Bug Summary

File:lib/CodeGen/CodeGenPrepare.cpp
Warning:line 5726, column 3
Undefined or garbage value returned to caller

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CodeGenPrepare.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/CodeGen -I /build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/CodeGen -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp -faddrsig
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/SmallPtrSet.h"
22#include "llvm/ADT/SmallVector.h"
23#include "llvm/ADT/Statistic.h"
24#include "llvm/Analysis/BlockFrequencyInfo.h"
25#include "llvm/Analysis/BranchProbabilityInfo.h"
26#include "llvm/Analysis/ConstantFolding.h"
27#include "llvm/Analysis/InstructionSimplify.h"
28#include "llvm/Analysis/LoopInfo.h"
29#include "llvm/Analysis/MemoryBuiltins.h"
30#include "llvm/Analysis/ProfileSummaryInfo.h"
31#include "llvm/Analysis/TargetLibraryInfo.h"
32#include "llvm/Analysis/TargetTransformInfo.h"
33#include "llvm/Transforms/Utils/Local.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/CodeGen/Analysis.h"
36#include "llvm/CodeGen/ISDOpcodes.h"
37#include "llvm/CodeGen/SelectionDAGNodes.h"
38#include "llvm/CodeGen/TargetLowering.h"
39#include "llvm/CodeGen/TargetPassConfig.h"
40#include "llvm/CodeGen/TargetSubtargetInfo.h"
41#include "llvm/CodeGen/ValueTypes.h"
42#include "llvm/Config/llvm-config.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/MachineValueType.h"
84#include "llvm/Support/MathExtras.h"
85#include "llvm/Support/raw_ostream.h"
86#include "llvm/Target/TargetMachine.h"
87#include "llvm/Target/TargetOptions.h"
88#include "llvm/Transforms/Utils/BasicBlockUtils.h"
89#include "llvm/Transforms/Utils/BypassSlowDivision.h"
90#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
91#include <algorithm>
92#include <cassert>
93#include <cstdint>
94#include <iterator>
95#include <limits>
96#include <memory>
97#include <utility>
98#include <vector>
99
100using namespace llvm;
101using namespace llvm::PatternMatch;
102
103#define DEBUG_TYPE"codegenprepare" "codegenprepare"
104
105STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim"
, "Number of blocks eliminated", {0}, {false}}
;
106STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim"
, "Number of trivial PHIs eliminated", {0}, {false}}
;
107STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim"
, "Number of GEPs converted to casts", {0}, {false}}
;
108STATISTIC(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}}
109 "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses"
, "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps"
, {0}, {false}}
;
110STATISTIC(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}}
111 "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses"
, "Number of uses of Cast expressions replaced with uses " "of sunken Casts"
, {0}, {false}}
;
112STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
, {0}, {false}}
113 "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
, {0}, {false}}
;
114STATISTIC(NumMemoryInstsPhiCreated,static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions", {0}, {false
}}
115 "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
}}
116 "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
}}
;
117STATISTIC(NumMemoryInstsSelectCreated,static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions", {0}, {false
}}
118 "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
}}
119 "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
}}
;
120STATISTIC(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}}
;
121STATISTIC(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}}
;
122STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded"
, "Number of and mask instructions added to form ext loads", {
0}, {false}}
123 "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}}
;
124STATISTIC(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}}
;
125STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup"
, "Number of return instructions duplicated", {0}, {false}}
;
126STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved"
, "Number of debug value instructions moved", {0}, {false}}
;
127STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare"
, "NumSelectsExpanded", "Number of selects turned into branches"
, {0}, {false}}
;
128STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare"
, "NumStoreExtractExposed", "Number of store(extractelement) exposed"
, {0}, {false}}
;
129
130static cl::opt<bool> DisableBranchOpts(
131 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
132 cl::desc("Disable branch optimizations in CodeGenPrepare"));
133
134static cl::opt<bool>
135 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
136 cl::desc("Disable GC optimizations in CodeGenPrepare"));
137
138static cl::opt<bool> DisableSelectToBranch(
139 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
140 cl::desc("Disable select to branch conversion."));
141
142static cl::opt<bool> AddrSinkUsingGEPs(
143 "addr-sink-using-gep", cl::Hidden, cl::init(true),
144 cl::desc("Address sinking in CGP using GEPs."));
145
146static cl::opt<bool> EnableAndCmpSinking(
147 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
148 cl::desc("Enable sinkinig and/cmp into branches."));
149
150static cl::opt<bool> DisableStoreExtract(
151 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
152 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
153
154static cl::opt<bool> StressStoreExtract(
155 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
156 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
157
158static cl::opt<bool> DisableExtLdPromotion(
159 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
160 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
161 "CodeGenPrepare"));
162
163static cl::opt<bool> StressExtLdPromotion(
164 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
165 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
166 "optimization in CodeGenPrepare"));
167
168static cl::opt<bool> DisablePreheaderProtect(
169 "disable-preheader-prot", cl::Hidden, cl::init(false),
170 cl::desc("Disable protection against removing loop preheaders"));
171
172static cl::opt<bool> ProfileGuidedSectionPrefix(
173 "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
174 cl::desc("Use profile info to add section prefix for hot/cold functions"));
175
176static cl::opt<unsigned> FreqRatioToSkipMerge(
177 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
178 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
179 "(frequency of destination block) is greater than this ratio"));
180
181static cl::opt<bool> ForceSplitStore(
182 "force-split-store", cl::Hidden, cl::init(false),
183 cl::desc("Force store splitting no matter what the target query says."));
184
185static cl::opt<bool>
186EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
187 cl::desc("Enable merging of redundant sexts when one is dominating"
188 " the other."), cl::init(true));
189
190static cl::opt<bool> DisableComplexAddrModes(
191 "disable-complex-addr-modes", cl::Hidden, cl::init(false),
192 cl::desc("Disables combining addressing modes with different parts "
193 "in optimizeMemoryInst."));
194
195static cl::opt<bool>
196AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
197 cl::desc("Allow creation of Phis in Address sinking."));
198
199static cl::opt<bool>
200AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
201 cl::desc("Allow creation of selects in Address sinking."));
202
203static cl::opt<bool> AddrSinkCombineBaseReg(
204 "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
205 cl::desc("Allow combining of BaseReg field in Address sinking."));
206
207static cl::opt<bool> AddrSinkCombineBaseGV(
208 "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
209 cl::desc("Allow combining of BaseGV field in Address sinking."));
210
211static cl::opt<bool> AddrSinkCombineBaseOffs(
212 "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
213 cl::desc("Allow combining of BaseOffs field in Address sinking."));
214
215static cl::opt<bool> AddrSinkCombineScaledReg(
216 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
217 cl::desc("Allow combining of ScaledReg field in Address sinking."));
218
219static cl::opt<bool>
220 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
221 cl::init(true),
222 cl::desc("Enable splitting large offset of GEP."));
223
224namespace {
225
226enum ExtType {
227 ZeroExtension, // Zero extension has been seen.
228 SignExtension, // Sign extension has been seen.
229 BothExtension // This extension type is used if we saw sext after
230 // ZeroExtension had been set, or if we saw zext after
231 // SignExtension had been set. It makes the type
232 // information of a promoted instruction invalid.
233};
234
235using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
236using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
237using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
238using SExts = SmallVector<Instruction *, 16>;
239using ValueToSExts = DenseMap<Value *, SExts>;
240
241class TypePromotionTransaction;
242
243 class CodeGenPrepare : public FunctionPass {
244 const TargetMachine *TM = nullptr;
245 const TargetSubtargetInfo *SubtargetInfo;
246 const TargetLowering *TLI = nullptr;
247 const TargetRegisterInfo *TRI;
248 const TargetTransformInfo *TTI = nullptr;
249 const TargetLibraryInfo *TLInfo;
250 const LoopInfo *LI;
251 std::unique_ptr<BlockFrequencyInfo> BFI;
252 std::unique_ptr<BranchProbabilityInfo> BPI;
253
254 /// As we scan instructions optimizing them, this is the next instruction
255 /// to optimize. Transforms that can invalidate this should update it.
256 BasicBlock::iterator CurInstIterator;
257
258 /// Keeps track of non-local addresses that have been sunk into a block.
259 /// This allows us to avoid inserting duplicate code for blocks with
260 /// multiple load/stores of the same address. The usage of WeakTrackingVH
261 /// enables SunkAddrs to be treated as a cache whose entries can be
262 /// invalidated if a sunken address computation has been erased.
263 ValueMap<Value*, WeakTrackingVH> SunkAddrs;
264
265 /// Keeps track of all instructions inserted for the current function.
266 SetOfInstrs InsertedInsts;
267
268 /// Keeps track of the type of the related instruction before their
269 /// promotion for the current function.
270 InstrToOrigTy PromotedInsts;
271
272 /// Keep track of instructions removed during promotion.
273 SetOfInstrs RemovedInsts;
274
275 /// Keep track of sext chains based on their initial value.
276 DenseMap<Value *, Instruction *> SeenChainsForSExt;
277
278 /// Keep track of GEPs accessing the same data structures such as structs or
279 /// arrays that are candidates to be split later because of their large
280 /// size.
281 MapVector<
282 AssertingVH<Value>,
283 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
284 LargeOffsetGEPMap;
285
286 /// Keep track of new GEP base after splitting the GEPs having large offset.
287 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
288
289 /// Map serial numbers to Large offset GEPs.
290 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
291
292 /// Keep track of SExt promoted.
293 ValueToSExts ValToSExtendedUses;
294
295 /// True if CFG is modified in any way.
296 bool ModifiedDT;
297
298 /// True if optimizing for size.
299 bool OptSize;
300
301 /// DataLayout for the Function being processed.
302 const DataLayout *DL = nullptr;
303
304 public:
305 static char ID; // Pass identification, replacement for typeid
306
307 CodeGenPrepare() : FunctionPass(ID) {
308 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
309 }
310
311 bool runOnFunction(Function &F) override;
312
313 StringRef getPassName() const override { return "CodeGen Prepare"; }
314
315 void getAnalysisUsage(AnalysisUsage &AU) const override {
316 // FIXME: When we can selectively preserve passes, preserve the domtree.
317 AU.addRequired<ProfileSummaryInfoWrapperPass>();
318 AU.addRequired<TargetLibraryInfoWrapperPass>();
319 AU.addRequired<TargetTransformInfoWrapperPass>();
320 AU.addRequired<LoopInfoWrapperPass>();
321 }
322
323 private:
324 template <typename F>
325 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
326 // Substituting can cause recursive simplifications, which can invalidate
327 // our iterator. Use a WeakTrackingVH to hold onto it in case this
328 // happens.
329 Value *CurValue = &*CurInstIterator;
330 WeakTrackingVH IterHandle(CurValue);
331
332 f();
333
334 // If the iterator instruction was recursively deleted, start over at the
335 // start of the block.
336 if (IterHandle != CurValue) {
337 CurInstIterator = BB->begin();
338 SunkAddrs.clear();
339 }
340 }
341
342 bool eliminateFallThrough(Function &F);
343 bool eliminateMostlyEmptyBlocks(Function &F);
344 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
345 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
346 void eliminateMostlyEmptyBlock(BasicBlock *BB);
347 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
348 bool isPreheader);
349 bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
350 bool optimizeInst(Instruction *I, bool &ModifiedDT);
351 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
352 Type *AccessTy, unsigned AddrSpace);
353 bool optimizeInlineAsmInst(CallInst *CS);
354 bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
355 bool optimizeExt(Instruction *&I);
356 bool optimizeExtUses(Instruction *I);
357 bool optimizeLoadExt(LoadInst *Load);
358 bool optimizeSelectInst(SelectInst *SI);
359 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
360 bool optimizeSwitchInst(SwitchInst *SI);
361 bool optimizeExtractElementInst(Instruction *Inst);
362 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
363 bool placeDbgValues(Function &F);
364 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
365 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
366 bool tryToPromoteExts(TypePromotionTransaction &TPT,
367 const SmallVectorImpl<Instruction *> &Exts,
368 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
369 unsigned CreatedInstsCost = 0);
370 bool mergeSExts(Function &F);
371 bool splitLargeGEPOffsets();
372 bool performAddressTypePromotion(
373 Instruction *&Inst,
374 bool AllowPromotionWithoutCommonHeader,
375 bool HasPromoted, TypePromotionTransaction &TPT,
376 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
377 bool splitBranchCondition(Function &F);
378 bool simplifyOffsetableRelocate(Instruction &I);
379 };
380
381} // end anonymous namespace
382
383char CodeGenPrepare::ID = 0;
384
385INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
386 "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
387INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry);
388INITIALIZE_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)); }
389 "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)); }
390
391FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
392
393bool CodeGenPrepare::runOnFunction(Function &F) {
394 if (skipFunction(F))
1
Assuming the condition is false
2
Taking false branch
395 return false;
396
397 DL = &F.getParent()->getDataLayout();
398
399 bool EverMadeChange = false;
400 // Clear per function information.
401 InsertedInsts.clear();
402 PromotedInsts.clear();
403
404 ModifiedDT = false;
405 if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
3
Assuming 'TPC' is null
4
Taking false branch
406 TM = &TPC->getTM<TargetMachine>();
407 SubtargetInfo = TM->getSubtargetImpl(F);
408 TLI = SubtargetInfo->getTargetLowering();
409 TRI = SubtargetInfo->getRegisterInfo();
410 }
411 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
412 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
413 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
414 BPI.reset(new BranchProbabilityInfo(F, *LI));
415 BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
416 OptSize = F.optForSize();
417
418 ProfileSummaryInfo *PSI =
419 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
420 if (ProfileGuidedSectionPrefix) {
5
Assuming the condition is false
6
Taking false branch
421 if (PSI->isFunctionHotInCallGraph(&F, *BFI))
422 F.setSectionPrefix(".hot");
423 else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
424 F.setSectionPrefix(".unlikely");
425 }
426
427 /// This optimization identifies DIV instructions that can be
428 /// profitably bypassed and carried out with a shorter, faster divide.
429 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
7
Assuming the condition is false
430 TLI->isSlowDivBypassed()) {
431 const DenseMap<unsigned int, unsigned int> &BypassWidths =
432 TLI->getBypassSlowDivWidths();
433 BasicBlock* BB = &*F.begin();
434 while (BB != nullptr) {
435 // bypassSlowDivision may create new BBs, but we don't want to reapply the
436 // optimization to those blocks.
437 BasicBlock* Next = BB->getNextNode();
438 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
439 BB = Next;
440 }
441 }
442
443 // Eliminate blocks that contain only PHI nodes and an
444 // unconditional branch.
445 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
446
447 if (!DisableBranchOpts)
8
Assuming the condition is false
9
Taking false branch
448 EverMadeChange |= splitBranchCondition(F);
449
450 // Split some critical edges where one of the sources is an indirect branch,
451 // to help generate sane code for PHIs involving such edges.
452 EverMadeChange |= SplitIndirectBrCriticalEdges(F);
453
454 bool MadeChange = true;
455 while (MadeChange) {
10
Loop condition is true. Entering loop body
456 MadeChange = false;
457 for (Function::iterator I = F.begin(); I != F.end(); ) {
11
Loop condition is true. Entering loop body
458 BasicBlock *BB = &*I++;
459 bool ModifiedDTOnIteration = false;
460 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
12
Calling 'CodeGenPrepare::optimizeBlock'
461
462 // Restart BB iteration if the dominator tree of the Function was changed
463 if (ModifiedDTOnIteration)
464 break;
465 }
466 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
467 MadeChange |= mergeSExts(F);
468 if (!LargeOffsetGEPMap.empty())
469 MadeChange |= splitLargeGEPOffsets();
470
471 // Really free removed instructions during promotion.
472 for (Instruction *I : RemovedInsts)
473 I->deleteValue();
474
475 EverMadeChange |= MadeChange;
476 SeenChainsForSExt.clear();
477 ValToSExtendedUses.clear();
478 RemovedInsts.clear();
479 LargeOffsetGEPMap.clear();
480 LargeOffsetGEPID.clear();
481 }
482
483 SunkAddrs.clear();
484
485 if (!DisableBranchOpts) {
486 MadeChange = false;
487 // Use a set vector to get deterministic iteration order. The order the
488 // blocks are removed may affect whether or not PHI nodes in successors
489 // are removed.
490 SmallSetVector<BasicBlock*, 8> WorkList;
491 for (BasicBlock &BB : F) {
492 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
493 MadeChange |= ConstantFoldTerminator(&BB, true);
494 if (!MadeChange) continue;
495
496 for (SmallVectorImpl<BasicBlock*>::iterator
497 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
498 if (pred_begin(*II) == pred_end(*II))
499 WorkList.insert(*II);
500 }
501
502 // Delete the dead blocks and any of their dead successors.
503 MadeChange |= !WorkList.empty();
504 while (!WorkList.empty()) {
505 BasicBlock *BB = WorkList.pop_back_val();
506 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
507
508 DeleteDeadBlock(BB);
509
510 for (SmallVectorImpl<BasicBlock*>::iterator
511 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
512 if (pred_begin(*II) == pred_end(*II))
513 WorkList.insert(*II);
514 }
515
516 // Merge pairs of basic blocks with unconditional branches, connected by
517 // a single edge.
518 if (EverMadeChange || MadeChange)
519 MadeChange |= eliminateFallThrough(F);
520
521 EverMadeChange |= MadeChange;
522 }
523
524 if (!DisableGCOpts) {
525 SmallVector<Instruction *, 2> Statepoints;
526 for (BasicBlock &BB : F)
527 for (Instruction &I : BB)
528 if (isStatepoint(I))
529 Statepoints.push_back(&I);
530 for (auto &I : Statepoints)
531 EverMadeChange |= simplifyOffsetableRelocate(*I);
532 }
533
534 // Do this last to clean up use-before-def scenarios introduced by other
535 // preparatory transforms.
536 EverMadeChange |= placeDbgValues(F);
537
538 return EverMadeChange;
539}
540
541/// Merge basic blocks which are connected by a single edge, where one of the
542/// basic blocks has a single successor pointing to the other basic block,
543/// which has a single predecessor.
544bool CodeGenPrepare::eliminateFallThrough(Function &F) {
545 bool Changed = false;
546 // Scan all of the blocks in the function, except for the entry block.
547 // Use a temporary array to avoid iterator being invalidated when
548 // deleting blocks.
549 SmallVector<WeakTrackingVH, 16> Blocks;
550 for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
551 Blocks.push_back(&Block);
552
553 for (auto &Block : Blocks) {
554 auto *BB = cast_or_null<BasicBlock>(Block);
555 if (!BB)
556 continue;
557 // If the destination block has a single pred, then this is a trivial
558 // edge, just collapse it.
559 BasicBlock *SinglePred = BB->getSinglePredecessor();
560
561 // Don't merge if BB's address is taken.
562 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
563
564 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
565 if (Term && !Term->isConditional()) {
566 Changed = true;
567 LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "To merge:\n" << *
BB << "\n\n\n"; } } while (false)
;
568
569 // Merge BB into SinglePred and delete it.
570 MergeBlockIntoPredecessor(BB);
571 }
572 }
573 return Changed;
574}
575
576/// Find a destination block from BB if BB is mergeable empty block.
577BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
578 // If this block doesn't end with an uncond branch, ignore it.
579 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
580 if (!BI || !BI->isUnconditional())
581 return nullptr;
582
583 // If the instruction before the branch (skipping debug info) isn't a phi
584 // node, then other stuff is happening here.
585 BasicBlock::iterator BBI = BI->getIterator();
586 if (BBI != BB->begin()) {
587 --BBI;
588 while (isa<DbgInfoIntrinsic>(BBI)) {
589 if (BBI == BB->begin())
590 break;
591 --BBI;
592 }
593 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
594 return nullptr;
595 }
596
597 // Do not break infinite loops.
598 BasicBlock *DestBB = BI->getSuccessor(0);
599 if (DestBB == BB)
600 return nullptr;
601
602 if (!canMergeBlocks(BB, DestBB))
603 DestBB = nullptr;
604
605 return DestBB;
606}
607
608/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
609/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
610/// edges in ways that are non-optimal for isel. Start by eliminating these
611/// blocks so we can split them the way we want them.
612bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
613 SmallPtrSet<BasicBlock *, 16> Preheaders;
614 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
615 while (!LoopList.empty()) {
616 Loop *L = LoopList.pop_back_val();
617 LoopList.insert(LoopList.end(), L->begin(), L->end());
618 if (BasicBlock *Preheader = L->getLoopPreheader())
619 Preheaders.insert(Preheader);
620 }
621
622 bool MadeChange = false;
623 // Copy blocks into a temporary array to avoid iterator invalidation issues
624 // as we remove them.
625 // Note that this intentionally skips the entry block.
626 SmallVector<WeakTrackingVH, 16> Blocks;
627 for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
628 Blocks.push_back(&Block);
629
630 for (auto &Block : Blocks) {
631 BasicBlock *BB = cast_or_null<BasicBlock>(Block);
632 if (!BB)
633 continue;
634 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
635 if (!DestBB ||
636 !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
637 continue;
638
639 eliminateMostlyEmptyBlock(BB);
640 MadeChange = true;
641 }
642 return MadeChange;
643}
644
645bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
646 BasicBlock *DestBB,
647 bool isPreheader) {
648 // Do not delete loop preheaders if doing so would create a critical edge.
649 // Loop preheaders can be good locations to spill registers. If the
650 // preheader is deleted and we create a critical edge, registers may be
651 // spilled in the loop body instead.
652 if (!DisablePreheaderProtect && isPreheader &&
653 !(BB->getSinglePredecessor() &&
654 BB->getSinglePredecessor()->getSingleSuccessor()))
655 return false;
656
657 // Try to skip merging if the unique predecessor of BB is terminated by a
658 // switch or indirect branch instruction, and BB is used as an incoming block
659 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
660 // add COPY instructions in the predecessor of BB instead of BB (if it is not
661 // merged). Note that the critical edge created by merging such blocks wont be
662 // split in MachineSink because the jump table is not analyzable. By keeping
663 // such empty block (BB), ISel will place COPY instructions in BB, not in the
664 // predecessor of BB.
665 BasicBlock *Pred = BB->getUniquePredecessor();
666 if (!Pred ||
667 !(isa<SwitchInst>(Pred->getTerminator()) ||
668 isa<IndirectBrInst>(Pred->getTerminator())))
669 return true;
670
671 if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
672 return true;
673
674 // We use a simple cost heuristic which determine skipping merging is
675 // profitable if the cost of skipping merging is less than the cost of
676 // merging : Cost(skipping merging) < Cost(merging BB), where the
677 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
678 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
679 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
680 // Freq(Pred) / Freq(BB) > 2.
681 // Note that if there are multiple empty blocks sharing the same incoming
682 // value for the PHIs in the DestBB, we consider them together. In such
683 // case, Cost(merging BB) will be the sum of their frequencies.
684
685 if (!isa<PHINode>(DestBB->begin()))
686 return true;
687
688 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
689
690 // Find all other incoming blocks from which incoming values of all PHIs in
691 // DestBB are the same as the ones from BB.
692 for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
693 ++PI) {
694 BasicBlock *DestBBPred = *PI;
695 if (DestBBPred == BB)
696 continue;
697
698 if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
699 return DestPN.getIncomingValueForBlock(BB) ==
700 DestPN.getIncomingValueForBlock(DestBBPred);
701 }))
702 SameIncomingValueBBs.insert(DestBBPred);
703 }
704
705 // See if all BB's incoming values are same as the value from Pred. In this
706 // case, no reason to skip merging because COPYs are expected to be place in
707 // Pred already.
708 if (SameIncomingValueBBs.count(Pred))
709 return true;
710
711 BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
712 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
713
714 for (auto SameValueBB : SameIncomingValueBBs)
715 if (SameValueBB->getUniquePredecessor() == Pred &&
716 DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
717 BBFreq += BFI->getBlockFreq(SameValueBB);
718
719 return PredFreq.getFrequency() <=
720 BBFreq.getFrequency() * FreqRatioToSkipMerge;
721}
722
723/// Return true if we can merge BB into DestBB if there is a single
724/// unconditional branch between them, and BB contains no other non-phi
725/// instructions.
726bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
727 const BasicBlock *DestBB) const {
728 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
729 // the successor. If there are more complex condition (e.g. preheaders),
730 // don't mess around with them.
731 for (const PHINode &PN : BB->phis()) {
732 for (const User *U : PN.users()) {
733 const Instruction *UI = cast<Instruction>(U);
734 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
735 return false;
736 // If User is inside DestBB block and it is a PHINode then check
737 // incoming value. If incoming value is not from BB then this is
738 // a complex condition (e.g. preheaders) we want to avoid here.
739 if (UI->getParent() == DestBB) {
740 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
741 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
742 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
743 if (Insn && Insn->getParent() == BB &&
744 Insn->getParent() != UPN->getIncomingBlock(I))
745 return false;
746 }
747 }
748 }
749 }
750
751 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
752 // and DestBB may have conflicting incoming values for the block. If so, we
753 // can't merge the block.
754 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
755 if (!DestBBPN) return true; // no conflict.
756
757 // Collect the preds of BB.
758 SmallPtrSet<const BasicBlock*, 16> BBPreds;
759 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
760 // It is faster to get preds from a PHI than with pred_iterator.
761 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
762 BBPreds.insert(BBPN->getIncomingBlock(i));
763 } else {
764 BBPreds.insert(pred_begin(BB), pred_end(BB));
765 }
766
767 // Walk the preds of DestBB.
768 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
769 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
770 if (BBPreds.count(Pred)) { // Common predecessor?
771 for (const PHINode &PN : DestBB->phis()) {
772 const Value *V1 = PN.getIncomingValueForBlock(Pred);
773 const Value *V2 = PN.getIncomingValueForBlock(BB);
774
775 // If V2 is a phi node in BB, look up what the mapped value will be.
776 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
777 if (V2PN->getParent() == BB)
778 V2 = V2PN->getIncomingValueForBlock(Pred);
779
780 // If there is a conflict, bail out.
781 if (V1 != V2) return false;
782 }
783 }
784 }
785
786 return true;
787}
788
789/// Eliminate a basic block that has only phi's and an unconditional branch in
790/// it.
791void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
792 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
793 BasicBlock *DestBB = BI->getSuccessor(0);
794
795 LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
<< *BB << *DestBB; } } while (false)
796 << *BB << *DestBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
<< *BB << *DestBB; } } while (false)
;
797
798 // If the destination block has a single pred, then this is a trivial edge,
799 // just collapse it.
800 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
801 if (SinglePred != DestBB) {
802 assert(SinglePred == BB &&((SinglePred == BB && "Single predecessor not the same as predecessor"
) ? static_cast<void> (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 803, __PRETTY_FUNCTION__))
803 "Single predecessor not the same as predecessor")((SinglePred == BB && "Single predecessor not the same as predecessor"
) ? static_cast<void> (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 803, __PRETTY_FUNCTION__))
;
804 // Merge DestBB into SinglePred/BB and delete it.
805 MergeBlockIntoPredecessor(DestBB);
806 // Note: BB(=SinglePred) will not be deleted on this path.
807 // DestBB(=its single successor) is the one that was deleted.
808 LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "AFTER:\n" << *SinglePred
<< "\n\n\n"; } } while (false)
;
809 return;
810 }
811 }
812
813 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
814 // to handle the new incoming edges it is about to have.
815 for (PHINode &PN : DestBB->phis()) {
816 // Remove the incoming value for BB, and remember it.
817 Value *InVal = PN.removeIncomingValue(BB, false);
818
819 // Two options: either the InVal is a phi node defined in BB or it is some
820 // value that dominates BB.
821 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
822 if (InValPhi && InValPhi->getParent() == BB) {
823 // Add all of the input values of the input PHI as inputs of this phi.
824 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
825 PN.addIncoming(InValPhi->getIncomingValue(i),
826 InValPhi->getIncomingBlock(i));
827 } else {
828 // Otherwise, add one instance of the dominating value for each edge that
829 // we will be adding.
830 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
831 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
832 PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
833 } else {
834 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
835 PN.addIncoming(InVal, *PI);
836 }
837 }
838 }
839
840 // The PHIs are now updated, change everything that refers to BB to use
841 // DestBB and remove BB.
842 BB->replaceAllUsesWith(DestBB);
843 BB->eraseFromParent();
844 ++NumBlocksElim;
845
846 LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "AFTER:\n" << *DestBB
<< "\n\n\n"; } } while (false)
;
847}
848
849// Computes a map of base pointer relocation instructions to corresponding
850// derived pointer relocation instructions given a vector of all relocate calls
851static void computeBaseDerivedRelocateMap(
852 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
853 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
854 &RelocateInstMap) {
855 // Collect information in two maps: one primarily for locating the base object
856 // while filling the second map; the second map is the final structure holding
857 // a mapping between Base and corresponding Derived relocate calls
858 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
859 for (auto *ThisRelocate : AllRelocateCalls) {
860 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
861 ThisRelocate->getDerivedPtrIndex());
862 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
863 }
864 for (auto &Item : RelocateIdxMap) {
865 std::pair<unsigned, unsigned> Key = Item.first;
866 if (Key.first == Key.second)
867 // Base relocation: nothing to insert
868 continue;
869
870 GCRelocateInst *I = Item.second;
871 auto BaseKey = std::make_pair(Key.first, Key.first);
872
873 // We're iterating over RelocateIdxMap so we cannot modify it.
874 auto MaybeBase = RelocateIdxMap.find(BaseKey);
875 if (MaybeBase == RelocateIdxMap.end())
876 // TODO: We might want to insert a new base object relocate and gep off
877 // that, if there are enough derived object relocates.
878 continue;
879
880 RelocateInstMap[MaybeBase->second].push_back(I);
881 }
882}
883
884// Accepts a GEP and extracts the operands into a vector provided they're all
885// small integer constants
886static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
887 SmallVectorImpl<Value *> &OffsetV) {
888 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
889 // Only accept small constant integer operands
890 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
891 if (!Op || Op->getZExtValue() > 20)
892 return false;
893 }
894
895 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
896 OffsetV.push_back(GEP->getOperand(i));
897 return true;
898}
899
900// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
901// replace, computes a replacement, and affects it.
902static bool
903simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
904 const SmallVectorImpl<GCRelocateInst *> &Targets) {
905 bool MadeChange = false;
906 // We must ensure the relocation of derived pointer is defined after
907 // relocation of base pointer. If we find a relocation corresponding to base
908 // defined earlier than relocation of base then we move relocation of base
909 // right before found relocation. We consider only relocation in the same
910 // basic block as relocation of base. Relocations from other basic block will
911 // be skipped by optimization and we do not care about them.
912 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
913 &*R != RelocatedBase; ++R)
914 if (auto RI = dyn_cast<GCRelocateInst>(R))
915 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
916 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
917 RelocatedBase->moveBefore(RI);
918 break;
919 }
920
921 for (GCRelocateInst *ToReplace : Targets) {
922 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&((ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex
() && "Not relocating a derived object of the original base object"
) ? static_cast<void> (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 923, __PRETTY_FUNCTION__))
923 "Not relocating a derived object of the original base object")((ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex
() && "Not relocating a derived object of the original base object"
) ? static_cast<void> (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 923, __PRETTY_FUNCTION__))
;
924 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
925 // A duplicate relocate call. TODO: coalesce duplicates.
926 continue;
927 }
928
929 if (RelocatedBase->getParent() != ToReplace->getParent()) {
930 // Base and derived relocates are in different basic blocks.
931 // In this case transform is only valid when base dominates derived
932 // relocate. However it would be too expensive to check dominance
933 // for each such relocate, so we skip the whole transformation.
934 continue;
935 }
936
937 Value *Base = ToReplace->getBasePtr();
938 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
939 if (!Derived || Derived->getPointerOperand() != Base)
940 continue;
941
942 SmallVector<Value *, 2> OffsetV;
943 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
944 continue;
945
946 // Create a Builder and replace the target callsite with a gep
947 assert(RelocatedBase->getNextNode() &&((RelocatedBase->getNextNode() && "Should always have one since it's not a terminator"
) ? static_cast<void> (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 948, __PRETTY_FUNCTION__))
948 "Should always have one since it's not a terminator")((RelocatedBase->getNextNode() && "Should always have one since it's not a terminator"
) ? static_cast<void> (0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 948, __PRETTY_FUNCTION__))
;
949
950 // Insert after RelocatedBase
951 IRBuilder<> Builder(RelocatedBase->getNextNode());
952 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
953
954 // If gc_relocate does not match the actual type, cast it to the right type.
955 // In theory, there must be a bitcast after gc_relocate if the type does not
956 // match, and we should reuse it to get the derived pointer. But it could be
957 // cases like this:
958 // bb1:
959 // ...
960 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
961 // br label %merge
962 //
963 // bb2:
964 // ...
965 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
966 // br label %merge
967 //
968 // merge:
969 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
970 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
971 //
972 // In this case, we can not find the bitcast any more. So we insert a new bitcast
973 // no matter there is already one or not. In this way, we can handle all cases, and
974 // the extra bitcast should be optimized away in later passes.
975 Value *ActualRelocatedBase = RelocatedBase;
976 if (RelocatedBase->getType() != Base->getType()) {
977 ActualRelocatedBase =
978 Builder.CreateBitCast(RelocatedBase, Base->getType());
979 }
980 Value *Replacement = Builder.CreateGEP(
981 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
982 Replacement->takeName(ToReplace);
983 // If the newly generated derived pointer's type does not match the original derived
984 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
985 Value *ActualReplacement = Replacement;
986 if (Replacement->getType() != ToReplace->getType()) {
987 ActualReplacement =
988 Builder.CreateBitCast(Replacement, ToReplace->getType());
989 }
990 ToReplace->replaceAllUsesWith(ActualReplacement);
991 ToReplace->eraseFromParent();
992
993 MadeChange = true;
994 }
995 return MadeChange;
996}
997
998// Turns this:
999//
1000// %base = ...
1001// %ptr = gep %base + 15
1002// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1003// %base' = relocate(%tok, i32 4, i32 4)
1004// %ptr' = relocate(%tok, i32 4, i32 5)
1005// %val = load %ptr'
1006//
1007// into this:
1008//
1009// %base = ...
1010// %ptr = gep %base + 15
1011// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1012// %base' = gc.relocate(%tok, i32 4, i32 4)
1013// %ptr' = gep %base' + 15
1014// %val = load %ptr'
1015bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1016 bool MadeChange = false;
1017 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1018
1019 for (auto *U : I.users())
1020 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1021 // Collect all the relocate calls associated with a statepoint
1022 AllRelocateCalls.push_back(Relocate);
1023
1024 // We need atleast one base pointer relocation + one derived pointer
1025 // relocation to mangle
1026 if (AllRelocateCalls.size() < 2)
1027 return false;
1028
1029 // RelocateInstMap is a mapping from the base relocate instruction to the
1030 // corresponding derived relocate instructions
1031 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1032 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1033 if (RelocateInstMap.empty())
1034 return false;
1035
1036 for (auto &Item : RelocateInstMap)
1037 // Item.first is the RelocatedBase to offset against
1038 // Item.second is the vector of Targets to replace
1039 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1040 return MadeChange;
1041}
1042
1043/// SinkCast - Sink the specified cast instruction into its user blocks
1044static bool SinkCast(CastInst *CI) {
1045 BasicBlock *DefBB = CI->getParent();
1046
1047 /// InsertedCasts - Only insert a cast in each block once.
1048 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1049
1050 bool MadeChange = false;
1051 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1052 UI != E; ) {
1053 Use &TheUse = UI.getUse();
1054 Instruction *User = cast<Instruction>(*UI);
1055
1056 // Figure out which BB this cast is used in. For PHI's this is the
1057 // appropriate predecessor block.
1058 BasicBlock *UserBB = User->getParent();
1059 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1060 UserBB = PN->getIncomingBlock(TheUse);
1061 }
1062
1063 // Preincrement use iterator so we don't invalidate it.
1064 ++UI;
1065
1066 // The first insertion point of a block containing an EH pad is after the
1067 // pad. If the pad is the user, we cannot sink the cast past the pad.
1068 if (User->isEHPad())
1069 continue;
1070
1071 // If the block selected to receive the cast is an EH pad that does not
1072 // allow non-PHI instructions before the terminator, we can't sink the
1073 // cast.
1074 if (UserBB->getTerminator()->isEHPad())
1075 continue;
1076
1077 // If this user is in the same block as the cast, don't change the cast.
1078 if (UserBB == DefBB) continue;
1079
1080 // If we have already inserted a cast into this block, use it.
1081 CastInst *&InsertedCast = InsertedCasts[UserBB];
1082
1083 if (!InsertedCast) {
1084 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1085 assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0)
: __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1085, __PRETTY_FUNCTION__))
;
1086 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1087 CI->getType(), "", &*InsertPt);
1088 InsertedCast->setDebugLoc(CI->getDebugLoc());
1089 }
1090
1091 // Replace a use of the cast with a use of the new cast.
1092 TheUse = InsertedCast;
1093 MadeChange = true;
1094 ++NumCastUses;
1095 }
1096
1097 // If we removed all uses, nuke the cast.
1098 if (CI->use_empty()) {
1099 salvageDebugInfo(*CI);
1100 CI->eraseFromParent();
1101 MadeChange = true;
1102 }
1103
1104 return MadeChange;
1105}
1106
1107/// If the specified cast instruction is a noop copy (e.g. it's casting from
1108/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1109/// reduce the number of virtual registers that must be created and coalesced.
1110///
1111/// Return true if any changes are made.
1112static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1113 const DataLayout &DL) {
1114 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1115 // than sinking only nop casts, but is helpful on some platforms.
1116 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1117 if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1118 ASC->getDestAddressSpace()))
1119 return false;
1120 }
1121
1122 // If this is a noop copy,
1123 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1124 EVT DstVT = TLI.getValueType(DL, CI->getType());
1125
1126 // This is an fp<->int conversion?
1127 if (SrcVT.isInteger() != DstVT.isInteger())
1128 return false;
1129
1130 // If this is an extension, it will be a zero or sign extension, which
1131 // isn't a noop.
1132 if (SrcVT.bitsLT(DstVT)) return false;
1133
1134 // If these values will be promoted, find out what they will be promoted
1135 // to. This helps us consider truncates on PPC as noop copies when they
1136 // are.
1137 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1138 TargetLowering::TypePromoteInteger)
1139 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1140 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1141 TargetLowering::TypePromoteInteger)
1142 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1143
1144 // If, after promotion, these are the same types, this is a noop copy.
1145 if (SrcVT != DstVT)
1146 return false;
1147
1148 return SinkCast(CI);
1149}
1150
1151/// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1152/// possible.
1153///
1154/// Return true if any changes were made.
1155static bool CombineUAddWithOverflow(CmpInst *CI) {
1156 Value *A, *B;
1157 Instruction *AddI;
1158 if (!match(CI,
1159 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1160 return false;
1161
1162 Type *Ty = AddI->getType();
1163 if (!isa<IntegerType>(Ty))
1164 return false;
1165
1166 // We don't want to move around uses of condition values this late, so we we
1167 // check if it is legal to create the call to the intrinsic in the basic
1168 // block containing the icmp:
1169
1170 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1171 return false;
1172
1173#ifndef NDEBUG
1174 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1175 // for now:
1176 if (AddI->hasOneUse())
1177 assert(*AddI->user_begin() == CI && "expected!")((*AddI->user_begin() == CI && "expected!") ? static_cast
<void> (0) : __assert_fail ("*AddI->user_begin() == CI && \"expected!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1177, __PRETTY_FUNCTION__))
;
1178#endif
1179
1180 Module *M = CI->getModule();
1181 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1182
1183 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1184
1185 DebugLoc Loc = CI->getDebugLoc();
1186 auto *UAddWithOverflow =
1187 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1188 UAddWithOverflow->setDebugLoc(Loc);
1189 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1190 UAdd->setDebugLoc(Loc);
1191 auto *Overflow =
1192 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1193 Overflow->setDebugLoc(Loc);
1194
1195 CI->replaceAllUsesWith(Overflow);
1196 AddI->replaceAllUsesWith(UAdd);
1197 CI->eraseFromParent();
1198 AddI->eraseFromParent();
1199 return true;
1200}
1201
1202/// Sink the given CmpInst into user blocks to reduce the number of virtual
1203/// registers that must be created and coalesced. This is a clear win except on
1204/// targets with multiple condition code registers (PowerPC), where it might
1205/// lose; some adjustment may be wanted there.
1206///
1207/// Return true if any changes are made.
1208static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1209 BasicBlock *DefBB = CI->getParent();
1210
1211 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1212 if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1213 return false;
1214
1215 // Only insert a cmp in each block once.
1216 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1217
1218 bool MadeChange = false;
1219 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1220 UI != E; ) {
1221 Use &TheUse = UI.getUse();
1222 Instruction *User = cast<Instruction>(*UI);
1223
1224 // Preincrement use iterator so we don't invalidate it.
1225 ++UI;
1226
1227 // Don't bother for PHI nodes.
1228 if (isa<PHINode>(User))
1229 continue;
1230
1231 // Figure out which BB this cmp is used in.
1232 BasicBlock *UserBB = User->getParent();
1233
1234 // If this user is in the same block as the cmp, don't change the cmp.
1235 if (UserBB == DefBB) continue;
1236
1237 // If we have already inserted a cmp into this block, use it.
1238 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1239
1240 if (!InsertedCmp) {
1241 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1242 assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0)
: __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1242, __PRETTY_FUNCTION__))
;
1243 InsertedCmp =
1244 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1245 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1246 // Propagate the debug info.
1247 InsertedCmp->setDebugLoc(CI->getDebugLoc());
1248 }
1249
1250 // Replace a use of the cmp with a use of the new cmp.
1251 TheUse = InsertedCmp;
1252 MadeChange = true;
1253 ++NumCmpUses;
1254 }
1255
1256 // If we removed all uses, nuke the cmp.
1257 if (CI->use_empty()) {
1258 CI->eraseFromParent();
1259 MadeChange = true;
1260 }
1261
1262 return MadeChange;
1263}
1264
1265static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1266 if (SinkCmpExpression(CI, TLI))
1267 return true;
1268
1269 if (CombineUAddWithOverflow(CI))
1270 return true;
1271
1272 return false;
1273}
1274
1275/// Duplicate and sink the given 'and' instruction into user blocks where it is
1276/// used in a compare to allow isel to generate better code for targets where
1277/// this operation can be combined.
1278///
1279/// Return true if any changes are made.
1280static bool sinkAndCmp0Expression(Instruction *AndI,
1281 const TargetLowering &TLI,
1282 SetOfInstrs &InsertedInsts) {
1283 // Double-check that we're not trying to optimize an instruction that was
1284 // already optimized by some other part of this pass.
1285 assert(!InsertedInsts.count(AndI) &&((!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction"
) ? static_cast<void> (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1286, __PRETTY_FUNCTION__))
1286 "Attempting to optimize already optimized and instruction")((!InsertedInsts.count(AndI) && "Attempting to optimize already optimized and instruction"
) ? static_cast<void> (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1286, __PRETTY_FUNCTION__))
;
1287 (void) InsertedInsts;
1288
1289 // Nothing to do for single use in same basic block.
1290 if (AndI->hasOneUse() &&
1291 AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1292 return false;
1293
1294 // Try to avoid cases where sinking/duplicating is likely to increase register
1295 // pressure.
1296 if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1297 !isa<ConstantInt>(AndI->getOperand(1)) &&
1298 AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1299 return false;
1300
1301 for (auto *U : AndI->users()) {
1302 Instruction *User = cast<Instruction>(U);
1303
1304 // Only sink for and mask feeding icmp with 0.
1305 if (!isa<ICmpInst>(User))
1306 return false;
1307
1308 auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1309 if (!CmpC || !CmpC->isZero())
1310 return false;
1311 }
1312
1313 if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1314 return false;
1315
1316 LLVM_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)
;
1317 LLVM_DEBUG(AndI->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { AndI->getParent()->dump(); } } while
(false)
;
1318
1319 // Push the 'and' into the same block as the icmp 0. There should only be
1320 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1321 // others, so we don't need to keep track of which BBs we insert into.
1322 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1323 UI != E; ) {
1324 Use &TheUse = UI.getUse();
1325 Instruction *User = cast<Instruction>(*UI);
1326
1327 // Preincrement use iterator so we don't invalidate it.
1328 ++UI;
1329
1330 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "sinking 'and' use: " <<
*User << "\n"; } } while (false)
;
1331
1332 // Keep the 'and' in the same place if the use is already in the same block.
1333 Instruction *InsertPt =
1334 User->getParent() == AndI->getParent() ? AndI : User;
1335 Instruction *InsertedAnd =
1336 BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1337 AndI->getOperand(1), "", InsertPt);
1338 // Propagate the debug info.
1339 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1340
1341 // Replace a use of the 'and' with a use of the new 'and'.
1342 TheUse = InsertedAnd;
1343 ++NumAndUses;
1344 LLVM_DEBUG(User->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { User->getParent()->dump(); } } while
(false)
;
1345 }
1346
1347 // We removed all uses, nuke the and.
1348 AndI->eraseFromParent();
1349 return true;
1350}
1351
1352/// Check if the candidates could be combined with a shift instruction, which
1353/// includes:
1354/// 1. Truncate instruction
1355/// 2. And instruction and the imm is a mask of the low bits:
1356/// imm & (imm+1) == 0
1357static bool isExtractBitsCandidateUse(Instruction *User) {
1358 if (!isa<TruncInst>(User)) {
1359 if (User->getOpcode() != Instruction::And ||
1360 !isa<ConstantInt>(User->getOperand(1)))
1361 return false;
1362
1363 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1364
1365 if ((Cimm & (Cimm + 1)).getBoolValue())
1366 return false;
1367 }
1368 return true;
1369}
1370
1371/// Sink both shift and truncate instruction to the use of truncate's BB.
1372static bool
1373SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1374 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1375 const TargetLowering &TLI, const DataLayout &DL) {
1376 BasicBlock *UserBB = User->getParent();
1377 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1378 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1379 bool MadeChange = false;
1380
1381 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1382 TruncE = TruncI->user_end();
1383 TruncUI != TruncE;) {
1384
1385 Use &TruncTheUse = TruncUI.getUse();
1386 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1387 // Preincrement use iterator so we don't invalidate it.
1388
1389 ++TruncUI;
1390
1391 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1392 if (!ISDOpcode)
1393 continue;
1394
1395 // If the use is actually a legal node, there will not be an
1396 // implicit truncate.
1397 // FIXME: always querying the result type is just an
1398 // approximation; some nodes' legality is determined by the
1399 // operand or other means. There's no good way to find out though.
1400 if (TLI.isOperationLegalOrCustom(
1401 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1402 continue;
1403
1404 // Don't bother for PHI nodes.
1405 if (isa<PHINode>(TruncUser))
1406 continue;
1407
1408 BasicBlock *TruncUserBB = TruncUser->getParent();
1409
1410 if (UserBB == TruncUserBB)
1411 continue;
1412
1413 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1414 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1415
1416 if (!InsertedShift && !InsertedTrunc) {
1417 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1418 assert(InsertPt != TruncUserBB->end())((InsertPt != TruncUserBB->end()) ? static_cast<void>
(0) : __assert_fail ("InsertPt != TruncUserBB->end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1418, __PRETTY_FUNCTION__))
;
1419 // Sink the shift
1420 if (ShiftI->getOpcode() == Instruction::AShr)
1421 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1422 "", &*InsertPt);
1423 else
1424 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1425 "", &*InsertPt);
1426 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1427
1428 // Sink the trunc
1429 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1430 TruncInsertPt++;
1431 assert(TruncInsertPt != TruncUserBB->end())((TruncInsertPt != TruncUserBB->end()) ? static_cast<void
> (0) : __assert_fail ("TruncInsertPt != TruncUserBB->end()"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1431, __PRETTY_FUNCTION__))
;
1432
1433 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1434 TruncI->getType(), "", &*TruncInsertPt);
1435 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1436
1437 MadeChange = true;
1438
1439 TruncTheUse = InsertedTrunc;
1440 }
1441 }
1442 return MadeChange;
1443}
1444
1445/// Sink the shift *right* instruction into user blocks if the uses could
1446/// potentially be combined with this shift instruction and generate BitExtract
1447/// instruction. It will only be applied if the architecture supports BitExtract
1448/// instruction. Here is an example:
1449/// BB1:
1450/// %x.extract.shift = lshr i64 %arg1, 32
1451/// BB2:
1452/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1453/// ==>
1454///
1455/// BB2:
1456/// %x.extract.shift.1 = lshr i64 %arg1, 32
1457/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1458///
1459/// CodeGen will recognize the pattern in BB2 and generate BitExtract
1460/// instruction.
1461/// Return true if any changes are made.
1462static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1463 const TargetLowering &TLI,
1464 const DataLayout &DL) {
1465 BasicBlock *DefBB = ShiftI->getParent();
1466
1467 /// Only insert instructions in each block once.
1468 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1469
1470 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1471
1472 bool MadeChange = false;
1473 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1474 UI != E;) {
1475 Use &TheUse = UI.getUse();
1476 Instruction *User = cast<Instruction>(*UI);
1477 // Preincrement use iterator so we don't invalidate it.
1478 ++UI;
1479
1480 // Don't bother for PHI nodes.
1481 if (isa<PHINode>(User))
1482 continue;
1483
1484 if (!isExtractBitsCandidateUse(User))
1485 continue;
1486
1487 BasicBlock *UserBB = User->getParent();
1488
1489 if (UserBB == DefBB) {
1490 // If the shift and truncate instruction are in the same BB. The use of
1491 // the truncate(TruncUse) may still introduce another truncate if not
1492 // legal. In this case, we would like to sink both shift and truncate
1493 // instruction to the BB of TruncUse.
1494 // for example:
1495 // BB1:
1496 // i64 shift.result = lshr i64 opnd, imm
1497 // trunc.result = trunc shift.result to i16
1498 //
1499 // BB2:
1500 // ----> We will have an implicit truncate here if the architecture does
1501 // not have i16 compare.
1502 // cmp i16 trunc.result, opnd2
1503 //
1504 if (isa<TruncInst>(User) && shiftIsLegal
1505 // If the type of the truncate is legal, no truncate will be
1506 // introduced in other basic blocks.
1507 &&
1508 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1509 MadeChange =
1510 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1511
1512 continue;
1513 }
1514 // If we have already inserted a shift into this block, use it.
1515 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1516
1517 if (!InsertedShift) {
1518 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1519 assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0)
: __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 1519, __PRETTY_FUNCTION__))
;
1520
1521 if (ShiftI->getOpcode() == Instruction::AShr)
1522 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1523 "", &*InsertPt);
1524 else
1525 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1526 "", &*InsertPt);
1527 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1528
1529 MadeChange = true;
1530 }
1531
1532 // Replace a use of the shift with a use of the new shift.
1533 TheUse = InsertedShift;
1534 }
1535
1536 // If we removed all uses, nuke the shift.
1537 if (ShiftI->use_empty()) {
1538 salvageDebugInfo(*ShiftI);
1539 ShiftI->eraseFromParent();
1540 }
1541
1542 return MadeChange;
1543}
1544
1545/// If counting leading or trailing zeros is an expensive operation and a zero
1546/// input is defined, add a check for zero to avoid calling the intrinsic.
1547///
1548/// We want to transform:
1549/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1550///
1551/// into:
1552/// entry:
1553/// %cmpz = icmp eq i64 %A, 0
1554/// br i1 %cmpz, label %cond.end, label %cond.false
1555/// cond.false:
1556/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1557/// br label %cond.end
1558/// cond.end:
1559/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1560///
1561/// If the transform is performed, return true and set ModifiedDT to true.
1562static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1563 const TargetLowering *TLI,
1564 const DataLayout *DL,
1565 bool &ModifiedDT) {
1566 if (!TLI || !DL)
1567 return false;
1568
1569 // If a zero input is undefined, it doesn't make sense to despeculate that.
1570 if (match(CountZeros->getOperand(1), m_One()))
1571 return false;
1572
1573 // If it's cheap to speculate, there's nothing to do.
1574 auto IntrinsicID = CountZeros->getIntrinsicID();
1575 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1576 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1577 return false;
1578
1579 // Only handle legal scalar cases. Anything else requires too much work.
1580 Type *Ty = CountZeros->getType();
1581 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1582 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1583 return false;
1584
1585 // The intrinsic will be sunk behind a compare against zero and branch.
1586 BasicBlock *StartBlock = CountZeros->getParent();
1587 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1588
1589 // Create another block after the count zero intrinsic. A PHI will be added
1590 // in this block to select the result of the intrinsic or the bit-width
1591 // constant if the input to the intrinsic is zero.
1592 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1593 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1594
1595 // Set up a builder to create a compare, conditional branch, and PHI.
1596 IRBuilder<> Builder(CountZeros->getContext());
1597 Builder.SetInsertPoint(StartBlock->getTerminator());
1598 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1599
1600 // Replace the unconditional branch that was created by the first split with
1601 // a compare against zero and a conditional branch.
1602 Value *Zero = Constant::getNullValue(Ty);
1603 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1604 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1605 StartBlock->getTerminator()->eraseFromParent();
1606
1607 // Create a PHI in the end block to select either the output of the intrinsic
1608 // or the bit width of the operand.
1609 Builder.SetInsertPoint(&EndBlock->front());
1610 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1611 CountZeros->replaceAllUsesWith(PN);
1612 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1613 PN->addIncoming(BitWidth, StartBlock);
1614 PN->addIncoming(CountZeros, CallBlock);
1615
1616 // We are explicitly handling the zero case, so we can set the intrinsic's
1617 // undefined zero argument to 'true'. This will also prevent reprocessing the
1618 // intrinsic; we only despeculate when a zero input is defined.
1619 CountZeros->setArgOperand(1, Builder.getTrue());
1620 ModifiedDT = true;
1621 return true;
1622}
1623
1624bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1625 BasicBlock *BB = CI->getParent();
1626
1627 // Lower inline assembly if we can.
1628 // If we found an inline asm expession, and if the target knows how to
1629 // lower it to normal LLVM code, do so now.
1630 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1631 if (TLI->ExpandInlineAsm(CI)) {
1632 // Avoid invalidating the iterator.
1633 CurInstIterator = BB->begin();
1634 // Avoid processing instructions out of order, which could cause
1635 // reuse before a value is defined.
1636 SunkAddrs.clear();
1637 return true;
1638 }
1639 // Sink address computing for memory operands into the block.
1640 if (optimizeInlineAsmInst(CI))
1641 return true;
1642 }
1643
1644 // Align the pointer arguments to this call if the target thinks it's a good
1645 // idea
1646 unsigned MinSize, PrefAlign;
1647 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1648 for (auto &Arg : CI->arg_operands()) {
1649 // We want to align both objects whose address is used directly and
1650 // objects whose address is used in casts and GEPs, though it only makes
1651 // sense for GEPs if the offset is a multiple of the desired alignment and
1652 // if size - offset meets the size threshold.
1653 if (!Arg->getType()->isPointerTy())
1654 continue;
1655 APInt Offset(DL->getIndexSizeInBits(
1656 cast<PointerType>(Arg->getType())->getAddressSpace()),
1657 0);
1658 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1659 uint64_t Offset2 = Offset.getLimitedValue();
1660 if ((Offset2 & (PrefAlign-1)) != 0)
1661 continue;
1662 AllocaInst *AI;
1663 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1664 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1665 AI->setAlignment(PrefAlign);
1666 // Global variables can only be aligned if they are defined in this
1667 // object (i.e. they are uniquely initialized in this object), and
1668 // over-aligning global variables that have an explicit section is
1669 // forbidden.
1670 GlobalVariable *GV;
1671 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1672 GV->getPointerAlignment(*DL) < PrefAlign &&
1673 DL->getTypeAllocSize(GV->getValueType()) >=
1674 MinSize + Offset2)
1675 GV->setAlignment(PrefAlign);
1676 }
1677 // If this is a memcpy (or similar) then we may be able to improve the
1678 // alignment
1679 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1680 unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
1681 if (DestAlign > MI->getDestAlignment())
1682 MI->setDestAlignment(DestAlign);
1683 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1684 unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1685 if (SrcAlign > MTI->getSourceAlignment())
1686 MTI->setSourceAlignment(SrcAlign);
1687 }
1688 }
1689 }
1690
1691 // If we have a cold call site, try to sink addressing computation into the
1692 // cold block. This interacts with our handling for loads and stores to
1693 // ensure that we can fold all uses of a potential addressing computation
1694 // into their uses. TODO: generalize this to work over profiling data
1695 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1696 for (auto &Arg : CI->arg_operands()) {
1697 if (!Arg->getType()->isPointerTy())
1698 continue;
1699 unsigned AS = Arg->getType()->getPointerAddressSpace();
1700 return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1701 }
1702
1703 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1704 if (II) {
1705 switch (II->getIntrinsicID()) {
1706 default: break;
1707 case Intrinsic::objectsize: {
1708 // Lower all uses of llvm.objectsize.*
1709 ConstantInt *RetVal =
1710 lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1711
1712 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1713 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1714 });
1715 return true;
1716 }
1717 case Intrinsic::is_constant: {
1718 // If is_constant hasn't folded away yet, lower it to false now.
1719 Constant *RetVal = ConstantInt::get(II->getType(), 0);
1720 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1721 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1722 });
1723 return true;
1724 }
1725 case Intrinsic::aarch64_stlxr:
1726 case Intrinsic::aarch64_stxr: {
1727 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1728 if (!ExtVal || !ExtVal->hasOneUse() ||
1729 ExtVal->getParent() == CI->getParent())
1730 return false;
1731 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1732 ExtVal->moveBefore(CI);
1733 // Mark this instruction as "inserted by CGP", so that other
1734 // optimizations don't touch it.
1735 InsertedInsts.insert(ExtVal);
1736 return true;
1737 }
1738 case Intrinsic::launder_invariant_group:
1739 case Intrinsic::strip_invariant_group: {
1740 Value *ArgVal = II->getArgOperand(0);
1741 auto it = LargeOffsetGEPMap.find(II);
1742 if (it != LargeOffsetGEPMap.end()) {
1743 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
1744 // Make sure not to have to deal with iterator invalidation
1745 // after possibly adding ArgVal to LargeOffsetGEPMap.
1746 auto GEPs = std::move(it->second);
1747 LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
1748 LargeOffsetGEPMap.erase(II);
1749 }
1750
1751 II->replaceAllUsesWith(ArgVal);
1752 II->eraseFromParent();
1753 return true;
1754 }
1755 case Intrinsic::cttz:
1756 case Intrinsic::ctlz:
1757 // If counting zeros is expensive, try to avoid it.
1758 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1759 }
1760
1761 if (TLI) {
1762 SmallVector<Value*, 2> PtrOps;
1763 Type *AccessTy;
1764 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1765 while (!PtrOps.empty()) {
1766 Value *PtrVal = PtrOps.pop_back_val();
1767 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1768 if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1769 return true;
1770 }
1771 }
1772 }
1773
1774 // From here on out we're working with named functions.
1775 if (!CI->getCalledFunction()) return false;
1776
1777 // Lower all default uses of _chk calls. This is very similar
1778 // to what InstCombineCalls does, but here we are only lowering calls
1779 // to fortified library functions (e.g. __memcpy_chk) that have the default
1780 // "don't know" as the objectsize. Anything else should be left alone.
1781 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1782 if (Value *V = Simplifier.optimizeCall(CI)) {
1783 CI->replaceAllUsesWith(V);
1784 CI->eraseFromParent();
1785 return true;
1786 }
1787
1788 return false;
1789}
1790
1791/// Look for opportunities to duplicate return instructions to the predecessor
1792/// to enable tail call optimizations. The case it is currently looking for is:
1793/// @code
1794/// bb0:
1795/// %tmp0 = tail call i32 @f0()
1796/// br label %return
1797/// bb1:
1798/// %tmp1 = tail call i32 @f1()
1799/// br label %return
1800/// bb2:
1801/// %tmp2 = tail call i32 @f2()
1802/// br label %return
1803/// return:
1804/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1805/// ret i32 %retval
1806/// @endcode
1807///
1808/// =>
1809///
1810/// @code
1811/// bb0:
1812/// %tmp0 = tail call i32 @f0()
1813/// ret i32 %tmp0
1814/// bb1:
1815/// %tmp1 = tail call i32 @f1()
1816/// ret i32 %tmp1
1817/// bb2:
1818/// %tmp2 = tail call i32 @f2()
1819/// ret i32 %tmp2
1820/// @endcode
1821bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1822 if (!TLI)
1823 return false;
1824
1825 ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1826 if (!RetI)
1827 return false;
1828
1829 PHINode *PN = nullptr;
1830 BitCastInst *BCI = nullptr;
1831 Value *V = RetI->getReturnValue();
1832 if (V) {
1833 BCI = dyn_cast<BitCastInst>(V);
1834 if (BCI)
1835 V = BCI->getOperand(0);
1836
1837 PN = dyn_cast<PHINode>(V);
1838 if (!PN)
1839 return false;
1840 }
1841
1842 if (PN && PN->getParent() != BB)
1843 return false;
1844
1845 // Make sure there are no instructions between the PHI and return, or that the
1846 // return is the first instruction in the block.
1847 if (PN) {
1848 BasicBlock::iterator BI = BB->begin();
1849 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1850 if (&*BI == BCI)
1851 // Also skip over the bitcast.
1852 ++BI;
1853 if (&*BI != RetI)
1854 return false;
1855 } else {
1856 BasicBlock::iterator BI = BB->begin();
1857 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1858 if (&*BI != RetI)
1859 return false;
1860 }
1861
1862 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1863 /// call.
1864 const Function *F = BB->getParent();
1865 SmallVector<CallInst*, 4> TailCalls;
1866 if (PN) {
1867 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1868 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1869 // Make sure the phi value is indeed produced by the tail call.
1870 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1871 TLI->mayBeEmittedAsTailCall(CI) &&
1872 attributesPermitTailCall(F, CI, RetI, *TLI))
1873 TailCalls.push_back(CI);
1874 }
1875 } else {
1876 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1877 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1878 if (!VisitedBBs.insert(*PI).second)
1879 continue;
1880
1881 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1882 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1883 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1884 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1885 if (RI == RE)
1886 continue;
1887
1888 CallInst *CI = dyn_cast<CallInst>(&*RI);
1889 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1890 attributesPermitTailCall(F, CI, RetI, *TLI))
1891 TailCalls.push_back(CI);
1892 }
1893 }
1894
1895 bool Changed = false;
1896 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1897 CallInst *CI = TailCalls[i];
1898 CallSite CS(CI);
1899
1900 // Make sure the call instruction is followed by an unconditional branch to
1901 // the return block.
1902 BasicBlock *CallBB = CI->getParent();
1903 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1904 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1905 continue;
1906
1907 // Duplicate the return into CallBB.
1908 (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1909 ModifiedDT = Changed = true;
1910 ++NumRetsDup;
1911 }
1912
1913 // If we eliminated all predecessors of the block, delete the block now.
1914 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1915 BB->eraseFromParent();
1916
1917 return Changed;
1918}
1919
1920//===----------------------------------------------------------------------===//
1921// Memory Optimization
1922//===----------------------------------------------------------------------===//
1923
1924namespace {
1925
1926/// This is an extended version of TargetLowering::AddrMode
1927/// which holds actual Value*'s for register values.
1928struct ExtAddrMode : public TargetLowering::AddrMode {
1929 Value *BaseReg = nullptr;
1930 Value *ScaledReg = nullptr;
1931 Value *OriginalValue = nullptr;
1932
1933 enum FieldName {
1934 NoField = 0x00,
1935 BaseRegField = 0x01,
1936 BaseGVField = 0x02,
1937 BaseOffsField = 0x04,
1938 ScaledRegField = 0x08,
1939 ScaleField = 0x10,
1940 MultipleFields = 0xff
1941 };
1942
1943 ExtAddrMode() = default;
1944
1945 void print(raw_ostream &OS) const;
1946 void dump() const;
1947
1948 FieldName compare(const ExtAddrMode &other) {
1949 // First check that the types are the same on each field, as differing types
1950 // is something we can't cope with later on.
1951 if (BaseReg && other.BaseReg &&
1952 BaseReg->getType() != other.BaseReg->getType())
1953 return MultipleFields;
1954 if (BaseGV && other.BaseGV &&
1955 BaseGV->getType() != other.BaseGV->getType())
1956 return MultipleFields;
1957 if (ScaledReg && other.ScaledReg &&
1958 ScaledReg->getType() != other.ScaledReg->getType())
1959 return MultipleFields;
1960
1961 // Check each field to see if it differs.
1962 unsigned Result = NoField;
1963 if (BaseReg != other.BaseReg)
1964 Result |= BaseRegField;
1965 if (BaseGV != other.BaseGV)
1966 Result |= BaseGVField;
1967 if (BaseOffs != other.BaseOffs)
1968 Result |= BaseOffsField;
1969 if (ScaledReg != other.ScaledReg)
1970 Result |= ScaledRegField;
1971 // Don't count 0 as being a different scale, because that actually means
1972 // unscaled (which will already be counted by having no ScaledReg).
1973 if (Scale && other.Scale && Scale != other.Scale)
1974 Result |= ScaleField;
1975
1976 if (countPopulation(Result) > 1)
1977 return MultipleFields;
1978 else
1979 return static_cast<FieldName>(Result);
1980 }
1981
1982 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1983 // with no offset.
1984 bool isTrivial() {
1985 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1986 // trivial if at most one of these terms is nonzero, except that BaseGV and
1987 // BaseReg both being zero actually means a null pointer value, which we
1988 // consider to be 'non-zero' here.
1989 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1990 }
1991
1992 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1993 switch (Field) {
1994 default:
1995 return nullptr;
1996 case BaseRegField:
1997 return BaseReg;
1998 case BaseGVField:
1999 return BaseGV;
2000 case ScaledRegField:
2001 return ScaledReg;
2002 case BaseOffsField:
2003 return ConstantInt::get(IntPtrTy, BaseOffs);
2004 }
2005 }
2006
2007 void SetCombinedField(FieldName Field, Value *V,
2008 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2009 switch (Field) {
2010 default:
2011 llvm_unreachable("Unhandled fields are expected to be rejected earlier")::llvm::llvm_unreachable_internal("Unhandled fields are expected to be rejected earlier"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2011)
;
2012 break;
2013 case ExtAddrMode::BaseRegField:
2014 BaseReg = V;
2015 break;
2016 case ExtAddrMode::BaseGVField:
2017 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2018 // in the BaseReg field.
2019 assert(BaseReg == nullptr)((BaseReg == nullptr) ? static_cast<void> (0) : __assert_fail
("BaseReg == nullptr", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2019, __PRETTY_FUNCTION__))
;
2020 BaseReg = V;
2021 BaseGV = nullptr;
2022 break;
2023 case ExtAddrMode::ScaledRegField:
2024 ScaledReg = V;
2025 // If we have a mix of scaled and unscaled addrmodes then we want scale
2026 // to be the scale and not zero.
2027 if (!Scale)
2028 for (const ExtAddrMode &AM : AddrModes)
2029 if (AM.Scale) {
2030 Scale = AM.Scale;
2031 break;
2032 }
2033 break;
2034 case ExtAddrMode::BaseOffsField:
2035 // The offset is no longer a constant, so it goes in ScaledReg with a
2036 // scale of 1.
2037 assert(ScaledReg == nullptr)((ScaledReg == nullptr) ? static_cast<void> (0) : __assert_fail
("ScaledReg == nullptr", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2037, __PRETTY_FUNCTION__))
;
2038 ScaledReg = V;
2039 Scale = 1;
2040 BaseOffs = 0;
2041 break;
2042 }
2043 }
2044};
2045
2046} // end anonymous namespace
2047
2048#ifndef NDEBUG
2049static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2050 AM.print(OS);
2051 return OS;
2052}
2053#endif
2054
2055#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2056void ExtAddrMode::print(raw_ostream &OS) const {
2057 bool NeedPlus = false;
2058 OS << "[";
2059 if (BaseGV) {
2060 OS << (NeedPlus ? " + " : "")
2061 << "GV:";
2062 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2063 NeedPlus = true;
2064 }
2065
2066 if (BaseOffs) {
2067 OS << (NeedPlus ? " + " : "")
2068 << BaseOffs;
2069 NeedPlus = true;
2070 }
2071
2072 if (BaseReg) {
2073 OS << (NeedPlus ? " + " : "")
2074 << "Base:";
2075 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2076 NeedPlus = true;
2077 }
2078 if (Scale) {
2079 OS << (NeedPlus ? " + " : "")
2080 << Scale << "*";
2081 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2082 }
2083
2084 OS << ']';
2085}
2086
2087LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void ExtAddrMode::dump() const {
2088 print(dbgs());
2089 dbgs() << '\n';
2090}
2091#endif
2092
2093namespace {
2094
2095/// This class provides transaction based operation on the IR.
2096/// Every change made through this class is recorded in the internal state and
2097/// can be undone (rollback) until commit is called.
2098class TypePromotionTransaction {
2099 /// This represents the common interface of the individual transaction.
2100 /// Each class implements the logic for doing one specific modification on
2101 /// the IR via the TypePromotionTransaction.
2102 class TypePromotionAction {
2103 protected:
2104 /// The Instruction modified.
2105 Instruction *Inst;
2106
2107 public:
2108 /// Constructor of the action.
2109 /// The constructor performs the related action on the IR.
2110 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2111
2112 virtual ~TypePromotionAction() = default;
2113
2114 /// Undo the modification done by this action.
2115 /// When this method is called, the IR must be in the same state as it was
2116 /// before this action was applied.
2117 /// \pre Undoing the action works if and only if the IR is in the exact same
2118 /// state as it was directly after this action was applied.
2119 virtual void undo() = 0;
2120
2121 /// Advocate every change made by this action.
2122 /// When the results on the IR of the action are to be kept, it is important
2123 /// to call this function, otherwise hidden information may be kept forever.
2124 virtual void commit() {
2125 // Nothing to be done, this action is not doing anything.
2126 }
2127 };
2128
2129 /// Utility to remember the position of an instruction.
2130 class InsertionHandler {
2131 /// Position of an instruction.
2132 /// Either an instruction:
2133 /// - Is the first in a basic block: BB is used.
2134 /// - Has a previous instruction: PrevInst is used.
2135 union {
2136 Instruction *PrevInst;
2137 BasicBlock *BB;
2138 } Point;
2139
2140 /// Remember whether or not the instruction had a previous instruction.
2141 bool HasPrevInstruction;
2142
2143 public:
2144 /// Record the position of \p Inst.
2145 InsertionHandler(Instruction *Inst) {
2146 BasicBlock::iterator It = Inst->getIterator();
2147 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2148 if (HasPrevInstruction)
2149 Point.PrevInst = &*--It;
2150 else
2151 Point.BB = Inst->getParent();
2152 }
2153
2154 /// Insert \p Inst at the recorded position.
2155 void insert(Instruction *Inst) {
2156 if (HasPrevInstruction) {
2157 if (Inst->getParent())
2158 Inst->removeFromParent();
2159 Inst->insertAfter(Point.PrevInst);
2160 } else {
2161 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2162 if (Inst->getParent())
2163 Inst->moveBefore(Position);
2164 else
2165 Inst->insertBefore(Position);
2166 }
2167 }
2168 };
2169
2170 /// Move an instruction before another.
2171 class InstructionMoveBefore : public TypePromotionAction {
2172 /// Original position of the instruction.
2173 InsertionHandler Position;
2174
2175 public:
2176 /// Move \p Inst before \p Before.
2177 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2178 : TypePromotionAction(Inst), Position(Inst) {
2179 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Beforedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: move: " << *
Inst << "\nbefore: " << *Before << "\n"; } }
while (false)
2180 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: move: " << *
Inst << "\nbefore: " << *Before << "\n"; } }
while (false)
;
2181 Inst->moveBefore(Before);
2182 }
2183
2184 /// Move the instruction back to its original position.
2185 void undo() override {
2186 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: moveBefore: " <<
*Inst << "\n"; } } while (false)
;
2187 Position.insert(Inst);
2188 }
2189 };
2190
2191 /// Set the operand of an instruction with a new value.
2192 class OperandSetter : public TypePromotionAction {
2193 /// Original operand of the instruction.
2194 Value *Origin;
2195
2196 /// Index of the modified instruction.
2197 unsigned Idx;
2198
2199 public:
2200 /// Set \p Idx operand of \p Inst with \p NewVal.
2201 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2202 : TypePromotionAction(Inst), Idx(Idx) {
2203 LLVM_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)
2204 << "for:" << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
2205 << "with:" << *NewVal << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
;
2206 Origin = Inst->getOperand(Idx);
2207 Inst->setOperand(Idx, NewVal);
2208 }
2209
2210 /// Restore the original value of the instruction.
2211 void undo() override {
2212 LLVM_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)
2213 << "for: " << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
2214 << "with: " << *Origin << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
;
2215 Inst->setOperand(Idx, Origin);
2216 }
2217 };
2218
2219 /// Hide the operands of an instruction.
2220 /// Do as if this instruction was not using any of its operands.
2221 class OperandsHider : public TypePromotionAction {
2222 /// The list of original operands.
2223 SmallVector<Value *, 4> OriginalValues;
2224
2225 public:
2226 /// Remove \p Inst from the uses of the operands of \p Inst.
2227 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2228 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: OperandsHider: " <<
*Inst << "\n"; } } while (false)
;
2229 unsigned NumOpnds = Inst->getNumOperands();
2230 OriginalValues.reserve(NumOpnds);
2231 for (unsigned It = 0; It < NumOpnds; ++It) {
2232 // Save the current operand.
2233 Value *Val = Inst->getOperand(It);
2234 OriginalValues.push_back(Val);
2235 // Set a dummy one.
2236 // We could use OperandSetter here, but that would imply an overhead
2237 // that we are not willing to pay.
2238 Inst->setOperand(It, UndefValue::get(Val->getType()));
2239 }
2240 }
2241
2242 /// Restore the original list of uses.
2243 void undo() override {
2244 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: OperandsHider: "
<< *Inst << "\n"; } } while (false)
;
2245 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2246 Inst->setOperand(It, OriginalValues[It]);
2247 }
2248 };
2249
2250 /// Build a truncate instruction.
2251 class TruncBuilder : public TypePromotionAction {
2252 Value *Val;
2253
2254 public:
2255 /// Build a truncate instruction of \p Opnd producing a \p Ty
2256 /// result.
2257 /// trunc Opnd to Ty.
2258 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2259 IRBuilder<> Builder(Opnd);
2260 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2261 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2262 }
2263
2264 /// Get the built value.
2265 Value *getBuiltValue() { return Val; }
2266
2267 /// Remove the built instruction.
2268 void undo() override {
2269 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2270 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2271 IVal->eraseFromParent();
2272 }
2273 };
2274
2275 /// Build a sign extension instruction.
2276 class SExtBuilder : public TypePromotionAction {
2277 Value *Val;
2278
2279 public:
2280 /// Build a sign extension instruction of \p Opnd producing a \p Ty
2281 /// result.
2282 /// sext Opnd to Ty.
2283 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2284 : TypePromotionAction(InsertPt) {
2285 IRBuilder<> Builder(InsertPt);
2286 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2287 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2288 }
2289
2290 /// Get the built value.
2291 Value *getBuiltValue() { return Val; }
2292
2293 /// Remove the built instruction.
2294 void undo() override {
2295 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2296 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2297 IVal->eraseFromParent();
2298 }
2299 };
2300
2301 /// Build a zero extension instruction.
2302 class ZExtBuilder : public TypePromotionAction {
2303 Value *Val;
2304
2305 public:
2306 /// Build a zero extension instruction of \p Opnd producing a \p Ty
2307 /// result.
2308 /// zext Opnd to Ty.
2309 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2310 : TypePromotionAction(InsertPt) {
2311 IRBuilder<> Builder(InsertPt);
2312 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2313 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2314 }
2315
2316 /// Get the built value.
2317 Value *getBuiltValue() { return Val; }
2318
2319 /// Remove the built instruction.
2320 void undo() override {
2321 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2322 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2323 IVal->eraseFromParent();
2324 }
2325 };
2326
2327 /// Mutate an instruction to another type.
2328 class TypeMutator : public TypePromotionAction {
2329 /// Record the original type.
2330 Type *OrigTy;
2331
2332 public:
2333 /// Mutate the type of \p Inst into \p NewTy.
2334 TypeMutator(Instruction *Inst, Type *NewTy)
2335 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2336 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
2337 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
;
2338 Inst->mutateType(NewTy);
2339 }
2340
2341 /// Mutate the instruction back to its original type.
2342 void undo() override {
2343 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
2344 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
;
2345 Inst->mutateType(OrigTy);
2346 }
2347 };
2348
2349 /// Replace the uses of an instruction by another instruction.
2350 class UsesReplacer : public TypePromotionAction {
2351 /// Helper structure to keep track of the replaced uses.
2352 struct InstructionAndIdx {
2353 /// The instruction using the instruction.
2354 Instruction *Inst;
2355
2356 /// The index where this instruction is used for Inst.
2357 unsigned Idx;
2358
2359 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2360 : Inst(Inst), Idx(Idx) {}
2361 };
2362
2363 /// Keep track of the original uses (pair Instruction, Index).
2364 SmallVector<InstructionAndIdx, 4> OriginalUses;
2365 /// Keep track of the debug users.
2366 SmallVector<DbgValueInst *, 1> DbgValues;
2367
2368 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2369
2370 public:
2371 /// Replace all the use of \p Inst by \p New.
2372 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2373 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
2374 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
;
2375 // Record the original uses.
2376 for (Use &U : Inst->uses()) {
2377 Instruction *UserI = cast<Instruction>(U.getUser());
2378 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2379 }
2380 // Record the debug uses separately. They are not in the instruction's
2381 // use list, but they are replaced by RAUW.
2382 findDbgValues(DbgValues, Inst);
2383
2384 // Now, we can replace the uses.
2385 Inst->replaceAllUsesWith(New);
2386 }
2387
2388 /// Reassign the original uses of Inst to Inst.
2389 void undo() override {
2390 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: UsersReplacer: "
<< *Inst << "\n"; } } while (false)
;
2391 for (use_iterator UseIt = OriginalUses.begin(),
2392 EndIt = OriginalUses.end();
2393 UseIt != EndIt; ++UseIt) {
2394 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2395 }
2396 // RAUW has replaced all original uses with references to the new value,
2397 // including the debug uses. Since we are undoing the replacements,
2398 // the original debug uses must also be reinstated to maintain the
2399 // correctness and utility of debug value instructions.
2400 for (auto *DVI: DbgValues) {
2401 LLVMContext &Ctx = Inst->getType()->getContext();
2402 auto *MV = MetadataAsValue::get(Ctx, ValueAsMetadata::get(Inst));
2403 DVI->setOperand(0, MV);
2404 }
2405 }
2406 };
2407
2408 /// Remove an instruction from the IR.
2409 class InstructionRemover : public TypePromotionAction {
2410 /// Original position of the instruction.
2411 InsertionHandler Inserter;
2412
2413 /// Helper structure to hide all the link to the instruction. In other
2414 /// words, this helps to do as if the instruction was removed.
2415 OperandsHider Hider;
2416
2417 /// Keep track of the uses replaced, if any.
2418 UsesReplacer *Replacer = nullptr;
2419
2420 /// Keep track of instructions removed.
2421 SetOfInstrs &RemovedInsts;
2422
2423 public:
2424 /// Remove all reference of \p Inst and optionally replace all its
2425 /// uses with New.
2426 /// \p RemovedInsts Keep track of the instructions removed by this Action.
2427 /// \pre If !Inst->use_empty(), then New != nullptr
2428 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2429 Value *New = nullptr)
2430 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2431 RemovedInsts(RemovedInsts) {
2432 if (New)
2433 Replacer = new UsesReplacer(Inst, New);
2434 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2435 RemovedInsts.insert(Inst);
2436 /// The instructions removed here will be freed after completing
2437 /// optimizeBlock() for all blocks as we need to keep track of the
2438 /// removed instructions during promotion.
2439 Inst->removeFromParent();
2440 }
2441
2442 ~InstructionRemover() override { delete Replacer; }
2443
2444 /// Resurrect the instruction and reassign it to the proper uses if
2445 /// new value was provided when build this action.
2446 void undo() override {
2447 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2448 Inserter.insert(Inst);
2449 if (Replacer)
2450 Replacer->undo();
2451 Hider.undo();
2452 RemovedInsts.erase(Inst);
2453 }
2454 };
2455
2456public:
2457 /// Restoration point.
2458 /// The restoration point is a pointer to an action instead of an iterator
2459 /// because the iterator may be invalidated but not the pointer.
2460 using ConstRestorationPt = const TypePromotionAction *;
2461
2462 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2463 : RemovedInsts(RemovedInsts) {}
2464
2465 /// Advocate every changes made in that transaction.
2466 void commit();
2467
2468 /// Undo all the changes made after the given point.
2469 void rollback(ConstRestorationPt Point);
2470
2471 /// Get the current restoration point.
2472 ConstRestorationPt getRestorationPoint() const;
2473
2474 /// \name API for IR modification with state keeping to support rollback.
2475 /// @{
2476 /// Same as Instruction::setOperand.
2477 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2478
2479 /// Same as Instruction::eraseFromParent.
2480 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2481
2482 /// Same as Value::replaceAllUsesWith.
2483 void replaceAllUsesWith(Instruction *Inst, Value *New);
2484
2485 /// Same as Value::mutateType.
2486 void mutateType(Instruction *Inst, Type *NewTy);
2487
2488 /// Same as IRBuilder::createTrunc.
2489 Value *createTrunc(Instruction *Opnd, Type *Ty);
2490
2491 /// Same as IRBuilder::createSExt.
2492 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2493
2494 /// Same as IRBuilder::createZExt.
2495 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2496
2497 /// Same as Instruction::moveBefore.
2498 void moveBefore(Instruction *Inst, Instruction *Before);
2499 /// @}
2500
2501private:
2502 /// The ordered list of actions made so far.
2503 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2504
2505 using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2506
2507 SetOfInstrs &RemovedInsts;
2508};
2509
2510} // end anonymous namespace
2511
2512void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2513 Value *NewVal) {
2514 Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2515 Inst, Idx, NewVal));
2516}
2517
2518void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2519 Value *NewVal) {
2520 Actions.push_back(
2521 llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2522 Inst, RemovedInsts, NewVal));
2523}
2524
2525void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2526 Value *New) {
2527 Actions.push_back(
2528 llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2529}
2530
2531void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2532 Actions.push_back(
2533 llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2534}
2535
2536Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2537 Type *Ty) {
2538 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2539 Value *Val = Ptr->getBuiltValue();
2540 Actions.push_back(std::move(Ptr));
2541 return Val;
2542}
2543
2544Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2545 Value *Opnd, Type *Ty) {
2546 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2547 Value *Val = Ptr->getBuiltValue();
2548 Actions.push_back(std::move(Ptr));
2549 return Val;
2550}
2551
2552Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2553 Value *Opnd, Type *Ty) {
2554 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2555 Value *Val = Ptr->getBuiltValue();
2556 Actions.push_back(std::move(Ptr));
2557 return Val;
2558}
2559
2560void TypePromotionTransaction::moveBefore(Instruction *Inst,
2561 Instruction *Before) {
2562 Actions.push_back(
2563 llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2564 Inst, Before));
2565}
2566
2567TypePromotionTransaction::ConstRestorationPt
2568TypePromotionTransaction::getRestorationPoint() const {
2569 return !Actions.empty() ? Actions.back().get() : nullptr;
2570}
2571
2572void TypePromotionTransaction::commit() {
2573 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2574 ++It)
2575 (*It)->commit();
2576 Actions.clear();
2577}
2578
2579void TypePromotionTransaction::rollback(
2580 TypePromotionTransaction::ConstRestorationPt Point) {
2581 while (!Actions.empty() && Point != Actions.back().get()) {
2582 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2583 Curr->undo();
2584 }
2585}
2586
2587namespace {
2588
2589/// A helper class for matching addressing modes.
2590///
2591/// This encapsulates the logic for matching the target-legal addressing modes.
2592class AddressingModeMatcher {
2593 SmallVectorImpl<Instruction*> &AddrModeInsts;
2594 const TargetLowering &TLI;
2595 const TargetRegisterInfo &TRI;
2596 const DataLayout &DL;
2597
2598 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2599 /// the memory instruction that we're computing this address for.
2600 Type *AccessTy;
2601 unsigned AddrSpace;
2602 Instruction *MemoryInst;
2603
2604 /// This is the addressing mode that we're building up. This is
2605 /// part of the return value of this addressing mode matching stuff.
2606 ExtAddrMode &AddrMode;
2607
2608 /// The instructions inserted by other CodeGenPrepare optimizations.
2609 const SetOfInstrs &InsertedInsts;
2610
2611 /// A map from the instructions to their type before promotion.
2612 InstrToOrigTy &PromotedInsts;
2613
2614 /// The ongoing transaction where every action should be registered.
2615 TypePromotionTransaction &TPT;
2616
2617 // A GEP which has too large offset to be folded into the addressing mode.
2618 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
2619
2620 /// This is set to true when we should not do profitability checks.
2621 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2622 bool IgnoreProfitability;
2623
2624 AddressingModeMatcher(
2625 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
2626 const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
2627 ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
2628 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
2629 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
2630 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2631 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2632 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2633 PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
2634 IgnoreProfitability = false;
2635 }
2636
2637public:
2638 /// Find the maximal addressing mode that a load/store of V can fold,
2639 /// give an access type of AccessTy. This returns a list of involved
2640 /// instructions in AddrModeInsts.
2641 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2642 /// optimizations.
2643 /// \p PromotedInsts maps the instructions to their type before promotion.
2644 /// \p The ongoing transaction where every action should be registered.
2645 static ExtAddrMode
2646 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
2647 SmallVectorImpl<Instruction *> &AddrModeInsts,
2648 const TargetLowering &TLI, const TargetRegisterInfo &TRI,
2649 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
2650 TypePromotionTransaction &TPT,
2651 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
2652 ExtAddrMode Result;
2653
2654 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
2655 MemoryInst, Result, InsertedInsts,
2656 PromotedInsts, TPT, LargeOffsetGEP)
2657 .matchAddr(V, 0);
2658 (void)Success; assert(Success && "Couldn't select *anything*?")((Success && "Couldn't select *anything*?") ? static_cast
<void> (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2658, __PRETTY_FUNCTION__))
;
2659 return Result;
2660 }
2661
2662private:
2663 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2664 bool matchAddr(Value *Addr, unsigned Depth);
2665 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
2666 bool *MovedAway = nullptr);
2667 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2668 ExtAddrMode &AMBefore,
2669 ExtAddrMode &AMAfter);
2670 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2671 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2672 Value *PromotedOperand) const;
2673};
2674
2675class PhiNodeSet;
2676
2677/// An iterator for PhiNodeSet.
2678class PhiNodeSetIterator {
2679 PhiNodeSet * const Set;
2680 size_t CurrentIndex = 0;
2681
2682public:
2683 /// The constructor. Start should point to either a valid element, or be equal
2684 /// to the size of the underlying SmallVector of the PhiNodeSet.
2685 PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start);
2686 PHINode * operator*() const;
2687 PhiNodeSetIterator& operator++();
2688 bool operator==(const PhiNodeSetIterator &RHS) const;
2689 bool operator!=(const PhiNodeSetIterator &RHS) const;
2690};
2691
2692/// Keeps a set of PHINodes.
2693///
2694/// This is a minimal set implementation for a specific use case:
2695/// It is very fast when there are very few elements, but also provides good
2696/// performance when there are many. It is similar to SmallPtrSet, but also
2697/// provides iteration by insertion order, which is deterministic and stable
2698/// across runs. It is also similar to SmallSetVector, but provides removing
2699/// elements in O(1) time. This is achieved by not actually removing the element
2700/// from the underlying vector, so comes at the cost of using more memory, but
2701/// that is fine, since PhiNodeSets are used as short lived objects.
2702class PhiNodeSet {
2703 friend class PhiNodeSetIterator;
2704
2705 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
2706 using iterator = PhiNodeSetIterator;
2707
2708 /// Keeps the elements in the order of their insertion in the underlying
2709 /// vector. To achieve constant time removal, it never deletes any element.
2710 SmallVector<PHINode *, 32> NodeList;
2711
2712 /// Keeps the elements in the underlying set implementation. This (and not the
2713 /// NodeList defined above) is the source of truth on whether an element
2714 /// is actually in the collection.
2715 MapType NodeMap;
2716
2717 /// Points to the first valid (not deleted) element when the set is not empty
2718 /// and the value is not zero. Equals to the size of the underlying vector
2719 /// when the set is empty. When the value is 0, as in the beginning, the
2720 /// first element may or may not be valid.
2721 size_t FirstValidElement = 0;
2722
2723public:
2724 /// Inserts a new element to the collection.
2725 /// \returns true if the element is actually added, i.e. was not in the
2726 /// collection before the operation.
2727 bool insert(PHINode *Ptr) {
2728 if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
2729 NodeList.push_back(Ptr);
2730 return true;
2731 }
2732 return false;
2733 }
2734
2735 /// Removes the element from the collection.
2736 /// \returns whether the element is actually removed, i.e. was in the
2737 /// collection before the operation.
2738 bool erase(PHINode *Ptr) {
2739 auto it = NodeMap.find(Ptr);
2740 if (it != NodeMap.end()) {
2741 NodeMap.erase(Ptr);
2742 SkipRemovedElements(FirstValidElement);
2743 return true;
2744 }
2745 return false;
2746 }
2747
2748 /// Removes all elements and clears the collection.
2749 void clear() {
2750 NodeMap.clear();
2751 NodeList.clear();
2752 FirstValidElement = 0;
2753 }
2754
2755 /// \returns an iterator that will iterate the elements in the order of
2756 /// insertion.
2757 iterator begin() {
2758 if (FirstValidElement == 0)
2759 SkipRemovedElements(FirstValidElement);
2760 return PhiNodeSetIterator(this, FirstValidElement);
2761 }
2762
2763 /// \returns an iterator that points to the end of the collection.
2764 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
2765
2766 /// Returns the number of elements in the collection.
2767 size_t size() const {
2768 return NodeMap.size();
2769 }
2770
2771 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
2772 size_t count(PHINode *Ptr) const {
2773 return NodeMap.count(Ptr);
2774 }
2775
2776private:
2777 /// Updates the CurrentIndex so that it will point to a valid element.
2778 ///
2779 /// If the element of NodeList at CurrentIndex is valid, it does not
2780 /// change it. If there are no more valid elements, it updates CurrentIndex
2781 /// to point to the end of the NodeList.
2782 void SkipRemovedElements(size_t &CurrentIndex) {
2783 while (CurrentIndex < NodeList.size()) {
2784 auto it = NodeMap.find(NodeList[CurrentIndex]);
2785 // If the element has been deleted and added again later, NodeMap will
2786 // point to a different index, so CurrentIndex will still be invalid.
2787 if (it != NodeMap.end() && it->second == CurrentIndex)
2788 break;
2789 ++CurrentIndex;
2790 }
2791 }
2792};
2793
2794PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
2795 : Set(Set), CurrentIndex(Start) {}
2796
2797PHINode * PhiNodeSetIterator::operator*() const {
2798 assert(CurrentIndex < Set->NodeList.size() &&((CurrentIndex < Set->NodeList.size() && "PhiNodeSet access out of range"
) ? static_cast<void> (0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2799, __PRETTY_FUNCTION__))
2799 "PhiNodeSet access out of range")((CurrentIndex < Set->NodeList.size() && "PhiNodeSet access out of range"
) ? static_cast<void> (0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2799, __PRETTY_FUNCTION__))
;
2800 return Set->NodeList[CurrentIndex];
2801}
2802
2803PhiNodeSetIterator& PhiNodeSetIterator::operator++() {
2804 assert(CurrentIndex < Set->NodeList.size() &&((CurrentIndex < Set->NodeList.size() && "PhiNodeSet access out of range"
) ? static_cast<void> (0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2805, __PRETTY_FUNCTION__))
2805 "PhiNodeSet access out of range")((CurrentIndex < Set->NodeList.size() && "PhiNodeSet access out of range"
) ? static_cast<void> (0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2805, __PRETTY_FUNCTION__))
;
2806 ++CurrentIndex;
2807 Set->SkipRemovedElements(CurrentIndex);
2808 return *this;
2809}
2810
2811bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
2812 return CurrentIndex == RHS.CurrentIndex;
2813}
2814
2815bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
2816 return !((*this) == RHS);
2817}
2818
2819/// Keep track of simplification of Phi nodes.
2820/// Accept the set of all phi nodes and erase phi node from this set
2821/// if it is simplified.
2822class SimplificationTracker {
2823 DenseMap<Value *, Value *> Storage;
2824 const SimplifyQuery &SQ;
2825 // Tracks newly created Phi nodes. The elements are iterated by insertion
2826 // order.
2827 PhiNodeSet AllPhiNodes;
2828 // Tracks newly created Select nodes.
2829 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
2830
2831public:
2832 SimplificationTracker(const SimplifyQuery &sq)
2833 : SQ(sq) {}
2834
2835 Value *Get(Value *V) {
2836 do {
2837 auto SV = Storage.find(V);
2838 if (SV == Storage.end())
2839 return V;
2840 V = SV->second;
2841 } while (true);
2842 }
2843
2844 Value *Simplify(Value *Val) {
2845 SmallVector<Value *, 32> WorkList;
2846 SmallPtrSet<Value *, 32> Visited;
2847 WorkList.push_back(Val);
2848 while (!WorkList.empty()) {
2849 auto P = WorkList.pop_back_val();
2850 if (!Visited.insert(P).second)
2851 continue;
2852 if (auto *PI = dyn_cast<Instruction>(P))
2853 if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2854 for (auto *U : PI->users())
2855 WorkList.push_back(cast<Value>(U));
2856 Put(PI, V);
2857 PI->replaceAllUsesWith(V);
2858 if (auto *PHI = dyn_cast<PHINode>(PI))
2859 AllPhiNodes.erase(PHI);
2860 if (auto *Select = dyn_cast<SelectInst>(PI))
2861 AllSelectNodes.erase(Select);
2862 PI->eraseFromParent();
2863 }
2864 }
2865 return Get(Val);
2866 }
2867
2868 void Put(Value *From, Value *To) {
2869 Storage.insert({ From, To });
2870 }
2871
2872 void ReplacePhi(PHINode *From, PHINode *To) {
2873 Value* OldReplacement = Get(From);
2874 while (OldReplacement != From) {
2875 From = To;
2876 To = dyn_cast<PHINode>(OldReplacement);
2877 OldReplacement = Get(From);
2878 }
2879 assert(Get(To) == To && "Replacement PHI node is already replaced.")((Get(To) == To && "Replacement PHI node is already replaced."
) ? static_cast<void> (0) : __assert_fail ("Get(To) == To && \"Replacement PHI node is already replaced.\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 2879, __PRETTY_FUNCTION__))
;
2880 Put(From, To);
2881 From->replaceAllUsesWith(To);
2882 AllPhiNodes.erase(From);
2883 From->eraseFromParent();
2884 }
2885
2886 PhiNodeSet& newPhiNodes() { return AllPhiNodes; }
2887
2888 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
2889
2890 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
2891
2892 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
2893
2894 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
2895
2896 void destroyNewNodes(Type *CommonType) {
2897 // For safe erasing, replace the uses with dummy value first.
2898 auto Dummy = UndefValue::get(CommonType);
2899 for (auto I : AllPhiNodes) {
2900 I->replaceAllUsesWith(Dummy);
2901 I->eraseFromParent();
2902 }
2903 AllPhiNodes.clear();
2904 for (auto I : AllSelectNodes) {
2905 I->replaceAllUsesWith(Dummy);
2906 I->eraseFromParent();
2907 }
2908 AllSelectNodes.clear();
2909 }
2910};
2911
2912/// A helper class for combining addressing modes.
2913class AddressingModeCombiner {
2914 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
2915 typedef std::pair<PHINode *, PHINode *> PHIPair;
2916
2917private:
2918 /// The addressing modes we've collected.
2919 SmallVector<ExtAddrMode, 16> AddrModes;
2920
2921 /// The field in which the AddrModes differ, when we have more than one.
2922 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2923
2924 /// Are the AddrModes that we have all just equal to their original values?
2925 bool AllAddrModesTrivial = true;
2926
2927 /// Common Type for all different fields in addressing modes.
2928 Type *CommonType;
2929
2930 /// SimplifyQuery for simplifyInstruction utility.
2931 const SimplifyQuery &SQ;
2932
2933 /// Original Address.
2934 Value *Original;
2935
2936public:
2937 AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
2938 : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2939
2940 /// Get the combined AddrMode
2941 const ExtAddrMode &getAddrMode() const {
2942 return AddrModes[0];
2943 }
2944
2945 /// Add a new AddrMode if it's compatible with the AddrModes we already
2946 /// have.
2947 /// \return True iff we succeeded in doing so.
2948 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2949 // Take note of if we have any non-trivial AddrModes, as we need to detect
2950 // when all AddrModes are trivial as then we would introduce a phi or select
2951 // which just duplicates what's already there.
2952 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2953
2954 // If this is the first addrmode then everything is fine.
2955 if (AddrModes.empty()) {
2956 AddrModes.emplace_back(NewAddrMode);
2957 return true;
2958 }
2959
2960 // Figure out how different this is from the other address modes, which we
2961 // can do just by comparing against the first one given that we only care
2962 // about the cumulative difference.
2963 ExtAddrMode::FieldName ThisDifferentField =
2964 AddrModes[0].compare(NewAddrMode);
2965 if (DifferentField == ExtAddrMode::NoField)
2966 DifferentField = ThisDifferentField;
2967 else if (DifferentField != ThisDifferentField)
2968 DifferentField = ExtAddrMode::MultipleFields;
2969
2970 // If NewAddrMode differs in more than one dimension we cannot handle it.
2971 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
2972
2973 // If Scale Field is different then we reject.
2974 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
2975
2976 // We also must reject the case when base offset is different and
2977 // scale reg is not null, we cannot handle this case due to merge of
2978 // different offsets will be used as ScaleReg.
2979 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
2980 !NewAddrMode.ScaledReg);
2981
2982 // We also must reject the case when GV is different and BaseReg installed
2983 // due to we want to use base reg as a merge of GV values.
2984 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
2985 !NewAddrMode.HasBaseReg);
2986
2987 // Even if NewAddMode is the same we still need to collect it due to
2988 // original value is different. And later we will need all original values
2989 // as anchors during finding the common Phi node.
2990 if (CanHandle)
2991 AddrModes.emplace_back(NewAddrMode);
2992 else
2993 AddrModes.clear();
2994
2995 return CanHandle;
2996 }
2997
2998 /// Combine the addressing modes we've collected into a single
2999 /// addressing mode.
3000 /// \return True iff we successfully combined them or we only had one so
3001 /// didn't need to combine them anyway.
3002 bool combineAddrModes() {
3003 // If we have no AddrModes then they can't be combined.
3004 if (AddrModes.size() == 0)
3005 return false;
3006
3007 // A single AddrMode can trivially be combined.
3008 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3009 return true;
3010
3011 // If the AddrModes we collected are all just equal to the value they are
3012 // derived from then combining them wouldn't do anything useful.
3013 if (AllAddrModesTrivial)
3014 return false;
3015
3016 if (!addrModeCombiningAllowed())
3017 return false;
3018
3019 // Build a map between <original value, basic block where we saw it> to
3020 // value of base register.
3021 // Bail out if there is no common type.
3022 FoldAddrToValueMapping Map;
3023 if (!initializeMap(Map))
3024 return false;
3025
3026 Value *CommonValue = findCommon(Map);
3027 if (CommonValue)
3028 AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3029 return CommonValue != nullptr;
3030 }
3031
3032private:
3033 /// Initialize Map with anchor values. For address seen
3034 /// we set the value of different field saw in this address.
3035 /// At the same time we find a common type for different field we will
3036 /// use to create new Phi/Select nodes. Keep it in CommonType field.
3037 /// Return false if there is no common type found.
3038 bool initializeMap(FoldAddrToValueMapping &Map) {
3039 // Keep track of keys where the value is null. We will need to replace it
3040 // with constant null when we know the common type.
3041 SmallVector<Value *, 2> NullValue;
3042 Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3043 for (auto &AM : AddrModes) {
3044 Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3045 if (DV) {
3046 auto *Type = DV->getType();
3047 if (CommonType && CommonType != Type)
3048 return false;
3049 CommonType = Type;
3050 Map[AM.OriginalValue] = DV;
3051 } else {
3052 NullValue.push_back(AM.OriginalValue);
3053 }
3054 }
3055 assert(CommonType && "At least one non-null value must be!")((CommonType && "At least one non-null value must be!"
) ? static_cast<void> (0) : __assert_fail ("CommonType && \"At least one non-null value must be!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3055, __PRETTY_FUNCTION__))
;
3056 for (auto *V : NullValue)
3057 Map[V] = Constant::getNullValue(CommonType);
3058 return true;
3059 }
3060
3061 /// We have mapping between value A and other value B where B was a field in
3062 /// addressing mode represented by A. Also we have an original value C
3063 /// representing an address we start with. Traversing from C through phi and
3064 /// selects we ended up with A's in a map. This utility function tries to find
3065 /// a value V which is a field in addressing mode C and traversing through phi
3066 /// nodes and selects we will end up in corresponded values B in a map.
3067 /// The utility will create a new Phi/Selects if needed.
3068 // The simple example looks as follows:
3069 // BB1:
3070 // p1 = b1 + 40
3071 // br cond BB2, BB3
3072 // BB2:
3073 // p2 = b2 + 40
3074 // br BB3
3075 // BB3:
3076 // p = phi [p1, BB1], [p2, BB2]
3077 // v = load p
3078 // Map is
3079 // p1 -> b1
3080 // p2 -> b2
3081 // Request is
3082 // p -> ?
3083 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
3084 Value *findCommon(FoldAddrToValueMapping &Map) {
3085 // Tracks the simplification of newly created phi nodes. The reason we use
3086 // this mapping is because we will add new created Phi nodes in AddrToBase.
3087 // Simplification of Phi nodes is recursive, so some Phi node may
3088 // be simplified after we added it to AddrToBase. In reality this
3089 // simplification is possible only if original phi/selects were not
3090 // simplified yet.
3091 // Using this mapping we can find the current value in AddrToBase.
3092 SimplificationTracker ST(SQ);
3093
3094 // First step, DFS to create PHI nodes for all intermediate blocks.
3095 // Also fill traverse order for the second step.
3096 SmallVector<Value *, 32> TraverseOrder;
3097 InsertPlaceholders(Map, TraverseOrder, ST);
3098
3099 // Second Step, fill new nodes by merged values and simplify if possible.
3100 FillPlaceholders(Map, TraverseOrder, ST);
3101
3102 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3103 ST.destroyNewNodes(CommonType);
3104 return nullptr;
3105 }
3106
3107 // Now we'd like to match New Phi nodes to existed ones.
3108 unsigned PhiNotMatchedCount = 0;
3109 if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3110 ST.destroyNewNodes(CommonType);
3111 return nullptr;
3112 }
3113
3114 auto *Result = ST.Get(Map.find(Original)->second);
3115 if (Result) {
3116 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3117 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3118 }
3119 return Result;
3120 }
3121
3122 /// Try to match PHI node to Candidate.
3123 /// Matcher tracks the matched Phi nodes.
3124 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3125 SmallSetVector<PHIPair, 8> &Matcher,
3126 PhiNodeSet &PhiNodesToMatch) {
3127 SmallVector<PHIPair, 8> WorkList;
3128 Matcher.insert({ PHI, Candidate });
3129 WorkList.push_back({ PHI, Candidate });
3130 SmallSet<PHIPair, 8> Visited;
3131 while (!WorkList.empty()) {
3132 auto Item = WorkList.pop_back_val();
3133 if (!Visited.insert(Item).second)
3134 continue;
3135 // We iterate over all incoming values to Phi to compare them.
3136 // If values are different and both of them Phi and the first one is a
3137 // Phi we added (subject to match) and both of them is in the same basic
3138 // block then we can match our pair if values match. So we state that
3139 // these values match and add it to work list to verify that.
3140 for (auto B : Item.first->blocks()) {
3141 Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3142 Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3143 if (FirstValue == SecondValue)
3144 continue;
3145
3146 PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3147 PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3148
3149 // One of them is not Phi or
3150 // The first one is not Phi node from the set we'd like to match or
3151 // Phi nodes from different basic blocks then
3152 // we will not be able to match.
3153 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3154 FirstPhi->getParent() != SecondPhi->getParent())
3155 return false;
3156
3157 // If we already matched them then continue.
3158 if (Matcher.count({ FirstPhi, SecondPhi }))
3159 continue;
3160 // So the values are different and does not match. So we need them to
3161 // match.
3162 Matcher.insert({ FirstPhi, SecondPhi });
3163 // But me must check it.
3164 WorkList.push_back({ FirstPhi, SecondPhi });
3165 }
3166 }
3167 return true;
3168 }
3169
3170 /// For the given set of PHI nodes (in the SimplificationTracker) try
3171 /// to find their equivalents.
3172 /// Returns false if this matching fails and creation of new Phi is disabled.
3173 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3174 unsigned &PhiNotMatchedCount) {
3175 // Matched and PhiNodesToMatch iterate their elements in a deterministic
3176 // order, so the replacements (ReplacePhi) are also done in a deterministic
3177 // order.
3178 SmallSetVector<PHIPair, 8> Matched;
3179 SmallPtrSet<PHINode *, 8> WillNotMatch;
3180 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3181 while (PhiNodesToMatch.size()) {
3182 PHINode *PHI = *PhiNodesToMatch.begin();
3183
3184 // Add us, if no Phi nodes in the basic block we do not match.
3185 WillNotMatch.clear();
3186 WillNotMatch.insert(PHI);
3187
3188 // Traverse all Phis until we found equivalent or fail to do that.
3189 bool IsMatched = false;
3190 for (auto &P : PHI->getParent()->phis()) {
3191 if (&P == PHI)
3192 continue;
3193 if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3194 break;
3195 // If it does not match, collect all Phi nodes from matcher.
3196 // if we end up with no match, them all these Phi nodes will not match
3197 // later.
3198 for (auto M : Matched)
3199 WillNotMatch.insert(M.first);
3200 Matched.clear();
3201 }
3202 if (IsMatched) {
3203 // Replace all matched values and erase them.
3204 for (auto MV : Matched)
3205 ST.ReplacePhi(MV.first, MV.second);
3206 Matched.clear();
3207 continue;
3208 }
3209 // If we are not allowed to create new nodes then bail out.
3210 if (!AllowNewPhiNodes)
3211 return false;
3212 // Just remove all seen values in matcher. They will not match anything.
3213 PhiNotMatchedCount += WillNotMatch.size();
3214 for (auto *P : WillNotMatch)
3215 PhiNodesToMatch.erase(P);
3216 }
3217 return true;
3218 }
3219 /// Fill the placeholders with values from predecessors and simplify them.
3220 void FillPlaceholders(FoldAddrToValueMapping &Map,
3221 SmallVectorImpl<Value *> &TraverseOrder,
3222 SimplificationTracker &ST) {
3223 while (!TraverseOrder.empty()) {
3224 Value *Current = TraverseOrder.pop_back_val();
3225 assert(Map.find(Current) != Map.end() && "No node to fill!!!")((Map.find(Current) != Map.end() && "No node to fill!!!"
) ? static_cast<void> (0) : __assert_fail ("Map.find(Current) != Map.end() && \"No node to fill!!!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3225, __PRETTY_FUNCTION__))
;
3226 Value *V = Map[Current];
3227
3228 if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3229 // CurrentValue also must be Select.
3230 auto *CurrentSelect = cast<SelectInst>(Current);
3231 auto *TrueValue = CurrentSelect->getTrueValue();
3232 assert(Map.find(TrueValue) != Map.end() && "No True Value!")((Map.find(TrueValue) != Map.end() && "No True Value!"
) ? static_cast<void> (0) : __assert_fail ("Map.find(TrueValue) != Map.end() && \"No True Value!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3232, __PRETTY_FUNCTION__))
;
3233 Select->setTrueValue(ST.Get(Map[TrueValue]));
3234 auto *FalseValue = CurrentSelect->getFalseValue();
3235 assert(Map.find(FalseValue) != Map.end() && "No False Value!")((Map.find(FalseValue) != Map.end() && "No False Value!"
) ? static_cast<void> (0) : __assert_fail ("Map.find(FalseValue) != Map.end() && \"No False Value!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3235, __PRETTY_FUNCTION__))
;
3236 Select->setFalseValue(ST.Get(Map[FalseValue]));
3237 } else {
3238 // Must be a Phi node then.
3239 PHINode *PHI = cast<PHINode>(V);
3240 auto *CurrentPhi = dyn_cast<PHINode>(Current);
3241 // Fill the Phi node with values from predecessors.
3242 for (auto B : predecessors(PHI->getParent())) {
3243 Value *PV = CurrentPhi->getIncomingValueForBlock(B);
3244 assert(Map.find(PV) != Map.end() && "No predecessor Value!")((Map.find(PV) != Map.end() && "No predecessor Value!"
) ? static_cast<void> (0) : __assert_fail ("Map.find(PV) != Map.end() && \"No predecessor Value!\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3244, __PRETTY_FUNCTION__))
;
3245 PHI->addIncoming(ST.Get(Map[PV]), B);
3246 }
3247 }
3248 Map[Current] = ST.Simplify(V);
3249 }
3250 }
3251
3252 /// Starting from original value recursively iterates over def-use chain up to
3253 /// known ending values represented in a map. For each traversed phi/select
3254 /// inserts a placeholder Phi or Select.
3255 /// Reports all new created Phi/Select nodes by adding them to set.
3256 /// Also reports and order in what values have been traversed.
3257 void InsertPlaceholders(FoldAddrToValueMapping &Map,
3258 SmallVectorImpl<Value *> &TraverseOrder,
3259 SimplificationTracker &ST) {
3260 SmallVector<Value *, 32> Worklist;
3261 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&(((isa<PHINode>(Original) || isa<SelectInst>(Original
)) && "Address must be a Phi or Select node") ? static_cast
<void> (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3262, __PRETTY_FUNCTION__))
3262 "Address must be a Phi or Select node")(((isa<PHINode>(Original) || isa<SelectInst>(Original
)) && "Address must be a Phi or Select node") ? static_cast
<void> (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3262, __PRETTY_FUNCTION__))
;
3263 auto *Dummy = UndefValue::get(CommonType);
3264 Worklist.push_back(Original);
3265 while (!Worklist.empty()) {
3266 Value *Current = Worklist.pop_back_val();
3267 // if it is already visited or it is an ending value then skip it.
3268 if (Map.find(Current) != Map.end())
3269 continue;
3270 TraverseOrder.push_back(Current);
3271
3272 // CurrentValue must be a Phi node or select. All others must be covered
3273 // by anchors.
3274 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3275 // Is it OK to get metadata from OrigSelect?!
3276 // Create a Select placeholder with dummy value.
3277 SelectInst *Select = SelectInst::Create(
3278 CurrentSelect->getCondition(), Dummy, Dummy,
3279 CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3280 Map[Current] = Select;
3281 ST.insertNewSelect(Select);
3282 // We are interested in True and False values.
3283 Worklist.push_back(CurrentSelect->getTrueValue());
3284 Worklist.push_back(CurrentSelect->getFalseValue());
3285 } else {
3286 // It must be a Phi node then.
3287 PHINode *CurrentPhi = cast<PHINode>(Current);
3288 unsigned PredCount = CurrentPhi->getNumIncomingValues();
3289 PHINode *PHI =
3290 PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3291 Map[Current] = PHI;
3292 ST.insertNewPhi(PHI);
3293 for (Value *P : CurrentPhi->incoming_values())
3294 Worklist.push_back(P);
3295 }
3296 }
3297 }
3298
3299 bool addrModeCombiningAllowed() {
3300 if (DisableComplexAddrModes)
3301 return false;
3302 switch (DifferentField) {
3303 default:
3304 return false;
3305 case ExtAddrMode::BaseRegField:
3306 return AddrSinkCombineBaseReg;
3307 case ExtAddrMode::BaseGVField:
3308 return AddrSinkCombineBaseGV;
3309 case ExtAddrMode::BaseOffsField:
3310 return AddrSinkCombineBaseOffs;
3311 case ExtAddrMode::ScaledRegField:
3312 return AddrSinkCombineScaledReg;
3313 }
3314 }
3315};
3316} // end anonymous namespace
3317
3318/// Try adding ScaleReg*Scale to the current addressing mode.
3319/// Return true and update AddrMode if this addr mode is legal for the target,
3320/// false if not.
3321bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3322 unsigned Depth) {
3323 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3324 // mode. Just process that directly.
3325 if (Scale == 1)
3326 return matchAddr(ScaleReg, Depth);
3327
3328 // If the scale is 0, it takes nothing to add this.
3329 if (Scale == 0)
3330 return true;
3331
3332 // If we already have a scale of this value, we can add to it, otherwise, we
3333 // need an available scale field.
3334 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3335 return false;
3336
3337 ExtAddrMode TestAddrMode = AddrMode;
3338
3339 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3340 // [A+B + A*7] -> [B+A*8].
3341 TestAddrMode.Scale += Scale;
3342 TestAddrMode.ScaledReg = ScaleReg;
3343
3344 // If the new address isn't legal, bail out.
3345 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3346 return false;
3347
3348 // It was legal, so commit it.
3349 AddrMode = TestAddrMode;
3350
3351 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3352 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3353 // X*Scale + C*Scale to addr mode.
3354 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3355 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3356 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3357 TestAddrMode.ScaledReg = AddLHS;
3358 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3359
3360 // If this addressing mode is legal, commit it and remember that we folded
3361 // this instruction.
3362 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3363 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3364 AddrMode = TestAddrMode;
3365 return true;
3366 }
3367 }
3368
3369 // Otherwise, not (x+c)*scale, just return what we have.
3370 return true;
3371}
3372
3373/// This is a little filter, which returns true if an addressing computation
3374/// involving I might be folded into a load/store accessing it.
3375/// This doesn't need to be perfect, but needs to accept at least
3376/// the set of instructions that MatchOperationAddr can.
3377static bool MightBeFoldableInst(Instruction *I) {
3378 switch (I->getOpcode()) {
3379 case Instruction::BitCast:
3380 case Instruction::AddrSpaceCast:
3381 // Don't touch identity bitcasts.
3382 if (I->getType() == I->getOperand(0)->getType())
3383 return false;
3384 return I->getType()->isIntOrPtrTy();
3385 case Instruction::PtrToInt:
3386 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3387 return true;
3388 case Instruction::IntToPtr:
3389 // We know the input is intptr_t, so this is foldable.
3390 return true;
3391 case Instruction::Add:
3392 return true;
3393 case Instruction::Mul:
3394 case Instruction::Shl:
3395 // Can only handle X*C and X << C.
3396 return isa<ConstantInt>(I->getOperand(1));
3397 case Instruction::GetElementPtr:
3398 return true;
3399 default:
3400 return false;
3401 }
3402}
3403
3404/// Check whether or not \p Val is a legal instruction for \p TLI.
3405/// \note \p Val is assumed to be the product of some type promotion.
3406/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3407/// to be legal, as the non-promoted value would have had the same state.
3408static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3409 const DataLayout &DL, Value *Val) {
3410 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3411 if (!PromotedInst)
3412 return false;
3413 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3414 // If the ISDOpcode is undefined, it was undefined before the promotion.
3415 if (!ISDOpcode)
3416 return true;
3417 // Otherwise, check if the promoted instruction is legal or not.
3418 return TLI.isOperationLegalOrCustom(
3419 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3420}
3421
3422namespace {
3423
3424/// Hepler class to perform type promotion.
3425class TypePromotionHelper {
3426 /// Utility function to add a promoted instruction \p ExtOpnd to
3427 /// \p PromotedInsts and record the type of extension we have seen.
3428 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
3429 Instruction *ExtOpnd,
3430 bool IsSExt) {
3431 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3432 InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
3433 if (It != PromotedInsts.end()) {
3434 // If the new extension is same as original, the information in
3435 // PromotedInsts[ExtOpnd] is still correct.
3436 if (It->second.getInt() == ExtTy)
3437 return;
3438
3439 // Now the new extension is different from old extension, we make
3440 // the type information invalid by setting extension type to
3441 // BothExtension.
3442 ExtTy = BothExtension;
3443 }
3444 PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
3445 }
3446
3447 /// Utility function to query the original type of instruction \p Opnd
3448 /// with a matched extension type. If the extension doesn't match, we
3449 /// cannot use the information we had on the original type.
3450 /// BothExtension doesn't match any extension type.
3451 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
3452 Instruction *Opnd,
3453 bool IsSExt) {
3454 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3455 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3456 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
3457 return It->second.getPointer();
3458 return nullptr;
3459 }
3460
3461 /// Utility function to check whether or not a sign or zero extension
3462 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3463 /// either using the operands of \p Inst or promoting \p Inst.
3464 /// The type of the extension is defined by \p IsSExt.
3465 /// In other words, check if:
3466 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3467 /// #1 Promotion applies:
3468 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3469 /// #2 Operand reuses:
3470 /// ext opnd1 to ConsideredExtType.
3471 /// \p PromotedInsts maps the instructions to their type before promotion.
3472 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3473 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3474
3475 /// Utility function to determine if \p OpIdx should be promoted when
3476 /// promoting \p Inst.
3477 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3478 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3479 }
3480
3481 /// Utility function to promote the operand of \p Ext when this
3482 /// operand is a promotable trunc or sext or zext.
3483 /// \p PromotedInsts maps the instructions to their type before promotion.
3484 /// \p CreatedInstsCost[out] contains the cost of all instructions
3485 /// created to promote the operand of Ext.
3486 /// Newly added extensions are inserted in \p Exts.
3487 /// Newly added truncates are inserted in \p Truncs.
3488 /// Should never be called directly.
3489 /// \return The promoted value which is used instead of Ext.
3490 static Value *promoteOperandForTruncAndAnyExt(
3491 Instruction *Ext, TypePromotionTransaction &TPT,
3492 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3493 SmallVectorImpl<Instruction *> *Exts,
3494 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3495
3496 /// Utility function to promote the operand of \p Ext when this
3497 /// operand is promotable and is not a supported trunc or sext.
3498 /// \p PromotedInsts maps the instructions to their type before promotion.
3499 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3500 /// created to promote the operand of Ext.
3501 /// Newly added extensions are inserted in \p Exts.
3502 /// Newly added truncates are inserted in \p Truncs.
3503 /// Should never be called directly.
3504 /// \return The promoted value which is used instead of Ext.
3505 static Value *promoteOperandForOther(Instruction *Ext,
3506 TypePromotionTransaction &TPT,
3507 InstrToOrigTy &PromotedInsts,
3508 unsigned &CreatedInstsCost,
3509 SmallVectorImpl<Instruction *> *Exts,
3510 SmallVectorImpl<Instruction *> *Truncs,
3511 const TargetLowering &TLI, bool IsSExt);
3512
3513 /// \see promoteOperandForOther.
3514 static Value *signExtendOperandForOther(
3515 Instruction *Ext, TypePromotionTransaction &TPT,
3516 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3517 SmallVectorImpl<Instruction *> *Exts,
3518 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3519 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3520 Exts, Truncs, TLI, true);
3521 }
3522
3523 /// \see promoteOperandForOther.
3524 static Value *zeroExtendOperandForOther(
3525 Instruction *Ext, TypePromotionTransaction &TPT,
3526 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3527 SmallVectorImpl<Instruction *> *Exts,
3528 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3529 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3530 Exts, Truncs, TLI, false);
3531 }
3532
3533public:
3534 /// Type for the utility function that promotes the operand of Ext.
3535 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3536 InstrToOrigTy &PromotedInsts,
3537 unsigned &CreatedInstsCost,
3538 SmallVectorImpl<Instruction *> *Exts,
3539 SmallVectorImpl<Instruction *> *Truncs,
3540 const TargetLowering &TLI);
3541
3542 /// Given a sign/zero extend instruction \p Ext, return the appropriate
3543 /// action to promote the operand of \p Ext instead of using Ext.
3544 /// \return NULL if no promotable action is possible with the current
3545 /// sign extension.
3546 /// \p InsertedInsts keeps track of all the instructions inserted by the
3547 /// other CodeGenPrepare optimizations. This information is important
3548 /// because we do not want to promote these instructions as CodeGenPrepare
3549 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3550 /// \p PromotedInsts maps the instructions to their type before promotion.
3551 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3552 const TargetLowering &TLI,
3553 const InstrToOrigTy &PromotedInsts);
3554};
3555
3556} // end anonymous namespace
3557
3558bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3559 Type *ConsideredExtType,
3560 const InstrToOrigTy &PromotedInsts,
3561 bool IsSExt) {
3562 // The promotion helper does not know how to deal with vector types yet.
3563 // To be able to fix that, we would need to fix the places where we
3564 // statically extend, e.g., constants and such.
3565 if (Inst->getType()->isVectorTy())
3566 return false;
3567
3568 // We can always get through zext.
3569 if (isa<ZExtInst>(Inst))
3570 return true;
3571
3572 // sext(sext) is ok too.
3573 if (IsSExt && isa<SExtInst>(Inst))
3574 return true;
3575
3576 // We can get through binary operator, if it is legal. In other words, the
3577 // binary operator must have a nuw or nsw flag.
3578 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3579 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3580 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3581 (IsSExt && BinOp->hasNoSignedWrap())))
3582 return true;
3583
3584 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
3585 if ((Inst->getOpcode() == Instruction::And ||
3586 Inst->getOpcode() == Instruction::Or))
3587 return true;
3588
3589 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
3590 if (Inst->getOpcode() == Instruction::Xor) {
3591 const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
3592 // Make sure it is not a NOT.
3593 if (Cst && !Cst->getValue().isAllOnesValue())
3594 return true;
3595 }
3596
3597 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
3598 // It may change a poisoned value into a regular value, like
3599 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
3600 // poisoned value regular value
3601 // It should be OK since undef covers valid value.
3602 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
3603 return true;
3604
3605 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
3606 // It may change a poisoned value into a regular value, like
3607 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
3608 // poisoned value regular value
3609 // It should be OK since undef covers valid value.
3610 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
3611 const Instruction *ExtInst =
3612 dyn_cast<const Instruction>(*Inst->user_begin());
3613 if (ExtInst->hasOneUse()) {
3614 const Instruction *AndInst =
3615 dyn_cast<const Instruction>(*ExtInst->user_begin());
3616 if (AndInst && AndInst->getOpcode() == Instruction::And) {
3617 const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
3618 if (Cst &&
3619 Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
3620 return true;
3621 }
3622 }
3623 }
3624
3625 // Check if we can do the following simplification.
3626 // ext(trunc(opnd)) --> ext(opnd)
3627 if (!isa<TruncInst>(Inst))
3628 return false;
3629
3630 Value *OpndVal = Inst->getOperand(0);
3631 // Check if we can use this operand in the extension.
3632 // If the type is larger than the result type of the extension, we cannot.
3633 if (!OpndVal->getType()->isIntegerTy() ||
3634 OpndVal->getType()->getIntegerBitWidth() >
3635 ConsideredExtType->getIntegerBitWidth())
3636 return false;
3637
3638 // If the operand of the truncate is not an instruction, we will not have
3639 // any information on the dropped bits.
3640 // (Actually we could for constant but it is not worth the extra logic).
3641 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3642 if (!Opnd)
3643 return false;
3644
3645 // Check if the source of the type is narrow enough.
3646 // I.e., check that trunc just drops extended bits of the same kind of
3647 // the extension.
3648 // #1 get the type of the operand and check the kind of the extended bits.
3649 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
3650 if (OpndType)
3651 ;
3652 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3653 OpndType = Opnd->getOperand(0)->getType();
3654 else
3655 return false;
3656
3657 // #2 check that the truncate just drops extended bits.
3658 return Inst->getType()->getIntegerBitWidth() >=
3659 OpndType->getIntegerBitWidth();
3660}
3661
3662TypePromotionHelper::Action TypePromotionHelper::getAction(
3663 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3664 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3665 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&(((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
"Unexpected instruction type") ? static_cast<void> (0)
: __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3666, __PRETTY_FUNCTION__))
3666 "Unexpected instruction type")(((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
"Unexpected instruction type") ? static_cast<void> (0)
: __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 3666, __PRETTY_FUNCTION__))
;
3667 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3668 Type *ExtTy = Ext->getType();
3669 bool IsSExt = isa<SExtInst>(Ext);
3670 // If the operand of the extension is not an instruction, we cannot
3671 // get through.
3672 // If it, check we can get through.
3673 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3674 return nullptr;
3675
3676 // Do not promote if the operand has been added by codegenprepare.
3677 // Otherwise, it means we are undoing an optimization that is likely to be
3678 // redone, thus causing potential infinite loop.
3679 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3680 return nullptr;
3681
3682 // SExt or Trunc instructions.
3683 // Return the related handler.
3684 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3685 isa<ZExtInst>(ExtOpnd))
3686 return promoteOperandForTruncAndAnyExt;
3687
3688 // Regular instruction.
3689 // Abort early if we will have to insert non-free instructions.
3690 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3691 return nullptr;
3692 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3693}
3694
3695Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3696 Instruction *SExt, TypePromotionTransaction &TPT,
3697 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3698 SmallVectorImpl<Instruction *> *Exts,
3699 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3700 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3701 // get through it and this method should not be called.
3702 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3703 Value *ExtVal = SExt;
3704 bool HasMergedNonFreeExt = false;
3705 if (isa<ZExtInst>(SExtOpnd)) {
3706 // Replace s|zext(zext(opnd))
3707 // => zext(opnd).
3708 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3709 Value *ZExt =
3710 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3711 TPT.replaceAllUsesWith(SExt, ZExt);
3712 TPT.eraseInstruction(SExt);
3713 ExtVal = ZExt;
3714 } else {
3715 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3716 // => z|sext(opnd).
3717 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3718 }
3719 CreatedInstsCost = 0;
3720
3721 // Remove dead code.
3722 if (SExtOpnd->use_empty())
3723 TPT.eraseInstruction(SExtOpnd);
3724
3725 // Check if the extension is still needed.
3726 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3727 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3728 if (ExtInst) {
3729 if (Exts)
3730 Exts->push_back(ExtInst);
3731 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3732 }
3733 return ExtVal;
3734 }
3735
3736 // At this point we have: ext ty opnd to ty.
3737 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3738 Value *NextVal = ExtInst->getOperand(0);
3739 TPT.eraseInstruction(ExtInst, NextVal);
3740 return NextVal;
3741}
3742
3743Value *TypePromotionHelper::promoteOperandForOther(
3744 Instruction *Ext, TypePromotionTransaction &TPT,
3745 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3746 SmallVectorImpl<Instruction *> *Exts,
3747 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3748 bool IsSExt) {
3749 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3750 // get through it and this method should not be called.
3751 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3752 CreatedInstsCost = 0;
3753 if (!ExtOpnd->hasOneUse()) {
3754 // ExtOpnd will be promoted.
3755 // All its uses, but Ext, will need to use a truncated value of the
3756 // promoted version.
3757 // Create the truncate now.
3758 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3759 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3760 // Insert it just after the definition.
3761 ITrunc->moveAfter(ExtOpnd);
3762 if (Truncs)
3763 Truncs->push_back(ITrunc);
3764 }
3765
3766 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3767 // Restore the operand of Ext (which has been replaced by the previous call
3768 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3769 TPT.setOperand(Ext, 0, ExtOpnd);
3770 }
3771
3772 // Get through the Instruction:
3773 // 1. Update its type.
3774 // 2. Replace the uses of Ext by Inst.
3775 // 3. Extend each operand that needs to be extended.
3776
3777 // Remember the original type of the instruction before promotion.
3778 // This is useful to know that the high bits are sign extended bits.
3779 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
3780 // Step #1.
3781 TPT.mutateType(ExtOpnd, Ext->getType());
3782 // Step #2.
3783 TPT.replaceAllUsesWith(Ext, ExtOpnd);
3784 // Step #3.
3785 Instruction *ExtForOpnd = Ext;
3786
3787 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Propagate Ext to operands\n"
; } } while (false)
;
3788 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3789 ++OpIdx) {
3790 LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Operand:\n" << *
(ExtOpnd->getOperand(OpIdx)) << '\n'; } } while (false
)
;
3791 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3792 !shouldExtOperand(ExtOpnd, OpIdx)) {
3793 LLVM_DEBUG(dbgs() << "No need to propagate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "No need to propagate\n"
; } } while (false)
;
3794 continue;
3795 }
3796 // Check if we can statically extend the operand.
3797 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3798 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3799 LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
3800 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3801 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3802 : Cst->getValue().zext(BitWidth);
3803 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3804 continue;
3805 }
3806 // UndefValue are typed, so we have to statically sign extend them.
3807 if (isa<UndefValue>(Opnd)) {
3808 LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
3809 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3810 continue;
3811 }
3812
3813 // Otherwise we have to explicitly sign extend the operand.
3814 // Check if Ext was reused to extend an operand.
3815 if (!ExtForOpnd) {
3816 // If yes, create a new one.
3817 LLVM_DEBUG(dbgs() << "More operands to ext\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "More operands to ext\n"
; } } while (false)
;
3818 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3819 : TPT.createZExt(Ext, Opnd, Ext->getType());
3820 if (!isa<Instruction>(ValForExtOpnd)) {
3821 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3822 continue;
3823 }
3824 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3825 }
3826 if (Exts)
3827 Exts->push_back(ExtForOpnd);
3828 TPT.setOperand(ExtForOpnd, 0, Opnd);
3829
3830 // Move the sign extension before the insertion point.
3831 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3832 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3833 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3834 // If more sext are required, new instructions will have to be created.
3835 ExtForOpnd = nullptr;
3836 }
3837 if (ExtForOpnd == Ext) {
3838 LLVM_DEBUG(dbgs() << "Extension is useless now\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Extension is useless now\n"
; } } while (false)
;
3839 TPT.eraseInstruction(Ext);
3840 }
3841 return ExtOpnd;
3842}
3843
3844/// Check whether or not promoting an instruction to a wider type is profitable.
3845/// \p NewCost gives the cost of extension instructions created by the
3846/// promotion.
3847/// \p OldCost gives the cost of extension instructions before the promotion
3848/// plus the number of instructions that have been
3849/// matched in the addressing mode the promotion.
3850/// \p PromotedOperand is the value that has been promoted.
3851/// \return True if the promotion is profitable, false otherwise.
3852bool AddressingModeMatcher::isPromotionProfitable(
3853 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3854 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "OldCost: " << OldCost
<< "\tNewCost: " << NewCost << '\n'; } } while
(false)
3855 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "OldCost: " << OldCost
<< "\tNewCost: " << NewCost << '\n'; } } while
(false)
;
3856 // The cost of the new extensions is greater than the cost of the
3857 // old extension plus what we folded.
3858 // This is not profitable.
3859 if (NewCost > OldCost)
3860 return false;
3861 if (NewCost < OldCost)
3862 return true;
3863 // The promotion is neutral but it may help folding the sign extension in
3864 // loads for instance.
3865 // Check that we did not create an illegal instruction.
3866 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3867}
3868
3869/// Given an instruction or constant expr, see if we can fold the operation
3870/// into the addressing mode. If so, update the addressing mode and return
3871/// true, otherwise return false without modifying AddrMode.
3872/// If \p MovedAway is not NULL, it contains the information of whether or
3873/// not AddrInst has to be folded into the addressing mode on success.
3874/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3875/// because it has been moved away.
3876/// Thus AddrInst must not be added in the matched instructions.
3877/// This state can happen when AddrInst is a sext, since it may be moved away.
3878/// Therefore, AddrInst may not be valid when MovedAway is true and it must
3879/// not be referenced anymore.
3880bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3881 unsigned Depth,
3882 bool *MovedAway) {
3883 // Avoid exponential behavior on extremely deep expression trees.
3884 if (Depth >= 5) return false;
3885
3886 // By default, all matched instructions stay in place.
3887 if (MovedAway)
3888 *MovedAway = false;
3889
3890 switch (Opcode) {
3891 case Instruction::PtrToInt:
3892 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3893 return matchAddr(AddrInst->getOperand(0), Depth);
3894 case Instruction::IntToPtr: {
3895 auto AS = AddrInst->getType()->getPointerAddressSpace();
3896 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3897 // This inttoptr is a no-op if the integer type is pointer sized.
3898 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3899 return matchAddr(AddrInst->getOperand(0), Depth);
3900 return false;
3901 }
3902 case Instruction::BitCast:
3903 // BitCast is always a noop, and we can handle it as long as it is
3904 // int->int or pointer->pointer (we don't want int<->fp or something).
3905 if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
3906 // Don't touch identity bitcasts. These were probably put here by LSR,
3907 // and we don't want to mess around with them. Assume it knows what it
3908 // is doing.
3909 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3910 return matchAddr(AddrInst->getOperand(0), Depth);
3911 return false;
3912 case Instruction::AddrSpaceCast: {
3913 unsigned SrcAS
3914 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3915 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3916 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3917 return matchAddr(AddrInst->getOperand(0), Depth);
3918 return false;
3919 }
3920 case Instruction::Add: {
3921 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3922 ExtAddrMode BackupAddrMode = AddrMode;
3923 unsigned OldSize = AddrModeInsts.size();
3924 // Start a transaction at this point.
3925 // The LHS may match but not the RHS.
3926 // Therefore, we need a higher level restoration point to undo partially
3927 // matched operation.
3928 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3929 TPT.getRestorationPoint();
3930
3931 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3932 matchAddr(AddrInst->getOperand(0), Depth+1))
3933 return true;
3934
3935 // Restore the old addr mode info.
3936 AddrMode = BackupAddrMode;
3937 AddrModeInsts.resize(OldSize);
3938 TPT.rollback(LastKnownGood);
3939
3940 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3941 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3942 matchAddr(AddrInst->getOperand(1), Depth+1))
3943 return true;
3944
3945 // Otherwise we definitely can't merge the ADD in.
3946 AddrMode = BackupAddrMode;
3947 AddrModeInsts.resize(OldSize);
3948 TPT.rollback(LastKnownGood);
3949 break;
3950 }
3951 //case Instruction::Or:
3952 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3953 //break;
3954 case Instruction::Mul:
3955 case Instruction::Shl: {
3956 // Can only handle X*C and X << C.
3957 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3958 if (!RHS || RHS->getBitWidth() > 64)
3959 return false;
3960 int64_t Scale = RHS->getSExtValue();
3961 if (Opcode == Instruction::Shl)
3962 Scale = 1LL << Scale;
3963
3964 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3965 }
3966 case Instruction::GetElementPtr: {
3967 // Scan the GEP. We check it if it contains constant offsets and at most
3968 // one variable offset.
3969 int VariableOperand = -1;
3970 unsigned VariableScale = 0;
3971
3972 int64_t ConstantOffset = 0;
3973 gep_type_iterator GTI = gep_type_begin(AddrInst);
3974 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3975 if (StructType *STy = GTI.getStructTypeOrNull()) {
3976 const StructLayout *SL = DL.getStructLayout(STy);
3977 unsigned Idx =
3978 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3979 ConstantOffset += SL->getElementOffset(Idx);
3980 } else {
3981 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3982 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3983 const APInt &CVal = CI->getValue();
3984 if (CVal.getMinSignedBits() <= 64) {
3985 ConstantOffset += CVal.getSExtValue() * TypeSize;
3986 continue;
3987 }
3988 }
3989 if (TypeSize) { // Scales of zero don't do anything.
3990 // We only allow one variable index at the moment.
3991 if (VariableOperand != -1)
3992 return false;
3993
3994 // Remember the variable index.
3995 VariableOperand = i;
3996 VariableScale = TypeSize;
3997 }
3998 }
3999 }
4000
4001 // A common case is for the GEP to only do a constant offset. In this case,
4002 // just add it to the disp field and check validity.
4003 if (VariableOperand == -1) {
4004 AddrMode.BaseOffs += ConstantOffset;
4005 if (ConstantOffset == 0 ||
4006 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4007 // Check to see if we can fold the base pointer in too.
4008 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4009 return true;
4010 } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4011 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4012 ConstantOffset > 0) {
4013 // Record GEPs with non-zero offsets as candidates for splitting in the
4014 // event that the offset cannot fit into the r+i addressing mode.
4015 // Simple and common case that only one GEP is used in calculating the
4016 // address for the memory access.
4017 Value *Base = AddrInst->getOperand(0);
4018 auto *BaseI = dyn_cast<Instruction>(Base);
4019 auto *GEP = cast<GetElementPtrInst>(AddrInst);
4020 if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4021 (BaseI && !isa<CastInst>(BaseI) &&
4022 !isa<GetElementPtrInst>(BaseI))) {
4023 // If the base is an instruction, make sure the GEP is not in the same
4024 // basic block as the base. If the base is an argument or global
4025 // value, make sure the GEP is not in the entry block. Otherwise,
4026 // instruction selection can undo the split. Also make sure the
4027 // parent block allows inserting non-PHI instructions before the
4028 // terminator.
4029 BasicBlock *Parent =
4030 BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4031 if (GEP->getParent() != Parent && !Parent->getTerminator()->isEHPad())
4032 LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4033 }
4034 }
4035 AddrMode.BaseOffs -= ConstantOffset;
4036 return false;
4037 }
4038
4039 // Save the valid addressing mode in case we can't match.
4040 ExtAddrMode BackupAddrMode = AddrMode;
4041 unsigned OldSize = AddrModeInsts.size();
4042
4043 // See if the scale and offset amount is valid for this target.
4044 AddrMode.BaseOffs += ConstantOffset;
4045
4046 // Match the base operand of the GEP.
4047 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4048 // If it couldn't be matched, just stuff the value in a register.
4049 if (AddrMode.HasBaseReg) {
4050 AddrMode = BackupAddrMode;
4051 AddrModeInsts.resize(OldSize);
4052 return false;
4053 }
4054 AddrMode.HasBaseReg = true;
4055 AddrMode.BaseReg = AddrInst->getOperand(0);
4056 }
4057
4058 // Match the remaining variable portion of the GEP.
4059 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4060 Depth)) {
4061 // If it couldn't be matched, try stuffing the base into a register
4062 // instead of matching it, and retrying the match of the scale.
4063 AddrMode = BackupAddrMode;
4064 AddrModeInsts.resize(OldSize);
4065 if (AddrMode.HasBaseReg)
4066 return false;
4067 AddrMode.HasBaseReg = true;
4068 AddrMode.BaseReg = AddrInst->getOperand(0);
4069 AddrMode.BaseOffs += ConstantOffset;
4070 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4071 VariableScale, Depth)) {
4072 // If even that didn't work, bail.
4073 AddrMode = BackupAddrMode;
4074 AddrModeInsts.resize(OldSize);
4075 return false;
4076 }
4077 }
4078
4079 return true;
4080 }
4081 case Instruction::SExt:
4082 case Instruction::ZExt: {
4083 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4084 if (!Ext)
4085 return false;
4086
4087 // Try to move this ext out of the way of the addressing mode.
4088 // Ask for a method for doing so.
4089 TypePromotionHelper::Action TPH =
4090 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4091 if (!TPH)
4092 return false;
4093
4094 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4095 TPT.getRestorationPoint();
4096 unsigned CreatedInstsCost = 0;
4097 unsigned ExtCost = !TLI.isExtFree(Ext);
4098 Value *PromotedOperand =
4099 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4100 // SExt has been moved away.
4101 // Thus either it will be rematched later in the recursive calls or it is
4102 // gone. Anyway, we must not fold it into the addressing mode at this point.
4103 // E.g.,
4104 // op = add opnd, 1
4105 // idx = ext op
4106 // addr = gep base, idx
4107 // is now:
4108 // promotedOpnd = ext opnd <- no match here
4109 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4110 // addr = gep base, op <- match
4111 if (MovedAway)
4112 *MovedAway = true;
4113
4114 assert(PromotedOperand &&((PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? static_cast<void> (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4115, __PRETTY_FUNCTION__))
4115 "TypePromotionHelper should have filtered out those cases")((PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? static_cast<void> (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4115, __PRETTY_FUNCTION__))
;
4116
4117 ExtAddrMode BackupAddrMode = AddrMode;
4118 unsigned OldSize = AddrModeInsts.size();
4119
4120 if (!matchAddr(PromotedOperand, Depth) ||
4121 // The total of the new cost is equal to the cost of the created
4122 // instructions.
4123 // The total of the old cost is equal to the cost of the extension plus
4124 // what we have saved in the addressing mode.
4125 !isPromotionProfitable(CreatedInstsCost,
4126 ExtCost + (AddrModeInsts.size() - OldSize),
4127 PromotedOperand)) {
4128 AddrMode = BackupAddrMode;
4129 AddrModeInsts.resize(OldSize);
4130 LLVM_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)
;
4131 TPT.rollback(LastKnownGood);
4132 return false;
4133 }
4134 return true;
4135 }
4136 }
4137 return false;
4138}
4139
4140/// If we can, try to add the value of 'Addr' into the current addressing mode.
4141/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4142/// unmodified. This assumes that Addr is either a pointer type or intptr_t
4143/// for the target.
4144///
4145bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4146 // Start a transaction at this point that we will rollback if the matching
4147 // fails.
4148 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4149 TPT.getRestorationPoint();
4150 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4151 // Fold in immediates if legal for the target.
4152 AddrMode.BaseOffs += CI->getSExtValue();
4153 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4154 return true;
4155 AddrMode.BaseOffs -= CI->getSExtValue();
4156 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4157 // If this is a global variable, try to fold it into the addressing mode.
4158 if (!AddrMode.BaseGV) {
4159 AddrMode.BaseGV = GV;
4160 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4161 return true;
4162 AddrMode.BaseGV = nullptr;
4163 }
4164 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4165 ExtAddrMode BackupAddrMode = AddrMode;
4166 unsigned OldSize = AddrModeInsts.size();
4167
4168 // Check to see if it is possible to fold this operation.
4169 bool MovedAway = false;
4170 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4171 // This instruction may have been moved away. If so, there is nothing
4172 // to check here.
4173 if (MovedAway)
4174 return true;
4175 // Okay, it's possible to fold this. Check to see if it is actually
4176 // *profitable* to do so. We use a simple cost model to avoid increasing
4177 // register pressure too much.
4178 if (I->hasOneUse() ||
4179 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4180 AddrModeInsts.push_back(I);
4181 return true;
4182 }
4183
4184 // It isn't profitable to do this, roll back.
4185 //cerr << "NOT FOLDING: " << *I;
4186 AddrMode = BackupAddrMode;
4187 AddrModeInsts.resize(OldSize);
4188 TPT.rollback(LastKnownGood);
4189 }
4190 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4191 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4192 return true;
4193 TPT.rollback(LastKnownGood);
4194 } else if (isa<ConstantPointerNull>(Addr)) {
4195 // Null pointer gets folded without affecting the addressing mode.
4196 return true;
4197 }
4198
4199 // Worse case, the target should support [reg] addressing modes. :)
4200 if (!AddrMode.HasBaseReg) {
4201 AddrMode.HasBaseReg = true;
4202 AddrMode.BaseReg = Addr;
4203 // Still check for legality in case the target supports [imm] but not [i+r].
4204 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4205 return true;
4206 AddrMode.HasBaseReg = false;
4207 AddrMode.BaseReg = nullptr;
4208 }
4209
4210 // If the base register is already taken, see if we can do [r+r].
4211 if (AddrMode.Scale == 0) {
4212 AddrMode.Scale = 1;
4213 AddrMode.ScaledReg = Addr;
4214 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4215 return true;
4216 AddrMode.Scale = 0;
4217 AddrMode.ScaledReg = nullptr;
4218 }
4219 // Couldn't match.
4220 TPT.rollback(LastKnownGood);
4221 return false;
4222}
4223
4224/// Check to see if all uses of OpVal by the specified inline asm call are due
4225/// to memory operands. If so, return true, otherwise return false.
4226static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4227 const TargetLowering &TLI,
4228 const TargetRegisterInfo &TRI) {
4229 const Function *F = CI->getFunction();
4230 TargetLowering::AsmOperandInfoVector TargetConstraints =
4231 TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
4232 ImmutableCallSite(CI));
4233
4234 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4235 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4236
4237 // Compute the constraint code and ConstraintType to use.
4238 TLI.ComputeConstraintToUse(OpInfo, SDValue());
4239
4240 // If this asm operand is our Value*, and if it isn't an indirect memory
4241 // operand, we can't fold it!
4242 if (OpInfo.CallOperandVal == OpVal &&
4243 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4244 !OpInfo.isIndirect))
4245 return false;
4246 }
4247
4248 return true;
4249}
4250
4251// Max number of memory uses to look at before aborting the search to conserve
4252// compile time.
4253static constexpr int MaxMemoryUsesToScan = 20;
4254
4255/// Recursively walk all the uses of I until we find a memory use.
4256/// If we find an obviously non-foldable instruction, return true.
4257/// Add the ultimately found memory instructions to MemoryUses.
4258static bool FindAllMemoryUses(
4259 Instruction *I,
4260 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4261 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4262 const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4263 // If we already considered this instruction, we're done.
4264 if (!ConsideredInsts.insert(I).second)
4265 return false;
4266
4267 // If this is an obviously unfoldable instruction, bail out.
4268 if (!MightBeFoldableInst(I))
4269 return true;
4270
4271 const bool OptSize = I->getFunction()->optForSize();
4272
4273 // Loop over all the uses, recursively processing them.
4274 for (Use &U : I->uses()) {
4275 // Conservatively return true if we're seeing a large number or a deep chain
4276 // of users. This avoids excessive compilation times in pathological cases.
4277 if (SeenInsts++ >= MaxMemoryUsesToScan)
4278 return true;
4279
4280 Instruction *UserI = cast<Instruction>(U.getUser());
4281 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4282 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4283 continue;
4284 }
4285
4286 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4287 unsigned opNo = U.getOperandNo();
4288 if (opNo != StoreInst::getPointerOperandIndex())
4289 return true; // Storing addr, not into addr.
4290 MemoryUses.push_back(std::make_pair(SI, opNo));
4291 continue;
4292 }
4293
4294 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4295 unsigned opNo = U.getOperandNo();
4296 if (opNo != AtomicRMWInst::getPointerOperandIndex())
4297 return true; // Storing addr, not into addr.
4298 MemoryUses.push_back(std::make_pair(RMW, opNo));
4299 continue;
4300 }
4301
4302 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4303 unsigned opNo = U.getOperandNo();
4304 if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
4305 return true; // Storing addr, not into addr.
4306 MemoryUses.push_back(std::make_pair(CmpX, opNo));
4307 continue;
4308 }
4309
4310 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4311 // If this is a cold call, we can sink the addressing calculation into
4312 // the cold path. See optimizeCallInst
4313 if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4314 continue;
4315
4316 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4317 if (!IA) return true;
4318
4319 // If this is a memory operand, we're cool, otherwise bail out.
4320 if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4321 return true;
4322 continue;
4323 }
4324
4325 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4326 SeenInsts))
4327 return true;
4328 }
4329
4330 return false;
4331}
4332
4333/// Return true if Val is already known to be live at the use site that we're
4334/// folding it into. If so, there is no cost to include it in the addressing
4335/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4336/// instruction already.
4337bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4338 Value *KnownLive2) {
4339 // If Val is either of the known-live values, we know it is live!
4340 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4341 return true;
4342
4343 // All values other than instructions and arguments (e.g. constants) are live.
4344 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4345
4346 // If Val is a constant sized alloca in the entry block, it is live, this is
4347 // true because it is just a reference to the stack/frame pointer, which is
4348 // live for the whole function.
4349 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4350 if (AI->isStaticAlloca())
4351 return true;
4352
4353 // Check to see if this value is already used in the memory instruction's
4354 // block. If so, it's already live into the block at the very least, so we
4355 // can reasonably fold it.
4356 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4357}
4358
4359/// It is possible for the addressing mode of the machine to fold the specified
4360/// instruction into a load or store that ultimately uses it.
4361/// However, the specified instruction has multiple uses.
4362/// Given this, it may actually increase register pressure to fold it
4363/// into the load. For example, consider this code:
4364///
4365/// X = ...
4366/// Y = X+1
4367/// use(Y) -> nonload/store
4368/// Z = Y+1
4369/// load Z
4370///
4371/// In this case, Y has multiple uses, and can be folded into the load of Z
4372/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4373/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4374/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4375/// number of computations either.
4376///
4377/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4378/// X was live across 'load Z' for other reasons, we actually *would* want to
4379/// fold the addressing mode in the Z case. This would make Y die earlier.
4380bool AddressingModeMatcher::
4381isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4382 ExtAddrMode &AMAfter) {
4383 if (IgnoreProfitability) return true;
4384
4385 // AMBefore is the addressing mode before this instruction was folded into it,
4386 // and AMAfter is the addressing mode after the instruction was folded. Get
4387 // the set of registers referenced by AMAfter and subtract out those
4388 // referenced by AMBefore: this is the set of values which folding in this
4389 // address extends the lifetime of.
4390 //
4391 // Note that there are only two potential values being referenced here,
4392 // BaseReg and ScaleReg (global addresses are always available, as are any
4393 // folded immediates).
4394 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4395
4396 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4397 // lifetime wasn't extended by adding this instruction.
4398 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4399 BaseReg = nullptr;
4400 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4401 ScaledReg = nullptr;
4402
4403 // If folding this instruction (and it's subexprs) didn't extend any live
4404 // ranges, we're ok with it.
4405 if (!BaseReg && !ScaledReg)
4406 return true;
4407
4408 // If all uses of this instruction can have the address mode sunk into them,
4409 // we can remove the addressing mode and effectively trade one live register
4410 // for another (at worst.) In this context, folding an addressing mode into
4411 // the use is just a particularly nice way of sinking it.
4412 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4413 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4414 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4415 return false; // Has a non-memory, non-foldable use!
4416
4417 // Now that we know that all uses of this instruction are part of a chain of
4418 // computation involving only operations that could theoretically be folded
4419 // into a memory use, loop over each of these memory operation uses and see
4420 // if they could *actually* fold the instruction. The assumption is that
4421 // addressing modes are cheap and that duplicating the computation involved
4422 // many times is worthwhile, even on a fastpath. For sinking candidates
4423 // (i.e. cold call sites), this serves as a way to prevent excessive code
4424 // growth since most architectures have some reasonable small and fast way to
4425 // compute an effective address. (i.e LEA on x86)
4426 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4427 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4428 Instruction *User = MemoryUses[i].first;
4429 unsigned OpNo = MemoryUses[i].second;
4430
4431 // Get the access type of this use. If the use isn't a pointer, we don't
4432 // know what it accesses.
4433 Value *Address = User->getOperand(OpNo);
4434 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4435 if (!AddrTy)
4436 return false;
4437 Type *AddressAccessTy = AddrTy->getElementType();
4438 unsigned AS = AddrTy->getAddressSpace();
4439
4440 // Do a match against the root of this address, ignoring profitability. This
4441 // will tell us if the addressing mode for the memory operation will
4442 // *actually* cover the shared instruction.
4443 ExtAddrMode Result;
4444 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4445 0);
4446 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4447 TPT.getRestorationPoint();
4448 AddressingModeMatcher Matcher(
4449 MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4450 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4451 Matcher.IgnoreProfitability = true;
4452 bool Success = Matcher.matchAddr(Address, 0);
4453 (void)Success; assert(Success && "Couldn't select *anything*?")((Success && "Couldn't select *anything*?") ? static_cast
<void> (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4453, __PRETTY_FUNCTION__))
;
4454
4455 // The match was to check the profitability, the changes made are not
4456 // part of the original matcher. Therefore, they should be dropped
4457 // otherwise the original matcher will not present the right state.
4458 TPT.rollback(LastKnownGood);
4459
4460 // If the match didn't cover I, then it won't be shared by it.
4461 if (!is_contained(MatchedAddrModeInsts, I))
4462 return false;
4463
4464 MatchedAddrModeInsts.clear();
4465 }
4466
4467 return true;
4468}
4469
4470/// Return true if the specified values are defined in a
4471/// different basic block than BB.
4472static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4473 if (Instruction *I = dyn_cast<Instruction>(V))
4474 return I->getParent() != BB;
4475 return false;
4476}
4477
4478/// Sink addressing mode computation immediate before MemoryInst if doing so
4479/// can be done without increasing register pressure. The need for the
4480/// register pressure constraint means this can end up being an all or nothing
4481/// decision for all uses of the same addressing computation.
4482///
4483/// Load and Store Instructions often have addressing modes that can do
4484/// significant amounts of computation. As such, instruction selection will try
4485/// to get the load or store to do as much computation as possible for the
4486/// program. The problem is that isel can only see within a single block. As
4487/// such, we sink as much legal addressing mode work into the block as possible.
4488///
4489/// This method is used to optimize both load/store and inline asms with memory
4490/// operands. It's also used to sink addressing computations feeding into cold
4491/// call sites into their (cold) basic block.
4492///
4493/// The motivation for handling sinking into cold blocks is that doing so can
4494/// both enable other address mode sinking (by satisfying the register pressure
4495/// constraint above), and reduce register pressure globally (by removing the
4496/// addressing mode computation from the fast path entirely.).
4497bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4498 Type *AccessTy, unsigned AddrSpace) {
4499 Value *Repl = Addr;
4500
4501 // Try to collapse single-value PHI nodes. This is necessary to undo
4502 // unprofitable PRE transformations.
4503 SmallVector<Value*, 8> worklist;
4504 SmallPtrSet<Value*, 16> Visited;
4505 worklist.push_back(Addr);
4506
4507 // Use a worklist to iteratively look through PHI and select nodes, and
4508 // ensure that the addressing mode obtained from the non-PHI/select roots of
4509 // the graph are compatible.
4510 bool PhiOrSelectSeen = false;
4511 SmallVector<Instruction*, 16> AddrModeInsts;
4512 const SimplifyQuery SQ(*DL, TLInfo);
4513 AddressingModeCombiner AddrModes(SQ, Addr);
4514 TypePromotionTransaction TPT(RemovedInsts);
4515 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4516 TPT.getRestorationPoint();
4517 while (!worklist.empty()) {
4518 Value *V = worklist.back();
4519 worklist.pop_back();
4520
4521 // We allow traversing cyclic Phi nodes.
4522 // In case of success after this loop we ensure that traversing through
4523 // Phi nodes ends up with all cases to compute address of the form
4524 // BaseGV + Base + Scale * Index + Offset
4525 // where Scale and Offset are constans and BaseGV, Base and Index
4526 // are exactly the same Values in all cases.
4527 // It means that BaseGV, Scale and Offset dominate our memory instruction
4528 // and have the same value as they had in address computation represented
4529 // as Phi. So we can safely sink address computation to memory instruction.
4530 if (!Visited.insert(V).second)
4531 continue;
4532
4533 // For a PHI node, push all of its incoming values.
4534 if (PHINode *P = dyn_cast<PHINode>(V)) {
4535 for (Value *IncValue : P->incoming_values())
4536 worklist.push_back(IncValue);
4537 PhiOrSelectSeen = true;
4538 continue;
4539 }
4540 // Similar for select.
4541 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4542 worklist.push_back(SI->getFalseValue());
4543 worklist.push_back(SI->getTrueValue());
4544 PhiOrSelectSeen = true;
4545 continue;
4546 }
4547
4548 // For non-PHIs, determine the addressing mode being computed. Note that
4549 // the result may differ depending on what other uses our candidate
4550 // addressing instructions might have.
4551 AddrModeInsts.clear();
4552 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4553 0);
4554 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4555 V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4556 InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4557
4558 GetElementPtrInst *GEP = LargeOffsetGEP.first;
4559 if (GEP && GEP->getParent() != MemoryInst->getParent() &&
4560 !NewGEPBases.count(GEP)) {
4561 // If splitting the underlying data structure can reduce the offset of a
4562 // GEP, collect the GEP. Skip the GEPs that are the new bases of
4563 // previously split data structures.
4564 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
4565 if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
4566 LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
4567 }
4568
4569 NewAddrMode.OriginalValue = V;
4570 if (!AddrModes.addNewAddrMode(NewAddrMode))
4571 break;
4572 }
4573
4574 // Try to combine the AddrModes we've collected. If we couldn't collect any,
4575 // or we have multiple but either couldn't combine them or combining them
4576 // wouldn't do anything useful, bail out now.
4577 if (!AddrModes.combineAddrModes()) {
4578 TPT.rollback(LastKnownGood);
4579 return false;
4580 }
4581 TPT.commit();
4582
4583 // Get the combined AddrMode (or the only AddrMode, if we only had one).
4584 ExtAddrMode AddrMode = AddrModes.getAddrMode();
4585
4586 // If all the instructions matched are already in this BB, don't do anything.
4587 // If we saw a Phi node then it is not local definitely, and if we saw a select
4588 // then we want to push the address calculation past it even if it's already
4589 // in this BB.
4590 if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4591 return IsNonLocalValue(V, MemoryInst->getParent());
4592 })) {
4593 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Found local addrmode: "
<< AddrMode << "\n"; } } while (false)
4594 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Found local addrmode: "
<< AddrMode << "\n"; } } while (false)
;
4595 return false;
4596 }
4597
4598 // Insert this computation right after this user. Since our caller is
4599 // scanning from the top of the BB to the bottom, reuse of the expr are
4600 // guaranteed to happen later.
4601 IRBuilder<> Builder(MemoryInst);
4602
4603 // Now that we determined the addressing expression we want to use and know
4604 // that we have to sink it into this block. Check to see if we have already
4605 // done this for some other load/store instr in this block. If so, reuse
4606 // the computation. Before attempting reuse, check if the address is valid
4607 // as it may have been erased.
4608
4609 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4610
4611 Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4612 if (SunkAddr) {
4613 LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4614 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4615 if (SunkAddr->getType() != Addr->getType())
4616 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4617 } else if (AddrSinkUsingGEPs ||
4618 (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4619 // By default, we use the GEP-based method when AA is used later. This
4620 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4621 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4622 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4623 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4624 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4625
4626 // First, find the pointer.
4627 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4628 ResultPtr = AddrMode.BaseReg;
4629 AddrMode.BaseReg = nullptr;
4630 }
4631
4632 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4633 // We can't add more than one pointer together, nor can we scale a
4634 // pointer (both of which seem meaningless).
4635 if (ResultPtr || AddrMode.Scale != 1)
4636 return false;
4637
4638 ResultPtr = AddrMode.ScaledReg;
4639 AddrMode.Scale = 0;
4640 }
4641
4642 // It is only safe to sign extend the BaseReg if we know that the math
4643 // required to create it did not overflow before we extend it. Since
4644 // the original IR value was tossed in favor of a constant back when
4645 // the AddrMode was created we need to bail out gracefully if widths
4646 // do not match instead of extending it.
4647 //
4648 // (See below for code to add the scale.)
4649 if (AddrMode.Scale) {
4650 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4651 if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4652 cast<IntegerType>(ScaledRegTy)->getBitWidth())
4653 return false;
4654 }
4655
4656 if (AddrMode.BaseGV) {
4657 if (ResultPtr)
4658 return false;
4659
4660 ResultPtr = AddrMode.BaseGV;
4661 }
4662
4663 // If the real base value actually came from an inttoptr, then the matcher
4664 // will look through it and provide only the integer value. In that case,
4665 // use it here.
4666 if (!DL->isNonIntegralPointerType(Addr->getType())) {
4667 if (!ResultPtr && AddrMode.BaseReg) {
4668 ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4669 "sunkaddr");
4670 AddrMode.BaseReg = nullptr;
4671 } else if (!ResultPtr && AddrMode.Scale == 1) {
4672 ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4673 "sunkaddr");
4674 AddrMode.Scale = 0;
4675 }
4676 }
4677
4678 if (!ResultPtr &&
4679 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4680 SunkAddr = Constant::getNullValue(Addr->getType());
4681 } else if (!ResultPtr) {
4682 return false;
4683 } else {
4684 Type *I8PtrTy =
4685 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4686 Type *I8Ty = Builder.getInt8Ty();
4687
4688 // Start with the base register. Do this first so that subsequent address
4689 // matching finds it last, which will prevent it from trying to match it
4690 // as the scaled value in case it happens to be a mul. That would be
4691 // problematic if we've sunk a different mul for the scale, because then
4692 // we'd end up sinking both muls.
4693 if (AddrMode.BaseReg) {
4694 Value *V = AddrMode.BaseReg;
4695 if (V->getType() != IntPtrTy)
4696 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4697
4698 ResultIndex = V;
4699 }
4700
4701 // Add the scale value.
4702 if (AddrMode.Scale) {
4703 Value *V = AddrMode.ScaledReg;
4704 if (V->getType() == IntPtrTy) {
4705 // done.
4706 } else {
4707 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <((cast<IntegerType>(IntPtrTy)->getBitWidth() < cast
<IntegerType>(V->getType())->getBitWidth() &&
"We can't transform if ScaledReg is too narrow") ? static_cast
<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-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4709, __PRETTY_FUNCTION__))
4708 cast<IntegerType>(V->getType())->getBitWidth() &&((cast<IntegerType>(IntPtrTy)->getBitWidth() < cast
<IntegerType>(V->getType())->getBitWidth() &&
"We can't transform if ScaledReg is too narrow") ? static_cast
<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-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4709, __PRETTY_FUNCTION__))
4709 "We can't transform if ScaledReg is too narrow")((cast<IntegerType>(IntPtrTy)->getBitWidth() < cast
<IntegerType>(V->getType())->getBitWidth() &&
"We can't transform if ScaledReg is too narrow") ? static_cast
<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-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4709, __PRETTY_FUNCTION__))
;
4710 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4711 }
4712
4713 if (AddrMode.Scale != 1)
4714 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4715 "sunkaddr");
4716 if (ResultIndex)
4717 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4718 else
4719 ResultIndex = V;
4720 }
4721
4722 // Add in the Base Offset if present.
4723 if (AddrMode.BaseOffs) {
4724 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4725 if (ResultIndex) {
4726 // We need to add this separately from the scale above to help with
4727 // SDAG consecutive load/store merging.
4728 if (ResultPtr->getType() != I8PtrTy)
4729 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4730 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4731 }
4732
4733 ResultIndex = V;
4734 }
4735
4736 if (!ResultIndex) {
4737 SunkAddr = ResultPtr;
4738 } else {
4739 if (ResultPtr->getType() != I8PtrTy)
4740 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4741 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4742 }
4743
4744 if (SunkAddr->getType() != Addr->getType())
4745 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4746 }
4747 } else {
4748 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4749 // non-integral pointers, so in that case bail out now.
4750 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4751 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4752 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4753 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4754 if (DL->isNonIntegralPointerType(Addr->getType()) ||
4755 (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4756 (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4757 (AddrMode.BaseGV &&
4758 DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4759 return false;
4760
4761 LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
4762 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
4763 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4764 Value *Result = nullptr;
4765
4766 // Start with the base register. Do this first so that subsequent address
4767 // matching finds it last, which will prevent it from trying to match it
4768 // as the scaled value in case it happens to be a mul. That would be
4769 // problematic if we've sunk a different mul for the scale, because then
4770 // we'd end up sinking both muls.
4771 if (AddrMode.BaseReg) {
4772 Value *V = AddrMode.BaseReg;
4773 if (V->getType()->isPointerTy())
4774 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4775 if (V->getType() != IntPtrTy)
4776 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4777 Result = V;
4778 }
4779
4780 // Add the scale value.
4781 if (AddrMode.Scale) {
4782 Value *V = AddrMode.ScaledReg;
4783 if (V->getType() == IntPtrTy) {
4784 // done.
4785 } else if (V->getType()->isPointerTy()) {
4786 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4787 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4788 cast<IntegerType>(V->getType())->getBitWidth()) {
4789 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4790 } else {
4791 // It is only safe to sign extend the BaseReg if we know that the math
4792 // required to create it did not overflow before we extend it. Since
4793 // the original IR value was tossed in favor of a constant back when
4794 // the AddrMode was created we need to bail out gracefully if widths
4795 // do not match instead of extending it.
4796 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4797 if (I && (Result != AddrMode.BaseReg))
4798 I->eraseFromParent();
4799 return false;
4800 }
4801 if (AddrMode.Scale != 1)
4802 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4803 "sunkaddr");
4804 if (Result)
4805 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4806 else
4807 Result = V;
4808 }
4809
4810 // Add in the BaseGV if present.
4811 if (AddrMode.BaseGV) {
4812 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4813 if (Result)
4814 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4815 else
4816 Result = V;
4817 }
4818
4819 // Add in the Base Offset if present.
4820 if (AddrMode.BaseOffs) {
4821 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4822 if (Result)
4823 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4824 else
4825 Result = V;
4826 }
4827
4828 if (!Result)
4829 SunkAddr = Constant::getNullValue(Addr->getType());
4830 else
4831 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4832 }
4833
4834 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4835 // Store the newly computed address into the cache. In the case we reused a
4836 // value, this should be idempotent.
4837 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4838
4839 // If we have no uses, recursively delete the value and all dead instructions
4840 // using it.
4841 if (Repl->use_empty()) {
4842 // This can cause recursive deletion, which can invalidate our iterator.
4843 // Use a WeakTrackingVH to hold onto it in case this happens.
4844 Value *CurValue = &*CurInstIterator;
4845 WeakTrackingVH IterHandle(CurValue);
4846 BasicBlock *BB = CurInstIterator->getParent();
4847
4848 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
4849
4850 if (IterHandle != CurValue) {
4851 // If the iterator instruction was recursively deleted, start over at the
4852 // start of the block.
4853 CurInstIterator = BB->begin();
4854 SunkAddrs.clear();
4855 }
4856 }
4857 ++NumMemoryInsts;
4858 return true;
4859}
4860
4861/// If there are any memory operands, use OptimizeMemoryInst to sink their
4862/// address computing into the block when possible / profitable.
4863bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4864 bool MadeChange = false;
4865
4866 const TargetRegisterInfo *TRI =
4867 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4868 TargetLowering::AsmOperandInfoVector TargetConstraints =
4869 TLI->ParseConstraints(*DL, TRI, CS);
4870 unsigned ArgNo = 0;
4871 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4872 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4873
4874 // Compute the constraint code and ConstraintType to use.
4875 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4876
4877 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4878 OpInfo.isIndirect) {
4879 Value *OpVal = CS->getArgOperand(ArgNo++);
4880 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4881 } else if (OpInfo.Type == InlineAsm::isInput)
4882 ArgNo++;
4883 }
4884
4885 return MadeChange;
4886}
4887
4888/// Check if all the uses of \p Val are equivalent (or free) zero or
4889/// sign extensions.
4890static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4891 assert(!Val->use_empty() && "Input must have at least one use")((!Val->use_empty() && "Input must have at least one use"
) ? static_cast<void> (0) : __assert_fail ("!Val->use_empty() && \"Input must have at least one use\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4891, __PRETTY_FUNCTION__))
;
4892 const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4893 bool IsSExt = isa<SExtInst>(FirstUser);
4894 Type *ExtTy = FirstUser->getType();
4895 for (const User *U : Val->users()) {
4896 const Instruction *UI = cast<Instruction>(U);
4897 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4898 return false;
4899 Type *CurTy = UI->getType();
4900 // Same input and output types: Same instruction after CSE.
4901 if (CurTy == ExtTy)
4902 continue;
4903
4904 // If IsSExt is true, we are in this situation:
4905 // a = Val
4906 // b = sext ty1 a to ty2
4907 // c = sext ty1 a to ty3
4908 // Assuming ty2 is shorter than ty3, this could be turned into:
4909 // a = Val
4910 // b = sext ty1 a to ty2
4911 // c = sext ty2 b to ty3
4912 // However, the last sext is not free.
4913 if (IsSExt)
4914 return false;
4915
4916 // This is a ZExt, maybe this is free to extend from one type to another.
4917 // In that case, we would not account for a different use.
4918 Type *NarrowTy;
4919 Type *LargeTy;
4920 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4921 CurTy->getScalarType()->getIntegerBitWidth()) {
4922 NarrowTy = CurTy;
4923 LargeTy = ExtTy;
4924 } else {
4925 NarrowTy = ExtTy;
4926 LargeTy = CurTy;
4927 }
4928
4929 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4930 return false;
4931 }
4932 // All uses are the same or can be derived from one another for free.
4933 return true;
4934}
4935
4936/// Try to speculatively promote extensions in \p Exts and continue
4937/// promoting through newly promoted operands recursively as far as doing so is
4938/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4939/// When some promotion happened, \p TPT contains the proper state to revert
4940/// them.
4941///
4942/// \return true if some promotion happened, false otherwise.
4943bool CodeGenPrepare::tryToPromoteExts(
4944 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4945 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4946 unsigned CreatedInstsCost) {
4947 bool Promoted = false;
4948
4949 // Iterate over all the extensions to try to promote them.
4950 for (auto I : Exts) {
4951 // Early check if we directly have ext(load).
4952 if (isa<LoadInst>(I->getOperand(0))) {
4953 ProfitablyMovedExts.push_back(I);
4954 continue;
4955 }
4956
4957 // Check whether or not we want to do any promotion. The reason we have
4958 // this check inside the for loop is to catch the case where an extension
4959 // is directly fed by a load because in such case the extension can be moved
4960 // up without any promotion on its operands.
4961 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4962 return false;
4963
4964 // Get the action to perform the promotion.
4965 TypePromotionHelper::Action TPH =
4966 TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4967 // Check if we can promote.
4968 if (!TPH) {
4969 // Save the current extension as we cannot move up through its operand.
4970 ProfitablyMovedExts.push_back(I);
4971 continue;
4972 }
4973
4974 // Save the current state.
4975 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4976 TPT.getRestorationPoint();
4977 SmallVector<Instruction *, 4> NewExts;
4978 unsigned NewCreatedInstsCost = 0;
4979 unsigned ExtCost = !TLI->isExtFree(I);
4980 // Promote.
4981 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4982 &NewExts, nullptr, *TLI);
4983 assert(PromotedVal &&((PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? static_cast<void> (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4984, __PRETTY_FUNCTION__))
4984 "TypePromotionHelper should have filtered out those cases")((PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? static_cast<void> (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 4984, __PRETTY_FUNCTION__))
;
4985
4986 // We would be able to merge only one extension in a load.
4987 // Therefore, if we have more than 1 new extension we heuristically
4988 // cut this search path, because it means we degrade the code quality.
4989 // With exactly 2, the transformation is neutral, because we will merge
4990 // one extension but leave one. However, we optimistically keep going,
4991 // because the new extension may be removed too.
4992 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4993 // FIXME: It would be possible to propagate a negative value instead of
4994 // conservatively ceiling it to 0.
4995 TotalCreatedInstsCost =
4996 std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4997 if (!StressExtLdPromotion &&
4998 (TotalCreatedInstsCost > 1 ||
4999 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5000 // This promotion is not profitable, rollback to the previous state, and
5001 // save the current extension in ProfitablyMovedExts as the latest
5002 // speculative promotion turned out to be unprofitable.
5003 TPT.rollback(LastKnownGood);
5004 ProfitablyMovedExts.push_back(I);
5005 continue;
5006 }
5007 // Continue promoting NewExts as far as doing so is profitable.
5008 SmallVector<Instruction *, 2> NewlyMovedExts;
5009 (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5010 bool NewPromoted = false;
5011 for (auto ExtInst : NewlyMovedExts) {
5012 Instruction *MovedExt = cast<Instruction>(ExtInst);
5013 Value *ExtOperand = MovedExt->getOperand(0);
5014 // If we have reached to a load, we need this extra profitability check
5015 // as it could potentially be merged into an ext(load).
5016 if (isa<LoadInst>(ExtOperand) &&
5017 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5018 (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5019 continue;
5020
5021 ProfitablyMovedExts.push_back(MovedExt);
5022 NewPromoted = true;
5023 }
5024
5025 // If none of speculative promotions for NewExts is profitable, rollback
5026 // and save the current extension (I) as the last profitable extension.
5027 if (!NewPromoted) {
5028 TPT.rollback(LastKnownGood);
5029 ProfitablyMovedExts.push_back(I);
5030 continue;
5031 }
5032 // The promotion is profitable.
5033 Promoted = true;
5034 }
5035 return Promoted;
5036}
5037
5038/// Merging redundant sexts when one is dominating the other.
5039bool CodeGenPrepare::mergeSExts(Function &F) {
5040 DominatorTree DT(F);
5041 bool Changed = false;
5042 for (auto &Entry : ValToSExtendedUses) {
5043 SExts &Insts = Entry.second;
5044 SExts CurPts;
5045 for (Instruction *Inst : Insts) {
5046 if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5047 Inst->getOperand(0) != Entry.first)
5048 continue;
5049 bool inserted = false;
5050 for (auto &Pt : CurPts) {
5051 if (DT.dominates(Inst, Pt)) {
5052 Pt->replaceAllUsesWith(Inst);
5053 RemovedInsts.insert(Pt);
5054 Pt->removeFromParent();
5055 Pt = Inst;
5056 inserted = true;
5057 Changed = true;
5058 break;
5059 }
5060 if (!DT.dominates(Pt, Inst))
5061 // Give up if we need to merge in a common dominator as the
5062 // experiments show it is not profitable.
5063 continue;
5064 Inst->replaceAllUsesWith(Pt);
5065 RemovedInsts.insert(Inst);
5066 Inst->removeFromParent();
5067 inserted = true;
5068 Changed = true;
5069 break;
5070 }
5071 if (!inserted)
5072 CurPts.push_back(Inst);
5073 }
5074 }
5075 return Changed;
5076}
5077
5078// Spliting large data structures so that the GEPs accessing them can have
5079// smaller offsets so that they can be sunk to the same blocks as their users.
5080// For example, a large struct starting from %base is splitted into two parts
5081// where the second part starts from %new_base.
5082//
5083// Before:
5084// BB0:
5085// %base =
5086//
5087// BB1:
5088// %gep0 = gep %base, off0
5089// %gep1 = gep %base, off1
5090// %gep2 = gep %base, off2
5091//
5092// BB2:
5093// %load1 = load %gep0
5094// %load2 = load %gep1
5095// %load3 = load %gep2
5096//
5097// After:
5098// BB0:
5099// %base =
5100// %new_base = gep %base, off0
5101//
5102// BB1:
5103// %new_gep0 = %new_base
5104// %new_gep1 = gep %new_base, off1 - off0
5105// %new_gep2 = gep %new_base, off2 - off0
5106//
5107// BB2:
5108// %load1 = load i32, i32* %new_gep0
5109// %load2 = load i32, i32* %new_gep1
5110// %load3 = load i32, i32* %new_gep2
5111//
5112// %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5113// their offsets are smaller enough to fit into the addressing mode.
5114bool CodeGenPrepare::splitLargeGEPOffsets() {
5115 bool Changed = false;
5116 for (auto &Entry : LargeOffsetGEPMap) {
5117 Value *OldBase = Entry.first;
5118 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
5119 &LargeOffsetGEPs = Entry.second;
5120 auto compareGEPOffset =
5121 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5122 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5123 if (LHS.first == RHS.first)
5124 return false;
5125 if (LHS.second != RHS.second)
5126 return LHS.second < RHS.second;
5127 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5128 };
5129 // Sorting all the GEPs of the same data structures based on the offsets.
5130 llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5131 LargeOffsetGEPs.erase(
5132 std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5133 LargeOffsetGEPs.end());
5134 // Skip if all the GEPs have the same offsets.
5135 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5136 continue;
5137 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5138 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5139 Value *NewBaseGEP = nullptr;
5140
5141 auto LargeOffsetGEP = LargeOffsetGEPs.begin();
5142 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5143 GetElementPtrInst *GEP = LargeOffsetGEP->first;
5144 int64_t Offset = LargeOffsetGEP->second;
5145 if (Offset != BaseOffset) {
5146 TargetLowering::AddrMode AddrMode;
5147 AddrMode.BaseOffs = Offset - BaseOffset;
5148 // The result type of the GEP might not be the type of the memory
5149 // access.
5150 if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5151 GEP->getResultElementType(),
5152 GEP->getAddressSpace())) {
5153 // We need to create a new base if the offset to the current base is
5154 // too large to fit into the addressing mode. So, a very large struct
5155 // may be splitted into several parts.
5156 BaseGEP = GEP;
5157 BaseOffset = Offset;
5158 NewBaseGEP = nullptr;
5159 }
5160 }
5161
5162 // Generate a new GEP to replace the current one.
5163 LLVMContext &Ctx = GEP->getContext();
5164 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5165 Type *I8PtrTy =
5166 Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace());
5167 Type *I8Ty = Type::getInt8Ty(Ctx);
5168
5169 if (!NewBaseGEP) {
5170 // Create a new base if we don't have one yet. Find the insertion
5171 // pointer for the new base first.
5172 BasicBlock::iterator NewBaseInsertPt;
5173 BasicBlock *NewBaseInsertBB;
5174 if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5175 // If the base of the struct is an instruction, the new base will be
5176 // inserted close to it.
5177 NewBaseInsertBB = BaseI->getParent();
5178 if (isa<PHINode>(BaseI))
5179 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5180 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5181 NewBaseInsertBB =
5182 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5183 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5184 } else
5185 NewBaseInsertPt = std::next(BaseI->getIterator());
5186 } else {
5187 // If the current base is an argument or global value, the new base
5188 // will be inserted to the entry block.
5189 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5190 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5191 }
5192 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5193 // Create a new base.
5194 Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5195 NewBaseGEP = OldBase;
5196 if (NewBaseGEP->getType() != I8PtrTy)
5197 NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5198 NewBaseGEP =
5199 NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5200 NewGEPBases.insert(NewBaseGEP);
5201 }
5202
5203 IRBuilder<> Builder(GEP);
5204 Value *NewGEP = NewBaseGEP;
5205 if (Offset == BaseOffset) {
5206 if (GEP->getType() != I8PtrTy)
5207 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5208 } else {
5209 // Calculate the new offset for the new GEP.
5210 Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5211 NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5212
5213 if (GEP->getType() != I8PtrTy)
5214 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5215 }
5216 GEP->replaceAllUsesWith(NewGEP);
5217 LargeOffsetGEPID.erase(GEP);
5218 LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5219 GEP->eraseFromParent();
5220 Changed = true;
5221 }
5222 }
5223 return Changed;
5224}
5225
5226/// Return true, if an ext(load) can be formed from an extension in
5227/// \p MovedExts.
5228bool CodeGenPrepare::canFormExtLd(
5229 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5230 Instruction *&Inst, bool HasPromoted) {
5231 for (auto *MovedExtInst : MovedExts) {
5232 if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5233 LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5234 Inst = MovedExtInst;
5235 break;
5236 }
5237 }
5238 if (!LI)
5239 return false;
5240
5241 // If they're already in the same block, there's nothing to do.
5242 // Make the cheap checks first if we did not promote.
5243 // If we promoted, we need to check if it is indeed profitable.
5244 if (!HasPromoted && LI->getParent() == Inst->getParent())
5245 return false;
5246
5247 return TLI->isExtLoad(LI, Inst, *DL);
5248}
5249
5250/// Move a zext or sext fed by a load into the same basic block as the load,
5251/// unless conditions are unfavorable. This allows SelectionDAG to fold the
5252/// extend into the load.
5253///
5254/// E.g.,
5255/// \code
5256/// %ld = load i32* %addr
5257/// %add = add nuw i32 %ld, 4
5258/// %zext = zext i32 %add to i64
5259// \endcode
5260/// =>
5261/// \code
5262/// %ld = load i32* %addr
5263/// %zext = zext i32 %ld to i64
5264/// %add = add nuw i64 %zext, 4
5265/// \encode
5266/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
5267/// allow us to match zext(load i32*) to i64.
5268///
5269/// Also, try to promote the computations used to obtain a sign extended
5270/// value used into memory accesses.
5271/// E.g.,
5272/// \code
5273/// a = add nsw i32 b, 3
5274/// d = sext i32 a to i64
5275/// e = getelementptr ..., i64 d
5276/// \endcode
5277/// =>
5278/// \code
5279/// f = sext i32 b to i64
5280/// a = add nsw i64 f, 3
5281/// e = getelementptr ..., i64 a
5282/// \endcode
5283///
5284/// \p Inst[in/out] the extension may be modified during the process if some
5285/// promotions apply.
5286bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
5287 // ExtLoad formation and address type promotion infrastructure requires TLI to
5288 // be effective.
5289 if (!TLI)
5290 return false;
5291
5292 bool AllowPromotionWithoutCommonHeader = false;
5293 /// See if it is an interesting sext operations for the address type
5294 /// promotion before trying to promote it, e.g., the ones with the right
5295 /// type and used in memory accesses.
5296 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
5297 *Inst, AllowPromotionWithoutCommonHeader);
5298 TypePromotionTransaction TPT(RemovedInsts);
5299 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5300 TPT.getRestorationPoint();
5301 SmallVector<Instruction *, 1> Exts;
5302 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
5303 Exts.push_back(Inst);
5304
5305 bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
5306
5307 // Look for a load being extended.
5308 LoadInst *LI = nullptr;
5309 Instruction *ExtFedByLoad;
5310
5311 // Try to promote a chain of computation if it allows to form an extended
5312 // load.
5313 if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
5314 assert(LI && ExtFedByLoad && "Expect a valid load and extension")((LI && ExtFedByLoad && "Expect a valid load and extension"
) ? static_cast<void> (0) : __assert_fail ("LI && ExtFedByLoad && \"Expect a valid load and extension\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5314, __PRETTY_FUNCTION__))
;
5315 TPT.commit();
5316 // Move the extend into the same block as the load
5317 ExtFedByLoad->moveAfter(LI);
5318 // CGP does not check if the zext would be speculatively executed when moved
5319 // to the same basic block as the load. Preserving its original location
5320 // would pessimize the debugging experience, as well as negatively impact
5321 // the quality of sample pgo. We don't want to use "line 0" as that has a
5322 // size cost in the line-table section and logically the zext can be seen as
5323 // part of the load. Therefore we conservatively reuse the same debug
5324 // location for the load and the zext.
5325 ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
5326 ++NumExtsMoved;
5327 Inst = ExtFedByLoad;
5328 return true;
5329 }
5330
5331 // Continue promoting SExts if known as considerable depending on targets.
5332 if (ATPConsiderable &&
5333 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
5334 HasPromoted, TPT, SpeculativelyMovedExts))
5335 return true;
5336
5337 TPT.rollback(LastKnownGood);
5338 return false;
5339}
5340
5341// Perform address type promotion if doing so is profitable.
5342// If AllowPromotionWithoutCommonHeader == false, we should find other sext
5343// instructions that sign extended the same initial value. However, if
5344// AllowPromotionWithoutCommonHeader == true, we expect promoting the
5345// extension is just profitable.
5346bool CodeGenPrepare::performAddressTypePromotion(
5347 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
5348 bool HasPromoted, TypePromotionTransaction &TPT,
5349 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
5350 bool Promoted = false;
5351 SmallPtrSet<Instruction *, 1> UnhandledExts;
5352 bool AllSeenFirst = true;
5353 for (auto I : SpeculativelyMovedExts) {
5354 Value *HeadOfChain = I->getOperand(0);
5355 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
5356 SeenChainsForSExt.find(HeadOfChain);
5357 // If there is an unhandled SExt which has the same header, try to promote
5358 // it as well.
5359 if (AlreadySeen != SeenChainsForSExt.end()) {
5360 if (AlreadySeen->second != nullptr)
5361 UnhandledExts.insert(AlreadySeen->second);
5362 AllSeenFirst = false;
5363 }
5364 }
5365
5366 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
5367 SpeculativelyMovedExts.size() == 1)) {
5368 TPT.commit();
5369 if (HasPromoted)
5370 Promoted = true;
5371 for (auto I : SpeculativelyMovedExts) {
5372 Value *HeadOfChain = I->getOperand(0);
5373 SeenChainsForSExt[HeadOfChain] = nullptr;
5374 ValToSExtendedUses[HeadOfChain].push_back(I);
5375 }
5376 // Update Inst as promotion happen.
5377 Inst = SpeculativelyMovedExts.pop_back_val();
5378 } else {
5379 // This is the first chain visited from the header, keep the current chain
5380 // as unhandled. Defer to promote this until we encounter another SExt
5381 // chain derived from the same header.
5382 for (auto I : SpeculativelyMovedExts) {
5383 Value *HeadOfChain = I->getOperand(0);
5384 SeenChainsForSExt[HeadOfChain] = Inst;
5385 }
5386 return false;
5387 }
5388
5389 if (!AllSeenFirst && !UnhandledExts.empty())
5390 for (auto VisitedSExt : UnhandledExts) {
5391 if (RemovedInsts.count(VisitedSExt))
5392 continue;
5393 TypePromotionTransaction TPT(RemovedInsts);
5394 SmallVector<Instruction *, 1> Exts;
5395 SmallVector<Instruction *, 2> Chains;
5396 Exts.push_back(VisitedSExt);
5397 bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5398 TPT.commit();
5399 if (HasPromoted)
5400 Promoted = true;
5401 for (auto I : Chains) {
5402 Value *HeadOfChain = I->getOperand(0);
5403 // Mark this as handled.
5404 SeenChainsForSExt[HeadOfChain] = nullptr;
5405 ValToSExtendedUses[HeadOfChain].push_back(I);
5406 }
5407 }
5408 return Promoted;
5409}
5410
5411bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5412 BasicBlock *DefBB = I->getParent();
5413
5414 // If the result of a {s|z}ext and its source are both live out, rewrite all
5415 // other uses of the source with result of extension.
5416 Value *Src = I->getOperand(0);
5417 if (Src->hasOneUse())
5418 return false;
5419
5420 // Only do this xform if truncating is free.
5421 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5422 return false;
5423
5424 // Only safe to perform the optimization if the source is also defined in
5425 // this block.
5426 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5427 return false;
5428
5429 bool DefIsLiveOut = false;
5430 for (User *U : I->users()) {
5431 Instruction *UI = cast<Instruction>(U);
5432
5433 // Figure out which BB this ext is used in.
5434 BasicBlock *UserBB = UI->getParent();
5435 if (UserBB == DefBB) continue;
5436 DefIsLiveOut = true;
5437 break;
5438 }
5439 if (!DefIsLiveOut)
5440 return false;
5441
5442 // Make sure none of the uses are PHI nodes.
5443 for (User *U : Src->users()) {
5444 Instruction *UI = cast<Instruction>(U);
5445 BasicBlock *UserBB = UI->getParent();
5446 if (UserBB == DefBB) continue;
5447 // Be conservative. We don't want this xform to end up introducing
5448 // reloads just before load / store instructions.
5449 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5450 return false;
5451 }
5452
5453 // InsertedTruncs - Only insert one trunc in each block once.
5454 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5455
5456 bool MadeChange = false;
5457 for (Use &U : Src->uses()) {
5458 Instruction *User = cast<Instruction>(U.getUser());
5459
5460 // Figure out which BB this ext is used in.
5461 BasicBlock *UserBB = User->getParent();
5462 if (UserBB == DefBB) continue;
5463
5464 // Both src and def are live in this block. Rewrite the use.
5465 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5466
5467 if (!InsertedTrunc) {
5468 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5469 assert(InsertPt != UserBB->end())((InsertPt != UserBB->end()) ? static_cast<void> (0)
: __assert_fail ("InsertPt != UserBB->end()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5469, __PRETTY_FUNCTION__))
;
5470 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5471 InsertedInsts.insert(InsertedTrunc);
5472 }
5473
5474 // Replace a use of the {s|z}ext source with a use of the result.
5475 U = InsertedTrunc;
5476 ++NumExtUses;
5477 MadeChange = true;
5478 }
5479
5480 return MadeChange;
5481}
5482
5483// Find loads whose uses only use some of the loaded value's bits. Add an "and"
5484// just after the load if the target can fold this into one extload instruction,
5485// with the hope of eliminating some of the other later "and" instructions using
5486// the loaded value. "and"s that are made trivially redundant by the insertion
5487// of the new "and" are removed by this function, while others (e.g. those whose
5488// path from the load goes through a phi) are left for isel to potentially
5489// remove.
5490//
5491// For example:
5492//
5493// b0:
5494// x = load i32
5495// ...
5496// b1:
5497// y = and x, 0xff
5498// z = use y
5499//
5500// becomes:
5501//
5502// b0:
5503// x = load i32
5504// x' = and x, 0xff
5505// ...
5506// b1:
5507// z = use x'
5508//
5509// whereas:
5510//
5511// b0:
5512// x1 = load i32
5513// ...
5514// b1:
5515// x2 = load i32
5516// ...
5517// b2:
5518// x = phi x1, x2
5519// y = and x, 0xff
5520//
5521// becomes (after a call to optimizeLoadExt for each load):
5522//
5523// b0:
5524// x1 = load i32
5525// x1' = and x1, 0xff
5526// ...
5527// b1:
5528// x2 = load i32
5529// x2' = and x2, 0xff
5530// ...
5531// b2:
5532// x = phi x1', x2'
5533// y = and x, 0xff
5534bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5535 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
5536 return false;
5537
5538 // Skip loads we've already transformed.
5539 if (Load->hasOneUse() &&
5540 InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5541 return false;
5542
5543 // Look at all uses of Load, looking through phis, to determine how many bits
5544 // of the loaded value are needed.
5545 SmallVector<Instruction *, 8> WorkList;
5546 SmallPtrSet<Instruction *, 16> Visited;
5547 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5548 for (auto *U : Load->users())
5549 WorkList.push_back(cast<Instruction>(U));
5550
5551 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5552 unsigned BitWidth = LoadResultVT.getSizeInBits();
5553 APInt DemandBits(BitWidth, 0);
5554 APInt WidestAndBits(BitWidth, 0);
5555
5556 while (!WorkList.empty()) {
5557 Instruction *I = WorkList.back();
5558 WorkList.pop_back();
5559
5560 // Break use-def graph loops.
5561 if (!Visited.insert(I).second)
5562 continue;
5563
5564 // For a PHI node, push all of its users.
5565 if (auto *Phi = dyn_cast<PHINode>(I)) {
5566 for (auto *U : Phi->users())
5567 WorkList.push_back(cast<Instruction>(U));
5568 continue;
5569 }
5570
5571 switch (I->getOpcode()) {
5572 case Instruction::And: {
5573 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5574 if (!AndC)
5575 return false;
5576 APInt AndBits = AndC->getValue();
5577 DemandBits |= AndBits;
5578 // Keep track of the widest and mask we see.
5579 if (AndBits.ugt(WidestAndBits))
5580 WidestAndBits = AndBits;
5581 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5582 AndsToMaybeRemove.push_back(I);
5583 break;
5584 }
5585
5586 case Instruction::Shl: {
5587 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5588 if (!ShlC)
5589 return false;
5590 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5591 DemandBits.setLowBits(BitWidth - ShiftAmt);
5592 break;
5593 }
5594
5595 case Instruction::Trunc: {
5596 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5597 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5598 DemandBits.setLowBits(TruncBitWidth);
5599 break;
5600 }
5601
5602 default:
5603 return false;
5604 }
5605 }
5606
5607 uint32_t ActiveBits = DemandBits.getActiveBits();
5608 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5609 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5610 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5611 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5612 // followed by an AND.
5613 // TODO: Look into removing this restriction by fixing backends to either
5614 // return false for isLoadExtLegal for i1 or have them select this pattern to
5615 // a single instruction.
5616 //
5617 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5618 // mask, since these are the only ands that will be removed by isel.
5619 if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5620 WidestAndBits != DemandBits)
5621 return false;
5622
5623 LLVMContext &Ctx = Load->getType()->getContext();
5624 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5625 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5626
5627 // Reject cases that won't be matched as extloads.
5628 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5629 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5630 return false;
5631
5632 IRBuilder<> Builder(Load->getNextNode());
5633 auto *NewAnd = dyn_cast<Instruction>(
5634 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5635 // Mark this instruction as "inserted by CGP", so that other
5636 // optimizations don't touch it.
5637 InsertedInsts.insert(NewAnd);
5638
5639 // Replace all uses of load with new and (except for the use of load in the
5640 // new and itself).
5641 Load->replaceAllUsesWith(NewAnd);
5642 NewAnd->setOperand(0, Load);
5643
5644 // Remove any and instructions that are now redundant.
5645 for (auto *And : AndsToMaybeRemove)
5646 // Check that the and mask is the same as the one we decided to put on the
5647 // new and.
5648 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5649 And->replaceAllUsesWith(NewAnd);
5650 if (&*CurInstIterator == And)
5651 CurInstIterator = std::next(And->getIterator());
5652 And->eraseFromParent();
5653 ++NumAndUses;
5654 }
5655
5656 ++NumAndsAdded;
5657 return true;
5658}
5659
5660/// Check if V (an operand of a select instruction) is an expensive instruction
5661/// that is only used once.
5662static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5663 auto *I = dyn_cast<Instruction>(V);
5664 // If it's safe to speculatively execute, then it should not have side
5665 // effects; therefore, it's safe to sink and possibly *not* execute.
5666 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5667 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5668}
5669
5670/// Returns true if a SelectInst should be turned into an explicit branch.
5671static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5672 const TargetLowering *TLI,
5673 SelectInst *SI) {
5674 // If even a predictable select is cheap, then a branch can't be cheaper.
5675 if (!TLI->isPredictableSelectExpensive())
5676 return false;
5677
5678 // FIXME: This should use the same heuristics as IfConversion to determine
5679 // whether a select is better represented as a branch.
5680
5681 // If metadata tells us that the select condition is obviously predictable,
5682 // then we want to replace the select with a branch.
5683 uint64_t TrueWeight, FalseWeight;
5684 if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5685 uint64_t Max = std::max(TrueWeight, FalseWeight);
5686 uint64_t Sum = TrueWeight + FalseWeight;
5687 if (Sum != 0) {
5688 auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5689 if (Probability > TLI->getPredictableBranchThreshold())
5690 return true;
5691 }
5692 }
5693
5694 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5695
5696 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5697 // comparison condition. If the compare has more than one use, there's
5698 // probably another cmov or setcc around, so it's not worth emitting a branch.
5699 if (!Cmp || !Cmp->hasOneUse())
5700 return false;
5701
5702 // If either operand of the select is expensive and only needed on one side
5703 // of the select, we should form a branch.
5704 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5705 sinkSelectOperand(TTI, SI->getFalseValue()))
5706 return true;
5707
5708 return false;
5709}
5710
5711/// If \p isTrue is true, return the true value of \p SI, otherwise return
5712/// false value of \p SI. If the true/false value of \p SI is defined by any
5713/// select instructions in \p Selects, look through the defining select
5714/// instruction until the true/false value is not defined in \p Selects.
5715static Value *getTrueOrFalseValue(
5716 SelectInst *SI, bool isTrue,
5717 const SmallPtrSet<const Instruction *, 2> &Selects) {
5718 Value *V;
45
'V' declared without an initial value
5719
5720 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
46
Loop condition is false. Execution continues on line 5726
5721 DefSI = dyn_cast<SelectInst>(V)) {
5722 assert(DefSI->getCondition() == SI->getCondition() &&((DefSI->getCondition() == SI->getCondition() &&
"The condition of DefSI does not match with SI") ? static_cast
<void> (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5723, __PRETTY_FUNCTION__))
5723 "The condition of DefSI does not match with SI")((DefSI->getCondition() == SI->getCondition() &&
"The condition of DefSI does not match with SI") ? static_cast
<void> (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5723, __PRETTY_FUNCTION__))
;
5724 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5725 }
5726 return V;
47
Undefined or garbage value returned to caller
5727}
5728
5729/// If we have a SelectInst that will likely profit from branch prediction,
5730/// turn it into a branch.
5731bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5732 // If branch conversion isn't desirable, exit early.
5733 if (DisableSelectToBranch || OptSize || !TLI)
28
Assuming the condition is false
29
Assuming the condition is false
30
Assuming the condition is false
31
Taking false branch
5734 return false;
5735
5736 // Find all consecutive select instructions that share the same condition.
5737 SmallVector<SelectInst *, 2> ASI;
5738 ASI.push_back(SI);
5739 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
32
Loop condition is false. Execution continues on line 5749
5740 It != SI->getParent()->end(); ++It) {
5741 SelectInst *I = dyn_cast<SelectInst>(&*It);
5742 if (I && SI->getCondition() == I->getCondition()) {
5743 ASI.push_back(I);
5744 } else {
5745 break;
5746 }
5747 }
5748
5749 SelectInst *LastSI = ASI.back();
5750 // Increment the current iterator to skip all the rest of select instructions
5751 // because they will be either "not lowered" or "all lowered" to branch.
5752 CurInstIterator = std::next(LastSI->getIterator());
5753
5754 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
33
Assuming the condition is false
5755
5756 // Can we convert the 'select' to CF ?
5757 if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
34
Assuming the condition is false
35
Taking false branch
5758 return false;
5759
5760 TargetLowering::SelectSupportKind SelectKind;
5761 if (VectorCond)
36
Taking false branch
5762 SelectKind = TargetLowering::VectorMaskSelect;
5763 else if (SI->getType()->isVectorTy())
37
Taking false branch
5764 SelectKind = TargetLowering::ScalarCondVectorVal;
5765 else
5766 SelectKind = TargetLowering::ScalarValSelect;
5767
5768 if (TLI->isSelectSupported(SelectKind) &&
38
Assuming the condition is false
5769 !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5770 return false;
5771
5772 ModifiedDT = true;
5773
5774 // Transform a sequence like this:
5775 // start:
5776 // %cmp = cmp uge i32 %a, %b
5777 // %sel = select i1 %cmp, i32 %c, i32 %d
5778 //
5779 // Into:
5780 // start:
5781 // %cmp = cmp uge i32 %a, %b
5782 // br i1 %cmp, label %select.true, label %select.false
5783 // select.true:
5784 // br label %select.end
5785 // select.false:
5786 // br label %select.end
5787 // select.end:
5788 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5789 //
5790 // In addition, we may sink instructions that produce %c or %d from
5791 // the entry block into the destination(s) of the new branch.
5792 // If the true or false blocks do not contain a sunken instruction, that
5793 // block and its branch may be optimized away. In that case, one side of the
5794 // first branch will point directly to select.end, and the corresponding PHI
5795 // predecessor block will be the start block.
5796
5797 // First, we split the block containing the select into 2 blocks.
5798 BasicBlock *StartBlock = SI->getParent();
5799 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5800 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5801
5802 // Delete the unconditional branch that was just created by the split.
5803 StartBlock->getTerminator()->eraseFromParent();
5804
5805 // These are the new basic blocks for the conditional branch.
5806 // At least one will become an actual new basic block.
5807 BasicBlock *TrueBlock = nullptr;
5808 BasicBlock *FalseBlock = nullptr;
5809 BranchInst *TrueBranch = nullptr;
5810 BranchInst *FalseBranch = nullptr;
5811
5812 // Sink expensive instructions into the conditional blocks to avoid executing
5813 // them speculatively.
5814 for (SelectInst *SI : ASI) {
39
Assuming '__begin1' is equal to '__end1'
5815 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5816 if (TrueBlock == nullptr) {
5817 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5818 EndBlock->getParent(), EndBlock);
5819 TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5820 TrueBranch->setDebugLoc(SI->getDebugLoc());
5821 }
5822 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5823 TrueInst->moveBefore(TrueBranch);
5824 }
5825 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5826 if (FalseBlock == nullptr) {
5827 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5828 EndBlock->getParent(), EndBlock);
5829 FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5830 FalseBranch->setDebugLoc(SI->getDebugLoc());
5831 }
5832 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5833 FalseInst->moveBefore(FalseBranch);
5834 }
5835 }
5836
5837 // If there was nothing to sink, then arbitrarily choose the 'false' side
5838 // for a new input value to the PHI.
5839 if (TrueBlock == FalseBlock) {
40
Taking true branch
5840 assert(TrueBlock == nullptr &&((TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? static_cast<void> (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5841, __PRETTY_FUNCTION__))
41
'?' condition is true
5841 "Unexpected basic block transform while optimizing select")((TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? static_cast<void> (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/CodeGen/CodeGenPrepare.cpp"
, 5841, __PRETTY_FUNCTION__))
;
5842