOrder, ST);
2935
2936 if (!AddrSinkNewSelects && NewSelectNodes.size() > 0) {
2937 DestroyNodes(NewPhiNodes);
2938 DestroyNodes(NewSelectNodes);
2939 return nullptr;
2940 }
2941
2942 // Now we'd like to match New Phi nodes to existed ones.
2943 unsigned PhiNotMatchedCount = 0;
2944 if (!MatchPhiSet(NewPhiNodes, ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2945 DestroyNodes(NewPhiNodes);
2946 DestroyNodes(NewSelectNodes);
2947 return nullptr;
2948 }
2949
2950 auto *Result = ST.Get(Map.find(Original)->second);
2951 if (Result) {
2952 NumMemoryInstsPhiCreated += NewPhiNodes.size() + PhiNotMatchedCount;
2953 NumMemoryInstsSelectCreated += NewSelectNodes.size();
2954 }
2955 return Result;
2956 }
2957
2958 /// \brief Destroy nodes from a set.
2959 template <typename T> void DestroyNodes(SmallPtrSetImpl<T *> &Instructions) {
2960 // For safe erasing, replace the Phi with dummy value first.
2961 auto Dummy = UndefValue::get(CommonType);
2962 for (auto I : Instructions) {
2963 I->replaceAllUsesWith(Dummy);
2964 I->eraseFromParent();
2965 }
2966 }
2967
2968 /// \brief Try to match PHI node to Candidate.
2969 /// Matcher tracks the matched Phi nodes.
2970 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2971 DenseSet<PHIPair> &Matcher,
2972 SmallPtrSetImpl<PHINode *> &PhiNodesToMatch) {
2973 SmallVector<PHIPair, 8> WorkList;
2974 Matcher.insert({ PHI, Candidate });
2975 WorkList.push_back({ PHI, Candidate });
2976 SmallSet<PHIPair, 8> Visited;
2977 while (!WorkList.empty()) {
2978 auto Item = WorkList.pop_back_val();
2979 if (!Visited.insert(Item).second)
2980 continue;
2981 // We iterate over all incoming values to Phi to compare them.
2982 // If values are different and both of them Phi and the first one is a
2983 // Phi we added (subject to match) and both of them is in the same basic
2984 // block then we can match our pair if values match. So we state that
2985 // these values match and add it to work list to verify that.
2986 for (auto B : Item.first->blocks()) {
2987 Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2988 Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2989 if (FirstValue == SecondValue)
2990 continue;
2991
2992 PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2993 PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2994
2995 // One of them is not Phi or
2996 // The first one is not Phi node from the set we'd like to match or
2997 // Phi nodes from different basic blocks then
2998 // we will not be able to match.
2999 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3000 FirstPhi->getParent() != SecondPhi->getParent())
3001 return false;
3002
3003 // If we already matched them then continue.
3004 if (Matcher.count({ FirstPhi, SecondPhi }))
3005 continue;
3006 // So the values are different and does not match. So we need them to
3007 // match.
3008 Matcher.insert({ FirstPhi, SecondPhi });
3009 // But me must check it.
3010 WorkList.push_back({ FirstPhi, SecondPhi });
3011 }
3012 }
3013 return true;
3014 }
3015
3016 /// \brief For the given set of PHI nodes try to find their equivalents.
3017 /// Returns false if this matching fails and creation of new Phi is disabled.
3018 bool MatchPhiSet(SmallPtrSetImpl<PHINode *> &PhiNodesToMatch,
3019 SimplificationTracker &ST, bool AllowNewPhiNodes,
3020 unsigned &PhiNotMatchedCount) {
3021 DenseSet<PHIPair> Matched;
3022 SmallPtrSet<PHINode *, 8> WillNotMatch;
3023 while (PhiNodesToMatch.size()) {
3024 PHINode *PHI = *PhiNodesToMatch.begin();
3025
3026 // Add us, if no Phi nodes in the basic block we do not match.
3027 WillNotMatch.clear();
3028 WillNotMatch.insert(PHI);
3029
3030 // Traverse all Phis until we found equivalent or fail to do that.
3031 bool IsMatched = false;
3032 for (auto &P : PHI->getParent()->phis()) {
3033 if (&P == PHI)
3034 continue;
3035 if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3036 break;
3037 // If it does not match, collect all Phi nodes from matcher.
3038 // if we end up with no match, them all these Phi nodes will not match
3039 // later.
3040 for (auto M : Matched)
3041 WillNotMatch.insert(M.first);
3042 Matched.clear();
3043 }
3044 if (IsMatched) {
3045 // Replace all matched values and erase them.
3046 for (auto MV : Matched) {
3047 MV.first->replaceAllUsesWith(MV.second);
3048 PhiNodesToMatch.erase(MV.first);
3049 ST.Put(MV.first, MV.second);
3050 MV.first->eraseFromParent();
3051 }
3052 Matched.clear();
3053 continue;
3054 }
3055 // If we are not allowed to create new nodes then bail out.
3056 if (!AllowNewPhiNodes)
3057 return false;
3058 // Just remove all seen values in matcher. They will not match anything.
3059 PhiNotMatchedCount += WillNotMatch.size();
3060 for (auto *P : WillNotMatch)
3061 PhiNodesToMatch.erase(P);
3062 }
3063 return true;
3064 }
3065 /// \brief Fill the placeholder with values from predecessors and simplify it.
3066 void FillPlaceholders(FoldAddrToValueMapping &Map,
3067 SmallVectorImpl<ValueInBB> &TraverseOrder,
3068 SimplificationTracker &ST) {
3069 while (!TraverseOrder.empty()) {
3070 auto Current = TraverseOrder.pop_back_val();
3071 assert(Map.find(Current) != Map.end() && "No node to fill!!!")(static_cast <bool> (Map.find(Current) != Map.end() &&
"No node to fill!!!") ? void (0) : __assert_fail ("Map.find(Current) != Map.end() && \"No node to fill!!!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3071, __extension__ __PRETTY_FUNCTION__))
;
3072 Value *CurrentValue = Current.first;
3073 BasicBlock *CurrentBlock = Current.second;
3074 Value *V = Map[Current];
3075
3076 if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3077 // CurrentValue also must be Select.
3078 auto *CurrentSelect = cast<SelectInst>(CurrentValue);
3079 auto *TrueValue = CurrentSelect->getTrueValue();
3080 ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
3081 ? CurrentBlock
3082 : nullptr };
3083 assert(Map.find(TrueItem) != Map.end() && "No True Value!")(static_cast <bool> (Map.find(TrueItem) != Map.end() &&
"No True Value!") ? void (0) : __assert_fail ("Map.find(TrueItem) != Map.end() && \"No True Value!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3083, __extension__ __PRETTY_FUNCTION__))
;
3084 Select->setTrueValue(Map[TrueItem]);
3085 auto *FalseValue = CurrentSelect->getFalseValue();
3086 ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
3087 ? CurrentBlock
3088 : nullptr };
3089 assert(Map.find(FalseItem) != Map.end() && "No False Value!")(static_cast <bool> (Map.find(FalseItem) != Map.end() &&
"No False Value!") ? void (0) : __assert_fail ("Map.find(FalseItem) != Map.end() && \"No False Value!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3089, __extension__ __PRETTY_FUNCTION__))
;
3090 Select->setFalseValue(Map[FalseItem]);
3091 } else {
3092 // Must be a Phi node then.
3093 PHINode *PHI = cast<PHINode>(V);
3094 // Fill the Phi node with values from predecessors.
3095 bool IsDefinedInThisBB =
3096 cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3097 auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3098 for (auto B : predecessors(CurrentBlock)) {
3099 Value *PV = IsDefinedInThisBB
3100 ? CurrentPhi->getIncomingValueForBlock(B)
3101 : CurrentValue;
3102 ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3103 assert(Map.find(item) != Map.end() && "No predecessor Value!")(static_cast <bool> (Map.find(item) != Map.end() &&
"No predecessor Value!") ? void (0) : __assert_fail ("Map.find(item) != Map.end() && \"No predecessor Value!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3103, __extension__ __PRETTY_FUNCTION__))
;
3104 PHI->addIncoming(ST.Get(Map[item]), B);
3105 }
3106 }
3107 // Simplify if possible.
3108 Map[Current] = ST.Simplify(V);
3109 }
3110 }
3111
3112 /// Starting from value recursively iterates over predecessors up to known
3113 /// ending values represented in a map. For each traversed block inserts
3114 /// a placeholder Phi or Select.
3115 /// Reports all new created Phi/Select nodes by adding them to set.
3116 /// Also reports and order in what basic blocks have been traversed.
3117 void InsertPlaceholders(FoldAddrToValueMapping &Map,
3118 SmallVectorImpl<ValueInBB> &TraverseOrder,
3119 SmallPtrSetImpl<PHINode *> &NewPhiNodes,
3120 SmallPtrSetImpl<SelectInst *> &NewSelectNodes) {
3121 SmallVector<ValueInBB, 32> Worklist;
3122 assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&(static_cast <bool> ((isa<PHINode>(Original.first
) || isa<SelectInst>(Original.first)) && "Address must be a Phi or Select node"
) ? void (0) : __assert_fail ("(isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) && \"Address must be a Phi or Select node\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3123, __extension__ __PRETTY_FUNCTION__))
3123 "Address must be a Phi or Select node")(static_cast <bool> ((isa<PHINode>(Original.first
) || isa<SelectInst>(Original.first)) && "Address must be a Phi or Select node"
) ? void (0) : __assert_fail ("(isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) && \"Address must be a Phi or Select node\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3123, __extension__ __PRETTY_FUNCTION__))
;
3124 auto *Dummy = UndefValue::get(CommonType);
3125 Worklist.push_back(Original);
3126 while (!Worklist.empty()) {
39
Loop condition is true. Entering loop body
3127 auto Current = Worklist.pop_back_val();
3128 // If value is not an instruction it is something global, constant,
3129 // parameter and we can say that this value is observable in any block.
3130 // Set block to null to denote it.
3131 // Also please take into account that it is how we build anchors.
3132 if (!isa<Instruction>(Current.first))
40
Taking true branch
3133 Current.second = nullptr;
41
Null pointer value stored to 'Current.second'
3134 // if it is already visited or it is an ending value then skip it.
3135 if (Map.find(Current) != Map.end())
42
Assuming the condition is false
43
Taking false branch
3136 continue;
3137 TraverseOrder.push_back(Current);
3138
3139 Value *CurrentValue = Current.first;
3140 BasicBlock *CurrentBlock = Current.second;
44
'CurrentBlock' initialized to a null pointer value
3141 // CurrentValue must be a Phi node or select. All others must be covered
3142 // by anchors.
3143 Instruction *CurrentI = cast<Instruction>(CurrentValue);
3144 bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
45
Assuming the condition is false
3145
3146 unsigned PredCount =
3147 std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
3148 // if Current Value is not defined in this basic block we are interested
3149 // in values in predecessors.
3150 if (!IsDefinedInThisBB) {
46
Taking true branch
3151 assert(PredCount && "Unreachable block?!")(static_cast <bool> (PredCount && "Unreachable block?!"
) ? void (0) : __assert_fail ("PredCount && \"Unreachable block?!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3151, __extension__ __PRETTY_FUNCTION__))
;
3152 PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3153 &CurrentBlock->front());
47
Called C++ object pointer is null
3154 Map[Current] = PHI;
3155 NewPhiNodes.insert(PHI);
3156 // Add all predecessors in work list.
3157 for (auto B : predecessors(CurrentBlock))
3158 Worklist.push_back({ CurrentValue, B });
3159 continue;
3160 }
3161 // Value is defined in this basic block.
3162 if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3163 // Is it OK to get metadata from OrigSelect?!
3164 // Create a Select placeholder with dummy value.
3165 SelectInst *Select =
3166 SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3167 OrigSelect->getName(), OrigSelect, OrigSelect);
3168 Map[Current] = Select;
3169 NewSelectNodes.insert(Select);
3170 // We are interested in True and False value in this basic block.
3171 Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3172 Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3173 } else {
3174 // It must be a Phi node then.
3175 auto *CurrentPhi = cast<PHINode>(CurrentI);
3176 // Create new Phi node for merge of bases.
3177 assert(PredCount && "Unreachable block?!")(static_cast <bool> (PredCount && "Unreachable block?!"
) ? void (0) : __assert_fail ("PredCount && \"Unreachable block?!\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3177, __extension__ __PRETTY_FUNCTION__))
;
3178 PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3179 &CurrentBlock->front());
3180 Map[Current] = PHI;
3181 NewPhiNodes.insert(PHI);
3182
3183 // Add all predecessors in work list.
3184 for (auto B : predecessors(CurrentBlock))
3185 Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3186 }
3187 }
3188 }
3189};
3190} // end anonymous namespace
3191
3192/// Try adding ScaleReg*Scale to the current addressing mode.
3193/// Return true and update AddrMode if this addr mode is legal for the target,
3194/// false if not.
3195bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3196 unsigned Depth) {
3197 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3198 // mode. Just process that directly.
3199 if (Scale == 1)
3200 return matchAddr(ScaleReg, Depth);
3201
3202 // If the scale is 0, it takes nothing to add this.
3203 if (Scale == 0)
3204 return true;
3205
3206 // If we already have a scale of this value, we can add to it, otherwise, we
3207 // need an available scale field.
3208 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3209 return false;
3210
3211 ExtAddrMode TestAddrMode = AddrMode;
3212
3213 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3214 // [A+B + A*7] -> [B+A*8].
3215 TestAddrMode.Scale += Scale;
3216 TestAddrMode.ScaledReg = ScaleReg;
3217
3218 // If the new address isn't legal, bail out.
3219 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3220 return false;
3221
3222 // It was legal, so commit it.
3223 AddrMode = TestAddrMode;
3224
3225 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3226 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3227 // X*Scale + C*Scale to addr mode.
3228 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3229 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3230 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3231 TestAddrMode.ScaledReg = AddLHS;
3232 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3233
3234 // If this addressing mode is legal, commit it and remember that we folded
3235 // this instruction.
3236 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3237 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3238 AddrMode = TestAddrMode;
3239 return true;
3240 }
3241 }
3242
3243 // Otherwise, not (x+c)*scale, just return what we have.
3244 return true;
3245}
3246
3247/// This is a little filter, which returns true if an addressing computation
3248/// involving I might be folded into a load/store accessing it.
3249/// This doesn't need to be perfect, but needs to accept at least
3250/// the set of instructions that MatchOperationAddr can.
3251static bool MightBeFoldableInst(Instruction *I) {
3252 switch (I->getOpcode()) {
3253 case Instruction::BitCast:
3254 case Instruction::AddrSpaceCast:
3255 // Don't touch identity bitcasts.
3256 if (I->getType() == I->getOperand(0)->getType())
3257 return false;
3258 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3259 case Instruction::PtrToInt:
3260 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3261 return true;
3262 case Instruction::IntToPtr:
3263 // We know the input is intptr_t, so this is foldable.
3264 return true;
3265 case Instruction::Add:
3266 return true;
3267 case Instruction::Mul:
3268 case Instruction::Shl:
3269 // Can only handle X*C and X << C.
3270 return isa<ConstantInt>(I->getOperand(1));
3271 case Instruction::GetElementPtr:
3272 return true;
3273 default:
3274 return false;
3275 }
3276}
3277
3278/// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3279/// \note \p Val is assumed to be the product of some type promotion.
3280/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3281/// to be legal, as the non-promoted value would have had the same state.
3282static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3283 const DataLayout &DL, Value *Val) {
3284 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3285 if (!PromotedInst)
3286 return false;
3287 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3288 // If the ISDOpcode is undefined, it was undefined before the promotion.
3289 if (!ISDOpcode)
3290 return true;
3291 // Otherwise, check if the promoted instruction is legal or not.
3292 return TLI.isOperationLegalOrCustom(
3293 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3294}
3295
3296namespace {
3297
3298/// \brief Hepler class to perform type promotion.
3299class TypePromotionHelper {
3300 /// \brief Utility function to check whether or not a sign or zero extension
3301 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3302 /// either using the operands of \p Inst or promoting \p Inst.
3303 /// The type of the extension is defined by \p IsSExt.
3304 /// In other words, check if:
3305 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3306 /// #1 Promotion applies:
3307 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3308 /// #2 Operand reuses:
3309 /// ext opnd1 to ConsideredExtType.
3310 /// \p PromotedInsts maps the instructions to their type before promotion.
3311 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3312 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3313
3314 /// \brief Utility function to determine if \p OpIdx should be promoted when
3315 /// promoting \p Inst.
3316 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3317 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3318 }
3319
3320 /// \brief Utility function to promote the operand of \p Ext when this
3321 /// operand is a promotable trunc or sext or zext.
3322 /// \p PromotedInsts maps the instructions to their type before promotion.
3323 /// \p CreatedInstsCost[out] contains the cost of all instructions
3324 /// created to promote the operand of Ext.
3325 /// Newly added extensions are inserted in \p Exts.
3326 /// Newly added truncates are inserted in \p Truncs.
3327 /// Should never be called directly.
3328 /// \return The promoted value which is used instead of Ext.
3329 static Value *promoteOperandForTruncAndAnyExt(
3330 Instruction *Ext, TypePromotionTransaction &TPT,
3331 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3332 SmallVectorImpl<Instruction *> *Exts,
3333 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3334
3335 /// \brief Utility function to promote the operand of \p Ext when this
3336 /// operand is promotable and is not a supported trunc or sext.
3337 /// \p PromotedInsts maps the instructions to their type before promotion.
3338 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3339 /// created to promote the operand of Ext.
3340 /// Newly added extensions are inserted in \p Exts.
3341 /// Newly added truncates are inserted in \p Truncs.
3342 /// Should never be called directly.
3343 /// \return The promoted value which is used instead of Ext.
3344 static Value *promoteOperandForOther(Instruction *Ext,
3345 TypePromotionTransaction &TPT,
3346 InstrToOrigTy &PromotedInsts,
3347 unsigned &CreatedInstsCost,
3348 SmallVectorImpl<Instruction *> *Exts,
3349 SmallVectorImpl<Instruction *> *Truncs,
3350 const TargetLowering &TLI, bool IsSExt);
3351
3352 /// \see promoteOperandForOther.
3353 static Value *signExtendOperandForOther(
3354 Instruction *Ext, TypePromotionTransaction &TPT,
3355 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3356 SmallVectorImpl<Instruction *> *Exts,
3357 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3358 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3359 Exts, Truncs, TLI, true);
3360 }
3361
3362 /// \see promoteOperandForOther.
3363 static Value *zeroExtendOperandForOther(
3364 Instruction *Ext, TypePromotionTransaction &TPT,
3365 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3366 SmallVectorImpl<Instruction *> *Exts,
3367 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3368 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3369 Exts, Truncs, TLI, false);
3370 }
3371
3372public:
3373 /// Type for the utility function that promotes the operand of Ext.
3374 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3375 InstrToOrigTy &PromotedInsts,
3376 unsigned &CreatedInstsCost,
3377 SmallVectorImpl<Instruction *> *Exts,
3378 SmallVectorImpl<Instruction *> *Truncs,
3379 const TargetLowering &TLI);
3380
3381 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3382 /// action to promote the operand of \p Ext instead of using Ext.
3383 /// \return NULL if no promotable action is possible with the current
3384 /// sign extension.
3385 /// \p InsertedInsts keeps track of all the instructions inserted by the
3386 /// other CodeGenPrepare optimizations. This information is important
3387 /// because we do not want to promote these instructions as CodeGenPrepare
3388 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3389 /// \p PromotedInsts maps the instructions to their type before promotion.
3390 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3391 const TargetLowering &TLI,
3392 const InstrToOrigTy &PromotedInsts);
3393};
3394
3395} // end anonymous namespace
3396
3397bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3398 Type *ConsideredExtType,
3399 const InstrToOrigTy &PromotedInsts,
3400 bool IsSExt) {
3401 // The promotion helper does not know how to deal with vector types yet.
3402 // To be able to fix that, we would need to fix the places where we
3403 // statically extend, e.g., constants and such.
3404 if (Inst->getType()->isVectorTy())
3405 return false;
3406
3407 // We can always get through zext.
3408 if (isa<ZExtInst>(Inst))
3409 return true;
3410
3411 // sext(sext) is ok too.
3412 if (IsSExt && isa<SExtInst>(Inst))
3413 return true;
3414
3415 // We can get through binary operator, if it is legal. In other words, the
3416 // binary operator must have a nuw or nsw flag.
3417 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3418 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3419 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3420 (IsSExt && BinOp->hasNoSignedWrap())))
3421 return true;
3422
3423 // Check if we can do the following simplification.
3424 // ext(trunc(opnd)) --> ext(opnd)
3425 if (!isa<TruncInst>(Inst))
3426 return false;
3427
3428 Value *OpndVal = Inst->getOperand(0);
3429 // Check if we can use this operand in the extension.
3430 // If the type is larger than the result type of the extension, we cannot.
3431 if (!OpndVal->getType()->isIntegerTy() ||
3432 OpndVal->getType()->getIntegerBitWidth() >
3433 ConsideredExtType->getIntegerBitWidth())
3434 return false;
3435
3436 // If the operand of the truncate is not an instruction, we will not have
3437 // any information on the dropped bits.
3438 // (Actually we could for constant but it is not worth the extra logic).
3439 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3440 if (!Opnd)
3441 return false;
3442
3443 // Check if the source of the type is narrow enough.
3444 // I.e., check that trunc just drops extended bits of the same kind of
3445 // the extension.
3446 // #1 get the type of the operand and check the kind of the extended bits.
3447 const Type *OpndType;
3448 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3449 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3450 OpndType = It->second.getPointer();
3451 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3452 OpndType = Opnd->getOperand(0)->getType();
3453 else
3454 return false;
3455
3456 // #2 check that the truncate just drops extended bits.
3457 return Inst->getType()->getIntegerBitWidth() >=
3458 OpndType->getIntegerBitWidth();
3459}
3460
3461TypePromotionHelper::Action TypePromotionHelper::getAction(
3462 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3463 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3464 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&(static_cast <bool> ((isa<SExtInst>(Ext) || isa<
ZExtInst>(Ext)) && "Unexpected instruction type") ?
void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3465, __extension__ __PRETTY_FUNCTION__))
3465 "Unexpected instruction type")(static_cast <bool> ((isa<SExtInst>(Ext) || isa<
ZExtInst>(Ext)) && "Unexpected instruction type") ?
void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3465, __extension__ __PRETTY_FUNCTION__))
;
3466 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3467 Type *ExtTy = Ext->getType();
3468 bool IsSExt = isa<SExtInst>(Ext);
3469 // If the operand of the extension is not an instruction, we cannot
3470 // get through.
3471 // If it, check we can get through.
3472 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3473 return nullptr;
3474
3475 // Do not promote if the operand has been added by codegenprepare.
3476 // Otherwise, it means we are undoing an optimization that is likely to be
3477 // redone, thus causing potential infinite loop.
3478 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3479 return nullptr;
3480
3481 // SExt or Trunc instructions.
3482 // Return the related handler.
3483 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3484 isa<ZExtInst>(ExtOpnd))
3485 return promoteOperandForTruncAndAnyExt;
3486
3487 // Regular instruction.
3488 // Abort early if we will have to insert non-free instructions.
3489 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3490 return nullptr;
3491 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3492}
3493
3494Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3495 Instruction *SExt, TypePromotionTransaction &TPT,
3496 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3497 SmallVectorImpl<Instruction *> *Exts,
3498 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3499 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3500 // get through it and this method should not be called.
3501 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3502 Value *ExtVal = SExt;
3503 bool HasMergedNonFreeExt = false;
3504 if (isa<ZExtInst>(SExtOpnd)) {
3505 // Replace s|zext(zext(opnd))
3506 // => zext(opnd).
3507 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3508 Value *ZExt =
3509 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3510 TPT.replaceAllUsesWith(SExt, ZExt);
3511 TPT.eraseInstruction(SExt);
3512 ExtVal = ZExt;
3513 } else {
3514 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3515 // => z|sext(opnd).
3516 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3517 }
3518 CreatedInstsCost = 0;
3519
3520 // Remove dead code.
3521 if (SExtOpnd->use_empty())
3522 TPT.eraseInstruction(SExtOpnd);
3523
3524 // Check if the extension is still needed.
3525 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3526 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3527 if (ExtInst) {
3528 if (Exts)
3529 Exts->push_back(ExtInst);
3530 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3531 }
3532 return ExtVal;
3533 }
3534
3535 // At this point we have: ext ty opnd to ty.
3536 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3537 Value *NextVal = ExtInst->getOperand(0);
3538 TPT.eraseInstruction(ExtInst, NextVal);
3539 return NextVal;
3540}
3541
3542Value *TypePromotionHelper::promoteOperandForOther(
3543 Instruction *Ext, TypePromotionTransaction &TPT,
3544 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3545 SmallVectorImpl<Instruction *> *Exts,
3546 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3547 bool IsSExt) {
3548 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3549 // get through it and this method should not be called.
3550 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3551 CreatedInstsCost = 0;
3552 if (!ExtOpnd->hasOneUse()) {
3553 // ExtOpnd will be promoted.
3554 // All its uses, but Ext, will need to use a truncated value of the
3555 // promoted version.
3556 // Create the truncate now.
3557 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3558 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3559 // Insert it just after the definition.
3560 ITrunc->moveAfter(ExtOpnd);
3561 if (Truncs)
3562 Truncs->push_back(ITrunc);
3563 }
3564
3565 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3566 // Restore the operand of Ext (which has been replaced by the previous call
3567 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3568 TPT.setOperand(Ext, 0, ExtOpnd);
3569 }
3570
3571 // Get through the Instruction:
3572 // 1. Update its type.
3573 // 2. Replace the uses of Ext by Inst.
3574 // 3. Extend each operand that needs to be extended.
3575
3576 // Remember the original type of the instruction before promotion.
3577 // This is useful to know that the high bits are sign extended bits.
3578 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3579 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3580 // Step #1.
3581 TPT.mutateType(ExtOpnd, Ext->getType());
3582 // Step #2.
3583 TPT.replaceAllUsesWith(Ext, ExtOpnd);
3584 // Step #3.
3585 Instruction *ExtForOpnd = Ext;
3586
3587 DEBUG(dbgs() << "Propagate Ext to operands\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Propagate Ext to operands\n"
; } } while (false)
;
3588 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3589 ++OpIdx) {
3590 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Operand:\n" << *
(ExtOpnd->getOperand(OpIdx)) << '\n'; } } while (false
)
;
3591 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3592 !shouldExtOperand(ExtOpnd, OpIdx)) {
3593 DEBUG(dbgs() << "No need to propagate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "No need to propagate\n"
; } } while (false)
;
3594 continue;
3595 }
3596 // Check if we can statically extend the operand.
3597 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3598 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3599 DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
3600 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3601 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3602 : Cst->getValue().zext(BitWidth);
3603 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3604 continue;
3605 }
3606 // UndefValue are typed, so we have to statically sign extend them.
3607 if (isa<UndefValue>(Opnd)) {
3608 DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
3609 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3610 continue;
3611 }
3612
3613 // Otherwise we have to explicity sign extend the operand.
3614 // Check if Ext was reused to extend an operand.
3615 if (!ExtForOpnd) {
3616 // If yes, create a new one.
3617 DEBUG(dbgs() << "More operands to ext\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "More operands to ext\n"
; } } while (false)
;
3618 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3619 : TPT.createZExt(Ext, Opnd, Ext->getType());
3620 if (!isa<Instruction>(ValForExtOpnd)) {
3621 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3622 continue;
3623 }
3624 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3625 }
3626 if (Exts)
3627 Exts->push_back(ExtForOpnd);
3628 TPT.setOperand(ExtForOpnd, 0, Opnd);
3629
3630 // Move the sign extension before the insertion point.
3631 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3632 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3633 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3634 // If more sext are required, new instructions will have to be created.
3635 ExtForOpnd = nullptr;
3636 }
3637 if (ExtForOpnd == Ext) {
3638 DEBUG(dbgs() << "Extension is useless now\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Extension is useless now\n"
; } } while (false)
;
3639 TPT.eraseInstruction(Ext);
3640 }
3641 return ExtOpnd;
3642}
3643
3644/// Check whether or not promoting an instruction to a wider type is profitable.
3645/// \p NewCost gives the cost of extension instructions created by the
3646/// promotion.
3647/// \p OldCost gives the cost of extension instructions before the promotion
3648/// plus the number of instructions that have been
3649/// matched in the addressing mode the promotion.
3650/// \p PromotedOperand is the value that has been promoted.
3651/// \return True if the promotion is profitable, false otherwise.
3652bool AddressingModeMatcher::isPromotionProfitable(
3653 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3654 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "OldCost: " << OldCost
<< "\tNewCost: " << NewCost << '\n'; } } while
(false)
;
3655 // The cost of the new extensions is greater than the cost of the
3656 // old extension plus what we folded.
3657 // This is not profitable.
3658 if (NewCost > OldCost)
3659 return false;
3660 if (NewCost < OldCost)
3661 return true;
3662 // The promotion is neutral but it may help folding the sign extension in
3663 // loads for instance.
3664 // Check that we did not create an illegal instruction.
3665 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3666}
3667
3668/// Given an instruction or constant expr, see if we can fold the operation
3669/// into the addressing mode. If so, update the addressing mode and return
3670/// true, otherwise return false without modifying AddrMode.
3671/// If \p MovedAway is not NULL, it contains the information of whether or
3672/// not AddrInst has to be folded into the addressing mode on success.
3673/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3674/// because it has been moved away.
3675/// Thus AddrInst must not be added in the matched instructions.
3676/// This state can happen when AddrInst is a sext, since it may be moved away.
3677/// Therefore, AddrInst may not be valid when MovedAway is true and it must
3678/// not be referenced anymore.
3679bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3680 unsigned Depth,
3681 bool *MovedAway) {
3682 // Avoid exponential behavior on extremely deep expression trees.
3683 if (Depth >= 5) return false;
3684
3685 // By default, all matched instructions stay in place.
3686 if (MovedAway)
3687 *MovedAway = false;
3688
3689 switch (Opcode) {
3690 case Instruction::PtrToInt:
3691 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3692 return matchAddr(AddrInst->getOperand(0), Depth);
3693 case Instruction::IntToPtr: {
3694 auto AS = AddrInst->getType()->getPointerAddressSpace();
3695 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3696 // This inttoptr is a no-op if the integer type is pointer sized.
3697 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3698 return matchAddr(AddrInst->getOperand(0), Depth);
3699 return false;
3700 }
3701 case Instruction::BitCast:
3702 // BitCast is always a noop, and we can handle it as long as it is
3703 // int->int or pointer->pointer (we don't want int<->fp or something).
3704 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3705 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3706 // Don't touch identity bitcasts. These were probably put here by LSR,
3707 // and we don't want to mess around with them. Assume it knows what it
3708 // is doing.
3709 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3710 return matchAddr(AddrInst->getOperand(0), Depth);
3711 return false;
3712 case Instruction::AddrSpaceCast: {
3713 unsigned SrcAS
3714 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3715 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3716 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3717 return matchAddr(AddrInst->getOperand(0), Depth);
3718 return false;
3719 }
3720 case Instruction::Add: {
3721 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3722 ExtAddrMode BackupAddrMode = AddrMode;
3723 unsigned OldSize = AddrModeInsts.size();
3724 // Start a transaction at this point.
3725 // The LHS may match but not the RHS.
3726 // Therefore, we need a higher level restoration point to undo partially
3727 // matched operation.
3728 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3729 TPT.getRestorationPoint();
3730
3731 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3732 matchAddr(AddrInst->getOperand(0), Depth+1))
3733 return true;
3734
3735 // Restore the old addr mode info.
3736 AddrMode = BackupAddrMode;
3737 AddrModeInsts.resize(OldSize);
3738 TPT.rollback(LastKnownGood);
3739
3740 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3741 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3742 matchAddr(AddrInst->getOperand(1), Depth+1))
3743 return true;
3744
3745 // Otherwise we definitely can't merge the ADD in.
3746 AddrMode = BackupAddrMode;
3747 AddrModeInsts.resize(OldSize);
3748 TPT.rollback(LastKnownGood);
3749 break;
3750 }
3751 //case Instruction::Or:
3752 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3753 //break;
3754 case Instruction::Mul:
3755 case Instruction::Shl: {
3756 // Can only handle X*C and X << C.
3757 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3758 if (!RHS || RHS->getBitWidth() > 64)
3759 return false;
3760 int64_t Scale = RHS->getSExtValue();
3761 if (Opcode == Instruction::Shl)
3762 Scale = 1LL << Scale;
3763
3764 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3765 }
3766 case Instruction::GetElementPtr: {
3767 // Scan the GEP. We check it if it contains constant offsets and at most
3768 // one variable offset.
3769 int VariableOperand = -1;
3770 unsigned VariableScale = 0;
3771
3772 int64_t ConstantOffset = 0;
3773 gep_type_iterator GTI = gep_type_begin(AddrInst);
3774 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3775 if (StructType *STy = GTI.getStructTypeOrNull()) {
3776 const StructLayout *SL = DL.getStructLayout(STy);
3777 unsigned Idx =
3778 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3779 ConstantOffset += SL->getElementOffset(Idx);
3780 } else {
3781 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3782 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3783 ConstantOffset += CI->getSExtValue()*TypeSize;
3784 } else if (TypeSize) { // Scales of zero don't do anything.
3785 // We only allow one variable index at the moment.
3786 if (VariableOperand != -1)
3787 return false;
3788
3789 // Remember the variable index.
3790 VariableOperand = i;
3791 VariableScale = TypeSize;
3792 }
3793 }
3794 }
3795
3796 // A common case is for the GEP to only do a constant offset. In this case,
3797 // just add it to the disp field and check validity.
3798 if (VariableOperand == -1) {
3799 AddrMode.BaseOffs += ConstantOffset;
3800 if (ConstantOffset == 0 ||
3801 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3802 // Check to see if we can fold the base pointer in too.
3803 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3804 return true;
3805 }
3806 AddrMode.BaseOffs -= ConstantOffset;
3807 return false;
3808 }
3809
3810 // Save the valid addressing mode in case we can't match.
3811 ExtAddrMode BackupAddrMode = AddrMode;
3812 unsigned OldSize = AddrModeInsts.size();
3813
3814 // See if the scale and offset amount is valid for this target.
3815 AddrMode.BaseOffs += ConstantOffset;
3816
3817 // Match the base operand of the GEP.
3818 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3819 // If it couldn't be matched, just stuff the value in a register.
3820 if (AddrMode.HasBaseReg) {
3821 AddrMode = BackupAddrMode;
3822 AddrModeInsts.resize(OldSize);
3823 return false;
3824 }
3825 AddrMode.HasBaseReg = true;
3826 AddrMode.BaseReg = AddrInst->getOperand(0);
3827 }
3828
3829 // Match the remaining variable portion of the GEP.
3830 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3831 Depth)) {
3832 // If it couldn't be matched, try stuffing the base into a register
3833 // instead of matching it, and retrying the match of the scale.
3834 AddrMode = BackupAddrMode;
3835 AddrModeInsts.resize(OldSize);
3836 if (AddrMode.HasBaseReg)
3837 return false;
3838 AddrMode.HasBaseReg = true;
3839 AddrMode.BaseReg = AddrInst->getOperand(0);
3840 AddrMode.BaseOffs += ConstantOffset;
3841 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3842 VariableScale, Depth)) {
3843 // If even that didn't work, bail.
3844 AddrMode = BackupAddrMode;
3845 AddrModeInsts.resize(OldSize);
3846 return false;
3847 }
3848 }
3849
3850 return true;
3851 }
3852 case Instruction::SExt:
3853 case Instruction::ZExt: {
3854 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3855 if (!Ext)
3856 return false;
3857
3858 // Try to move this ext out of the way of the addressing mode.
3859 // Ask for a method for doing so.
3860 TypePromotionHelper::Action TPH =
3861 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3862 if (!TPH)
3863 return false;
3864
3865 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3866 TPT.getRestorationPoint();
3867 unsigned CreatedInstsCost = 0;
3868 unsigned ExtCost = !TLI.isExtFree(Ext);
3869 Value *PromotedOperand =
3870 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3871 // SExt has been moved away.
3872 // Thus either it will be rematched later in the recursive calls or it is
3873 // gone. Anyway, we must not fold it into the addressing mode at this point.
3874 // E.g.,
3875 // op = add opnd, 1
3876 // idx = ext op
3877 // addr = gep base, idx
3878 // is now:
3879 // promotedOpnd = ext opnd <- no match here
3880 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3881 // addr = gep base, op <- match
3882 if (MovedAway)
3883 *MovedAway = true;
3884
3885 assert(PromotedOperand &&(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3886, __extension__ __PRETTY_FUNCTION__))
3886 "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 3886, __extension__ __PRETTY_FUNCTION__))
;
3887
3888 ExtAddrMode BackupAddrMode = AddrMode;
3889 unsigned OldSize = AddrModeInsts.size();
3890
3891 if (!matchAddr(PromotedOperand, Depth) ||
3892 // The total of the new cost is equal to the cost of the created
3893 // instructions.
3894 // The total of the old cost is equal to the cost of the extension plus
3895 // what we have saved in the addressing mode.
3896 !isPromotionProfitable(CreatedInstsCost,
3897 ExtCost + (AddrModeInsts.size() - OldSize),
3898 PromotedOperand)) {
3899 AddrMode = BackupAddrMode;
3900 AddrModeInsts.resize(OldSize);
3901 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Sign extension does not pay off: rollback\n"
; } } while (false)
;
3902 TPT.rollback(LastKnownGood);
3903 return false;
3904 }
3905 return true;
3906 }
3907 }
3908 return false;
3909}
3910
3911/// If we can, try to add the value of 'Addr' into the current addressing mode.
3912/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3913/// unmodified. This assumes that Addr is either a pointer type or intptr_t
3914/// for the target.
3915///
3916bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3917 // Start a transaction at this point that we will rollback if the matching
3918 // fails.
3919 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3920 TPT.getRestorationPoint();
3921 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3922 // Fold in immediates if legal for the target.
3923 AddrMode.BaseOffs += CI->getSExtValue();
3924 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3925 return true;
3926 AddrMode.BaseOffs -= CI->getSExtValue();
3927 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3928 // If this is a global variable, try to fold it into the addressing mode.
3929 if (!AddrMode.BaseGV) {
3930 AddrMode.BaseGV = GV;
3931 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3932 return true;
3933 AddrMode.BaseGV = nullptr;
3934 }
3935 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3936 ExtAddrMode BackupAddrMode = AddrMode;
3937 unsigned OldSize = AddrModeInsts.size();
3938
3939 // Check to see if it is possible to fold this operation.
3940 bool MovedAway = false;
3941 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3942 // This instruction may have been moved away. If so, there is nothing
3943 // to check here.
3944 if (MovedAway)
3945 return true;
3946 // Okay, it's possible to fold this. Check to see if it is actually
3947 // *profitable* to do so. We use a simple cost model to avoid increasing
3948 // register pressure too much.
3949 if (I->hasOneUse() ||
3950 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3951 AddrModeInsts.push_back(I);
3952 return true;
3953 }
3954
3955 // It isn't profitable to do this, roll back.
3956 //cerr << "NOT FOLDING: " << *I;
3957 AddrMode = BackupAddrMode;
3958 AddrModeInsts.resize(OldSize);
3959 TPT.rollback(LastKnownGood);
3960 }
3961 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3962 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3963 return true;
3964 TPT.rollback(LastKnownGood);
3965 } else if (isa<ConstantPointerNull>(Addr)) {
3966 // Null pointer gets folded without affecting the addressing mode.
3967 return true;
3968 }
3969
3970 // Worse case, the target should support [reg] addressing modes. :)
3971 if (!AddrMode.HasBaseReg) {
3972 AddrMode.HasBaseReg = true;
3973 AddrMode.BaseReg = Addr;
3974 // Still check for legality in case the target supports [imm] but not [i+r].
3975 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3976 return true;
3977 AddrMode.HasBaseReg = false;
3978 AddrMode.BaseReg = nullptr;
3979 }
3980
3981 // If the base register is already taken, see if we can do [r+r].
3982 if (AddrMode.Scale == 0) {
3983 AddrMode.Scale = 1;
3984 AddrMode.ScaledReg = Addr;
3985 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3986 return true;
3987 AddrMode.Scale = 0;
3988 AddrMode.ScaledReg = nullptr;
3989 }
3990 // Couldn't match.
3991 TPT.rollback(LastKnownGood);
3992 return false;
3993}
3994
3995/// Check to see if all uses of OpVal by the specified inline asm call are due
3996/// to memory operands. If so, return true, otherwise return false.
3997static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3998 const TargetLowering &TLI,
3999 const TargetRegisterInfo &TRI) {
4000 const Function *F = CI->getFunction();
4001 TargetLowering::AsmOperandInfoVector TargetConstraints =
4002 TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
4003 ImmutableCallSite(CI));
4004
4005 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4006 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4007
4008 // Compute the constraint code and ConstraintType to use.
4009 TLI.ComputeConstraintToUse(OpInfo, SDValue());
4010
4011 // If this asm operand is our Value*, and if it isn't an indirect memory
4012 // operand, we can't fold it!
4013 if (OpInfo.CallOperandVal == OpVal &&
4014 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4015 !OpInfo.isIndirect))
4016 return false;
4017 }
4018
4019 return true;
4020}
4021
4022// Max number of memory uses to look at before aborting the search to conserve
4023// compile time.
4024static constexpr int MaxMemoryUsesToScan = 20;
4025
4026/// Recursively walk all the uses of I until we find a memory use.
4027/// If we find an obviously non-foldable instruction, return true.
4028/// Add the ultimately found memory instructions to MemoryUses.
4029static bool FindAllMemoryUses(
4030 Instruction *I,
4031 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4032 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4033 const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4034 // If we already considered this instruction, we're done.
4035 if (!ConsideredInsts.insert(I).second)
4036 return false;
4037
4038 // If this is an obviously unfoldable instruction, bail out.
4039 if (!MightBeFoldableInst(I))
4040 return true;
4041
4042 const bool OptSize = I->getFunction()->optForSize();
4043
4044 // Loop over all the uses, recursively processing them.
4045 for (Use &U : I->uses()) {
4046 // Conservatively return true if we're seeing a large number or a deep chain
4047 // of users. This avoids excessive compilation times in pathological cases.
4048 if (SeenInsts++ >= MaxMemoryUsesToScan)
4049 return true;
4050
4051 Instruction *UserI = cast<Instruction>(U.getUser());
4052 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4053 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4054 continue;
4055 }
4056
4057 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4058 unsigned opNo = U.getOperandNo();
4059 if (opNo != StoreInst::getPointerOperandIndex())
4060 return true; // Storing addr, not into addr.
4061 MemoryUses.push_back(std::make_pair(SI, opNo));
4062 continue;
4063 }
4064
4065 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4066 unsigned opNo = U.getOperandNo();
4067 if (opNo != AtomicRMWInst::getPointerOperandIndex())
4068 return true; // Storing addr, not into addr.
4069 MemoryUses.push_back(std::make_pair(RMW, opNo));
4070 continue;
4071 }
4072
4073 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4074 unsigned opNo = U.getOperandNo();
4075 if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
4076 return true; // Storing addr, not into addr.
4077 MemoryUses.push_back(std::make_pair(CmpX, opNo));
4078 continue;
4079 }
4080
4081 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4082 // If this is a cold call, we can sink the addressing calculation into
4083 // the cold path. See optimizeCallInst
4084 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4085 continue;
4086
4087 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4088 if (!IA) return true;
4089
4090 // If this is a memory operand, we're cool, otherwise bail out.
4091 if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4092 return true;
4093 continue;
4094 }
4095
4096 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4097 SeenInsts))
4098 return true;
4099 }
4100
4101 return false;
4102}
4103
4104/// Return true if Val is already known to be live at the use site that we're
4105/// folding it into. If so, there is no cost to include it in the addressing
4106/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4107/// instruction already.
4108bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4109 Value *KnownLive2) {
4110 // If Val is either of the known-live values, we know it is live!
4111 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4112 return true;
4113
4114 // All values other than instructions and arguments (e.g. constants) are live.
4115 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4116
4117 // If Val is a constant sized alloca in the entry block, it is live, this is
4118 // true because it is just a reference to the stack/frame pointer, which is
4119 // live for the whole function.
4120 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4121 if (AI->isStaticAlloca())
4122 return true;
4123
4124 // Check to see if this value is already used in the memory instruction's
4125 // block. If so, it's already live into the block at the very least, so we
4126 // can reasonably fold it.
4127 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4128}
4129
4130/// It is possible for the addressing mode of the machine to fold the specified
4131/// instruction into a load or store that ultimately uses it.
4132/// However, the specified instruction has multiple uses.
4133/// Given this, it may actually increase register pressure to fold it
4134/// into the load. For example, consider this code:
4135///
4136/// X = ...
4137/// Y = X+1
4138/// use(Y) -> nonload/store
4139/// Z = Y+1
4140/// load Z
4141///
4142/// In this case, Y has multiple uses, and can be folded into the load of Z
4143/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4144/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4145/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4146/// number of computations either.
4147///
4148/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4149/// X was live across 'load Z' for other reasons, we actually *would* want to
4150/// fold the addressing mode in the Z case. This would make Y die earlier.
4151bool AddressingModeMatcher::
4152isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4153 ExtAddrMode &AMAfter) {
4154 if (IgnoreProfitability) return true;
4155
4156 // AMBefore is the addressing mode before this instruction was folded into it,
4157 // and AMAfter is the addressing mode after the instruction was folded. Get
4158 // the set of registers referenced by AMAfter and subtract out those
4159 // referenced by AMBefore: this is the set of values which folding in this
4160 // address extends the lifetime of.
4161 //
4162 // Note that there are only two potential values being referenced here,
4163 // BaseReg and ScaleReg (global addresses are always available, as are any
4164 // folded immediates).
4165 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4166
4167 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4168 // lifetime wasn't extended by adding this instruction.
4169 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4170 BaseReg = nullptr;
4171 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4172 ScaledReg = nullptr;
4173
4174 // If folding this instruction (and it's subexprs) didn't extend any live
4175 // ranges, we're ok with it.
4176 if (!BaseReg && !ScaledReg)
4177 return true;
4178
4179 // If all uses of this instruction can have the address mode sunk into them,
4180 // we can remove the addressing mode and effectively trade one live register
4181 // for another (at worst.) In this context, folding an addressing mode into
4182 // the use is just a particularly nice way of sinking it.
4183 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4184 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4185 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4186 return false; // Has a non-memory, non-foldable use!
4187
4188 // Now that we know that all uses of this instruction are part of a chain of
4189 // computation involving only operations that could theoretically be folded
4190 // into a memory use, loop over each of these memory operation uses and see
4191 // if they could *actually* fold the instruction. The assumption is that
4192 // addressing modes are cheap and that duplicating the computation involved
4193 // many times is worthwhile, even on a fastpath. For sinking candidates
4194 // (i.e. cold call sites), this serves as a way to prevent excessive code
4195 // growth since most architectures have some reasonable small and fast way to
4196 // compute an effective address. (i.e LEA on x86)
4197 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4198 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4199 Instruction *User = MemoryUses[i].first;
4200 unsigned OpNo = MemoryUses[i].second;
4201
4202 // Get the access type of this use. If the use isn't a pointer, we don't
4203 // know what it accesses.
4204 Value *Address = User->getOperand(OpNo);
4205 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4206 if (!AddrTy)
4207 return false;
4208 Type *AddressAccessTy = AddrTy->getElementType();
4209 unsigned AS = AddrTy->getAddressSpace();
4210
4211 // Do a match against the root of this address, ignoring profitability. This
4212 // will tell us if the addressing mode for the memory operation will
4213 // *actually* cover the shared instruction.
4214 ExtAddrMode Result;
4215 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4216 TPT.getRestorationPoint();
4217 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
4218 AddressAccessTy, AS,
4219 MemoryInst, Result, InsertedInsts,
4220 PromotedInsts, TPT);
4221 Matcher.IgnoreProfitability = true;
4222 bool Success = Matcher.matchAddr(Address, 0);
4223 (void)Success; assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?"
) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4223, __extension__ __PRETTY_FUNCTION__))
;
4224
4225 // The match was to check the profitability, the changes made are not
4226 // part of the original matcher. Therefore, they should be dropped
4227 // otherwise the original matcher will not present the right state.
4228 TPT.rollback(LastKnownGood);
4229
4230 // If the match didn't cover I, then it won't be shared by it.
4231 if (!is_contained(MatchedAddrModeInsts, I))
4232 return false;
4233
4234 MatchedAddrModeInsts.clear();
4235 }
4236
4237 return true;
4238}
4239
4240/// Return true if the specified values are defined in a
4241/// different basic block than BB.
4242static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4243 if (Instruction *I = dyn_cast<Instruction>(V))
4244 return I->getParent() != BB;
4245 return false;
4246}
4247
4248/// Sink addressing mode computation immediate before MemoryInst if doing so
4249/// can be done without increasing register pressure. The need for the
4250/// register pressure constraint means this can end up being an all or nothing
4251/// decision for all uses of the same addressing computation.
4252///
4253/// Load and Store Instructions often have addressing modes that can do
4254/// significant amounts of computation. As such, instruction selection will try
4255/// to get the load or store to do as much computation as possible for the
4256/// program. The problem is that isel can only see within a single block. As
4257/// such, we sink as much legal addressing mode work into the block as possible.
4258///
4259/// This method is used to optimize both load/store and inline asms with memory
4260/// operands. It's also used to sink addressing computations feeding into cold
4261/// call sites into their (cold) basic block.
4262///
4263/// The motivation for handling sinking into cold blocks is that doing so can
4264/// both enable other address mode sinking (by satisfying the register pressure
4265/// constraint above), and reduce register pressure globally (by removing the
4266/// addressing mode computation from the fast path entirely.).
4267bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4268 Type *AccessTy, unsigned AddrSpace) {
4269 Value *Repl = Addr;
4270
4271 // Try to collapse single-value PHI nodes. This is necessary to undo
4272 // unprofitable PRE transformations.
4273 SmallVector<Value*, 8> worklist;
4274 SmallPtrSet<Value*, 16> Visited;
4275 worklist.push_back(Addr);
4276
4277 // Use a worklist to iteratively look through PHI and select nodes, and
4278 // ensure that the addressing mode obtained from the non-PHI/select roots of
4279 // the graph are compatible.
4280 bool PhiOrSelectSeen = false;
4281 SmallVector<Instruction*, 16> AddrModeInsts;
4282 const SimplifyQuery SQ(*DL, TLInfo);
4283 AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4284 TypePromotionTransaction TPT(RemovedInsts);
4285 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4286 TPT.getRestorationPoint();
4287 while (!worklist.empty()) {
1
Loop condition is true. Entering loop body
7
Loop condition is true. Entering loop body
13
Loop condition is true. Entering loop body
19
Loop condition is true. Entering loop body
4288 Value *V = worklist.back();
4289 worklist.pop_back();
4290
4291 // We allow traversing cyclic Phi nodes.
4292 // In case of success after this loop we ensure that traversing through
4293 // Phi nodes ends up with all cases to compute address of the form
4294 // BaseGV + Base + Scale * Index + Offset
4295 // where Scale and Offset are constans and BaseGV, Base and Index
4296 // are exactly the same Values in all cases.
4297 // It means that BaseGV, Scale and Offset dominate our memory instruction
4298 // and have the same value as they had in address computation represented
4299 // as Phi. So we can safely sink address computation to memory instruction.
4300 if (!Visited.insert(V).second)
2
Assuming the condition is false
3
Taking false branch
8
Assuming the condition is false
9
Taking false branch
14
Assuming the condition is false
15
Taking false branch
20
Assuming the condition is false
21
Taking false branch
4301 continue;
4302
4303 // For a PHI node, push all of its incoming values.
4304 if (PHINode *P = dyn_cast<PHINode>(V)) {
4
Taking false branch
10
Taking false branch
16
Taking false branch
22
Taking false branch
4305 for (Value *IncValue : P->incoming_values())
4306 worklist.push_back(IncValue);
4307 PhiOrSelectSeen = true;
4308 continue;
4309 }
4310 // Similar for select.
4311 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
5
Taking false branch
11
Taking false branch
17
Taking false branch
23
Taking false branch
4312 worklist.push_back(SI->getFalseValue());
4313 worklist.push_back(SI->getTrueValue());
4314 PhiOrSelectSeen = true;
4315 continue;
4316 }
4317
4318 // For non-PHIs, determine the addressing mode being computed. Note that
4319 // the result may differ depending on what other uses our candidate
4320 // addressing instructions might have.
4321 AddrModeInsts.clear();
4322 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4323 V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4324 InsertedInsts, PromotedInsts, TPT);
4325 NewAddrMode.OriginalValue = V;
4326
4327 if (!AddrModes.addNewAddrMode(NewAddrMode))
6
Taking false branch
12
Taking false branch
18
Taking false branch
24
Taking true branch
4328 break;
25
Execution continues on line 4334
4329 }
4330
4331 // Try to combine the AddrModes we've collected. If we couldn't collect any,
4332 // or we have multiple but either couldn't combine them or combining them
4333 // wouldn't do anything useful, bail out now.
4334 if (!AddrModes.combineAddrModes()) {
26
Calling 'AddressingModeCombiner::combineAddrModes'
4335 TPT.rollback(LastKnownGood);
4336 return false;
4337 }
4338 TPT.commit();
4339
4340 // Get the combined AddrMode (or the only AddrMode, if we only had one).
4341 ExtAddrMode AddrMode = AddrModes.getAddrMode();
4342
4343 // If all the instructions matched are already in this BB, don't do anything.
4344 // If we saw a Phi node then it is not local definitely, and if we saw a select
4345 // then we want to push the address calculation past it even if it's already
4346 // in this BB.
4347 if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4348 return IsNonLocalValue(V, MemoryInst->getParent());
4349 })) {
4350 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Found local addrmode: "
<< AddrMode << "\n"; } } while (false)
;
4351 return false;
4352 }
4353
4354 // Insert this computation right after this user. Since our caller is
4355 // scanning from the top of the BB to the bottom, reuse of the expr are
4356 // guaranteed to happen later.
4357 IRBuilder<> Builder(MemoryInst);
4358
4359 // Now that we determined the addressing expression we want to use and know
4360 // that we have to sink it into this block. Check to see if we have already
4361 // done this for some other load/store instr in this block. If so, reuse the
4362 // computation.
4363 Value *&SunkAddr = SunkAddrs[Addr];
4364 if (SunkAddr) {
4365 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4366 << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4367 if (SunkAddr->getType() != Addr->getType())
4368 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4369 } else if (AddrSinkUsingGEPs ||
4370 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4371 SubtargetInfo->useAA())) {
4372 // By default, we use the GEP-based method when AA is used later. This
4373 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4374 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4375 << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4376 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4377 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4378
4379 // First, find the pointer.
4380 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4381 ResultPtr = AddrMode.BaseReg;
4382 AddrMode.BaseReg = nullptr;
4383 }
4384
4385 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4386 // We can't add more than one pointer together, nor can we scale a
4387 // pointer (both of which seem meaningless).
4388 if (ResultPtr || AddrMode.Scale != 1)
4389 return false;
4390
4391 ResultPtr = AddrMode.ScaledReg;
4392 AddrMode.Scale = 0;
4393 }
4394
4395 // It is only safe to sign extend the BaseReg if we know that the math
4396 // required to create it did not overflow before we extend it. Since
4397 // the original IR value was tossed in favor of a constant back when
4398 // the AddrMode was created we need to bail out gracefully if widths
4399 // do not match instead of extending it.
4400 //
4401 // (See below for code to add the scale.)
4402 if (AddrMode.Scale) {
4403 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4404 if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4405 cast<IntegerType>(ScaledRegTy)->getBitWidth())
4406 return false;
4407 }
4408
4409 if (AddrMode.BaseGV) {
4410 if (ResultPtr)
4411 return false;
4412
4413 ResultPtr = AddrMode.BaseGV;
4414 }
4415
4416 // If the real base value actually came from an inttoptr, then the matcher
4417 // will look through it and provide only the integer value. In that case,
4418 // use it here.
4419 if (!DL->isNonIntegralPointerType(Addr->getType())) {
4420 if (!ResultPtr && AddrMode.BaseReg) {
4421 ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4422 "sunkaddr");
4423 AddrMode.BaseReg = nullptr;
4424 } else if (!ResultPtr && AddrMode.Scale == 1) {
4425 ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4426 "sunkaddr");
4427 AddrMode.Scale = 0;
4428 }
4429 }
4430
4431 if (!ResultPtr &&
4432 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4433 SunkAddr = Constant::getNullValue(Addr->getType());
4434 } else if (!ResultPtr) {
4435 return false;
4436 } else {
4437 Type *I8PtrTy =
4438 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4439 Type *I8Ty = Builder.getInt8Ty();
4440
4441 // Start with the base register. Do this first so that subsequent address
4442 // matching finds it last, which will prevent it from trying to match it
4443 // as the scaled value in case it happens to be a mul. That would be
4444 // problematic if we've sunk a different mul for the scale, because then
4445 // we'd end up sinking both muls.
4446 if (AddrMode.BaseReg) {
4447 Value *V = AddrMode.BaseReg;
4448 if (V->getType() != IntPtrTy)
4449 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4450
4451 ResultIndex = V;
4452 }
4453
4454 // Add the scale value.
4455 if (AddrMode.Scale) {
4456 Value *V = AddrMode.ScaledReg;
4457 if (V->getType() == IntPtrTy) {
4458 // done.
4459 } else {
4460 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4462, __extension__ __PRETTY_FUNCTION__))
4461 cast<IntegerType>(V->getType())->getBitWidth() &&(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4462, __extension__ __PRETTY_FUNCTION__))
4462 "We can't transform if ScaledReg is too narrow")(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4462, __extension__ __PRETTY_FUNCTION__))
;
4463 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4464 }
4465
4466 if (AddrMode.Scale != 1)
4467 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4468 "sunkaddr");
4469 if (ResultIndex)
4470 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4471 else
4472 ResultIndex = V;
4473 }
4474
4475 // Add in the Base Offset if present.
4476 if (AddrMode.BaseOffs) {
4477 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4478 if (ResultIndex) {
4479 // We need to add this separately from the scale above to help with
4480 // SDAG consecutive load/store merging.
4481 if (ResultPtr->getType() != I8PtrTy)
4482 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4483 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4484 }
4485
4486 ResultIndex = V;
4487 }
4488
4489 if (!ResultIndex) {
4490 SunkAddr = ResultPtr;
4491 } else {
4492 if (ResultPtr->getType() != I8PtrTy)
4493 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4494 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4495 }
4496
4497 if (SunkAddr->getType() != Addr->getType())
4498 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4499 }
4500 } else {
4501 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4502 // non-integral pointers, so in that case bail out now.
4503 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4504 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4505 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4506 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4507 if (DL->isNonIntegralPointerType(Addr->getType()) ||
4508 (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4509 (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4510 (AddrMode.BaseGV &&
4511 DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4512 return false;
4513
4514 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4515 << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4516 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4517 Value *Result = nullptr;
4518
4519 // Start with the base register. Do this first so that subsequent address
4520 // matching finds it last, which will prevent it from trying to match it
4521 // as the scaled value in case it happens to be a mul. That would be
4522 // problematic if we've sunk a different mul for the scale, because then
4523 // we'd end up sinking both muls.
4524 if (AddrMode.BaseReg) {
4525 Value *V = AddrMode.BaseReg;
4526 if (V->getType()->isPointerTy())
4527 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4528 if (V->getType() != IntPtrTy)
4529 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4530 Result = V;
4531 }
4532
4533 // Add the scale value.
4534 if (AddrMode.Scale) {
4535 Value *V = AddrMode.ScaledReg;
4536 if (V->getType() == IntPtrTy) {
4537 // done.
4538 } else if (V->getType()->isPointerTy()) {
4539 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4540 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4541 cast<IntegerType>(V->getType())->getBitWidth()) {
4542 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4543 } else {
4544 // It is only safe to sign extend the BaseReg if we know that the math
4545 // required to create it did not overflow before we extend it. Since
4546 // the original IR value was tossed in favor of a constant back when
4547 // the AddrMode was created we need to bail out gracefully if widths
4548 // do not match instead of extending it.
4549 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4550 if (I && (Result != AddrMode.BaseReg))
4551 I->eraseFromParent();
4552 return false;
4553 }
4554 if (AddrMode.Scale != 1)
4555 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4556 "sunkaddr");
4557 if (Result)
4558 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4559 else
4560 Result = V;
4561 }
4562
4563 // Add in the BaseGV if present.
4564 if (AddrMode.BaseGV) {
4565 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4566 if (Result)
4567 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4568 else
4569 Result = V;
4570 }
4571
4572 // Add in the Base Offset if present.
4573 if (AddrMode.BaseOffs) {
4574 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4575 if (Result)
4576 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4577 else
4578 Result = V;
4579 }
4580
4581 if (!Result)
4582 SunkAddr = Constant::getNullValue(Addr->getType());
4583 else
4584 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4585 }
4586
4587 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4588
4589 // If we have no uses, recursively delete the value and all dead instructions
4590 // using it.
4591 if (Repl->use_empty()) {
4592 // This can cause recursive deletion, which can invalidate our iterator.
4593 // Use a WeakTrackingVH to hold onto it in case this happens.
4594 Value *CurValue = &*CurInstIterator;
4595 WeakTrackingVH IterHandle(CurValue);
4596 BasicBlock *BB = CurInstIterator->getParent();
4597
4598 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
4599
4600 if (IterHandle != CurValue) {
4601 // If the iterator instruction was recursively deleted, start over at the
4602 // start of the block.
4603 CurInstIterator = BB->begin();
4604 SunkAddrs.clear();
4605 }
4606 }
4607 ++NumMemoryInsts;
4608 return true;
4609}
4610
4611/// If there are any memory operands, use OptimizeMemoryInst to sink their
4612/// address computing into the block when possible / profitable.
4613bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4614 bool MadeChange = false;
4615
4616 const TargetRegisterInfo *TRI =
4617 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4618 TargetLowering::AsmOperandInfoVector TargetConstraints =
4619 TLI->ParseConstraints(*DL, TRI, CS);
4620 unsigned ArgNo = 0;
4621 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4622 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4623
4624 // Compute the constraint code and ConstraintType to use.
4625 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4626
4627 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4628 OpInfo.isIndirect) {
4629 Value *OpVal = CS->getArgOperand(ArgNo++);
4630 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4631 } else if (OpInfo.Type == InlineAsm::isInput)
4632 ArgNo++;
4633 }
4634
4635 return MadeChange;
4636}
4637
4638/// \brief Check if all the uses of \p Val are equivalent (or free) zero or
4639/// sign extensions.
4640static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4641 assert(!Val->use_empty() && "Input must have at least one use")(static_cast <bool> (!Val->use_empty() && "Input must have at least one use"
) ? void (0) : __assert_fail ("!Val->use_empty() && \"Input must have at least one use\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4641, __extension__ __PRETTY_FUNCTION__))
;
4642 const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4643 bool IsSExt = isa<SExtInst>(FirstUser);
4644 Type *ExtTy = FirstUser->getType();
4645 for (const User *U : Val->users()) {
4646 const Instruction *UI = cast<Instruction>(U);
4647 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4648 return false;
4649 Type *CurTy = UI->getType();
4650 // Same input and output types: Same instruction after CSE.
4651 if (CurTy == ExtTy)
4652 continue;
4653
4654 // If IsSExt is true, we are in this situation:
4655 // a = Val
4656 // b = sext ty1 a to ty2
4657 // c = sext ty1 a to ty3
4658 // Assuming ty2 is shorter than ty3, this could be turned into:
4659 // a = Val
4660 // b = sext ty1 a to ty2
4661 // c = sext ty2 b to ty3
4662 // However, the last sext is not free.
4663 if (IsSExt)
4664 return false;
4665
4666 // This is a ZExt, maybe this is free to extend from one type to another.
4667 // In that case, we would not account for a different use.
4668 Type *NarrowTy;
4669 Type *LargeTy;
4670 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4671 CurTy->getScalarType()->getIntegerBitWidth()) {
4672 NarrowTy = CurTy;
4673 LargeTy = ExtTy;
4674 } else {
4675 NarrowTy = ExtTy;
4676 LargeTy = CurTy;
4677 }
4678
4679 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4680 return false;
4681 }
4682 // All uses are the same or can be derived from one another for free.
4683 return true;
4684}
4685
4686/// \brief Try to speculatively promote extensions in \p Exts and continue
4687/// promoting through newly promoted operands recursively as far as doing so is
4688/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4689/// When some promotion happened, \p TPT contains the proper state to revert
4690/// them.
4691///
4692/// \return true if some promotion happened, false otherwise.
4693bool CodeGenPrepare::tryToPromoteExts(
4694 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4695 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4696 unsigned CreatedInstsCost) {
4697 bool Promoted = false;
4698
4699 // Iterate over all the extensions to try to promote them.
4700 for (auto I : Exts) {
4701 // Early check if we directly have ext(load).
4702 if (isa<LoadInst>(I->getOperand(0))) {
4703 ProfitablyMovedExts.push_back(I);
4704 continue;
4705 }
4706
4707 // Check whether or not we want to do any promotion. The reason we have
4708 // this check inside the for loop is to catch the case where an extension
4709 // is directly fed by a load because in such case the extension can be moved
4710 // up without any promotion on its operands.
4711 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4712 return false;
4713
4714 // Get the action to perform the promotion.
4715 TypePromotionHelper::Action TPH =
4716 TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4717 // Check if we can promote.
4718 if (!TPH) {
4719 // Save the current extension as we cannot move up through its operand.
4720 ProfitablyMovedExts.push_back(I);
4721 continue;
4722 }
4723
4724 // Save the current state.
4725 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4726 TPT.getRestorationPoint();
4727 SmallVector<Instruction *, 4> NewExts;
4728 unsigned NewCreatedInstsCost = 0;
4729 unsigned ExtCost = !TLI->isExtFree(I);
4730 // Promote.
4731 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4732 &NewExts, nullptr, *TLI);
4733 assert(PromotedVal &&(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4734, __extension__ __PRETTY_FUNCTION__))
4734 "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4734, __extension__ __PRETTY_FUNCTION__))
;
4735
4736 // We would be able to merge only one extension in a load.
4737 // Therefore, if we have more than 1 new extension we heuristically
4738 // cut this search path, because it means we degrade the code quality.
4739 // With exactly 2, the transformation is neutral, because we will merge
4740 // one extension but leave one. However, we optimistically keep going,
4741 // because the new extension may be removed too.
4742 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4743 // FIXME: It would be possible to propagate a negative value instead of
4744 // conservatively ceiling it to 0.
4745 TotalCreatedInstsCost =
4746 std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4747 if (!StressExtLdPromotion &&
4748 (TotalCreatedInstsCost > 1 ||
4749 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4750 // This promotion is not profitable, rollback to the previous state, and
4751 // save the current extension in ProfitablyMovedExts as the latest
4752 // speculative promotion turned out to be unprofitable.
4753 TPT.rollback(LastKnownGood);
4754 ProfitablyMovedExts.push_back(I);
4755 continue;
4756 }
4757 // Continue promoting NewExts as far as doing so is profitable.
4758 SmallVector<Instruction *, 2> NewlyMovedExts;
4759 (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4760 bool NewPromoted = false;
4761 for (auto ExtInst : NewlyMovedExts) {
4762 Instruction *MovedExt = cast<Instruction>(ExtInst);
4763 Value *ExtOperand = MovedExt->getOperand(0);
4764 // If we have reached to a load, we need this extra profitability check
4765 // as it could potentially be merged into an ext(load).
4766 if (isa<LoadInst>(ExtOperand) &&
4767 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4768 (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4769 continue;
4770
4771 ProfitablyMovedExts.push_back(MovedExt);
4772 NewPromoted = true;
4773 }
4774
4775 // If none of speculative promotions for NewExts is profitable, rollback
4776 // and save the current extension (I) as the last profitable extension.
4777 if (!NewPromoted) {
4778 TPT.rollback(LastKnownGood);
4779 ProfitablyMovedExts.push_back(I);
4780 continue;
4781 }
4782 // The promotion is profitable.
4783 Promoted = true;
4784 }
4785 return Promoted;
4786}
4787
4788/// Merging redundant sexts when one is dominating the other.
4789bool CodeGenPrepare::mergeSExts(Function &F) {
4790 DominatorTree DT(F);
4791 bool Changed = false;
4792 for (auto &Entry : ValToSExtendedUses) {
4793 SExts &Insts = Entry.second;
4794 SExts CurPts;
4795 for (Instruction *Inst : Insts) {
4796 if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4797 Inst->getOperand(0) != Entry.first)
4798 continue;
4799 bool inserted = false;
4800 for (auto &Pt : CurPts) {
4801 if (DT.dominates(Inst, Pt)) {
4802 Pt->replaceAllUsesWith(Inst);
4803 RemovedInsts.insert(Pt);
4804 Pt->removeFromParent();
4805 Pt = Inst;
4806 inserted = true;
4807 Changed = true;
4808 break;
4809 }
4810 if (!DT.dominates(Pt, Inst))
4811 // Give up if we need to merge in a common dominator as the
4812 // expermients show it is not profitable.
4813 continue;
4814 Inst->replaceAllUsesWith(Pt);
4815 RemovedInsts.insert(Inst);
4816 Inst->removeFromParent();
4817 inserted = true;
4818 Changed = true;
4819 break;
4820 }
4821 if (!inserted)
4822 CurPts.push_back(Inst);
4823 }
4824 }
4825 return Changed;
4826}
4827
4828/// Return true, if an ext(load) can be formed from an extension in
4829/// \p MovedExts.
4830bool CodeGenPrepare::canFormExtLd(
4831 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
4832 Instruction *&Inst, bool HasPromoted) {
4833 for (auto *MovedExtInst : MovedExts) {
4834 if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
4835 LI = cast<LoadInst>(MovedExtInst->getOperand(0));
4836 Inst = MovedExtInst;
4837 break;
4838 }
4839 }
4840 if (!LI)
4841 return false;
4842
4843 // If they're already in the same block, there's nothing to do.
4844 // Make the cheap checks first if we did not promote.
4845 // If we promoted, we need to check if it is indeed profitable.
4846 if (!HasPromoted && LI->getParent() == Inst->getParent())
4847 return false;
4848
4849 return TLI->isExtLoad(LI, Inst, *DL);
4850}
4851
4852/// Move a zext or sext fed by a load into the same basic block as the load,
4853/// unless conditions are unfavorable. This allows SelectionDAG to fold the
4854/// extend into the load.
4855///
4856/// E.g.,
4857/// \code
4858/// %ld = load i32* %addr
4859/// %add = add nuw i32 %ld, 4
4860/// %zext = zext i32 %add to i64
4861// \endcode
4862/// =>
4863/// \code
4864/// %ld = load i32* %addr
4865/// %zext = zext i32 %ld to i64
4866/// %add = add nuw i64 %zext, 4
4867/// \encode
4868/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
4869/// allow us to match zext(load i32*) to i64.
4870///
4871/// Also, try to promote the computations used to obtain a sign extended
4872/// value used into memory accesses.
4873/// E.g.,
4874/// \code
4875/// a = add nsw i32 b, 3
4876/// d = sext i32 a to i64
4877/// e = getelementptr ..., i64 d
4878/// \endcode
4879/// =>
4880/// \code
4881/// f = sext i32 b to i64
4882/// a = add nsw i64 f, 3
4883/// e = getelementptr ..., i64 a
4884/// \endcode
4885///
4886/// \p Inst[in/out] the extension may be modified during the process if some
4887/// promotions apply.
4888bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
4889 // ExtLoad formation and address type promotion infrastructure requires TLI to
4890 // be effective.
4891 if (!TLI)
4892 return false;
4893
4894 bool AllowPromotionWithoutCommonHeader = false;
4895 /// See if it is an interesting sext operations for the address type
4896 /// promotion before trying to promote it, e.g., the ones with the right
4897 /// type and used in memory accesses.
4898 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
4899 *Inst, AllowPromotionWithoutCommonHeader);
4900 TypePromotionTransaction TPT(RemovedInsts);
4901 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4902 TPT.getRestorationPoint();
4903 SmallVector<Instruction *, 1> Exts;
4904 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
4905 Exts.push_back(Inst);
4906
4907 bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
4908
4909 // Look for a load being extended.
4910 LoadInst *LI = nullptr;
4911 Instruction *ExtFedByLoad;
4912
4913 // Try to promote a chain of computation if it allows to form an extended
4914 // load.
4915 if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
4916 assert(LI && ExtFedByLoad && "Expect a valid load and extension")(static_cast <bool> (LI && ExtFedByLoad &&
"Expect a valid load and extension") ? void (0) : __assert_fail
("LI && ExtFedByLoad && \"Expect a valid load and extension\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 4916, __extension__ __PRETTY_FUNCTION__))
;
4917 TPT.commit();
4918 // Move the extend into the same block as the load
4919 ExtFedByLoad->moveAfter(LI);
4920 // CGP does not check if the zext would be speculatively executed when moved
4921 // to the same basic block as the load. Preserving its original location
4922 // would pessimize the debugging experience, as well as negatively impact
4923 // the quality of sample pgo. We don't want to use "line 0" as that has a
4924 // size cost in the line-table section and logically the zext can be seen as
4925 // part of the load. Therefore we conservatively reuse the same debug
4926 // location for the load and the zext.
4927 ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
4928 ++NumExtsMoved;
4929 Inst = ExtFedByLoad;
4930 return true;
4931 }
4932
4933 // Continue promoting SExts if known as considerable depending on targets.
4934 if (ATPConsiderable &&
4935 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
4936 HasPromoted, TPT, SpeculativelyMovedExts))
4937 return true;
4938
4939 TPT.rollback(LastKnownGood);
4940 return false;
4941}
4942
4943// Perform address type promotion if doing so is profitable.
4944// If AllowPromotionWithoutCommonHeader == false, we should find other sext
4945// instructions that sign extended the same initial value. However, if
4946// AllowPromotionWithoutCommonHeader == true, we expect promoting the
4947// extension is just profitable.
4948bool CodeGenPrepare::performAddressTypePromotion(
4949 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
4950 bool HasPromoted, TypePromotionTransaction &TPT,
4951 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
4952 bool Promoted = false;
4953 SmallPtrSet<Instruction *, 1> UnhandledExts;
4954 bool AllSeenFirst = true;
4955 for (auto I : SpeculativelyMovedExts) {
4956 Value *HeadOfChain = I->getOperand(0);
4957 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
4958 SeenChainsForSExt.find(HeadOfChain);
4959 // If there is an unhandled SExt which has the same header, try to promote
4960 // it as well.
4961 if (AlreadySeen != SeenChainsForSExt.end()) {
4962 if (AlreadySeen->second != nullptr)
4963 UnhandledExts.insert(AlreadySeen->second);
4964 AllSeenFirst = false;
4965 }
4966 }
4967
4968 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
4969 SpeculativelyMovedExts.size() == 1)) {
4970 TPT.commit();
4971 if (HasPromoted)
4972 Promoted = true;
4973 for (auto I : SpeculativelyMovedExts) {
4974 Value *HeadOfChain = I->getOperand(0);
4975 SeenChainsForSExt[HeadOfChain] = nullptr;
4976 ValToSExtendedUses[HeadOfChain].push_back(I);
4977 }
4978 // Update Inst as promotion happen.
4979 Inst = SpeculativelyMovedExts.pop_back_val();
4980 } else {
4981 // This is the first chain visited from the header, keep the current chain
4982 // as unhandled. Defer to promote this until we encounter another SExt
4983 // chain derived from the same header.
4984 for (auto I : SpeculativelyMovedExts) {
4985 Value *HeadOfChain = I->getOperand(0);
4986 SeenChainsForSExt[HeadOfChain] = Inst;
4987 }
4988 return false;
4989 }
4990
4991 if (!AllSeenFirst && !UnhandledExts.empty())
4992 for (auto VisitedSExt : UnhandledExts) {
4993 if (RemovedInsts.count(VisitedSExt))
4994 continue;
4995 TypePromotionTransaction TPT(RemovedInsts);
4996 SmallVector<Instruction *, 1> Exts;
4997 SmallVector<Instruction *, 2> Chains;
4998 Exts.push_back(VisitedSExt);
4999 bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5000 TPT.commit();
5001 if (HasPromoted)
5002 Promoted = true;
5003 for (auto I : Chains) {
5004 Value *HeadOfChain = I->getOperand(0);
5005 // Mark this as handled.
5006 SeenChainsForSExt[HeadOfChain] = nullptr;
5007 ValToSExtendedUses[HeadOfChain].push_back(I);
5008 }
5009 }
5010 return Promoted;
5011}
5012
5013bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5014 BasicBlock *DefBB = I->getParent();
5015
5016 // If the result of a {s|z}ext and its source are both live out, rewrite all
5017 // other uses of the source with result of extension.
5018 Value *Src = I->getOperand(0);
5019 if (Src->hasOneUse())
5020 return false;
5021
5022 // Only do this xform if truncating is free.
5023 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5024 return false;
5025
5026 // Only safe to perform the optimization if the source is also defined in
5027 // this block.
5028 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5029 return false;
5030
5031 bool DefIsLiveOut = false;
5032 for (User *U : I->users()) {
5033 Instruction *UI = cast<Instruction>(U);
5034
5035 // Figure out which BB this ext is used in.
5036 BasicBlock *UserBB = UI->getParent();
5037 if (UserBB == DefBB) continue;
5038 DefIsLiveOut = true;
5039 break;
5040 }
5041 if (!DefIsLiveOut)
5042 return false;
5043
5044 // Make sure none of the uses are PHI nodes.
5045 for (User *U : Src->users()) {
5046 Instruction *UI = cast<Instruction>(U);
5047 BasicBlock *UserBB = UI->getParent();
5048 if (UserBB == DefBB) continue;
5049 // Be conservative. We don't want this xform to end up introducing
5050 // reloads just before load / store instructions.
5051 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5052 return false;
5053 }
5054
5055 // InsertedTruncs - Only insert one trunc in each block once.
5056 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5057
5058 bool MadeChange = false;
5059 for (Use &U : Src->uses()) {
5060 Instruction *User = cast<Instruction>(U.getUser());
5061
5062 // Figure out which BB this ext is used in.
5063 BasicBlock *UserBB = User->getParent();
5064 if (UserBB == DefBB) continue;
5065
5066 // Both src and def are live in this block. Rewrite the use.
5067 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5068
5069 if (!InsertedTrunc) {
5070 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5071 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5071, __extension__ __PRETTY_FUNCTION__))
;
5072 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5073 InsertedInsts.insert(InsertedTrunc);
5074 }
5075
5076 // Replace a use of the {s|z}ext source with a use of the result.
5077 U = InsertedTrunc;
5078 ++NumExtUses;
5079 MadeChange = true;
5080 }
5081
5082 return MadeChange;
5083}
5084
5085// Find loads whose uses only use some of the loaded value's bits. Add an "and"
5086// just after the load if the target can fold this into one extload instruction,
5087// with the hope of eliminating some of the other later "and" instructions using
5088// the loaded value. "and"s that are made trivially redundant by the insertion
5089// of the new "and" are removed by this function, while others (e.g. those whose
5090// path from the load goes through a phi) are left for isel to potentially
5091// remove.
5092//
5093// For example:
5094//
5095// b0:
5096// x = load i32
5097// ...
5098// b1:
5099// y = and x, 0xff
5100// z = use y
5101//
5102// becomes:
5103//
5104// b0:
5105// x = load i32
5106// x' = and x, 0xff
5107// ...
5108// b1:
5109// z = use x'
5110//
5111// whereas:
5112//
5113// b0:
5114// x1 = load i32
5115// ...
5116// b1:
5117// x2 = load i32
5118// ...
5119// b2:
5120// x = phi x1, x2
5121// y = and x, 0xff
5122//
5123// becomes (after a call to optimizeLoadExt for each load):
5124//
5125// b0:
5126// x1 = load i32
5127// x1' = and x1, 0xff
5128// ...
5129// b1:
5130// x2 = load i32
5131// x2' = and x2, 0xff
5132// ...
5133// b2:
5134// x = phi x1', x2'
5135// y = and x, 0xff
5136bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5137 if (!Load->isSimple() ||
5138 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5139 return false;
5140
5141 // Skip loads we've already transformed.
5142 if (Load->hasOneUse() &&
5143 InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5144 return false;
5145
5146 // Look at all uses of Load, looking through phis, to determine how many bits
5147 // of the loaded value are needed.
5148 SmallVector<Instruction *, 8> WorkList;
5149 SmallPtrSet<Instruction *, 16> Visited;
5150 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5151 for (auto *U : Load->users())
5152 WorkList.push_back(cast<Instruction>(U));
5153
5154 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5155 unsigned BitWidth = LoadResultVT.getSizeInBits();
5156 APInt DemandBits(BitWidth, 0);
5157 APInt WidestAndBits(BitWidth, 0);
5158
5159 while (!WorkList.empty()) {
5160 Instruction *I = WorkList.back();
5161 WorkList.pop_back();
5162
5163 // Break use-def graph loops.
5164 if (!Visited.insert(I).second)
5165 continue;
5166
5167 // For a PHI node, push all of its users.
5168 if (auto *Phi = dyn_cast<PHINode>(I)) {
5169 for (auto *U : Phi->users())
5170 WorkList.push_back(cast<Instruction>(U));
5171 continue;
5172 }
5173
5174 switch (I->getOpcode()) {
5175 case Instruction::And: {
5176 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5177 if (!AndC)
5178 return false;
5179 APInt AndBits = AndC->getValue();
5180 DemandBits |= AndBits;
5181 // Keep track of the widest and mask we see.
5182 if (AndBits.ugt(WidestAndBits))
5183 WidestAndBits = AndBits;
5184 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5185 AndsToMaybeRemove.push_back(I);
5186 break;
5187 }
5188
5189 case Instruction::Shl: {
5190 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5191 if (!ShlC)
5192 return false;
5193 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5194 DemandBits.setLowBits(BitWidth - ShiftAmt);
5195 break;
5196 }
5197
5198 case Instruction::Trunc: {
5199 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5200 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5201 DemandBits.setLowBits(TruncBitWidth);
5202 break;
5203 }
5204
5205 default:
5206 return false;
5207 }
5208 }
5209
5210 uint32_t ActiveBits = DemandBits.getActiveBits();
5211 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5212 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5213 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5214 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5215 // followed by an AND.
5216 // TODO: Look into removing this restriction by fixing backends to either
5217 // return false for isLoadExtLegal for i1 or have them select this pattern to
5218 // a single instruction.
5219 //
5220 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5221 // mask, since these are the only ands that will be removed by isel.
5222 if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5223 WidestAndBits != DemandBits)
5224 return false;
5225
5226 LLVMContext &Ctx = Load->getType()->getContext();
5227 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5228 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5229
5230 // Reject cases that won't be matched as extloads.
5231 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5232 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5233 return false;
5234
5235 IRBuilder<> Builder(Load->getNextNode());
5236 auto *NewAnd = dyn_cast<Instruction>(
5237 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5238 // Mark this instruction as "inserted by CGP", so that other
5239 // optimizations don't touch it.
5240 InsertedInsts.insert(NewAnd);
5241
5242 // Replace all uses of load with new and (except for the use of load in the
5243 // new and itself).
5244 Load->replaceAllUsesWith(NewAnd);
5245 NewAnd->setOperand(0, Load);
5246
5247 // Remove any and instructions that are now redundant.
5248 for (auto *And : AndsToMaybeRemove)
5249 // Check that the and mask is the same as the one we decided to put on the
5250 // new and.
5251 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5252 And->replaceAllUsesWith(NewAnd);
5253 if (&*CurInstIterator == And)
5254 CurInstIterator = std::next(And->getIterator());
5255 And->eraseFromParent();
5256 ++NumAndUses;
5257 }
5258
5259 ++NumAndsAdded;
5260 return true;
5261}
5262
5263/// Check if V (an operand of a select instruction) is an expensive instruction
5264/// that is only used once.
5265static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5266 auto *I = dyn_cast<Instruction>(V);
5267 // If it's safe to speculatively execute, then it should not have side
5268 // effects; therefore, it's safe to sink and possibly *not* execute.
5269 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5270 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5271}
5272
5273/// Returns true if a SelectInst should be turned into an explicit branch.
5274static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5275 const TargetLowering *TLI,
5276 SelectInst *SI) {
5277 // If even a predictable select is cheap, then a branch can't be cheaper.
5278 if (!TLI->isPredictableSelectExpensive())
5279 return false;
5280
5281 // FIXME: This should use the same heuristics as IfConversion to determine
5282 // whether a select is better represented as a branch.
5283
5284 // If metadata tells us that the select condition is obviously predictable,
5285 // then we want to replace the select with a branch.
5286 uint64_t TrueWeight, FalseWeight;
5287 if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5288 uint64_t Max = std::max(TrueWeight, FalseWeight);
5289 uint64_t Sum = TrueWeight + FalseWeight;
5290 if (Sum != 0) {
5291 auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5292 if (Probability > TLI->getPredictableBranchThreshold())
5293 return true;
5294 }
5295 }
5296
5297 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5298
5299 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5300 // comparison condition. If the compare has more than one use, there's
5301 // probably another cmov or setcc around, so it's not worth emitting a branch.
5302 if (!Cmp || !Cmp->hasOneUse())
5303 return false;
5304
5305 // If either operand of the select is expensive and only needed on one side
5306 // of the select, we should form a branch.
5307 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5308 sinkSelectOperand(TTI, SI->getFalseValue()))
5309 return true;
5310
5311 return false;
5312}
5313
5314/// If \p isTrue is true, return the true value of \p SI, otherwise return
5315/// false value of \p SI. If the true/false value of \p SI is defined by any
5316/// select instructions in \p Selects, look through the defining select
5317/// instruction until the true/false value is not defined in \p Selects.
5318static Value *getTrueOrFalseValue(
5319 SelectInst *SI, bool isTrue,
5320 const SmallPtrSet<const Instruction *, 2> &Selects) {
5321 Value *V;
5322
5323 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5324 DefSI = dyn_cast<SelectInst>(V)) {
5325 assert(DefSI->getCondition() == SI->getCondition() &&(static_cast <bool> (DefSI->getCondition() == SI->
getCondition() && "The condition of DefSI does not match with SI"
) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5326, __extension__ __PRETTY_FUNCTION__))
5326 "The condition of DefSI does not match with SI")(static_cast <bool> (DefSI->getCondition() == SI->
getCondition() && "The condition of DefSI does not match with SI"
) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5326, __extension__ __PRETTY_FUNCTION__))
;
5327 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5328 }
5329 return V;
5330}
5331
5332/// If we have a SelectInst that will likely profit from branch prediction,
5333/// turn it into a branch.
5334bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5335 // Find all consecutive select instructions that share the same condition.
5336 SmallVector<SelectInst *, 2> ASI;
5337 ASI.push_back(SI);
5338 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
5339 It != SI->getParent()->end(); ++It) {
5340 SelectInst *I = dyn_cast<SelectInst>(&*It);
5341 if (I && SI->getCondition() == I->getCondition()) {
5342 ASI.push_back(I);
5343 } else {
5344 break;
5345 }
5346 }
5347
5348 SelectInst *LastSI = ASI.back();
5349 // Increment the current iterator to skip all the rest of select instructions
5350 // because they will be either "not lowered" or "all lowered" to branch.
5351 CurInstIterator = std::next(LastSI->getIterator());
5352
5353 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5354
5355 // Can we convert the 'select' to CF ?
5356 if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5357 SI->getMetadata(LLVMContext::MD_unpredictable))
5358 return false;
5359
5360 TargetLowering::SelectSupportKind SelectKind;
5361 if (VectorCond)
5362 SelectKind = TargetLowering::VectorMaskSelect;
5363 else if (SI->getType()->isVectorTy())
5364 SelectKind = TargetLowering::ScalarCondVectorVal;
5365 else
5366 SelectKind = TargetLowering::ScalarValSelect;
5367
5368 if (TLI->isSelectSupported(SelectKind) &&
5369 !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5370 return false;
5371
5372 ModifiedDT = true;
5373
5374 // Transform a sequence like this:
5375 // start:
5376 // %cmp = cmp uge i32 %a, %b
5377 // %sel = select i1 %cmp, i32 %c, i32 %d
5378 //
5379 // Into:
5380 // start:
5381 // %cmp = cmp uge i32 %a, %b
5382 // br i1 %cmp, label %select.true, label %select.false
5383 // select.true:
5384 // br label %select.end
5385 // select.false:
5386 // br label %select.end
5387 // select.end:
5388 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5389 //
5390 // In addition, we may sink instructions that produce %c or %d from
5391 // the entry block into the destination(s) of the new branch.
5392 // If the true or false blocks do not contain a sunken instruction, that
5393 // block and its branch may be optimized away. In that case, one side of the
5394 // first branch will point directly to select.end, and the corresponding PHI
5395 // predecessor block will be the start block.
5396
5397 // First, we split the block containing the select into 2 blocks.
5398 BasicBlock *StartBlock = SI->getParent();
5399 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5400 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5401
5402 // Delete the unconditional branch that was just created by the split.
5403 StartBlock->getTerminator()->eraseFromParent();
5404
5405 // These are the new basic blocks for the conditional branch.
5406 // At least one will become an actual new basic block.
5407 BasicBlock *TrueBlock = nullptr;
5408 BasicBlock *FalseBlock = nullptr;
5409 BranchInst *TrueBranch = nullptr;
5410 BranchInst *FalseBranch = nullptr;
5411
5412 // Sink expensive instructions into the conditional blocks to avoid executing
5413 // them speculatively.
5414 for (SelectInst *SI : ASI) {
5415 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5416 if (TrueBlock == nullptr) {
5417 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5418 EndBlock->getParent(), EndBlock);
5419 TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5420 }
5421 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5422 TrueInst->moveBefore(TrueBranch);
5423 }
5424 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5425 if (FalseBlock == nullptr) {
5426 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5427 EndBlock->getParent(), EndBlock);
5428 FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5429 }
5430 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5431 FalseInst->moveBefore(FalseBranch);
5432 }
5433 }
5434
5435 // If there was nothing to sink, then arbitrarily choose the 'false' side
5436 // for a new input value to the PHI.
5437 if (TrueBlock == FalseBlock) {
5438 assert(TrueBlock == nullptr &&(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5439, __extension__ __PRETTY_FUNCTION__))
5439 "Unexpected basic block transform while optimizing select")(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5439, __extension__ __PRETTY_FUNCTION__))
;
5440
5441 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5442 EndBlock->getParent(), EndBlock);
5443 BranchInst::Create(EndBlock, FalseBlock);
5444 }
5445
5446 // Insert the real conditional branch based on the original condition.
5447 // If we did not create a new block for one of the 'true' or 'false' paths
5448 // of the condition, it means that side of the branch goes to the end block
5449 // directly and the path originates from the start block from the point of
5450 // view of the new PHI.
5451 BasicBlock *TT, *FT;
5452 if (TrueBlock == nullptr) {
5453 TT = EndBlock;
5454 FT = FalseBlock;
5455 TrueBlock = StartBlock;
5456 } else if (FalseBlock == nullptr) {
5457 TT = TrueBlock;
5458 FT = EndBlock;
5459 FalseBlock = StartBlock;
5460 } else {
5461 TT = TrueBlock;
5462 FT = FalseBlock;
5463 }
5464 IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5465
5466 SmallPtrSet<const Instruction *, 2> INS;
5467 INS.insert(ASI.begin(), ASI.end());
5468 // Use reverse iterator because later select may use the value of the
5469 // earlier select, and we need to propagate value through earlier select
5470 // to get the PHI operand.
5471 for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5472 SelectInst *SI = *It;
5473 // The select itself is replaced with a PHI Node.
5474 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5475 PN->takeName(SI);
5476 PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5477 PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5478
5479 SI->replaceAllUsesWith(PN);
5480 SI->eraseFromParent();
5481 INS.erase(SI);
5482 ++NumSelectsExpanded;
5483 }
5484
5485 // Instruct OptimizeBlock to skip to the next block.
5486 CurInstIterator = StartBlock->end();
5487 return true;
5488}
5489
5490static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5491 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5492 int SplatElem = -1;
5493 for (unsigned i = 0; i < Mask.size(); ++i) {
5494 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5495 return false;
5496 SplatElem = Mask[i];
5497 }
5498
5499 return true;
5500}
5501
5502/// Some targets have expensive vector shifts if the lanes aren't all the same
5503/// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5504/// it's often worth sinking a shufflevector splat down to its use so that
5505/// codegen can spot all lanes are identical.
5506bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5507 BasicBlock *DefBB = SVI->getParent();
5508
5509 // Only do this xform if variable vector shifts are particularly expensive.
5510 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5511 return false;
5512
5513 // We only expect better codegen by sinking a shuffle if we can recognise a
5514 // constant splat.
5515 if (!isBroadcastShuffle(SVI))
5516 return false;
5517
5518 // InsertedShuffles - Only insert a shuffle in each block once.
5519 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5520
5521 bool MadeChange = false;
5522 for (User *U : SVI->users()) {
5523 Instruction *UI = cast<Instruction>(U);
5524
5525 // Figure out which BB this ext is used in.
5526 BasicBlock *UserBB = UI->getParent();
5527 if (UserBB == DefBB) continue;
5528
5529 // For now only apply this when the splat is used by a shift instruction.
5530 if (!UI->isShift()) continue;
5531
5532 // Everything checks out, sink the shuffle if the user's block doesn't
5533 // already have a copy.
5534 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5535
5536 if (!InsertedShuffle) {
5537 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5538 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5538, __extension__ __PRETTY_FUNCTION__))
;
5539 InsertedShuffle =
5540 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5541 SVI->getOperand(2), "", &*InsertPt);
5542 }
5543
5544 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5545 MadeChange = true;
5546 }
5547
5548 // If we removed all uses, nuke the shuffle.
5549 if (SVI->use_empty()) {
5550 SVI->eraseFromParent();
5551 MadeChange = true;
5552 }
5553
5554 return MadeChange;
5555}
5556
5557bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5558 if (!TLI || !DL)
5559 return false;
5560
5561 Value *Cond = SI->getCondition();
5562 Type *OldType = Cond->getType();
5563 LLVMContext &Context = Cond->getContext();
5564 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5565 unsigned RegWidth = RegType.getSizeInBits();
5566
5567 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5568 return false;
5569
5570 // If the register width is greater than the type width, expand the condition
5571 // of the switch instruction and each case constant to the width of the
5572 // register. By widening the type of the switch condition, subsequent
5573 // comparisons (for case comparisons) will not need to be extended to the
5574 // preferred register width, so we will potentially eliminate N-1 extends,
5575 // where N is the number of cases in the switch.
5576 auto *NewType = Type::getIntNTy(Context, RegWidth);
5577
5578 // Zero-extend the switch condition and case constants unless the switch
5579 // condition is a function argument that is already being sign-extended.
5580 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5581 // everything instead.
5582 Instruction::CastOps ExtType = Instruction::ZExt;
5583 if (auto *Arg = dyn_cast<Argument>(Cond))
5584 if (Arg->hasSExtAttr())
5585 ExtType = Instruction::SExt;
5586
5587 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5588 ExtInst->insertBefore(SI);
5589 SI->setCondition(ExtInst);
5590 for (auto Case : SI->cases()) {
5591 APInt NarrowConst = Case.getCaseValue()->getValue();
5592 APInt WideConst = (ExtType == Instruction::ZExt) ?
5593 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5594 Case.setValue(ConstantInt::get(Context, WideConst));
5595 }
5596
5597 return true;
5598}
5599
5600
5601namespace {
5602
5603/// \brief Helper class to promote a scalar operation to a vector one.
5604/// This class is used to move downward extractelement transition.
5605/// E.g.,
5606/// a = vector_op <2 x i32>
5607/// b = extractelement <2 x i32> a, i32 0
5608/// c = scalar_op b
5609/// store c
5610///
5611/// =>
5612/// a = vector_op <2 x i32>
5613/// c = vector_op a (equivalent to scalar_op on the related lane)
5614/// * d = extractelement <2 x i32> c, i32 0
5615/// * store d
5616/// Assuming both extractelement and store can be combine, we get rid of the
5617/// transition.
5618class VectorPromoteHelper {
5619 /// DataLayout associated with the current module.
5620 const DataLayout &DL;
5621
5622 /// Used to perform some checks on the legality of vector operations.
5623 const TargetLowering &TLI;
5624
5625 /// Used to estimated the cost of the promoted chain.
5626 const TargetTransformInfo &TTI;
5627
5628 /// The transition being moved downwards.
5629 Instruction *Transition;
5630
5631 /// The sequence of instructions to be promoted.
5632 SmallVector<Instruction *, 4> InstsToBePromoted;
5633
5634 /// Cost of combining a store and an extract.
5635 unsigned StoreExtractCombineCost;
5636
5637 /// Instruction that will be combined with the transition.
5638 Instruction *CombineInst = nullptr;
5639
5640 /// \brief The instruction that represents the current end of the transition.
5641 /// Since we are faking the promotion until we reach the end of the chain
5642 /// of computation, we need a way to get the current end of the transition.
5643 Instruction *getEndOfTransition() const {
5644 if (InstsToBePromoted.empty())
5645 return Transition;
5646 return InstsToBePromoted.back();
5647 }
5648
5649 /// \brief Return the index of the original value in the transition.
5650 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5651 /// c, is at index 0.
5652 unsigned getTransitionOriginalValueIdx() const {
5653 assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5654, __extension__ __PRETTY_FUNCTION__))
5654 "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5654, __extension__ __PRETTY_FUNCTION__))
;
5655 return 0;
5656 }
5657
5658 /// \brief Return the index of the index in the transition.
5659 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5660 /// is at index 1.
5661 unsigned getTransitionIdx() const {
5662 assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5663, __extension__ __PRETTY_FUNCTION__))
5663 "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5663, __extension__ __PRETTY_FUNCTION__))
;
5664 return 1;
5665 }
5666
5667 /// \brief Get the type of the transition.
5668 /// This is the type of the original value.
5669 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5670 /// transition is <2 x i32>.
5671 Type *getTransitionType() const {
5672 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5673 }
5674
5675 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5676 /// I.e., we have the following sequence:
5677 /// Def = Transition <ty1> a to <ty2>
5678 /// b = ToBePromoted <ty2> Def, ...
5679 /// =>
5680 /// b = ToBePromoted <ty1> a, ...
5681 /// Def = Transition <ty1> ToBePromoted to <ty2>
5682 void promoteImpl(Instruction *ToBePromoted);
5683
5684 /// \brief Check whether or not it is profitable to promote all the
5685 /// instructions enqueued to be promoted.
5686 bool isProfitableToPromote() {
5687 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5688 unsigned Index = isa<ConstantInt>(ValIdx)
5689 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5690 : -1;
5691 Type *PromotedType = getTransitionType();
5692
5693 StoreInst *ST = cast<StoreInst>(CombineInst);
5694 unsigned AS = ST->getPointerAddressSpace();
5695 unsigned Align = ST->getAlignment();
5696 // Check if this store is supported.
5697 if (!TLI.allowsMisalignedMemoryAccesses(
5698 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5699 Align)) {
5700 // If this is not supported, there is no way we can combine
5701 // the extract with the store.
5702 return false;
5703 }
5704
5705 // The scalar chain of computation has to pay for the transition
5706 // scalar to vector.
5707 // The vector chain has to account for the combining cost.
5708 uint64_t ScalarCost =
5709 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5710 uint64_t VectorCost = StoreExtractCombineCost;
5711 for (const auto &Inst : InstsToBePromoted) {
5712 // Compute the cost.
5713 // By construction, all instructions being promoted are arithmetic ones.
5714 // Moreover, one argument is a constant that can be viewed as a splat
5715 // constant.
5716 Value *Arg0 = Inst->getOperand(0);
5717 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5718 isa<ConstantFP>(Arg0);
5719 TargetTransformInfo::OperandValueKind Arg0OVK =
5720 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5721 : TargetTransformInfo::OK_AnyValue;
5722 TargetTransformInfo::OperandValueKind Arg1OVK =
5723 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5724 : TargetTransformInfo::OK_AnyValue;
5725 ScalarCost += TTI.getArithmeticInstrCost(
5726 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5727 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5728 Arg0OVK, Arg1OVK);
5729 }
5730 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
<< ScalarCost << "\nVector: " << VectorCost
<< '\n'; } } while (false)
5731 << ScalarCost << "\nVector: " << VectorCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
<< ScalarCost << "\nVector: " << VectorCost
<< '\n'; } } while (false)
;
5732 return ScalarCost > VectorCost;
5733 }
5734
5735 /// \brief Generate a constant vector with \p Val with the same
5736 /// number of elements as the transition.
5737 /// \p UseSplat defines whether or not \p Val should be replicated
5738 /// across the whole vector.
5739 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5740 /// otherwise we generate a vector with as many undef as possible:
5741 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5742 /// used at the index of the extract.
5743 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5744 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
5745 if (!UseSplat) {
5746 // If we cannot determine where the constant must be, we have to
5747 // use a splat constant.
5748 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5749 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5750 ExtractIdx = CstVal->getSExtValue();
5751 else
5752 UseSplat = true;
5753 }
5754
5755 unsigned End = getTransitionType()->getVectorNumElements();
5756 if (UseSplat)
5757 return ConstantVector::getSplat(End, Val);
5758
5759 SmallVector<Constant *, 4> ConstVec;
5760 UndefValue *UndefVal = UndefValue::get(Val->getType());
5761 for (unsigned Idx = 0; Idx != End; ++Idx) {
5762 if (Idx == ExtractIdx)
5763 ConstVec.push_back(Val);
5764 else
5765 ConstVec.push_back(UndefVal);
5766 }
5767 return ConstantVector::get(ConstVec);
5768 }
5769
5770 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5771 /// in \p Use can trigger undefined behavior.
5772 static bool canCauseUndefinedBehavior(const Instruction *Use,
5773 unsigned OperandIdx) {
5774 // This is not safe to introduce undef when the operand is on
5775 // the right hand side of a division-like instruction.
5776 if (OperandIdx != 1)
5777 return false;
5778 switch (Use->getOpcode()) {
5779 default:
5780 return false;
5781 case Instruction::SDiv:
5782 case Instruction::UDiv:
5783 case Instruction::SRem:
5784 case Instruction::URem:
5785 return true;
5786 case Instruction::FDiv:
5787 case Instruction::FRem:
5788 return !Use->hasNoNaNs();
5789 }
5790 llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5790)
;
5791 }
5792
5793public:
5794 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5795 const TargetTransformInfo &TTI, Instruction *Transition,
5796 unsigned CombineCost)
5797 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5798 StoreExtractCombineCost(CombineCost) {
5799 assert(Transition && "Do not know how to promote null")(static_cast <bool> (Transition && "Do not know how to promote null"
) ? void (0) : __assert_fail ("Transition && \"Do not know how to promote null\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5799, __extension__ __PRETTY_FUNCTION__))
;
5800 }
5801
5802 /// \brief Check if we can promote \p ToBePromoted to \p Type.
5803 bool canPromote(const Instruction *ToBePromoted) const {
5804 // We could support CastInst too.
5805 return isa<BinaryOperator>(ToBePromoted);
5806 }
5807
5808 /// \brief Check if it is profitable to promote \p ToBePromoted
5809 /// by moving downward the transition through.
5810 bool shouldPromote(const Instruction *ToBePromoted) const {
5811 // Promote only if all the operands can be statically expanded.
5812 // Indeed, we do not want to introduce any new kind of transitions.
5813 for (const Use &U : ToBePromoted->operands()) {
5814 const Value *Val = U.get();
5815 if (Val == getEndOfTransition()) {
5816 // If the use is a division and the transition is on the rhs,
5817 // we cannot promote the operation, otherwise we may create a
5818 // division by zero.
5819 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5820 return false;
5821 continue;
5822 }
5823 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5824 !isa<ConstantFP>(Val))
5825 return false;
5826 }
5827 // Check that the resulting operation is legal.
5828 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5829 if (!ISDOpcode)
5830 return false;
5831 return StressStoreExtract ||
5832 TLI.isOperationLegalOrCustom(
5833 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5834 }
5835
5836 /// \brief Check whether or not \p Use can be combined
5837 /// with the transition.
5838 /// I.e., is it possible to do Use(Transition) => AnotherUse?
5839 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5840
5841 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5842 void enqueueForPromotion(Instruction *ToBePromoted) {
5843 InstsToBePromoted.push_back(ToBePromoted);
5844 }
5845
5846 /// \brief Set the instruction that will be combined with the transition.
5847 void recordCombineInstruction(Instruction *ToBeCombined) {
5848 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine")(static_cast <bool> (canCombine(ToBeCombined) &&
"Unsupported instruction to combine") ? void (0) : __assert_fail
("canCombine(ToBeCombined) && \"Unsupported instruction to combine\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5848, __extension__ __PRETTY_FUNCTION__))
;
5849 CombineInst = ToBeCombined;
5850 }
5851
5852 /// \brief Promote all the instructions enqueued for promotion if it is
5853 /// is profitable.
5854 /// \return True if the promotion happened, false otherwise.
5855 bool promote() {
5856 // Check if there is something to promote.
5857 // Right now, if we do not have anything to combine with,
5858 // we assume the promotion is not profitable.
5859 if (InstsToBePromoted.empty() || !CombineInst)
5860 return false;
5861
5862 // Check cost.
5863 if (!StressStoreExtract && !isProfitableToPromote())
5864 return false;
5865
5866 // Promote.
5867 for (auto &ToBePromoted : InstsToBePromoted)
5868 promoteImpl(ToBePromoted);
5869 InstsToBePromoted.clear();
5870 return true;
5871 }
5872};
5873
5874} // end anonymous namespace
5875
5876void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
5877 // At this point, we know that all the operands of ToBePromoted but Def
5878 // can be statically promoted.
5879 // For Def, we need to use its parameter in ToBePromoted:
5880 // b = ToBePromoted ty1 a
5881 // Def = Transition ty1 b to ty2
5882 // Move the transition down.
5883 // 1. Replace all uses of the promoted operation by the transition.
5884 // = ... b => = ... Def.
5885 assert(ToBePromoted->getType() == Transition->getType() &&(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5887, __extension__ __PRETTY_FUNCTION__))
5886 "The type of the result of the transition does not match "(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5887, __extension__ __PRETTY_FUNCTION__))
5887 "the final type")(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5887, __extension__ __PRETTY_FUNCTION__))
;
5888 ToBePromoted->replaceAllUsesWith(Transition);
5889 // 2. Update the type of the uses.
5890 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5891 Type *TransitionTy = getTransitionType();
5892 ToBePromoted->mutateType(TransitionTy);
5893 // 3. Update all the operands of the promoted operation with promoted
5894 // operands.
5895 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5896 for (Use &U : ToBePromoted->operands()) {
5897 Value *Val = U.get();
5898 Value *NewVal = nullptr;
5899 if (Val == Transition)
5900 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5901 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5902 isa<ConstantFP>(Val)) {
5903 // Use a splat constant if it is not safe to use undef.
5904 NewVal = getConstantVector(
5905 cast<Constant>(Val),
5906 isa<UndefValue>(Val) ||
5907 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5908 } else
5909 llvm_unreachable("Did you modified shouldPromote and forgot to update "::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update "
"this?", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5910)
5910 "this?")::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update "
"this?", "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 5910)
;
5911 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5912 }
5913 Transition->moveAfter(ToBePromoted);
5914 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5915}
5916
5917/// Some targets can do store(extractelement) with one instruction.
5918/// Try to push the extractelement towards the stores when the target
5919/// has this feature and this is profitable.
5920bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5921 unsigned CombineCost = std::numeric_limits<unsigned>::max();
5922 if (DisableStoreExtract || !TLI ||
5923 (!StressStoreExtract &&
5924 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5925 Inst->getOperand(1), CombineCost)))
5926 return false;
5927
5928 // At this point we know that Inst is a vector to scalar transition.
5929 // Try to move it down the def-use chain, until:
5930 // - We can combine the transition with its single use
5931 // => we got rid of the transition.
5932 // - We escape the current basic block
5933 // => we would need to check that we are moving it at a cheaper place and
5934 // we do not do that for now.
5935 BasicBlock *Parent = Inst->getParent();
5936 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Found an interesting transition: "
<< *Inst << '\n'; } } while (false)
;
5937 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5938 // If the transition has more than one use, assume this is not going to be
5939 // beneficial.
5940 while (Inst->hasOneUse()) {
5941 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5942 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Use: " << *ToBePromoted
<< '\n'; } } while (false)
;
5943
5944 if (ToBePromoted->getParent() != Parent) {
5945 DEBUG(dbgs() << "Instruction to promote is in a different block ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
5946 << ToBePromoted->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
5947 << ") than the transition (" << Parent->getName() << ").\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
;
5948 return false;
5949 }
5950
5951 if (VPH.canCombine(ToBePromoted)) {
5952 DEBUG(dbgs() << "Assume " << *Inst << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Assume " << *Inst
<< '\n' << "will be combined with: " << *ToBePromoted
<< '\n'; } } while (false)
5953 << "will be combined with: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Assume " << *Inst
<< '\n' << "will be combined with: " << *ToBePromoted
<< '\n'; } } while (false)
;
5954 VPH.recordCombineInstruction(ToBePromoted);
5955 bool Changed = VPH.promote();
5956 NumStoreExtractExposed += Changed;
5957 return Changed;
5958 }
5959
5960 DEBUG(dbgs() << "Try promoting.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Try promoting.\n"; } }
while (false)
;
5961 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5962 return false;
5963
5964 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Promoting is possible... Enqueue for promotion!\n"
; } } while (false)
;
5965
5966 VPH.enqueueForPromotion(ToBePromoted);
5967 Inst = ToBePromoted;
5968 }
5969 return false;
5970}
5971
5972/// For the instruction sequence of store below, F and I values
5973/// are bundled together as an i64 value before being stored into memory.
5974/// Sometimes it is more efficent to generate separate stores for F and I,
5975/// which can remove the bitwise instructions or sink them to colder places.
5976///
5977/// (store (or (zext (bitcast F to i32) to i64),
5978/// (shl (zext I to i64), 32)), addr) -->
5979/// (store F, addr) and (store I, addr+4)
5980///
5981/// Similarly, splitting for other merged store can also be beneficial, like:
5982/// For pair of {i32, i32}, i64 store --> two i32 stores.
5983/// For pair of {i32, i16}, i64 store --> two i32 stores.
5984/// For pair of {i16, i16}, i32 store --> two i16 stores.
5985/// For pair of {i16, i8}, i32 store --> two i16 stores.
5986/// For pair of {i8, i8}, i16 store --> two i8 stores.
5987///
5988/// We allow each target to determine specifically which kind of splitting is
5989/// supported.
5990///
5991/// The store patterns are commonly seen from the simple code snippet below
5992/// if only std::make_pair(...) is sroa transformed before inlined into hoo.
5993/// void goo(const std::pair<int, float> &);
5994/// hoo() {
5995/// ...
5996/// goo(std::make_pair(tmp, ftmp));
5997/// ...
5998/// }
5999///
6000/// Although we already have similar splitting in DAG Combine, we duplicate
6001/// it in CodeGenPrepare to catch the case in which pattern is across
6002/// multiple BBs. The logic in DAG Combine is kept to catch case generated
6003/// during code expansion.
6004static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
6005 const TargetLowering &TLI) {
6006 // Handle simple but common cases only.
6007 Type *StoreType = SI.getValueOperand()->getType();
6008 if (DL.getTypeStoreSizeInBits(StoreType) != DL.getTypeSizeInBits(StoreType) ||
6009 DL.getTypeSizeInBits(StoreType) == 0)
6010 return false;
6011
6012 unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
6013 Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
6014 if (DL.getTypeStoreSizeInBits(SplitStoreType) !=
6015 DL.getTypeSizeInBits(SplitStoreType))
6016 return false;
6017
6018 // Match the following patterns:
6019 // (store (or (zext LValue to i64),
6020 // (shl (zext HValue to i64), 32)), HalfValBitSize)
6021 // or
6022 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
6023 // (zext LValue to i64),
6024 // Expect both operands of OR and the first operand of SHL have only
6025 // one use.
6026 Value *LValue, *HValue;
6027 if (!match(SI.getValueOperand(),
6028 m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
6029 m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
6030 m_SpecificInt(HalfValBitSize))))))
6031 return false;
6032
6033 // Check LValue and HValue are int with size less or equal than 32.
6034 if (!LValue->getType()->isIntegerTy() ||
6035 DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
6036 !HValue->getType()->isIntegerTy() ||
6037 DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
6038 return false;
6039
6040 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
6041 // as the input of target query.
6042 auto *LBC = dyn_cast<BitCastInst>(LValue);
6043 auto *HBC = dyn_cast<BitCastInst>(HValue);
6044 EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
6045 : EVT::getEVT(LValue->getType());
6046 EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
6047 : EVT::getEVT(HValue->getType());
6048 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
6049 return false;
6050
6051 // Start to split store.
6052 IRBuilder<> Builder(SI.getContext());
6053 Builder.SetInsertPoint(&SI);
6054
6055 // If LValue/HValue is a bitcast in another BB, create a new one in current
6056 // BB so it may be merged with the splitted stores by dag combiner.
6057 if (LBC && LBC->getParent() != SI.getParent())
6058 LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
6059 if (HBC && HBC->getParent() != SI.getParent())
6060 HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
6061
6062 auto CreateSplitStore = [&](Value *V, bool Upper) {
6063 V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
6064 Value *Addr = Builder.CreateBitCast(
6065 SI.getOperand(1),
6066 SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
6067 if (Upper)
6068 Addr = Builder.CreateGEP(
6069 SplitStoreType, Addr,
6070 ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
6071 Builder.CreateAlignedStore(
6072 V, Addr, Upper ? SI.getAlignment() / 2 : SI.getAlignment());
6073 };
6074
6075 CreateSplitStore(LValue, false);
6076 CreateSplitStore(HValue, true);
6077
6078 // Delete the old store.
6079 SI.eraseFromParent();
6080 return true;
6081}
6082
6083// Return true if the GEP has two operands, the first operand is of a sequential
6084// type, and the second operand is a constant.
6085static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
6086 gep_type_iterator I = gep_type_begin(*GEP);
6087 return GEP->getNumOperands() == 2 &&
6088 I.isSequential() &&
6089 isa<ConstantInt>(GEP->getOperand(1));
6090}
6091
6092// Try unmerging GEPs to reduce liveness interference (register pressure) across
6093// IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
6094// reducing liveness interference across those edges benefits global register
6095// allocation. Currently handles only certain cases.
6096//
6097// For example, unmerge %GEPI and %UGEPI as below.
6098//
6099// ---------- BEFORE ----------
6100// SrcBlock:
6101// ...
6102// %GEPIOp = ...
6103// ...
6104// %GEPI = gep %GEPIOp, Idx
6105// ...
6106// indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
6107// (* %GEPI is alive on the indirectbr edges due to other uses ahead)
6108// (* %GEPIOp is alive on the indirectbr edges only because of it's used by
6109// %UGEPI)
6110//
6111// DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
6112// DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
6113// ...
6114//
6115// DstBi:
6116// ...
6117// %UGEPI = gep %GEPIOp, UIdx
6118// ...
6119// ---------------------------
6120//
6121// ---------- AFTER ----------
6122// SrcBlock:
6123// ... (same as above)
6124// (* %GEPI is still alive on the indirectbr edges)
6125// (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
6126// unmerging)
6127// ...
6128//
6129// DstBi:
6130// ...
6131// %UGEPI = gep %GEPI, (UIdx-Idx)
6132// ...
6133// ---------------------------
6134//
6135// The register pressure on the IndirectBr edges is reduced because %GEPIOp is
6136// no longer alive on them.
6137//
6138// We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
6139// of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
6140// not to disable further simplications and optimizations as a result of GEP
6141// merging.
6142//
6143// Note this unmerging may increase the length of the data flow critical path
6144// (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
6145// between the register pressure and the length of data-flow critical
6146// path. Restricting this to the uncommon IndirectBr case would minimize the
6147// impact of potentially longer critical path, if any, and the impact on compile
6148// time.
6149static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
6150 const TargetTransformInfo *TTI) {
6151 BasicBlock *SrcBlock = GEPI->getParent();
6152 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
6153 // (non-IndirectBr) cases exit early here.
6154 if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
6155 return false;
6156 // Check that GEPI is a simple gep with a single constant index.
6157 if (!GEPSequentialConstIndexed(GEPI))
6158 return false;
6159 ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
6160 // Check that GEPI is a cheap one.
6161 if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType())
6162 > TargetTransformInfo::TCC_Basic)
6163 return false;
6164 Value *GEPIOp = GEPI->getOperand(0);
6165 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
6166 if (!isa<Instruction>(GEPIOp))
6167 return false;
6168 auto *GEPIOpI = cast<Instruction>(GEPIOp);
6169 if (GEPIOpI->getParent() != SrcBlock)
6170 return false;
6171 // Check that GEP is used outside the block, meaning it's alive on the
6172 // IndirectBr edge(s).
6173 if (find_if(GEPI->users(), [&](User *Usr) {
6174 if (auto *I = dyn_cast<Instruction>(Usr)) {
6175 if (I->getParent() != SrcBlock) {
6176 return true;
6177 }
6178 }
6179 return false;
6180 }) == GEPI->users().end())
6181 return false;
6182 // The second elements of the GEP chains to be unmerged.
6183 std::vector<GetElementPtrInst *> UGEPIs;
6184 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
6185 // on IndirectBr edges.
6186 for (User *Usr : GEPIOp->users()) {
6187 if (Usr == GEPI) continue;
6188 // Check if Usr is an Instruction. If not, give up.
6189 if (!isa<Instruction>(Usr))
6190 return false;
6191 auto *UI = cast<Instruction>(Usr);
6192 // Check if Usr in the same block as GEPIOp, which is fine, skip.
6193 if (UI->getParent() == SrcBlock)
6194 continue;
6195 // Check if Usr is a GEP. If not, give up.
6196 if (!isa<GetElementPtrInst>(Usr))
6197 return false;
6198 auto *UGEPI = cast<GetElementPtrInst>(Usr);
6199 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
6200 // the pointer operand to it. If so, record it in the vector. If not, give
6201 // up.
6202 if (!GEPSequentialConstIndexed(UGEPI))
6203 return false;
6204 if (UGEPI->getOperand(0) != GEPIOp)
6205 return false;
6206 if (GEPIIdx->getType() !=
6207 cast<ConstantInt>(UGEPI->getOperand(1))->getType())
6208 return false;
6209 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6210 if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType())
6211 > TargetTransformInfo::TCC_Basic)
6212 return false;
6213 UGEPIs.push_back(UGEPI);
6214 }
6215 if (UGEPIs.size() == 0)
6216 return false;
6217 // Check the materializing cost of (Uidx-Idx).
6218 for (GetElementPtrInst *UGEPI : UGEPIs) {
6219 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6220 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
6221 unsigned ImmCost = TTI->getIntImmCost(NewIdx, GEPIIdx->getType());
6222 if (ImmCost > TargetTransformInfo::TCC_Basic)
6223 return false;
6224 }
6225 // Now unmerge between GEPI and UGEPIs.
6226 for (GetElementPtrInst *UGEPI : UGEPIs) {
6227 UGEPI->setOperand(0, GEPI);
6228 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
6229 Constant *NewUGEPIIdx =
6230 ConstantInt::get(GEPIIdx->getType(),
6231 UGEPIIdx->getValue() - GEPIIdx->getValue());
6232 UGEPI->setOperand(1, NewUGEPIIdx);
6233 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
6234 // inbounds to avoid UB.
6235 if (!GEPI->isInBounds()) {
6236 UGEPI->setIsInBounds(false);
6237 }
6238 }
6239 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
6240 // alive on IndirectBr edges).
6241 assert(find_if(GEPIOp->users(), [&](User *Usr) {(static_cast <bool> (find_if(GEPIOp->users(), [&
](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 6243, __extension__ __PRETTY_FUNCTION__))
6242 return cast<Instruction>(Usr)->getParent() != SrcBlock;(static_cast <bool> (find_if(GEPIOp->users(), [&
](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 6243, __extension__ __PRETTY_FUNCTION__))
6243 }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock")(static_cast <bool> (find_if(GEPIOp->users(), [&
](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("find_if(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) == GEPIOp->users().end() && \"GEPIOp is used outside SrcBlock\""
, "/build/llvm-toolchain-snapshot-6.0~svn318631/lib/CodeGen/CodeGenPrepare.cpp"
, 6243, __extension__ __PRETTY_FUNCTION__))
;
6244 return true;
6245}
6246
6247bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
6248 // Bail out if we inserted the instruction to prevent optimizations from
6249 // stepping on each other's toes.
6250 if (InsertedInsts.count(I))
6251 return false;
6252
6253 if (PHINode *P = dyn_cast<PHINode>(I)) {
6254 // It is possible for very late stage optimizations (such as SimplifyCFG)
6255 // to introduce PHI nodes too late to be cleaned up. If we detect such a
6256 // trivial PHI, go ahead and zap it here.
6257 if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) {
6258 P->replaceAllUsesWith(V);
6259 P->eraseFromParent();
6260 ++NumPHIsElim;
6261 return true;
6262 }
6263 return false;
6264 }
6265
6266 if (CastInst *CI = dyn_cast<CastInst>(I)) {
6267 // If the source of the cast is a constant, then this should have
6268 // already been constant folded. The only reason NOT to constant fold
6269 // it is if something (e.g. LSR) was careful to place the constant
6270 // evaluation in a block other than then one that uses it (e.g. to hoist
6271 // the address of globals out of a loop). If this is the case, we don't
6272 // want to forward-subst the cast.
6273 if (isa<Constant>(CI->getOperand(0)))
6274 return false;
6275
6276 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6277 return true;
6278
6279 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6280 /// Sink a zext or sext into its user blocks if the target type doesn't
6281 /// fit in one register
6282 if (TLI &&
6283 TLI->getTypeAction(CI->getContext(),
6284 TLI->getValueType(*DL, CI->getType())) ==
6285 TargetLowering::TypeExpandInteger) {
6286 return SinkCast(CI);
6287 } else {
6288 bool MadeChange = optimizeExt(I);
6289 return MadeChange | optimizeExtUses(I);
6290 }
6291 }
6292 return false;
6293 }
6294
6295 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6296 if (!TLI || !TLI->hasMultipleConditionRegisters())
6297 return OptimizeCmpExpression(CI, TLI);
6298
6299 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6300 LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
6301 if (TLI) {
6302 bool Modified = optimizeLoadExt(LI);
6303 unsigned AS = LI->getPointerAddressSpace();
6304 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6305 return Modified;
6306 }
6307 return false;
6308 }
6309
6310 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6311 if (TLI && splitMergedValStore(*SI, *DL, *TLI))
6312 return true;
6313 SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
6314 if (TLI) {
6315 unsigned AS = SI->getPointerAddressSpace();
6316 return optimizeMemoryInst(I, SI->getOperand(1),
6317 SI->getOperand(0)->getType(), AS);
6318 }
6319 return false;
6320 }
6321
6322 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
6323 unsigned AS = RMW->getPointerAddressSpace();
6324 return optimizeMemoryInst(I, RMW->getPointerOperand(),
6325 RMW->getType(), AS);
6326 }
6327
6328 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
6329 unsigned AS = CmpX->getPointerAddressSpace();
6330 return optimizeMemoryInst(I, CmpX->getPointerOperand(),
6331 CmpX->getCompareOperand()->getType(), AS);
6332 }
6333
6334 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6335
6336 if (BinOp && (BinOp->getOpcode() == Instruction::And) &&
6337 EnableAndCmpSinking && TLI)
6338 return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts);
6339
6340 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6341 BinOp->getOpcode() == Instruction::LShr)) {
6342 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6343 if (TLI && CI && TLI->hasExtractBitsInsn())
6344 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6345
6346 return false;
6347 }
6348
6349 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6350 if (GEPI->hasAllZeroIndices()) {
6351 /// The GEP operand must be a pointer, so must its result -> BitCast
6352 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6353 GEPI->getName(), GEPI);
6354 GEPI->replaceAllUsesWith(NC);
6355 GEPI->eraseFromParent();
6356 ++NumGEPsElim;
6357 optimizeInst(NC, ModifiedDT);
6358 return true;
6359 }
6360 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
6361 return true;
6362 }
6363 return false;
6364 }
6365
6366 if (CallInst *CI = dyn_cast<CallInst>(I))
6367 return optimizeCallInst(CI, ModifiedDT);
6368
6369 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6370 return optimizeSelectInst(SI);
6371
6372 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6373 return optimizeShuffleVectorInst(SVI);
6374
6375 if (auto *Switch = dyn_cast<SwitchInst>(I))
6376 return optimizeSwitchInst(Switch);
6377
6378 if (isa<ExtractElementInst>(I))
6379 return optimizeExtractElementInst(I);
6380
6381 return false;
6382}
6383
6384/// Given an OR instruction, check to see if this is a bitreverse
6385/// idiom. If so, insert the new intrinsic and return true.
6386static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6387 const TargetLowering &TLI) {
6388 if (!I.getType()->isIntegerTy() ||
6389 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6390 TLI.getValueType(DL, I.getType(), true)))
6391 return false;
6392
6393 SmallVector<Instruction*, 4> Insts;
6394 if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
6395 return false;
6396 Instruction *LastInst = Insts.back();
6397 I.replaceAllUsesWith(LastInst);
6398 RecursivelyDeleteTriviallyDeadInstructions(&I);
6399 return true;
6400}
6401
6402// In this pass we look for GEP and cast instructions that are used
6403// across basic blocks and rewrite them to improve basic-block-at-a-time
6404// selection.
6405bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
6406 SunkAddrs.clear();
6407 bool MadeChange = false;
6408
6409 CurInstIterator = BB.begin();
6410 while (CurInstIterator != BB.end()) {
6411 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6412 if (ModifiedDT)
6413 return true;
6414 }
6415
6416 bool MadeBitReverse = true;
6417 while (TLI && MadeBitReverse) {
6418 MadeBitReverse = false;
6419 for (auto &I : reverse(BB)) {
6420 if (makeBitReverse(I, *DL, *TLI)) {
6421 MadeBitReverse = MadeChange = true;
6422 ModifiedDT = true;
6423 break;
6424 }
6425 }
6426 }
6427 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6428
6429 return MadeChange;
6430}
6431
6432// llvm.dbg.value is far away from the value then iSel may not be able
6433// handle it properly. iSel will drop llvm.dbg.value if it can not
6434// find a node corresponding to the value.
6435bool CodeGenPrepare::placeDbgValues(Function &F) {
6436 bool MadeChange = false;
6437 for (BasicBlock &BB : F) {
6438 Instruction *PrevNonDbgInst = nullptr;
6439 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6440 Instruction *Insn = &*BI++;
6441 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6442 // Leave dbg.values that refer to an alloca alone. These
6443 // intrinsics describe the address of a variable (= the alloca)
6444 // being taken. They should not be moved next to the alloca
6445 // (and to the beginning of the scope), but rather stay close to
6446 // where said address is used.
6447 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6448 PrevNonDbgInst = Insn;
6449 continue;
6450 }
6451
6452 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6453 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6454 // If VI is a phi in a block with an EHPad terminator, we can't insert
6455 // after it.
6456 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6457 continue;
6458 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Moving Debug Value before :\n"
<< *DVI << ' ' << *VI; } } while (false)
;
6459 DVI->removeFromParent();
6460 if (isa<PHINode>(VI))
6461 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6462 else
6463 DVI->insertAfter(VI);
6464 MadeChange = true;
6465 ++NumDbgValueMoved;
6466 }
6467 }
6468 }
6469 return MadeChange;
6470}
6471
6472/// \brief Scale down both weights to fit into uint32_t.
6473static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6474 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6475 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
6476 NewTrue = NewTrue / Scale;
6477 NewFalse = NewFalse / Scale;
6478}
6479
6480/// \brief Some targets prefer to split a conditional branch like:
6481/// \code
6482/// %0 = icmp ne i32 %a, 0
6483/// %1 = icmp ne i32 %b, 0
6484/// %or.cond = or i1 %0, %1
6485/// br i1 %or.cond, label %TrueBB, label %FalseBB
6486/// \endcode
6487/// into multiple branch instructions like:
6488/// \code
6489/// bb1:
6490/// %0 = icmp ne i32 %a, 0
6491/// br i1 %0, label %TrueBB, label %bb2
6492/// bb2:
6493/// %1 = icmp ne i32 %b, 0
6494/// br i1 %1, label %TrueBB, label %FalseBB
6495/// \endcode
6496/// This usually allows instruction selection to do even further optimizations
6497/// and combine the compare with the branch instruction. Currently this is
6498/// applied for targets which have "cheap" jump instructions.
6499///
6500/// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6501///
6502bool CodeGenPrepare::splitBranchCondition(Function &F) {
6503 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6504 return false;
6505
6506 bool MadeChange = false;
6507 for (auto &BB : F) {
6508 // Does this BB end with the following?
6509 // %cond1 = icmp|fcmp|binary instruction ...
6510 // %cond2 = icmp|fcmp|binary instruction ...
6511 // %cond.or = or|and i1 %cond1, cond2
6512 // br i1 %cond.or label %dest1, label %dest2"
6513 BinaryOperator *LogicOp;
6514 BasicBlock *TBB, *FBB;
6515 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6516 continue;
6517
6518 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6519 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6520 continue;
6521
6522 unsigned Opc;
6523 Value *Cond1, *Cond2;
6524 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6525 m_OneUse(m_Value(Cond2)))))
6526 Opc = Instruction::And;
6527 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6528 m_OneUse(m_Value(Cond2)))))
6529 Opc = Instruction::Or;
6530 else
6531 continue;
6532
6533 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6534 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6535 continue;
6536
6537 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Before branch condition splitting\n"
; BB.dump(); } } while (false)
;
6538
6539 // Create a new BB.
6540 auto TmpBB =
6541 BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
6542 BB.getParent(), BB.getNextNode());
6543
6544 // Update original basic block by using the first condition directly by the
6545 // branch instruction and removing the no longer needed and/or instruction.
6546 Br1->setCondition(Cond1);
6547 LogicOp->eraseFromParent();
6548
6549 // Depending on the conditon we have to either replace the true or the false
6550 // successor of the original branch instruction.
6551 if (Opc == Instruction::And)
6552 Br1->setSuccessor(0, TmpBB);
6553 else
6554 Br1->setSuccessor(1, TmpBB);
6555
6556 // Fill in the new basic block.
6557 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6558 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6559 I->removeFromParent();
6560 I->insertBefore(Br2);
6561 }
6562
6563 // Update PHI nodes in both successors. The original BB needs to be
6564 // replaced in one successor's PHI nodes, because the branch comes now from
6565 // the newly generated BB (NewBB). In the other successor we need to add one
6566 // incoming edge to the PHI nodes, because both branch instructions target
6567 // now the same successor. Depending on the original branch condition
6568 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6569 // we perform the correct update for the PHI nodes.
6570 // This doesn't change the successor order of the just created branch
6571 // instruction (or any other instruction).
6572 if (Opc == Instruction::Or)
6573 std::swap(TBB, FBB);
6574
6575 // Replace the old BB with the new BB.
6576 for (auto &I : *TBB) {
6577 PHINode *PN = dyn_cast<PHINode>(&I);
6578 if (!PN)
6579 break;
6580 int i;
6581 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6582 PN->setIncomingBlock(i, TmpBB);
6583 }
6584
6585 // Add another incoming edge form the new BB.
6586 for (auto &I : *FBB) {
6587 PHINode *PN = dyn_cast<PHINode>(&I);
6588 if (!PN)
6589 break;
6590 auto *Val = PN->getIncomingValueForBlock(&BB);
6591 PN->addIncoming(Val, TmpBB);
6592 }
6593
6594 // Update the branch weights (from SelectionDAGBuilder::
6595 // FindMergedConditions).
6596 if (Opc == Instruction::Or) {
6597 // Codegen X | Y as:
6598 // BB1:
6599 // jmp_if_X TBB
6600 // jmp TmpBB
6601 // TmpBB:
6602 // jmp_if_Y TBB
6603 // jmp FBB
6604 //
6605
6606 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6607 // The requirement is that
6608 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6609 // = TrueProb for orignal BB.
6610 // Assuming the orignal weights are A and B, one choice is to set BB1's
6611 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6612 // assumes that
6613 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6614 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6615 // TmpBB, but the math is more complicated.
6616 uint64_t TrueWeight, FalseWeight;
6617 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
6618 uint64_t NewTrueWeight = TrueWeight;
6619 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6620 scaleWeights(NewTrueWeight, NewFalseWeight);
6621 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6622 .createBranchWeights(TrueWeight, FalseWeight));
6623
6624 NewTrueWeight = TrueWeight;
6625 NewFalseWeight = 2 * FalseWeight;
6626 scaleWeights(NewTrueWeight, NewFalseWeight);
6627 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6628 .createBranchWeights(TrueWeight, FalseWeight));
6629 }
6630 } else {
6631 // Codegen X & Y as:
6632 // BB1:
6633 // jmp_if_X TmpBB
6634 // jmp FBB
6635 // TmpBB:
6636 // jmp_if_Y TBB
6637 // jmp FBB
6638 //
6639 // This requires creation of TmpBB after CurBB.
6640
6641 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6642 // The requirement is that
6643 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6644 // = FalseProb for orignal BB.
6645 // Assuming the orignal weights are A and B, one choice is to set BB1's
6646 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6647 // assumes that
6648 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6649 uint64_t TrueWeight, FalseWeight;
6650 if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) {
6651 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6652 uint64_t NewFalseWeight = FalseWeight;
6653 scaleWeights(NewTrueWeight, NewFalseWeight);
6654 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6655 .createBranchWeights(TrueWeight, FalseWeight));
6656
6657 NewTrueWeight = 2 * TrueWeight;
6658 NewFalseWeight = FalseWeight;
6659 scaleWeights(NewTrueWeight, NewFalseWeight);
6660 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6661 .createBranchWeights(TrueWeight, FalseWeight));
6662 }
6663 }
6664
6665 // Note: No point in getting fancy here, since the DT info is never
6666 // available to CodeGenPrepare.
6667 ModifiedDT = true;
6668
6669 MadeChange = true;
6670
6671 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "After branch condition splitting\n"
; BB.dump(); TmpBB->dump(); } } while (false)
6672 TmpBB->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "After branch condition splitting\n"
; BB.dump(); TmpBB->dump(); } } while (false)
;
6673 }
6674 return MadeChange;
6675}