Bug Summary

File:build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/CodeGen/CodeGenPrepare.cpp
Warning:line 2386, column 10
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name CodeGenPrepare.cpp -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 -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm -resource-dir /usr/lib/llvm-16/lib/clang/16.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/CodeGen -I /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/CodeGen -I include -I /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-16/lib/clang/16.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fmacro-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fcoverage-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -Wno-misleading-indentation -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/build-llvm=build-llvm -fdebug-prefix-map=/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2022-09-04-125545-48738-1 -x c++ /build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/CodeGen/CodeGenPrepare.cpp

/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/lib/CodeGen/CodeGenPrepare.cpp

1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass munges the code in the input function to better prepare it for
10// SelectionDAG-based code generation. This works around limitations in it's
11// basic-block-at-a-time approach. It should eventually be removed.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/ArrayRef.h"
17#include "llvm/ADT/DenseMap.h"
18#include "llvm/ADT/MapVector.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/InstructionSimplify.h"
27#include "llvm/Analysis/LoopInfo.h"
28#include "llvm/Analysis/ProfileSummaryInfo.h"
29#include "llvm/Analysis/TargetLibraryInfo.h"
30#include "llvm/Analysis/TargetTransformInfo.h"
31#include "llvm/Analysis/ValueTracking.h"
32#include "llvm/Analysis/VectorUtils.h"
33#include "llvm/CodeGen/Analysis.h"
34#include "llvm/CodeGen/BasicBlockSectionsProfileReader.h"
35#include "llvm/CodeGen/ISDOpcodes.h"
36#include "llvm/CodeGen/SelectionDAGNodes.h"
37#include "llvm/CodeGen/TargetLowering.h"
38#include "llvm/CodeGen/TargetPassConfig.h"
39#include "llvm/CodeGen/TargetSubtargetInfo.h"
40#include "llvm/CodeGen/ValueTypes.h"
41#include "llvm/Config/llvm-config.h"
42#include "llvm/IR/Argument.h"
43#include "llvm/IR/Attributes.h"
44#include "llvm/IR/BasicBlock.h"
45#include "llvm/IR/Constant.h"
46#include "llvm/IR/Constants.h"
47#include "llvm/IR/DataLayout.h"
48#include "llvm/IR/DebugInfo.h"
49#include "llvm/IR/DerivedTypes.h"
50#include "llvm/IR/Dominators.h"
51#include "llvm/IR/Function.h"
52#include "llvm/IR/GetElementPtrTypeIterator.h"
53#include "llvm/IR/GlobalValue.h"
54#include "llvm/IR/GlobalVariable.h"
55#include "llvm/IR/IRBuilder.h"
56#include "llvm/IR/InlineAsm.h"
57#include "llvm/IR/InstrTypes.h"
58#include "llvm/IR/Instruction.h"
59#include "llvm/IR/Instructions.h"
60#include "llvm/IR/IntrinsicInst.h"
61#include "llvm/IR/Intrinsics.h"
62#include "llvm/IR/IntrinsicsAArch64.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/ProfDataUtils.h"
69#include "llvm/IR/Statepoint.h"
70#include "llvm/IR/Type.h"
71#include "llvm/IR/Use.h"
72#include "llvm/IR/User.h"
73#include "llvm/IR/Value.h"
74#include "llvm/IR/ValueHandle.h"
75#include "llvm/IR/ValueMap.h"
76#include "llvm/InitializePasses.h"
77#include "llvm/Pass.h"
78#include "llvm/Support/BlockFrequency.h"
79#include "llvm/Support/BranchProbability.h"
80#include "llvm/Support/Casting.h"
81#include "llvm/Support/CommandLine.h"
82#include "llvm/Support/Compiler.h"
83#include "llvm/Support/Debug.h"
84#include "llvm/Support/ErrorHandling.h"
85#include "llvm/Support/MachineValueType.h"
86#include "llvm/Support/MathExtras.h"
87#include "llvm/Support/raw_ostream.h"
88#include "llvm/Target/TargetMachine.h"
89#include "llvm/Target/TargetOptions.h"
90#include "llvm/Transforms/Utils/BasicBlockUtils.h"
91#include "llvm/Transforms/Utils/BypassSlowDivision.h"
92#include "llvm/Transforms/Utils/Local.h"
93#include "llvm/Transforms/Utils/SimplifyLibCalls.h"
94#include "llvm/Transforms/Utils/SizeOpts.h"
95#include <algorithm>
96#include <cassert>
97#include <cstdint>
98#include <iterator>
99#include <limits>
100#include <memory>
101#include <utility>
102#include <vector>
103
104using namespace llvm;
105using namespace llvm::PatternMatch;
106
107#define DEBUG_TYPE"codegenprepare" "codegenprepare"
108
109STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim"
, "Number of blocks eliminated"}
;
110STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim"
, "Number of trivial PHIs eliminated"}
;
111STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim"
, "Number of GEPs converted to casts"}
;
112STATISTIC(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"
}
113 "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses"
, "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps"
}
;
114STATISTIC(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"
}
115 "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses"
, "Number of uses of Cast expressions replaced with uses " "of sunken Casts"
}
;
116STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
}
117 "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts"
, "Number of memory instructions whose address " "computations were sunk"
}
;
118STATISTIC(NumMemoryInstsPhiCreated,static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions"}
119 "Number of phis created when address "static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions"}
120 "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare"
, "NumMemoryInstsPhiCreated", "Number of phis created when address "
"computations were sunk to memory instructions"}
;
121STATISTIC(NumMemoryInstsSelectCreated,static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions"}
122 "Number of select created when address "static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions"}
123 "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare"
, "NumMemoryInstsSelectCreated", "Number of select created when address "
"computations were sunk to memory instructions"}
;
124STATISTIC(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"}
;
125STATISTIC(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"}
;
126STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded"
, "Number of and mask instructions added to form ext loads"}
127 "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"}
;
128STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized")static llvm::Statistic NumAndUses = {"codegenprepare", "NumAndUses"
, "Number of uses of and mask instructions optimized"}
;
129STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup"
, "Number of return instructions duplicated"}
;
130STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved"
, "Number of debug value instructions moved"}
;
131STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare"
, "NumSelectsExpanded", "Number of selects turned into branches"
}
;
132STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare"
, "NumStoreExtractExposed", "Number of store(extractelement) exposed"
}
;
133
134static cl::opt<bool> DisableBranchOpts(
135 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
136 cl::desc("Disable branch optimizations in CodeGenPrepare"));
137
138static cl::opt<bool>
139 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
140 cl::desc("Disable GC optimizations in CodeGenPrepare"));
141
142static cl::opt<bool>
143 DisableSelectToBranch("disable-cgp-select2branch", cl::Hidden,
144 cl::init(false),
145 cl::desc("Disable select to branch conversion."));
146
147static cl::opt<bool>
148 AddrSinkUsingGEPs("addr-sink-using-gep", cl::Hidden, cl::init(true),
149 cl::desc("Address sinking in CGP using GEPs."));
150
151static cl::opt<bool>
152 EnableAndCmpSinking("enable-andcmp-sinking", cl::Hidden, cl::init(true),
153 cl::desc("Enable sinkinig and/cmp into branches."));
154
155static cl::opt<bool> DisableStoreExtract(
156 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
157 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
158
159static cl::opt<bool> StressStoreExtract(
160 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
161 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
162
163static cl::opt<bool> DisableExtLdPromotion(
164 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
165 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
166 "CodeGenPrepare"));
167
168static cl::opt<bool> StressExtLdPromotion(
169 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
170 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
171 "optimization in CodeGenPrepare"));
172
173static cl::opt<bool> DisablePreheaderProtect(
174 "disable-preheader-prot", cl::Hidden, cl::init(false),
175 cl::desc("Disable protection against removing loop preheaders"));
176
177static cl::opt<bool> ProfileGuidedSectionPrefix(
178 "profile-guided-section-prefix", cl::Hidden, cl::init(true),
179 cl::desc("Use profile info to add section prefix for hot/cold functions"));
180
181static cl::opt<bool> ProfileUnknownInSpecialSection(
182 "profile-unknown-in-special-section", cl::Hidden,
183 cl::desc("In profiling mode like sampleFDO, if a function doesn't have "
184 "profile, we cannot tell the function is cold for sure because "
185 "it may be a function newly added without ever being sampled. "
186 "With the flag enabled, compiler can put such profile unknown "
187 "functions into a special section, so runtime system can choose "
188 "to handle it in a different way than .text section, to save "
189 "RAM for example. "));
190
191static cl::opt<bool> BBSectionsGuidedSectionPrefix(
192 "bbsections-guided-section-prefix", cl::Hidden, cl::init(true),
193 cl::desc("Use the basic-block-sections profile to determine the text "
194 "section prefix for hot functions. Functions with "
195 "basic-block-sections profile will be placed in `.text.hot` "
196 "regardless of their FDO profile info. Other functions won't be "
197 "impacted, i.e., their prefixes will be decided by FDO/sampleFDO "
198 "profiles."));
199
200static cl::opt<unsigned> FreqRatioToSkipMerge(
201 "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
202 cl::desc("Skip merging empty blocks if (frequency of empty block) / "
203 "(frequency of destination block) is greater than this ratio"));
204
205static cl::opt<bool> ForceSplitStore(
206 "force-split-store", cl::Hidden, cl::init(false),
207 cl::desc("Force store splitting no matter what the target query says."));
208
209static cl::opt<bool> EnableTypePromotionMerge(
210 "cgp-type-promotion-merge", cl::Hidden,
211 cl::desc("Enable merging of redundant sexts when one is dominating"
212 " the other."),
213 cl::init(true));
214
215static cl::opt<bool> DisableComplexAddrModes(
216 "disable-complex-addr-modes", cl::Hidden, cl::init(false),
217 cl::desc("Disables combining addressing modes with different parts "
218 "in optimizeMemoryInst."));
219
220static cl::opt<bool>
221 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
222 cl::desc("Allow creation of Phis in Address sinking."));
223
224static cl::opt<bool> AddrSinkNewSelects(
225 "addr-sink-new-select", cl::Hidden, cl::init(true),
226 cl::desc("Allow creation of selects in Address sinking."));
227
228static cl::opt<bool> AddrSinkCombineBaseReg(
229 "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
230 cl::desc("Allow combining of BaseReg field in Address sinking."));
231
232static cl::opt<bool> AddrSinkCombineBaseGV(
233 "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
234 cl::desc("Allow combining of BaseGV field in Address sinking."));
235
236static cl::opt<bool> AddrSinkCombineBaseOffs(
237 "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
238 cl::desc("Allow combining of BaseOffs field in Address sinking."));
239
240static cl::opt<bool> AddrSinkCombineScaledReg(
241 "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
242 cl::desc("Allow combining of ScaledReg field in Address sinking."));
243
244static cl::opt<bool>
245 EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
246 cl::init(true),
247 cl::desc("Enable splitting large offset of GEP."));
248
249static cl::opt<bool> EnableICMP_EQToICMP_ST(
250 "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(false),
251 cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion."));
252
253static cl::opt<bool>
254 VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(false),
255 cl::desc("Enable BFI update verification for "
256 "CodeGenPrepare."));
257
258static cl::opt<bool>
259 OptimizePhiTypes("cgp-optimize-phi-types", cl::Hidden, cl::init(false),
260 cl::desc("Enable converting phi types in CodeGenPrepare"));
261
262namespace {
263
264enum ExtType {
265 ZeroExtension, // Zero extension has been seen.
266 SignExtension, // Sign extension has been seen.
267 BothExtension // This extension type is used if we saw sext after
268 // ZeroExtension had been set, or if we saw zext after
269 // SignExtension had been set. It makes the type
270 // information of a promoted instruction invalid.
271};
272
273using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
274using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
275using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
276using SExts = SmallVector<Instruction *, 16>;
277using ValueToSExts = DenseMap<Value *, SExts>;
278
279class TypePromotionTransaction;
280
281class CodeGenPrepare : public FunctionPass {
282 const TargetMachine *TM = nullptr;
283 const TargetSubtargetInfo *SubtargetInfo;
284 const TargetLowering *TLI = nullptr;
285 const TargetRegisterInfo *TRI;
286 const TargetTransformInfo *TTI = nullptr;
287 const BasicBlockSectionsProfileReader *BBSectionsProfileReader = nullptr;
288 const TargetLibraryInfo *TLInfo;
289 const LoopInfo *LI;
290 std::unique_ptr<BlockFrequencyInfo> BFI;
291 std::unique_ptr<BranchProbabilityInfo> BPI;
292 ProfileSummaryInfo *PSI;
293
294 /// As we scan instructions optimizing them, this is the next instruction
295 /// to optimize. Transforms that can invalidate this should update it.
296 BasicBlock::iterator CurInstIterator;
297
298 /// Keeps track of non-local addresses that have been sunk into a block.
299 /// This allows us to avoid inserting duplicate code for blocks with
300 /// multiple load/stores of the same address. The usage of WeakTrackingVH
301 /// enables SunkAddrs to be treated as a cache whose entries can be
302 /// invalidated if a sunken address computation has been erased.
303 ValueMap<Value *, WeakTrackingVH> SunkAddrs;
304
305 /// Keeps track of all instructions inserted for the current function.
306 SetOfInstrs InsertedInsts;
307
308 /// Keeps track of the type of the related instruction before their
309 /// promotion for the current function.
310 InstrToOrigTy PromotedInsts;
311
312 /// Keep track of instructions removed during promotion.
313 SetOfInstrs RemovedInsts;
314
315 /// Keep track of sext chains based on their initial value.
316 DenseMap<Value *, Instruction *> SeenChainsForSExt;
317
318 /// Keep track of GEPs accessing the same data structures such as structs or
319 /// arrays that are candidates to be split later because of their large
320 /// size.
321 MapVector<AssertingVH<Value>,
322 SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>>
323 LargeOffsetGEPMap;
324
325 /// Keep track of new GEP base after splitting the GEPs having large offset.
326 SmallSet<AssertingVH<Value>, 2> NewGEPBases;
327
328 /// Map serial numbers to Large offset GEPs.
329 DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
330
331 /// Keep track of SExt promoted.
332 ValueToSExts ValToSExtendedUses;
333
334 /// True if the function has the OptSize attribute.
335 bool OptSize;
336
337 /// DataLayout for the Function being processed.
338 const DataLayout *DL = nullptr;
339
340 /// Building the dominator tree can be expensive, so we only build it
341 /// lazily and update it when required.
342 std::unique_ptr<DominatorTree> DT;
343
344public:
345 static char ID; // Pass identification, replacement for typeid
346
347 CodeGenPrepare() : FunctionPass(ID) {
348 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
349 }
350
351 bool runOnFunction(Function &F) override;
352
353 StringRef getPassName() const override { return "CodeGen Prepare"; }
354
355 void getAnalysisUsage(AnalysisUsage &AU) const override {
356 // FIXME: When we can selectively preserve passes, preserve the domtree.
357 AU.addRequired<ProfileSummaryInfoWrapperPass>();
358 AU.addRequired<TargetLibraryInfoWrapperPass>();
359 AU.addRequired<TargetPassConfig>();
360 AU.addRequired<TargetTransformInfoWrapperPass>();
361 AU.addRequired<LoopInfoWrapperPass>();
362 AU.addUsedIfAvailable<BasicBlockSectionsProfileReader>();
363 }
364
365private:
366 template <typename F>
367 void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
368 // Substituting can cause recursive simplifications, which can invalidate
369 // our iterator. Use a WeakTrackingVH to hold onto it in case this
370 // happens.
371 Value *CurValue = &*CurInstIterator;
372 WeakTrackingVH IterHandle(CurValue);
373
374 f();
375
376 // If the iterator instruction was recursively deleted, start over at the
377 // start of the block.
378 if (IterHandle != CurValue) {
379 CurInstIterator = BB->begin();
380 SunkAddrs.clear();
381 }
382 }
383
384 // Get the DominatorTree, building if necessary.
385 DominatorTree &getDT(Function &F) {
386 if (!DT)
387 DT = std::make_unique<DominatorTree>(F);
388 return *DT;
389 }
390
391 void removeAllAssertingVHReferences(Value *V);
392 bool eliminateAssumptions(Function &F);
393 bool eliminateFallThrough(Function &F);
394 bool eliminateMostlyEmptyBlocks(Function &F);
395 BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
396 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
397 void eliminateMostlyEmptyBlock(BasicBlock *BB);
398 bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
399 bool isPreheader);
400 bool makeBitReverse(Instruction &I);
401 bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
402 bool optimizeInst(Instruction *I, bool &ModifiedDT);
403 bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, Type *AccessTy,
404 unsigned AddrSpace);
405 bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
406 bool optimizeInlineAsmInst(CallInst *CS);
407 bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
408 bool optimizeExt(Instruction *&I);
409 bool optimizeExtUses(Instruction *I);
410 bool optimizeLoadExt(LoadInst *Load);
411 bool optimizeShiftInst(BinaryOperator *BO);
412 bool optimizeFunnelShift(IntrinsicInst *Fsh);
413 bool optimizeSelectInst(SelectInst *SI);
414 bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
415 bool optimizeSwitchType(SwitchInst *SI);
416 bool optimizeSwitchPhiConstants(SwitchInst *SI);
417 bool optimizeSwitchInst(SwitchInst *SI);
418 bool optimizeExtractElementInst(Instruction *Inst);
419 bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT);
420 bool fixupDbgValue(Instruction *I);
421 bool placeDbgValues(Function &F);
422 bool placePseudoProbes(Function &F);
423 bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
424 LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
425 bool tryToPromoteExts(TypePromotionTransaction &TPT,
426 const SmallVectorImpl<Instruction *> &Exts,
427 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
428 unsigned CreatedInstsCost = 0);
429 bool mergeSExts(Function &F);
430 bool splitLargeGEPOffsets();
431 bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited,
432 SmallPtrSetImpl<Instruction *> &DeletedInstrs);
433 bool optimizePhiTypes(Function &F);
434 bool performAddressTypePromotion(
435 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
436 bool HasPromoted, TypePromotionTransaction &TPT,
437 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
438 bool splitBranchCondition(Function &F, bool &ModifiedDT);
439 bool simplifyOffsetableRelocate(GCStatepointInst &I);
440
441 bool tryToSinkFreeOperands(Instruction *I);
442 bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, Value *Arg1,
443 CmpInst *Cmp, Intrinsic::ID IID);
444 bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT);
445 bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
446 bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
447 void verifyBFIUpdates(Function &F);
448};
449
450} // end anonymous namespace
451
452char CodeGenPrepare::ID = 0;
453
454INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
455 "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry &
Registry) {
456INITIALIZE_PASS_DEPENDENCY(BasicBlockSectionsProfileReader)initializeBasicBlockSectionsProfileReaderPass(Registry);
457INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry);
458INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry);
459INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
460INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)initializeTargetPassConfigPass(Registry);
461INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
462INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE, "Optimize for code generation",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)); }
463 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)); }
464
465FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
466
467bool CodeGenPrepare::runOnFunction(Function &F) {
468 if (skipFunction(F))
469 return false;
470
471 DL = &F.getParent()->getDataLayout();
472
473 bool EverMadeChange = false;
474 // Clear per function information.
475 InsertedInsts.clear();
476 PromotedInsts.clear();
477
478 TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
479 SubtargetInfo = TM->getSubtargetImpl(F);
480 TLI = SubtargetInfo->getTargetLowering();
481 TRI = SubtargetInfo->getRegisterInfo();
482 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
483 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
484 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
485 BPI.reset(new BranchProbabilityInfo(F, *LI));
486 BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
487 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
488 BBSectionsProfileReader =
489 getAnalysisIfAvailable<BasicBlockSectionsProfileReader>();
490 OptSize = F.hasOptSize();
491 // Use the basic-block-sections profile to promote hot functions to .text.hot
492 // if requested.
493 if (BBSectionsGuidedSectionPrefix && BBSectionsProfileReader &&
494 BBSectionsProfileReader->isFunctionHot(F.getName())) {
495 F.setSectionPrefix("hot");
496 } else if (ProfileGuidedSectionPrefix) {
497 // The hot attribute overwrites profile count based hotness while profile
498 // counts based hotness overwrite the cold attribute.
499 // This is a conservative behabvior.
500 if (F.hasFnAttribute(Attribute::Hot) ||
501 PSI->isFunctionHotInCallGraph(&F, *BFI))
502 F.setSectionPrefix("hot");
503 // If PSI shows this function is not hot, we will placed the function
504 // into unlikely section if (1) PSI shows this is a cold function, or
505 // (2) the function has a attribute of cold.
506 else if (PSI->isFunctionColdInCallGraph(&F, *BFI) ||
507 F.hasFnAttribute(Attribute::Cold))
508 F.setSectionPrefix("unlikely");
509 else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() &&
510 PSI->isFunctionHotnessUnknown(F))
511 F.setSectionPrefix("unknown");
512 }
513
514 /// This optimization identifies DIV instructions that can be
515 /// profitably bypassed and carried out with a shorter, faster divide.
516 if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) {
517 const DenseMap<unsigned int, unsigned int> &BypassWidths =
518 TLI->getBypassSlowDivWidths();
519 BasicBlock *BB = &*F.begin();
520 while (BB != nullptr) {
521 // bypassSlowDivision may create new BBs, but we don't want to reapply the
522 // optimization to those blocks.
523 BasicBlock *Next = BB->getNextNode();
524 // F.hasOptSize is already checked in the outer if statement.
525 if (!llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
526 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
527 BB = Next;
528 }
529 }
530
531 // Get rid of @llvm.assume builtins before attempting to eliminate empty
532 // blocks, since there might be blocks that only contain @llvm.assume calls
533 // (plus arguments that we can get rid of).
534 EverMadeChange |= eliminateAssumptions(F);
535
536 // Eliminate blocks that contain only PHI nodes and an
537 // unconditional branch.
538 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
539
540 bool ModifiedDT = false;
541 if (!DisableBranchOpts)
542 EverMadeChange |= splitBranchCondition(F, ModifiedDT);
543
544 // Split some critical edges where one of the sources is an indirect branch,
545 // to help generate sane code for PHIs involving such edges.
546 EverMadeChange |=
547 SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/true);
548
549 bool MadeChange = true;
550 while (MadeChange) {
551 MadeChange = false;
552 DT.reset();
553 for (BasicBlock &BB : llvm::make_early_inc_range(F)) {
554 bool ModifiedDTOnIteration = false;
555 MadeChange |= optimizeBlock(BB, ModifiedDTOnIteration);
556
557 // Restart BB iteration if the dominator tree of the Function was changed
558 if (ModifiedDTOnIteration)
559 break;
560 }
561 if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
562 MadeChange |= mergeSExts(F);
563 if (!LargeOffsetGEPMap.empty())
564 MadeChange |= splitLargeGEPOffsets();
565 MadeChange |= optimizePhiTypes(F);
566
567 if (MadeChange)
568 eliminateFallThrough(F);
569
570 // Really free removed instructions during promotion.
571 for (Instruction *I : RemovedInsts)
572 I->deleteValue();
573
574 EverMadeChange |= MadeChange;
575 SeenChainsForSExt.clear();
576 ValToSExtendedUses.clear();
577 RemovedInsts.clear();
578 LargeOffsetGEPMap.clear();
579 LargeOffsetGEPID.clear();
580 }
581
582 NewGEPBases.clear();
583 SunkAddrs.clear();
584
585 if (!DisableBranchOpts) {
586 MadeChange = false;
587 // Use a set vector to get deterministic iteration order. The order the
588 // blocks are removed may affect whether or not PHI nodes in successors
589 // are removed.
590 SmallSetVector<BasicBlock *, 8> WorkList;
591 for (BasicBlock &BB : F) {
592 SmallVector<BasicBlock *, 2> Successors(successors(&BB));
593 MadeChange |= ConstantFoldTerminator(&BB, true);
594 if (!MadeChange)
595 continue;
596
597 for (BasicBlock *Succ : Successors)
598 if (pred_empty(Succ))
599 WorkList.insert(Succ);
600 }
601
602 // Delete the dead blocks and any of their dead successors.
603 MadeChange |= !WorkList.empty();
604 while (!WorkList.empty()) {
605 BasicBlock *BB = WorkList.pop_back_val();
606 SmallVector<BasicBlock *, 2> Successors(successors(BB));
607
608 DeleteDeadBlock(BB);
609
610 for (BasicBlock *Succ : Successors)
611 if (pred_empty(Succ))
612 WorkList.insert(Succ);
613 }
614
615 // Merge pairs of basic blocks with unconditional branches, connected by
616 // a single edge.
617 if (EverMadeChange || MadeChange)
618 MadeChange |= eliminateFallThrough(F);
619
620 EverMadeChange |= MadeChange;
621 }
622
623 if (!DisableGCOpts) {
624 SmallVector<GCStatepointInst *, 2> Statepoints;
625 for (BasicBlock &BB : F)
626 for (Instruction &I : BB)
627 if (auto *SP = dyn_cast<GCStatepointInst>(&I))
628 Statepoints.push_back(SP);
629 for (auto &I : Statepoints)
630 EverMadeChange |= simplifyOffsetableRelocate(*I);
631 }
632
633 // Do this last to clean up use-before-def scenarios introduced by other
634 // preparatory transforms.
635 EverMadeChange |= placeDbgValues(F);
636 EverMadeChange |= placePseudoProbes(F);
637
638#ifndef NDEBUG
639 if (VerifyBFIUpdates)
640 verifyBFIUpdates(F);
641#endif
642
643 return EverMadeChange;
644}
645
646bool CodeGenPrepare::eliminateAssumptions(Function &F) {
647 bool MadeChange = false;
648 for (BasicBlock &BB : F) {
649 CurInstIterator = BB.begin();
650 while (CurInstIterator != BB.end()) {
651 Instruction *I = &*(CurInstIterator++);
652 if (auto *Assume = dyn_cast<AssumeInst>(I)) {
653 MadeChange = true;
654 Value *Operand = Assume->getOperand(0);
655 Assume->eraseFromParent();
656
657 resetIteratorIfInvalidatedWhileCalling(&BB, [&]() {
658 RecursivelyDeleteTriviallyDeadInstructions(Operand, TLInfo, nullptr);
659 });
660 }
661 }
662 }
663 return MadeChange;
664}
665
666/// An instruction is about to be deleted, so remove all references to it in our
667/// GEP-tracking data strcutures.
668void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) {
669 LargeOffsetGEPMap.erase(V);
670 NewGEPBases.erase(V);
671
672 auto GEP = dyn_cast<GetElementPtrInst>(V);
673 if (!GEP)
674 return;
675
676 LargeOffsetGEPID.erase(GEP);
677
678 auto VecI = LargeOffsetGEPMap.find(GEP->getPointerOperand());
679 if (VecI == LargeOffsetGEPMap.end())
680 return;
681
682 auto &GEPVector = VecI->second;
683 llvm::erase_if(GEPVector, [=](auto &Elt) { return Elt.first == GEP; });
684
685 if (GEPVector.empty())
686 LargeOffsetGEPMap.erase(VecI);
687}
688
689// Verify BFI has been updated correctly by recomputing BFI and comparing them.
690void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) CodeGenPrepare::verifyBFIUpdates(Function &F) {
691 DominatorTree NewDT(F);
692 LoopInfo NewLI(NewDT);
693 BranchProbabilityInfo NewBPI(F, NewLI, TLInfo);
694 BlockFrequencyInfo NewBFI(F, NewBPI, NewLI);
695 NewBFI.verifyMatch(*BFI);
696}
697
698/// Merge basic blocks which are connected by a single edge, where one of the
699/// basic blocks has a single successor pointing to the other basic block,
700/// which has a single predecessor.
701bool CodeGenPrepare::eliminateFallThrough(Function &F) {
702 bool Changed = false;
703 // Scan all of the blocks in the function, except for the entry block.
704 // Use a temporary array to avoid iterator being invalidated when
705 // deleting blocks.
706 SmallVector<WeakTrackingVH, 16> Blocks;
707 for (auto &Block : llvm::drop_begin(F))
708 Blocks.push_back(&Block);
709
710 SmallSet<WeakTrackingVH, 16> Preds;
711 for (auto &Block : Blocks) {
712 auto *BB = cast_or_null<BasicBlock>(Block);
713 if (!BB)
714 continue;
715 // If the destination block has a single pred, then this is a trivial
716 // edge, just collapse it.
717 BasicBlock *SinglePred = BB->getSinglePredecessor();
718
719 // Don't merge if BB's address is taken.
720 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken())
721 continue;
722
723 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
724 if (Term && !Term->isConditional()) {
725 Changed = true;
726 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)
;
727
728 // Merge BB into SinglePred and delete it.
729 MergeBlockIntoPredecessor(BB);
730 Preds.insert(SinglePred);
731 }
732 }
733
734 // (Repeatedly) merging blocks into their predecessors can create redundant
735 // debug intrinsics.
736 for (const auto &Pred : Preds)
737 if (auto *BB = cast_or_null<BasicBlock>(Pred))
738 RemoveRedundantDbgInstrs(BB);
739
740 return Changed;
741}
742
743/// Find a destination block from BB if BB is mergeable empty block.
744BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
745 // If this block doesn't end with an uncond branch, ignore it.
746 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
747 if (!BI || !BI->isUnconditional())
748 return nullptr;
749
750 // If the instruction before the branch (skipping debug info) isn't a phi
751 // node, then other stuff is happening here.
752 BasicBlock::iterator BBI = BI->getIterator();
753 if (BBI != BB->begin()) {
754 --BBI;
755 while (isa<DbgInfoIntrinsic>(BBI)) {
756 if (BBI == BB->begin())
757 break;
758 --BBI;
759 }
760 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
761 return nullptr;
762 }
763
764 // Do not break infinite loops.
765 BasicBlock *DestBB = BI->getSuccessor(0);
766 if (DestBB == BB)
767 return nullptr;
768
769 if (!canMergeBlocks(BB, DestBB))
770 DestBB = nullptr;
771
772 return DestBB;
773}
774
775/// Eliminate blocks that contain only PHI nodes, debug info directives, and an
776/// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
777/// edges in ways that are non-optimal for isel. Start by eliminating these
778/// blocks so we can split them the way we want them.
779bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
780 SmallPtrSet<BasicBlock *, 16> Preheaders;
781 SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
782 while (!LoopList.empty()) {
783 Loop *L = LoopList.pop_back_val();
784 llvm::append_range(LoopList, *L);
785 if (BasicBlock *Preheader = L->getLoopPreheader())
786 Preheaders.insert(Preheader);
787 }
788
789 bool MadeChange = false;
790 // Copy blocks into a temporary array to avoid iterator invalidation issues
791 // as we remove them.
792 // Note that this intentionally skips the entry block.
793 SmallVector<WeakTrackingVH, 16> Blocks;
794 for (auto &Block : llvm::drop_begin(F))
795 Blocks.push_back(&Block);
796
797 for (auto &Block : Blocks) {
798 BasicBlock *BB = cast_or_null<BasicBlock>(Block);
799 if (!BB)
800 continue;
801 BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
802 if (!DestBB ||
803 !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
804 continue;
805
806 eliminateMostlyEmptyBlock(BB);
807 MadeChange = true;
808 }
809 return MadeChange;
810}
811
812bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
813 BasicBlock *DestBB,
814 bool isPreheader) {
815 // Do not delete loop preheaders if doing so would create a critical edge.
816 // Loop preheaders can be good locations to spill registers. If the
817 // preheader is deleted and we create a critical edge, registers may be
818 // spilled in the loop body instead.
819 if (!DisablePreheaderProtect && isPreheader &&
820 !(BB->getSinglePredecessor() &&
821 BB->getSinglePredecessor()->getSingleSuccessor()))
822 return false;
823
824 // Skip merging if the block's successor is also a successor to any callbr
825 // that leads to this block.
826 // FIXME: Is this really needed? Is this a correctness issue?
827 for (BasicBlock *Pred : predecessors(BB)) {
828 if (auto *CBI = dyn_cast<CallBrInst>((Pred)->getTerminator()))
829 for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i)
830 if (DestBB == CBI->getSuccessor(i))
831 return false;
832 }
833
834 // Try to skip merging if the unique predecessor of BB is terminated by a
835 // switch or indirect branch instruction, and BB is used as an incoming block
836 // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
837 // add COPY instructions in the predecessor of BB instead of BB (if it is not
838 // merged). Note that the critical edge created by merging such blocks wont be
839 // split in MachineSink because the jump table is not analyzable. By keeping
840 // such empty block (BB), ISel will place COPY instructions in BB, not in the
841 // predecessor of BB.
842 BasicBlock *Pred = BB->getUniquePredecessor();
843 if (!Pred || !(isa<SwitchInst>(Pred->getTerminator()) ||
844 isa<IndirectBrInst>(Pred->getTerminator())))
845 return true;
846
847 if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
848 return true;
849
850 // We use a simple cost heuristic which determine skipping merging is
851 // profitable if the cost of skipping merging is less than the cost of
852 // merging : Cost(skipping merging) < Cost(merging BB), where the
853 // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
854 // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
855 // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
856 // Freq(Pred) / Freq(BB) > 2.
857 // Note that if there are multiple empty blocks sharing the same incoming
858 // value for the PHIs in the DestBB, we consider them together. In such
859 // case, Cost(merging BB) will be the sum of their frequencies.
860
861 if (!isa<PHINode>(DestBB->begin()))
862 return true;
863
864 SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
865
866 // Find all other incoming blocks from which incoming values of all PHIs in
867 // DestBB are the same as the ones from BB.
868 for (BasicBlock *DestBBPred : predecessors(DestBB)) {
869 if (DestBBPred == BB)
870 continue;
871
872 if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
873 return DestPN.getIncomingValueForBlock(BB) ==
874 DestPN.getIncomingValueForBlock(DestBBPred);
875 }))
876 SameIncomingValueBBs.insert(DestBBPred);
877 }
878
879 // See if all BB's incoming values are same as the value from Pred. In this
880 // case, no reason to skip merging because COPYs are expected to be place in
881 // Pred already.
882 if (SameIncomingValueBBs.count(Pred))
883 return true;
884
885 BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
886 BlockFrequency BBFreq = BFI->getBlockFreq(BB);
887
888 for (auto *SameValueBB : SameIncomingValueBBs)
889 if (SameValueBB->getUniquePredecessor() == Pred &&
890 DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
891 BBFreq += BFI->getBlockFreq(SameValueBB);
892
893 return PredFreq.getFrequency() <=
894 BBFreq.getFrequency() * FreqRatioToSkipMerge;
895}
896
897/// Return true if we can merge BB into DestBB if there is a single
898/// unconditional branch between them, and BB contains no other non-phi
899/// instructions.
900bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
901 const BasicBlock *DestBB) const {
902 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
903 // the successor. If there are more complex condition (e.g. preheaders),
904 // don't mess around with them.
905 for (const PHINode &PN : BB->phis()) {
906 for (const User *U : PN.users()) {
907 const Instruction *UI = cast<Instruction>(U);
908 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
909 return false;
910 // If User is inside DestBB block and it is a PHINode then check
911 // incoming value. If incoming value is not from BB then this is
912 // a complex condition (e.g. preheaders) we want to avoid here.
913 if (UI->getParent() == DestBB) {
914 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
915 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
916 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
917 if (Insn && Insn->getParent() == BB &&
918 Insn->getParent() != UPN->getIncomingBlock(I))
919 return false;
920 }
921 }
922 }
923 }
924
925 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
926 // and DestBB may have conflicting incoming values for the block. If so, we
927 // can't merge the block.
928 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
929 if (!DestBBPN)
930 return true; // no conflict.
931
932 // Collect the preds of BB.
933 SmallPtrSet<const BasicBlock *, 16> BBPreds;
934 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
935 // It is faster to get preds from a PHI than with pred_iterator.
936 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
937 BBPreds.insert(BBPN->getIncomingBlock(i));
938 } else {
939 BBPreds.insert(pred_begin(BB), pred_end(BB));
940 }
941
942 // Walk the preds of DestBB.
943 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
944 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
945 if (BBPreds.count(Pred)) { // Common predecessor?
946 for (const PHINode &PN : DestBB->phis()) {
947 const Value *V1 = PN.getIncomingValueForBlock(Pred);
948 const Value *V2 = PN.getIncomingValueForBlock(BB);
949
950 // If V2 is a phi node in BB, look up what the mapped value will be.
951 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
952 if (V2PN->getParent() == BB)
953 V2 = V2PN->getIncomingValueForBlock(Pred);
954
955 // If there is a conflict, bail out.
956 if (V1 != V2)
957 return false;
958 }
959 }
960 }
961
962 return true;
963}
964
965/// Eliminate a basic block that has only phi's and an unconditional branch in
966/// it.
967void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
968 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
969 BasicBlock *DestBB = BI->getSuccessor(0);
970
971 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)
972 << *BB << *DestBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
<< *BB << *DestBB; } } while (false)
;
973
974 // If the destination block has a single pred, then this is a trivial edge,
975 // just collapse it.
976 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
977 if (SinglePred != DestBB) {
978 assert(SinglePred == BB &&(static_cast <bool> (SinglePred == BB && "Single predecessor not the same as predecessor"
) ? void (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 979, __extension__ __PRETTY_FUNCTION__
))
979 "Single predecessor not the same as predecessor")(static_cast <bool> (SinglePred == BB && "Single predecessor not the same as predecessor"
) ? void (0) : __assert_fail ("SinglePred == BB && \"Single predecessor not the same as predecessor\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 979, __extension__ __PRETTY_FUNCTION__
))
;
980 // Merge DestBB into SinglePred/BB and delete it.
981 MergeBlockIntoPredecessor(DestBB);
982 // Note: BB(=SinglePred) will not be deleted on this path.
983 // DestBB(=its single successor) is the one that was deleted.
984 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)
;
985 return;
986 }
987 }
988
989 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
990 // to handle the new incoming edges it is about to have.
991 for (PHINode &PN : DestBB->phis()) {
992 // Remove the incoming value for BB, and remember it.
993 Value *InVal = PN.removeIncomingValue(BB, false);
994
995 // Two options: either the InVal is a phi node defined in BB or it is some
996 // value that dominates BB.
997 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
998 if (InValPhi && InValPhi->getParent() == BB) {
999 // Add all of the input values of the input PHI as inputs of this phi.
1000 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1001 PN.addIncoming(InValPhi->getIncomingValue(i),
1002 InValPhi->getIncomingBlock(i));
1003 } else {
1004 // Otherwise, add one instance of the dominating value for each edge that
1005 // we will be adding.
1006 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1007 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1008 PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
1009 } else {
1010 for (BasicBlock *Pred : predecessors(BB))
1011 PN.addIncoming(InVal, Pred);
1012 }
1013 }
1014 }
1015
1016 // The PHIs are now updated, change everything that refers to BB to use
1017 // DestBB and remove BB.
1018 BB->replaceAllUsesWith(DestBB);
1019 BB->eraseFromParent();
1020 ++NumBlocksElim;
1021
1022 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)
;
1023}
1024
1025// Computes a map of base pointer relocation instructions to corresponding
1026// derived pointer relocation instructions given a vector of all relocate calls
1027static void computeBaseDerivedRelocateMap(
1028 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1029 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1030 &RelocateInstMap) {
1031 // Collect information in two maps: one primarily for locating the base object
1032 // while filling the second map; the second map is the final structure holding
1033 // a mapping between Base and corresponding Derived relocate calls
1034 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1035 for (auto *ThisRelocate : AllRelocateCalls) {
1036 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1037 ThisRelocate->getDerivedPtrIndex());
1038 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1039 }
1040 for (auto &Item : RelocateIdxMap) {
1041 std::pair<unsigned, unsigned> Key = Item.first;
1042 if (Key.first == Key.second)
1043 // Base relocation: nothing to insert
1044 continue;
1045
1046 GCRelocateInst *I = Item.second;
1047 auto BaseKey = std::make_pair(Key.first, Key.first);
1048
1049 // We're iterating over RelocateIdxMap so we cannot modify it.
1050 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1051 if (MaybeBase == RelocateIdxMap.end())
1052 // TODO: We might want to insert a new base object relocate and gep off
1053 // that, if there are enough derived object relocates.
1054 continue;
1055
1056 RelocateInstMap[MaybeBase->second].push_back(I);
1057 }
1058}
1059
1060// Accepts a GEP and extracts the operands into a vector provided they're all
1061// small integer constants
1062static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1063 SmallVectorImpl<Value *> &OffsetV) {
1064 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1065 // Only accept small constant integer operands
1066 auto *Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1067 if (!Op || Op->getZExtValue() > 20)
1068 return false;
1069 }
1070
1071 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1072 OffsetV.push_back(GEP->getOperand(i));
1073 return true;
1074}
1075
1076// Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1077// replace, computes a replacement, and affects it.
1078static bool
1079simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1080 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1081 bool MadeChange = false;
1082 // We must ensure the relocation of derived pointer is defined after
1083 // relocation of base pointer. If we find a relocation corresponding to base
1084 // defined earlier than relocation of base then we move relocation of base
1085 // right before found relocation. We consider only relocation in the same
1086 // basic block as relocation of base. Relocations from other basic block will
1087 // be skipped by optimization and we do not care about them.
1088 for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
1089 &*R != RelocatedBase; ++R)
1090 if (auto *RI = dyn_cast<GCRelocateInst>(R))
1091 if (RI->getStatepoint() == RelocatedBase->getStatepoint())
1092 if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
1093 RelocatedBase->moveBefore(RI);
1094 break;
1095 }
1096
1097 for (GCRelocateInst *ToReplace : Targets) {
1098 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&(static_cast <bool> (ToReplace->getBasePtrIndex() ==
RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object"
) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1099, __extension__ __PRETTY_FUNCTION__
))
1099 "Not relocating a derived object of the original base object")(static_cast <bool> (ToReplace->getBasePtrIndex() ==
RelocatedBase->getBasePtrIndex() && "Not relocating a derived object of the original base object"
) ? void (0) : __assert_fail ("ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() && \"Not relocating a derived object of the original base object\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1099, __extension__ __PRETTY_FUNCTION__
))
;
1100 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1101 // A duplicate relocate call. TODO: coalesce duplicates.
1102 continue;
1103 }
1104
1105 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1106 // Base and derived relocates are in different basic blocks.
1107 // In this case transform is only valid when base dominates derived
1108 // relocate. However it would be too expensive to check dominance
1109 // for each such relocate, so we skip the whole transformation.
1110 continue;
1111 }
1112
1113 Value *Base = ToReplace->getBasePtr();
1114 auto *Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1115 if (!Derived || Derived->getPointerOperand() != Base)
1116 continue;
1117
1118 SmallVector<Value *, 2> OffsetV;
1119 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1120 continue;
1121
1122 // Create a Builder and replace the target callsite with a gep
1123 assert(RelocatedBase->getNextNode() &&(static_cast <bool> (RelocatedBase->getNextNode() &&
"Should always have one since it's not a terminator") ? void
(0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1124, __extension__ __PRETTY_FUNCTION__
))
1124 "Should always have one since it's not a terminator")(static_cast <bool> (RelocatedBase->getNextNode() &&
"Should always have one since it's not a terminator") ? void
(0) : __assert_fail ("RelocatedBase->getNextNode() && \"Should always have one since it's not a terminator\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1124, __extension__ __PRETTY_FUNCTION__
))
;
1125
1126 // Insert after RelocatedBase
1127 IRBuilder<> Builder(RelocatedBase->getNextNode());
1128 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1129
1130 // If gc_relocate does not match the actual type, cast it to the right type.
1131 // In theory, there must be a bitcast after gc_relocate if the type does not
1132 // match, and we should reuse it to get the derived pointer. But it could be
1133 // cases like this:
1134 // bb1:
1135 // ...
1136 // %g1 = call coldcc i8 addrspace(1)*
1137 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1138 //
1139 // bb2:
1140 // ...
1141 // %g2 = call coldcc i8 addrspace(1)*
1142 // @llvm.experimental.gc.relocate.p1i8(...) br label %merge
1143 //
1144 // merge:
1145 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1146 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1147 //
1148 // In this case, we can not find the bitcast any more. So we insert a new
1149 // bitcast no matter there is already one or not. In this way, we can handle
1150 // all cases, and the extra bitcast should be optimized away in later
1151 // passes.
1152 Value *ActualRelocatedBase = RelocatedBase;
1153 if (RelocatedBase->getType() != Base->getType()) {
1154 ActualRelocatedBase =
1155 Builder.CreateBitCast(RelocatedBase, Base->getType());
1156 }
1157 Value *Replacement =
1158 Builder.CreateGEP(Derived->getSourceElementType(), ActualRelocatedBase,
1159 makeArrayRef(OffsetV));
1160 Replacement->takeName(ToReplace);
1161 // If the newly generated derived pointer's type does not match the original
1162 // derived pointer's type, cast the new derived pointer to match it. Same
1163 // reasoning as above.
1164 Value *ActualReplacement = Replacement;
1165 if (Replacement->getType() != ToReplace->getType()) {
1166 ActualReplacement =
1167 Builder.CreateBitCast(Replacement, ToReplace->getType());
1168 }
1169 ToReplace->replaceAllUsesWith(ActualReplacement);
1170 ToReplace->eraseFromParent();
1171
1172 MadeChange = true;
1173 }
1174 return MadeChange;
1175}
1176
1177// Turns this:
1178//
1179// %base = ...
1180// %ptr = gep %base + 15
1181// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1182// %base' = relocate(%tok, i32 4, i32 4)
1183// %ptr' = relocate(%tok, i32 4, i32 5)
1184// %val = load %ptr'
1185//
1186// into this:
1187//
1188// %base = ...
1189// %ptr = gep %base + 15
1190// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1191// %base' = gc.relocate(%tok, i32 4, i32 4)
1192// %ptr' = gep %base' + 15
1193// %val = load %ptr'
1194bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) {
1195 bool MadeChange = false;
1196 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1197 for (auto *U : I.users())
1198 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1199 // Collect all the relocate calls associated with a statepoint
1200 AllRelocateCalls.push_back(Relocate);
1201
1202 // We need at least one base pointer relocation + one derived pointer
1203 // relocation to mangle
1204 if (AllRelocateCalls.size() < 2)
1205 return false;
1206
1207 // RelocateInstMap is a mapping from the base relocate instruction to the
1208 // corresponding derived relocate instructions
1209 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1210 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1211 if (RelocateInstMap.empty())
1212 return false;
1213
1214 for (auto &Item : RelocateInstMap)
1215 // Item.first is the RelocatedBase to offset against
1216 // Item.second is the vector of Targets to replace
1217 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1218 return MadeChange;
1219}
1220
1221/// Sink the specified cast instruction into its user blocks.
1222static bool SinkCast(CastInst *CI) {
1223 BasicBlock *DefBB = CI->getParent();
1224
1225 /// InsertedCasts - Only insert a cast in each block once.
1226 DenseMap<BasicBlock *, CastInst *> InsertedCasts;
1227
1228 bool MadeChange = false;
1229 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1230 UI != E;) {
1231 Use &TheUse = UI.getUse();
1232 Instruction *User = cast<Instruction>(*UI);
1233
1234 // Figure out which BB this cast is used in. For PHI's this is the
1235 // appropriate predecessor block.
1236 BasicBlock *UserBB = User->getParent();
1237 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1238 UserBB = PN->getIncomingBlock(TheUse);
1239 }
1240
1241 // Preincrement use iterator so we don't invalidate it.
1242 ++UI;
1243
1244 // The first insertion point of a block containing an EH pad is after the
1245 // pad. If the pad is the user, we cannot sink the cast past the pad.
1246 if (User->isEHPad())
1247 continue;
1248
1249 // If the block selected to receive the cast is an EH pad that does not
1250 // allow non-PHI instructions before the terminator, we can't sink the
1251 // cast.
1252 if (UserBB->getTerminator()->isEHPad())
1253 continue;
1254
1255 // If this user is in the same block as the cast, don't change the cast.
1256 if (UserBB == DefBB)
1257 continue;
1258
1259 // If we have already inserted a cast into this block, use it.
1260 CastInst *&InsertedCast = InsertedCasts[UserBB];
1261
1262 if (!InsertedCast) {
1263 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1264 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 1264, __extension__ __PRETTY_FUNCTION__))
;
1265 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1266 CI->getType(), "", &*InsertPt);
1267 InsertedCast->setDebugLoc(CI->getDebugLoc());
1268 }
1269
1270 // Replace a use of the cast with a use of the new cast.
1271 TheUse = InsertedCast;
1272 MadeChange = true;
1273 ++NumCastUses;
1274 }
1275
1276 // If we removed all uses, nuke the cast.
1277 if (CI->use_empty()) {
1278 salvageDebugInfo(*CI);
1279 CI->eraseFromParent();
1280 MadeChange = true;
1281 }
1282
1283 return MadeChange;
1284}
1285
1286/// If the specified cast instruction is a noop copy (e.g. it's casting from
1287/// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1288/// reduce the number of virtual registers that must be created and coalesced.
1289///
1290/// Return true if any changes are made.
1291static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1292 const DataLayout &DL) {
1293 // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1294 // than sinking only nop casts, but is helpful on some platforms.
1295 if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1296 if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(),
1297 ASC->getDestAddressSpace()))
1298 return false;
1299 }
1300
1301 // If this is a noop copy,
1302 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1303 EVT DstVT = TLI.getValueType(DL, CI->getType());
1304
1305 // This is an fp<->int conversion?
1306 if (SrcVT.isInteger() != DstVT.isInteger())
1307 return false;
1308
1309 // If this is an extension, it will be a zero or sign extension, which
1310 // isn't a noop.
1311 if (SrcVT.bitsLT(DstVT))
1312 return false;
1313
1314 // If these values will be promoted, find out what they will be promoted
1315 // to. This helps us consider truncates on PPC as noop copies when they
1316 // are.
1317 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1318 TargetLowering::TypePromoteInteger)
1319 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1320 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1321 TargetLowering::TypePromoteInteger)
1322 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1323
1324 // If, after promotion, these are the same types, this is a noop copy.
1325 if (SrcVT != DstVT)
1326 return false;
1327
1328 return SinkCast(CI);
1329}
1330
1331// Match a simple increment by constant operation. Note that if a sub is
1332// matched, the step is negated (as if the step had been canonicalized to
1333// an add, even though we leave the instruction alone.)
1334bool matchIncrement(const Instruction *IVInc, Instruction *&LHS,
1335 Constant *&Step) {
1336 if (match(IVInc, m_Add(m_Instruction(LHS), m_Constant(Step))) ||
1337 match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::uadd_with_overflow>(
1338 m_Instruction(LHS), m_Constant(Step)))))
1339 return true;
1340 if (match(IVInc, m_Sub(m_Instruction(LHS), m_Constant(Step))) ||
1341 match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::usub_with_overflow>(
1342 m_Instruction(LHS), m_Constant(Step))))) {
1343 Step = ConstantExpr::getNeg(Step);
1344 return true;
1345 }
1346 return false;
1347}
1348
1349/// If given \p PN is an inductive variable with value IVInc coming from the
1350/// backedge, and on each iteration it gets increased by Step, return pair
1351/// <IVInc, Step>. Otherwise, return None.
1352static Optional<std::pair<Instruction *, Constant *>>
1353getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
1354 const Loop *L = LI->getLoopFor(PN->getParent());
1355 if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch())
1356 return None;
1357 auto *IVInc =
1358 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
1359 if (!IVInc || LI->getLoopFor(IVInc->getParent()) != L)
1360 return None;
1361 Instruction *LHS = nullptr;
1362 Constant *Step = nullptr;
1363 if (matchIncrement(IVInc, LHS, Step) && LHS == PN)
1364 return std::make_pair(IVInc, Step);
1365 return None;
1366}
1367
1368static bool isIVIncrement(const Value *V, const LoopInfo *LI) {
1369 auto *I = dyn_cast<Instruction>(V);
1370 if (!I)
1371 return false;
1372 Instruction *LHS = nullptr;
1373 Constant *Step = nullptr;
1374 if (!matchIncrement(I, LHS, Step))
1375 return false;
1376 if (auto *PN = dyn_cast<PHINode>(LHS))
1377 if (auto IVInc = getIVIncrement(PN, LI))
1378 return IVInc->first == I;
1379 return false;
1380}
1381
1382bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1383 Value *Arg0, Value *Arg1,
1384 CmpInst *Cmp,
1385 Intrinsic::ID IID) {
1386 auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) {
1387 if (!isIVIncrement(BO, LI))
1388 return false;
1389 const Loop *L = LI->getLoopFor(BO->getParent());
1390 assert(L && "L should not be null after isIVIncrement()")(static_cast <bool> (L && "L should not be null after isIVIncrement()"
) ? void (0) : __assert_fail ("L && \"L should not be null after isIVIncrement()\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1390, __extension__ __PRETTY_FUNCTION__
))
;
1391 // Do not risk on moving increment into a child loop.
1392 if (LI->getLoopFor(Cmp->getParent()) != L)
1393 return false;
1394
1395 // Finally, we need to ensure that the insert point will dominate all
1396 // existing uses of the increment.
1397
1398 auto &DT = getDT(*BO->getParent()->getParent());
1399 if (DT.dominates(Cmp->getParent(), BO->getParent()))
1400 // If we're moving up the dom tree, all uses are trivially dominated.
1401 // (This is the common case for code produced by LSR.)
1402 return true;
1403
1404 // Otherwise, special case the single use in the phi recurrence.
1405 return BO->hasOneUse() && DT.dominates(Cmp->getParent(), L->getLoopLatch());
1406 };
1407 if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) {
1408 // We used to use a dominator tree here to allow multi-block optimization.
1409 // But that was problematic because:
1410 // 1. It could cause a perf regression by hoisting the math op into the
1411 // critical path.
1412 // 2. It could cause a perf regression by creating a value that was live
1413 // across multiple blocks and increasing register pressure.
1414 // 3. Use of a dominator tree could cause large compile-time regression.
1415 // This is because we recompute the DT on every change in the main CGP
1416 // run-loop. The recomputing is probably unnecessary in many cases, so if
1417 // that was fixed, using a DT here would be ok.
1418 //
1419 // There is one important particular case we still want to handle: if BO is
1420 // the IV increment. Important properties that make it profitable:
1421 // - We can speculate IV increment anywhere in the loop (as long as the
1422 // indvar Phi is its only user);
1423 // - Upon computing Cmp, we effectively compute something equivalent to the
1424 // IV increment (despite it loops differently in the IR). So moving it up
1425 // to the cmp point does not really increase register pressure.
1426 return false;
1427 }
1428
1429 // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1430 if (BO->getOpcode() == Instruction::Add &&
1431 IID == Intrinsic::usub_with_overflow) {
1432 assert(isa<Constant>(Arg1) && "Unexpected input for usubo")(static_cast <bool> (isa<Constant>(Arg1) &&
"Unexpected input for usubo") ? void (0) : __assert_fail ("isa<Constant>(Arg1) && \"Unexpected input for usubo\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1432, __extension__ __PRETTY_FUNCTION__
))
;
1433 Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1));
1434 }
1435
1436 // Insert at the first instruction of the pair.
1437 Instruction *InsertPt = nullptr;
1438 for (Instruction &Iter : *Cmp->getParent()) {
1439 // If BO is an XOR, it is not guaranteed that it comes after both inputs to
1440 // the overflow intrinsic are defined.
1441 if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) {
1442 InsertPt = &Iter;
1443 break;
1444 }
1445 }
1446 assert(InsertPt != nullptr && "Parent block did not contain cmp or binop")(static_cast <bool> (InsertPt != nullptr && "Parent block did not contain cmp or binop"
) ? void (0) : __assert_fail ("InsertPt != nullptr && \"Parent block did not contain cmp or binop\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1446, __extension__ __PRETTY_FUNCTION__
))
;
1447
1448 IRBuilder<> Builder(InsertPt);
1449 Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1);
1450 if (BO->getOpcode() != Instruction::Xor) {
1451 Value *Math = Builder.CreateExtractValue(MathOV, 0, "math");
1452 BO->replaceAllUsesWith(Math);
1453 } else
1454 assert(BO->hasOneUse() &&(static_cast <bool> (BO->hasOneUse() && "Patterns with XOr should use the BO only in the compare"
) ? void (0) : __assert_fail ("BO->hasOneUse() && \"Patterns with XOr should use the BO only in the compare\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1455, __extension__ __PRETTY_FUNCTION__
))
1455 "Patterns with XOr should use the BO only in the compare")(static_cast <bool> (BO->hasOneUse() && "Patterns with XOr should use the BO only in the compare"
) ? void (0) : __assert_fail ("BO->hasOneUse() && \"Patterns with XOr should use the BO only in the compare\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1455, __extension__ __PRETTY_FUNCTION__
))
;
1456 Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov");
1457 Cmp->replaceAllUsesWith(OV);
1458 Cmp->eraseFromParent();
1459 BO->eraseFromParent();
1460 return true;
1461}
1462
1463/// Match special-case patterns that check for unsigned add overflow.
1464static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp,
1465 BinaryOperator *&Add) {
1466 // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1467 // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1468 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1469
1470 // We are not expecting non-canonical/degenerate code. Just bail out.
1471 if (isa<Constant>(A))
1472 return false;
1473
1474 ICmpInst::Predicate Pred = Cmp->getPredicate();
1475 if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes()))
1476 B = ConstantInt::get(B->getType(), 1);
1477 else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt()))
1478 B = ConstantInt::get(B->getType(), -1);
1479 else
1480 return false;
1481
1482 // Check the users of the variable operand of the compare looking for an add
1483 // with the adjusted constant.
1484 for (User *U : A->users()) {
1485 if (match(U, m_Add(m_Specific(A), m_Specific(B)))) {
1486 Add = cast<BinaryOperator>(U);
1487 return true;
1488 }
1489 }
1490 return false;
1491}
1492
1493/// Try to combine the compare into a call to the llvm.uadd.with.overflow
1494/// intrinsic. Return true if any changes were made.
1495bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT) {
1496 Value *A, *B;
1497 BinaryOperator *Add;
1498 if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add)))) {
1499 if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add))
1500 return false;
1501 // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases.
1502 A = Add->getOperand(0);
1503 B = Add->getOperand(1);
1504 }
1505
1506 if (!TLI->shouldFormOverflowOp(ISD::UADDO,
1507 TLI->getValueType(*DL, Add->getType()),
1508 Add->hasNUsesOrMore(2)))
1509 return false;
1510
1511 // We don't want to move around uses of condition values this late, so we
1512 // check if it is legal to create the call to the intrinsic in the basic
1513 // block containing the icmp.
1514 if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1515 return false;
1516
1517 if (!replaceMathCmpWithIntrinsic(Add, A, B, Cmp,
1518 Intrinsic::uadd_with_overflow))
1519 return false;
1520
1521 // Reset callers - do not crash by iterating over a dead instruction.
1522 ModifiedDT = true;
1523 return true;
1524}
1525
1526bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT) {
1527 // We are not expecting non-canonical/degenerate code. Just bail out.
1528 Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1529 if (isa<Constant>(A) && isa<Constant>(B))
1530 return false;
1531
1532 // Convert (A u> B) to (A u< B) to simplify pattern matching.
1533 ICmpInst::Predicate Pred = Cmp->getPredicate();
1534 if (Pred == ICmpInst::ICMP_UGT) {
1535 std::swap(A, B);
1536 Pred = ICmpInst::ICMP_ULT;
1537 }
1538 // Convert special-case: (A == 0) is the same as (A u< 1).
1539 if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) {
1540 B = ConstantInt::get(B->getType(), 1);
1541 Pred = ICmpInst::ICMP_ULT;
1542 }
1543 // Convert special-case: (A != 0) is the same as (0 u< A).
1544 if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) {
1545 std::swap(A, B);
1546 Pred = ICmpInst::ICMP_ULT;
1547 }
1548 if (Pred != ICmpInst::ICMP_ULT)
1549 return false;
1550
1551 // Walk the users of a variable operand of a compare looking for a subtract or
1552 // add with that same operand. Also match the 2nd operand of the compare to
1553 // the add/sub, but that may be a negated constant operand of an add.
1554 Value *CmpVariableOperand = isa<Constant>(A) ? B : A;
1555 BinaryOperator *Sub = nullptr;
1556 for (User *U : CmpVariableOperand->users()) {
1557 // A - B, A u< B --> usubo(A, B)
1558 if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) {
1559 Sub = cast<BinaryOperator>(U);
1560 break;
1561 }
1562
1563 // A + (-C), A u< C (canonicalized form of (sub A, C))
1564 const APInt *CmpC, *AddC;
1565 if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) &&
1566 match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) {
1567 Sub = cast<BinaryOperator>(U);
1568 break;
1569 }
1570 }
1571 if (!Sub)
1572 return false;
1573
1574 if (!TLI->shouldFormOverflowOp(ISD::USUBO,
1575 TLI->getValueType(*DL, Sub->getType()),
1576 Sub->hasNUsesOrMore(2)))
1577 return false;
1578
1579 if (!replaceMathCmpWithIntrinsic(Sub, Sub->getOperand(0), Sub->getOperand(1),
1580 Cmp, Intrinsic::usub_with_overflow))
1581 return false;
1582
1583 // Reset callers - do not crash by iterating over a dead instruction.
1584 ModifiedDT = true;
1585 return true;
1586}
1587
1588/// Sink the given CmpInst into user blocks to reduce the number of virtual
1589/// registers that must be created and coalesced. This is a clear win except on
1590/// targets with multiple condition code registers (PowerPC), where it might
1591/// lose; some adjustment may be wanted there.
1592///
1593/// Return true if any changes are made.
1594static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) {
1595 if (TLI.hasMultipleConditionRegisters())
1596 return false;
1597
1598 // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1599 if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp))
1600 return false;
1601
1602 // Only insert a cmp in each block once.
1603 DenseMap<BasicBlock *, CmpInst *> InsertedCmps;
1604
1605 bool MadeChange = false;
1606 for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1607 UI != E;) {
1608 Use &TheUse = UI.getUse();
1609 Instruction *User = cast<Instruction>(*UI);
1610
1611 // Preincrement use iterator so we don't invalidate it.
1612 ++UI;
1613
1614 // Don't bother for PHI nodes.
1615 if (isa<PHINode>(User))
1616 continue;
1617
1618 // Figure out which BB this cmp is used in.
1619 BasicBlock *UserBB = User->getParent();
1620 BasicBlock *DefBB = Cmp->getParent();
1621
1622 // If this user is in the same block as the cmp, don't change the cmp.
1623 if (UserBB == DefBB)
1624 continue;
1625
1626 // If we have already inserted a cmp into this block, use it.
1627 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1628
1629 if (!InsertedCmp) {
1630 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1631 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 1631, __extension__ __PRETTY_FUNCTION__))
;
1632 InsertedCmp = CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(),
1633 Cmp->getOperand(0), Cmp->getOperand(1), "",
1634 &*InsertPt);
1635 // Propagate the debug info.
1636 InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1637 }
1638
1639 // Replace a use of the cmp with a use of the new cmp.
1640 TheUse = InsertedCmp;
1641 MadeChange = true;
1642 ++NumCmpUses;
1643 }
1644
1645 // If we removed all uses, nuke the cmp.
1646 if (Cmp->use_empty()) {
1647 Cmp->eraseFromParent();
1648 MadeChange = true;
1649 }
1650
1651 return MadeChange;
1652}
1653
1654/// For pattern like:
1655///
1656/// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB)
1657/// ...
1658/// DomBB:
1659/// ...
1660/// br DomCond, TrueBB, CmpBB
1661/// CmpBB: (with DomBB being the single predecessor)
1662/// ...
1663/// Cmp = icmp eq CmpOp0, CmpOp1
1664/// ...
1665///
1666/// It would use two comparison on targets that lowering of icmp sgt/slt is
1667/// different from lowering of icmp eq (PowerPC). This function try to convert
1668/// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'.
1669/// After that, DomCond and Cmp can use the same comparison so reduce one
1670/// comparison.
1671///
1672/// Return true if any changes are made.
1673static bool foldICmpWithDominatingICmp(CmpInst *Cmp,
1674 const TargetLowering &TLI) {
1675 if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp())
1676 return false;
1677
1678 ICmpInst::Predicate Pred = Cmp->getPredicate();
1679 if (Pred != ICmpInst::ICMP_EQ)
1680 return false;
1681
1682 // If icmp eq has users other than BranchInst and SelectInst, converting it to
1683 // icmp slt/sgt would introduce more redundant LLVM IR.
1684 for (User *U : Cmp->users()) {
1685 if (isa<BranchInst>(U))
1686 continue;
1687 if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == Cmp)
1688 continue;
1689 return false;
1690 }
1691
1692 // This is a cheap/incomplete check for dominance - just match a single
1693 // predecessor with a conditional branch.
1694 BasicBlock *CmpBB = Cmp->getParent();
1695 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1696 if (!DomBB)
1697 return false;
1698
1699 // We want to ensure that the only way control gets to the comparison of
1700 // interest is that a less/greater than comparison on the same operands is
1701 // false.
1702 Value *DomCond;
1703 BasicBlock *TrueBB, *FalseBB;
1704 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1705 return false;
1706 if (CmpBB != FalseBB)
1707 return false;
1708
1709 Value *CmpOp0 = Cmp->getOperand(0), *CmpOp1 = Cmp->getOperand(1);
1710 ICmpInst::Predicate DomPred;
1711 if (!match(DomCond, m_ICmp(DomPred, m_Specific(CmpOp0), m_Specific(CmpOp1))))
1712 return false;
1713 if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT)
1714 return false;
1715
1716 // Convert the equality comparison to the opposite of the dominating
1717 // comparison and swap the direction for all branch/select users.
1718 // We have conceptually converted:
1719 // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>;
1720 // to
1721 // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>;
1722 // And similarly for branches.
1723 for (User *U : Cmp->users()) {
1724 if (auto *BI = dyn_cast<BranchInst>(U)) {
1725 assert(BI->isConditional() && "Must be conditional")(static_cast <bool> (BI->isConditional() && "Must be conditional"
) ? void (0) : __assert_fail ("BI->isConditional() && \"Must be conditional\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1725, __extension__ __PRETTY_FUNCTION__
))
;
1726 BI->swapSuccessors();
1727 continue;
1728 }
1729 if (auto *SI = dyn_cast<SelectInst>(U)) {
1730 // Swap operands
1731 SI->swapValues();
1732 SI->swapProfMetadata();
1733 continue;
1734 }
1735 llvm_unreachable("Must be a branch or a select")::llvm::llvm_unreachable_internal("Must be a branch or a select"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1735)
;
1736 }
1737 Cmp->setPredicate(CmpInst::getSwappedPredicate(DomPred));
1738 return true;
1739}
1740
1741bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) {
1742 if (sinkCmpExpression(Cmp, *TLI))
1743 return true;
1744
1745 if (combineToUAddWithOverflow(Cmp, ModifiedDT))
1746 return true;
1747
1748 if (combineToUSubWithOverflow(Cmp, ModifiedDT))
1749 return true;
1750
1751 if (foldICmpWithDominatingICmp(Cmp, *TLI))
1752 return true;
1753
1754 return false;
1755}
1756
1757/// Duplicate and sink the given 'and' instruction into user blocks where it is
1758/// used in a compare to allow isel to generate better code for targets where
1759/// this operation can be combined.
1760///
1761/// Return true if any changes are made.
1762static bool sinkAndCmp0Expression(Instruction *AndI, const TargetLowering &TLI,
1763 SetOfInstrs &InsertedInsts) {
1764 // Double-check that we're not trying to optimize an instruction that was
1765 // already optimized by some other part of this pass.
1766 assert(!InsertedInsts.count(AndI) &&(static_cast <bool> (!InsertedInsts.count(AndI) &&
"Attempting to optimize already optimized and instruction") ?
void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1767, __extension__ __PRETTY_FUNCTION__
))
1767 "Attempting to optimize already optimized and instruction")(static_cast <bool> (!InsertedInsts.count(AndI) &&
"Attempting to optimize already optimized and instruction") ?
void (0) : __assert_fail ("!InsertedInsts.count(AndI) && \"Attempting to optimize already optimized and instruction\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1767, __extension__ __PRETTY_FUNCTION__
))
;
1768 (void)InsertedInsts;
1769
1770 // Nothing to do for single use in same basic block.
1771 if (AndI->hasOneUse() &&
1772 AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1773 return false;
1774
1775 // Try to avoid cases where sinking/duplicating is likely to increase register
1776 // pressure.
1777 if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1778 !isa<ConstantInt>(AndI->getOperand(1)) &&
1779 AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1780 return false;
1781
1782 for (auto *U : AndI->users()) {
1783 Instruction *User = cast<Instruction>(U);
1784
1785 // Only sink 'and' feeding icmp with 0.
1786 if (!isa<ICmpInst>(User))
1787 return false;
1788
1789 auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1790 if (!CmpC || !CmpC->isZero())
1791 return false;
1792 }
1793
1794 if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1795 return false;
1796
1797 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)
;
1798 LLVM_DEBUG(AndI->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { AndI->getParent()->dump(); } } while
(false)
;
1799
1800 // Push the 'and' into the same block as the icmp 0. There should only be
1801 // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1802 // others, so we don't need to keep track of which BBs we insert into.
1803 for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1804 UI != E;) {
1805 Use &TheUse = UI.getUse();
1806 Instruction *User = cast<Instruction>(*UI);
1807
1808 // Preincrement use iterator so we don't invalidate it.
1809 ++UI;
1810
1811 LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "sinking 'and' use: " <<
*User << "\n"; } } while (false)
;
1812
1813 // Keep the 'and' in the same place if the use is already in the same block.
1814 Instruction *InsertPt =
1815 User->getParent() == AndI->getParent() ? AndI : User;
1816 Instruction *InsertedAnd =
1817 BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1818 AndI->getOperand(1), "", InsertPt);
1819 // Propagate the debug info.
1820 InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1821
1822 // Replace a use of the 'and' with a use of the new 'and'.
1823 TheUse = InsertedAnd;
1824 ++NumAndUses;
1825 LLVM_DEBUG(User->getParent()->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { User->getParent()->dump(); } } while
(false)
;
1826 }
1827
1828 // We removed all uses, nuke the and.
1829 AndI->eraseFromParent();
1830 return true;
1831}
1832
1833/// Check if the candidates could be combined with a shift instruction, which
1834/// includes:
1835/// 1. Truncate instruction
1836/// 2. And instruction and the imm is a mask of the low bits:
1837/// imm & (imm+1) == 0
1838static bool isExtractBitsCandidateUse(Instruction *User) {
1839 if (!isa<TruncInst>(User)) {
1840 if (User->getOpcode() != Instruction::And ||
1841 !isa<ConstantInt>(User->getOperand(1)))
1842 return false;
1843
1844 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1845
1846 if ((Cimm & (Cimm + 1)).getBoolValue())
1847 return false;
1848 }
1849 return true;
1850}
1851
1852/// Sink both shift and truncate instruction to the use of truncate's BB.
1853static bool
1854SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1855 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1856 const TargetLowering &TLI, const DataLayout &DL) {
1857 BasicBlock *UserBB = User->getParent();
1858 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1859 auto *TruncI = cast<TruncInst>(User);
1860 bool MadeChange = false;
1861
1862 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1863 TruncE = TruncI->user_end();
1864 TruncUI != TruncE;) {
1865
1866 Use &TruncTheUse = TruncUI.getUse();
1867 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1868 // Preincrement use iterator so we don't invalidate it.
1869
1870 ++TruncUI;
1871
1872 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1873 if (!ISDOpcode)
1874 continue;
1875
1876 // If the use is actually a legal node, there will not be an
1877 // implicit truncate.
1878 // FIXME: always querying the result type is just an
1879 // approximation; some nodes' legality is determined by the
1880 // operand or other means. There's no good way to find out though.
1881 if (TLI.isOperationLegalOrCustom(
1882 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1883 continue;
1884
1885 // Don't bother for PHI nodes.
1886 if (isa<PHINode>(TruncUser))
1887 continue;
1888
1889 BasicBlock *TruncUserBB = TruncUser->getParent();
1890
1891 if (UserBB == TruncUserBB)
1892 continue;
1893
1894 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1895 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1896
1897 if (!InsertedShift && !InsertedTrunc) {
1898 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1899 assert(InsertPt != TruncUserBB->end())(static_cast <bool> (InsertPt != TruncUserBB->end())
? void (0) : __assert_fail ("InsertPt != TruncUserBB->end()"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1899, __extension__ __PRETTY_FUNCTION__
))
;
1900 // Sink the shift
1901 if (ShiftI->getOpcode() == Instruction::AShr)
1902 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1903 "", &*InsertPt);
1904 else
1905 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1906 "", &*InsertPt);
1907 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1908
1909 // Sink the trunc
1910 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1911 TruncInsertPt++;
1912 assert(TruncInsertPt != TruncUserBB->end())(static_cast <bool> (TruncInsertPt != TruncUserBB->end
()) ? void (0) : __assert_fail ("TruncInsertPt != TruncUserBB->end()"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 1912, __extension__ __PRETTY_FUNCTION__
))
;
1913
1914 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1915 TruncI->getType(), "", &*TruncInsertPt);
1916 InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1917
1918 MadeChange = true;
1919
1920 TruncTheUse = InsertedTrunc;
1921 }
1922 }
1923 return MadeChange;
1924}
1925
1926/// Sink the shift *right* instruction into user blocks if the uses could
1927/// potentially be combined with this shift instruction and generate BitExtract
1928/// instruction. It will only be applied if the architecture supports BitExtract
1929/// instruction. Here is an example:
1930/// BB1:
1931/// %x.extract.shift = lshr i64 %arg1, 32
1932/// BB2:
1933/// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1934/// ==>
1935///
1936/// BB2:
1937/// %x.extract.shift.1 = lshr i64 %arg1, 32
1938/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1939///
1940/// CodeGen will recognize the pattern in BB2 and generate BitExtract
1941/// instruction.
1942/// Return true if any changes are made.
1943static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1944 const TargetLowering &TLI,
1945 const DataLayout &DL) {
1946 BasicBlock *DefBB = ShiftI->getParent();
1947
1948 /// Only insert instructions in each block once.
1949 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1950
1951 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1952
1953 bool MadeChange = false;
1954 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1955 UI != E;) {
1956 Use &TheUse = UI.getUse();
1957 Instruction *User = cast<Instruction>(*UI);
1958 // Preincrement use iterator so we don't invalidate it.
1959 ++UI;
1960
1961 // Don't bother for PHI nodes.
1962 if (isa<PHINode>(User))
1963 continue;
1964
1965 if (!isExtractBitsCandidateUse(User))
1966 continue;
1967
1968 BasicBlock *UserBB = User->getParent();
1969
1970 if (UserBB == DefBB) {
1971 // If the shift and truncate instruction are in the same BB. The use of
1972 // the truncate(TruncUse) may still introduce another truncate if not
1973 // legal. In this case, we would like to sink both shift and truncate
1974 // instruction to the BB of TruncUse.
1975 // for example:
1976 // BB1:
1977 // i64 shift.result = lshr i64 opnd, imm
1978 // trunc.result = trunc shift.result to i16
1979 //
1980 // BB2:
1981 // ----> We will have an implicit truncate here if the architecture does
1982 // not have i16 compare.
1983 // cmp i16 trunc.result, opnd2
1984 //
1985 if (isa<TruncInst>(User) &&
1986 shiftIsLegal
1987 // If the type of the truncate is legal, no truncate will be
1988 // introduced in other basic blocks.
1989 && (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1990 MadeChange =
1991 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1992
1993 continue;
1994 }
1995 // If we have already inserted a shift into this block, use it.
1996 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1997
1998 if (!InsertedShift) {
1999 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
2000 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 2000, __extension__ __PRETTY_FUNCTION__))
;
2001
2002 if (ShiftI->getOpcode() == Instruction::AShr)
2003 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
2004 "", &*InsertPt);
2005 else
2006 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
2007 "", &*InsertPt);
2008 InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
2009
2010 MadeChange = true;
2011 }
2012
2013 // Replace a use of the shift with a use of the new shift.
2014 TheUse = InsertedShift;
2015 }
2016
2017 // If we removed all uses, or there are none, nuke the shift.
2018 if (ShiftI->use_empty()) {
2019 salvageDebugInfo(*ShiftI);
2020 ShiftI->eraseFromParent();
2021 MadeChange = true;
2022 }
2023
2024 return MadeChange;
2025}
2026
2027/// If counting leading or trailing zeros is an expensive operation and a zero
2028/// input is defined, add a check for zero to avoid calling the intrinsic.
2029///
2030/// We want to transform:
2031/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2032///
2033/// into:
2034/// entry:
2035/// %cmpz = icmp eq i64 %A, 0
2036/// br i1 %cmpz, label %cond.end, label %cond.false
2037/// cond.false:
2038/// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2039/// br label %cond.end
2040/// cond.end:
2041/// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2042///
2043/// If the transform is performed, return true and set ModifiedDT to true.
2044static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2045 const TargetLowering *TLI,
2046 const DataLayout *DL, bool &ModifiedDT) {
2047 // If a zero input is undefined, it doesn't make sense to despeculate that.
2048 if (match(CountZeros->getOperand(1), m_One()))
2049 return false;
2050
2051 // If it's cheap to speculate, there's nothing to do.
2052 Type *Ty = CountZeros->getType();
2053 auto IntrinsicID = CountZeros->getIntrinsicID();
2054 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz(Ty)) ||
2055 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz(Ty)))
2056 return false;
2057
2058 // Only handle legal scalar cases. Anything else requires too much work.
2059 unsigned SizeInBits = Ty->getScalarSizeInBits();
2060 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
2061 return false;
2062
2063 // Bail if the value is never zero.
2064 Use &Op = CountZeros->getOperandUse(0);
2065 if (isKnownNonZero(Op, *DL))
2066 return false;
2067
2068 // The intrinsic will be sunk behind a compare against zero and branch.
2069 BasicBlock *StartBlock = CountZeros->getParent();
2070 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2071
2072 // Create another block after the count zero intrinsic. A PHI will be added
2073 // in this block to select the result of the intrinsic or the bit-width
2074 // constant if the input to the intrinsic is zero.
2075 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2076 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2077
2078 // Set up a builder to create a compare, conditional branch, and PHI.
2079 IRBuilder<> Builder(CountZeros->getContext());
2080 Builder.SetInsertPoint(StartBlock->getTerminator());
2081 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2082
2083 // Replace the unconditional branch that was created by the first split with
2084 // a compare against zero and a conditional branch.
2085 Value *Zero = Constant::getNullValue(Ty);
2086 // Avoid introducing branch on poison. This also replaces the ctz operand.
2087 if (!isGuaranteedNotToBeUndefOrPoison(Op))
2088 Op = Builder.CreateFreeze(Op, Op->getName() + ".fr");
2089 Value *Cmp = Builder.CreateICmpEQ(Op, Zero, "cmpz");
2090 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2091 StartBlock->getTerminator()->eraseFromParent();
2092
2093 // Create a PHI in the end block to select either the output of the intrinsic
2094 // or the bit width of the operand.
2095 Builder.SetInsertPoint(&EndBlock->front());
2096 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2097 CountZeros->replaceAllUsesWith(PN);
2098 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2099 PN->addIncoming(BitWidth, StartBlock);
2100 PN->addIncoming(CountZeros, CallBlock);
2101
2102 // We are explicitly handling the zero case, so we can set the intrinsic's
2103 // undefined zero argument to 'true'. This will also prevent reprocessing the
2104 // intrinsic; we only despeculate when a zero input is defined.
2105 CountZeros->setArgOperand(1, Builder.getTrue());
2106 ModifiedDT = true;
2107 return true;
2108}
2109
2110bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
2111 BasicBlock *BB = CI->getParent();
2112
2113 // Lower inline assembly if we can.
2114 // If we found an inline asm expession, and if the target knows how to
2115 // lower it to normal LLVM code, do so now.
2116 if (CI->isInlineAsm()) {
2117 if (TLI->ExpandInlineAsm(CI)) {
2118 // Avoid invalidating the iterator.
2119 CurInstIterator = BB->begin();
2120 // Avoid processing instructions out of order, which could cause
2121 // reuse before a value is defined.
2122 SunkAddrs.clear();
2123 return true;
2124 }
2125 // Sink address computing for memory operands into the block.
2126 if (optimizeInlineAsmInst(CI))
2127 return true;
2128 }
2129
2130 // Align the pointer arguments to this call if the target thinks it's a good
2131 // idea
2132 unsigned MinSize;
2133 Align PrefAlign;
2134 if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2135 for (auto &Arg : CI->args()) {
2136 // We want to align both objects whose address is used directly and
2137 // objects whose address is used in casts and GEPs, though it only makes
2138 // sense for GEPs if the offset is a multiple of the desired alignment and
2139 // if size - offset meets the size threshold.
2140 if (!Arg->getType()->isPointerTy())
2141 continue;
2142 APInt Offset(DL->getIndexSizeInBits(
2143 cast<PointerType>(Arg->getType())->getAddressSpace()),
2144 0);
2145 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2146 uint64_t Offset2 = Offset.getLimitedValue();
2147 if (!isAligned(PrefAlign, Offset2))
2148 continue;
2149 AllocaInst *AI;
2150 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlign() < PrefAlign &&
2151 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2152 AI->setAlignment(PrefAlign);
2153 // Global variables can only be aligned if they are defined in this
2154 // object (i.e. they are uniquely initialized in this object), and
2155 // over-aligning global variables that have an explicit section is
2156 // forbidden.
2157 GlobalVariable *GV;
2158 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2159 GV->getPointerAlignment(*DL) < PrefAlign &&
2160 DL->getTypeAllocSize(GV->getValueType()) >= MinSize + Offset2)
2161 GV->setAlignment(PrefAlign);
2162 }
2163 // If this is a memcpy (or similar) then we may be able to improve the
2164 // alignment
2165 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2166 Align DestAlign = getKnownAlignment(MI->getDest(), *DL);
2167 MaybeAlign MIDestAlign = MI->getDestAlign();
2168 if (!MIDestAlign || DestAlign > *MIDestAlign)
2169 MI->setDestAlignment(DestAlign);
2170 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
2171 MaybeAlign MTISrcAlign = MTI->getSourceAlign();
2172 Align SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
2173 if (!MTISrcAlign || SrcAlign > *MTISrcAlign)
2174 MTI->setSourceAlignment(SrcAlign);
2175 }
2176 }
2177 }
2178
2179 // If we have a cold call site, try to sink addressing computation into the
2180 // cold block. This interacts with our handling for loads and stores to
2181 // ensure that we can fold all uses of a potential addressing computation
2182 // into their uses. TODO: generalize this to work over profiling data
2183 if (CI->hasFnAttr(Attribute::Cold) && !OptSize &&
2184 !llvm::shouldOptimizeForSize(BB, PSI, BFI.get()))
2185 for (auto &Arg : CI->args()) {
2186 if (!Arg->getType()->isPointerTy())
2187 continue;
2188 unsigned AS = Arg->getType()->getPointerAddressSpace();
2189 return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
2190 }
2191
2192 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2193 if (II) {
2194 switch (II->getIntrinsicID()) {
2195 default:
2196 break;
2197 case Intrinsic::assume:
2198 llvm_unreachable("llvm.assume should have been removed already")::llvm::llvm_unreachable_internal("llvm.assume should have been removed already"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2198)
;
2199 case Intrinsic::experimental_widenable_condition: {
2200 // Give up on future widening oppurtunties so that we can fold away dead
2201 // paths and merge blocks before going into block-local instruction
2202 // selection.
2203 if (II->use_empty()) {
2204 II->eraseFromParent();
2205 return true;
2206 }
2207 Constant *RetVal = ConstantInt::getTrue(II->getContext());
2208 resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
2209 replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
2210 });
2211 return true;
2212 }
2213 case Intrinsic::objectsize:
2214 llvm_unreachable("llvm.objectsize.* should have been lowered already")::llvm::llvm_unreachable_internal("llvm.objectsize.* should have been lowered already"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2214)
;
2215 case Intrinsic::is_constant:
2216 llvm_unreachable("llvm.is.constant.* should have been lowered already")::llvm::llvm_unreachable_internal("llvm.is.constant.* should have been lowered already"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2216)
;
2217 case Intrinsic::aarch64_stlxr:
2218 case Intrinsic::aarch64_stxr: {
2219 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2220 if (!ExtVal || !ExtVal->hasOneUse() ||
2221 ExtVal->getParent() == CI->getParent())
2222 return false;
2223 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2224 ExtVal->moveBefore(CI);
2225 // Mark this instruction as "inserted by CGP", so that other
2226 // optimizations don't touch it.
2227 InsertedInsts.insert(ExtVal);
2228 return true;
2229 }
2230
2231 case Intrinsic::launder_invariant_group:
2232 case Intrinsic::strip_invariant_group: {
2233 Value *ArgVal = II->getArgOperand(0);
2234 auto it = LargeOffsetGEPMap.find(II);
2235 if (it != LargeOffsetGEPMap.end()) {
2236 // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
2237 // Make sure not to have to deal with iterator invalidation
2238 // after possibly adding ArgVal to LargeOffsetGEPMap.
2239 auto GEPs = std::move(it->second);
2240 LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
2241 LargeOffsetGEPMap.erase(II);
2242 }
2243
2244 II->replaceAllUsesWith(ArgVal);
2245 II->eraseFromParent();
2246 return true;
2247 }
2248 case Intrinsic::cttz:
2249 case Intrinsic::ctlz:
2250 // If counting zeros is expensive, try to avoid it.
2251 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2252 case Intrinsic::fshl:
2253 case Intrinsic::fshr:
2254 return optimizeFunnelShift(II);
2255 case Intrinsic::dbg_value:
2256 return fixupDbgValue(II);
2257 case Intrinsic::vscale: {
2258 // If datalayout has no special restrictions on vector data layout,
2259 // replace `llvm.vscale` by an equivalent constant expression
2260 // to benefit from cheap constant propagation.
2261 Type *ScalableVectorTy =
2262 VectorType::get(Type::getInt8Ty(II->getContext()), 1, true);
2263 if (DL->getTypeAllocSize(ScalableVectorTy).getKnownMinSize() == 8) {
2264 auto *Null = Constant::getNullValue(ScalableVectorTy->getPointerTo());
2265 auto *One = ConstantInt::getSigned(II->getType(), 1);
2266 auto *CGep =
2267 ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One);
2268 II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType()));
2269 II->eraseFromParent();
2270 return true;
2271 }
2272 break;
2273 }
2274 case Intrinsic::masked_gather:
2275 return optimizeGatherScatterInst(II, II->getArgOperand(0));
2276 case Intrinsic::masked_scatter:
2277 return optimizeGatherScatterInst(II, II->getArgOperand(1));
2278 }
2279
2280 SmallVector<Value *, 2> PtrOps;
2281 Type *AccessTy;
2282 if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
2283 while (!PtrOps.empty()) {
2284 Value *PtrVal = PtrOps.pop_back_val();
2285 unsigned AS = PtrVal->getType()->getPointerAddressSpace();
2286 if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
2287 return true;
2288 }
2289 }
2290
2291 // From here on out we're working with named functions.
2292 if (!CI->getCalledFunction())
2293 return false;
2294
2295 // Lower all default uses of _chk calls. This is very similar
2296 // to what InstCombineCalls does, but here we are only lowering calls
2297 // to fortified library functions (e.g. __memcpy_chk) that have the default
2298 // "don't know" as the objectsize. Anything else should be left alone.
2299 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2300 IRBuilder<> Builder(CI);
2301 if (Value *V = Simplifier.optimizeCall(CI, Builder)) {
2302 CI->replaceAllUsesWith(V);
2303 CI->eraseFromParent();
2304 return true;
2305 }
2306
2307 return false;
2308}
2309
2310/// Look for opportunities to duplicate return instructions to the predecessor
2311/// to enable tail call optimizations. The case it is currently looking for is:
2312/// @code
2313/// bb0:
2314/// %tmp0 = tail call i32 @f0()
2315/// br label %return
2316/// bb1:
2317/// %tmp1 = tail call i32 @f1()
2318/// br label %return
2319/// bb2:
2320/// %tmp2 = tail call i32 @f2()
2321/// br label %return
2322/// return:
2323/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2324/// ret i32 %retval
2325/// @endcode
2326///
2327/// =>
2328///
2329/// @code
2330/// bb0:
2331/// %tmp0 = tail call i32 @f0()
2332/// ret i32 %tmp0
2333/// bb1:
2334/// %tmp1 = tail call i32 @f1()
2335/// ret i32 %tmp1
2336/// bb2:
2337/// %tmp2 = tail call i32 @f2()
2338/// ret i32 %tmp2
2339/// @endcode
2340bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB,
2341 bool &ModifiedDT) {
2342 ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1
Assuming the object is a 'CastReturnType'
2343 if (!RetI
1.1
'RetI' is non-null
1.1
'RetI' is non-null
)
2
Taking false branch
2344 return false;
2345
2346 PHINode *PN = nullptr;
2347 ExtractValueInst *EVI = nullptr;
2348 BitCastInst *BCI = nullptr;
2349 Value *V = RetI->getReturnValue();
3
Calling 'ReturnInst::getReturnValue'
7
Returning from 'ReturnInst::getReturnValue'
2350 if (V
7.1
'V' is null
7.1
'V' is null
) {
2351 BCI = dyn_cast<BitCastInst>(V);
2352 if (BCI)
2353 V = BCI->getOperand(0);
2354
2355 EVI = dyn_cast<ExtractValueInst>(V);
2356 if (EVI) {
2357 V = EVI->getOperand(0);
2358 if (!llvm::all_of(EVI->indices(), [](unsigned idx) { return idx == 0; }))
2359 return false;
2360 }
2361
2362 PN = dyn_cast<PHINode>(V);
2363 if (!PN)
2364 return false;
2365 }
2366
2367 if (PN
7.2
'PN' is null
7.2
'PN' is null
&& PN->getParent() != BB)
2368 return false;
2369
2370 auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) {
2371 const BitCastInst *BC = dyn_cast<BitCastInst>(Inst);
2372 if (BC && BC->hasOneUse())
2373 Inst = BC->user_back();
2374
2375 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
2376 return II->getIntrinsicID() == Intrinsic::lifetime_end;
2377 return false;
2378 };
2379
2380 // Make sure there are no instructions between the first instruction
2381 // and return.
2382 const Instruction *BI = BB->getFirstNonPHI();
8
'BI' initialized here
2383 // Skip over debug and the bitcast.
2384 while (isa<DbgInfoIntrinsic>(BI) || BI == BCI || BI == EVI ||
9
Assuming 'BI' is not a 'DbgInfoIntrinsic'
10
Assuming 'BI' is equal to 'BCI'
2385 isa<PseudoProbeInst>(BI) || isLifetimeEndOrBitCastFor(BI))
2386 BI = BI->getNextNode();
11
Called C++ object pointer is null
2387 if (BI != RetI)
2388 return false;
2389
2390 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2391 /// call.
2392 const Function *F = BB->getParent();
2393 SmallVector<BasicBlock *, 4> TailCallBBs;
2394 if (PN) {
2395 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2396 // Look through bitcasts.
2397 Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts();
2398 CallInst *CI = dyn_cast<CallInst>(IncomingVal);
2399 BasicBlock *PredBB = PN->getIncomingBlock(I);
2400 // Make sure the phi value is indeed produced by the tail call.
2401 if (CI && CI->hasOneUse() && CI->getParent() == PredBB &&
2402 TLI->mayBeEmittedAsTailCall(CI) &&
2403 attributesPermitTailCall(F, CI, RetI, *TLI))
2404 TailCallBBs.push_back(PredBB);
2405 }
2406 } else {
2407 SmallPtrSet<BasicBlock *, 4> VisitedBBs;
2408 for (BasicBlock *Pred : predecessors(BB)) {
2409 if (!VisitedBBs.insert(Pred).second)
2410 continue;
2411 if (Instruction *I = Pred->rbegin()->getPrevNonDebugInstruction(true)) {
2412 CallInst *CI = dyn_cast<CallInst>(I);
2413 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
2414 attributesPermitTailCall(F, CI, RetI, *TLI))
2415 TailCallBBs.push_back(Pred);
2416 }
2417 }
2418 }
2419
2420 bool Changed = false;
2421 for (auto const &TailCallBB : TailCallBBs) {
2422 // Make sure the call instruction is followed by an unconditional branch to
2423 // the return block.
2424 BranchInst *BI = dyn_cast<BranchInst>(TailCallBB->getTerminator());
2425 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2426 continue;
2427
2428 // Duplicate the return into TailCallBB.
2429 (void)FoldReturnIntoUncondBranch(RetI, BB, TailCallBB);
2430 assert(!VerifyBFIUpdates ||(static_cast <bool> (!VerifyBFIUpdates || BFI->getBlockFreq
(BB) >= BFI->getBlockFreq(TailCallBB)) ? void (0) : __assert_fail
("!VerifyBFIUpdates || BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB)"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2431, __extension__ __PRETTY_FUNCTION__
))
2431 BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB))(static_cast <bool> (!VerifyBFIUpdates || BFI->getBlockFreq
(BB) >= BFI->getBlockFreq(TailCallBB)) ? void (0) : __assert_fail
("!VerifyBFIUpdates || BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB)"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2431, __extension__ __PRETTY_FUNCTION__
))
;
2432 BFI->setBlockFreq(
2433 BB,
2434 (BFI->getBlockFreq(BB) - BFI->getBlockFreq(TailCallBB)).getFrequency());
2435 ModifiedDT = Changed = true;
2436 ++NumRetsDup;
2437 }
2438
2439 // If we eliminated all predecessors of the block, delete the block now.
2440 if (Changed && !BB->hasAddressTaken() && pred_empty(BB))
2441 BB->eraseFromParent();
2442
2443 return Changed;
2444}
2445
2446//===----------------------------------------------------------------------===//
2447// Memory Optimization
2448//===----------------------------------------------------------------------===//
2449
2450namespace {
2451
2452/// This is an extended version of TargetLowering::AddrMode
2453/// which holds actual Value*'s for register values.
2454struct ExtAddrMode : public TargetLowering::AddrMode {
2455 Value *BaseReg = nullptr;
2456 Value *ScaledReg = nullptr;
2457 Value *OriginalValue = nullptr;
2458 bool InBounds = true;
2459
2460 enum FieldName {
2461 NoField = 0x00,
2462 BaseRegField = 0x01,
2463 BaseGVField = 0x02,
2464 BaseOffsField = 0x04,
2465 ScaledRegField = 0x08,
2466 ScaleField = 0x10,
2467 MultipleFields = 0xff
2468 };
2469
2470 ExtAddrMode() = default;
2471
2472 void print(raw_ostream &OS) const;
2473 void dump() const;
2474
2475 FieldName compare(const ExtAddrMode &other) {
2476 // First check that the types are the same on each field, as differing types
2477 // is something we can't cope with later on.
2478 if (BaseReg && other.BaseReg &&
2479 BaseReg->getType() != other.BaseReg->getType())
2480 return MultipleFields;
2481 if (BaseGV && other.BaseGV && BaseGV->getType() != other.BaseGV->getType())
2482 return MultipleFields;
2483 if (ScaledReg && other.ScaledReg &&
2484 ScaledReg->getType() != other.ScaledReg->getType())
2485 return MultipleFields;
2486
2487 // Conservatively reject 'inbounds' mismatches.
2488 if (InBounds != other.InBounds)
2489 return MultipleFields;
2490
2491 // Check each field to see if it differs.
2492 unsigned Result = NoField;
2493 if (BaseReg != other.BaseReg)
2494 Result |= BaseRegField;
2495 if (BaseGV != other.BaseGV)
2496 Result |= BaseGVField;
2497 if (BaseOffs != other.BaseOffs)
2498 Result |= BaseOffsField;
2499 if (ScaledReg != other.ScaledReg)
2500 Result |= ScaledRegField;
2501 // Don't count 0 as being a different scale, because that actually means
2502 // unscaled (which will already be counted by having no ScaledReg).
2503 if (Scale && other.Scale && Scale != other.Scale)
2504 Result |= ScaleField;
2505
2506 if (countPopulation(Result) > 1)
2507 return MultipleFields;
2508 else
2509 return static_cast<FieldName>(Result);
2510 }
2511
2512 // An AddrMode is trivial if it involves no calculation i.e. it is just a base
2513 // with no offset.
2514 bool isTrivial() {
2515 // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
2516 // trivial if at most one of these terms is nonzero, except that BaseGV and
2517 // BaseReg both being zero actually means a null pointer value, which we
2518 // consider to be 'non-zero' here.
2519 return !BaseOffs && !Scale && !(BaseGV && BaseReg);
2520 }
2521
2522 Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
2523 switch (Field) {
2524 default:
2525 return nullptr;
2526 case BaseRegField:
2527 return BaseReg;
2528 case BaseGVField:
2529 return BaseGV;
2530 case ScaledRegField:
2531 return ScaledReg;
2532 case BaseOffsField:
2533 return ConstantInt::get(IntPtrTy, BaseOffs);
2534 }
2535 }
2536
2537 void SetCombinedField(FieldName Field, Value *V,
2538 const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2539 switch (Field) {
2540 default:
2541 llvm_unreachable("Unhandled fields are expected to be rejected earlier")::llvm::llvm_unreachable_internal("Unhandled fields are expected to be rejected earlier"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 2541)
;
2542 break;
2543 case ExtAddrMode::BaseRegField:
2544 BaseReg = V;
2545 break;
2546 case ExtAddrMode::BaseGVField:
2547 // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2548 // in the BaseReg field.
2549 assert(BaseReg == nullptr)(static_cast <bool> (BaseReg == nullptr) ? void (0) : __assert_fail
("BaseReg == nullptr", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 2549, __extension__ __PRETTY_FUNCTION__))
;
2550 BaseReg = V;
2551 BaseGV = nullptr;
2552 break;
2553 case ExtAddrMode::ScaledRegField:
2554 ScaledReg = V;
2555 // If we have a mix of scaled and unscaled addrmodes then we want scale
2556 // to be the scale and not zero.
2557 if (!Scale)
2558 for (const ExtAddrMode &AM : AddrModes)
2559 if (AM.Scale) {
2560 Scale = AM.Scale;
2561 break;
2562 }
2563 break;
2564 case ExtAddrMode::BaseOffsField:
2565 // The offset is no longer a constant, so it goes in ScaledReg with a
2566 // scale of 1.
2567 assert(ScaledReg == nullptr)(static_cast <bool> (ScaledReg == nullptr) ? void (0) :
__assert_fail ("ScaledReg == nullptr", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 2567, __extension__ __PRETTY_FUNCTION__))
;
2568 ScaledReg = V;
2569 Scale = 1;
2570 BaseOffs = 0;
2571 break;
2572 }
2573 }
2574};
2575
2576#ifndef NDEBUG
2577static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2578 AM.print(OS);
2579 return OS;
2580}
2581#endif
2582
2583#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2584void ExtAddrMode::print(raw_ostream &OS) const {
2585 bool NeedPlus = false;
2586 OS << "[";
2587 if (InBounds)
2588 OS << "inbounds ";
2589 if (BaseGV) {
2590 OS << (NeedPlus ? " + " : "") << "GV:";
2591 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2592 NeedPlus = true;
2593 }
2594
2595 if (BaseOffs) {
2596 OS << (NeedPlus ? " + " : "") << BaseOffs;
2597 NeedPlus = true;
2598 }
2599
2600 if (BaseReg) {
2601 OS << (NeedPlus ? " + " : "") << "Base:";
2602 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2603 NeedPlus = true;
2604 }
2605 if (Scale) {
2606 OS << (NeedPlus ? " + " : "") << Scale << "*";
2607 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2608 }
2609
2610 OS << ']';
2611}
2612
2613LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void ExtAddrMode::dump() const {
2614 print(dbgs());
2615 dbgs() << '\n';
2616}
2617#endif
2618
2619} // end anonymous namespace
2620
2621namespace {
2622
2623/// This class provides transaction based operation on the IR.
2624/// Every change made through this class is recorded in the internal state and
2625/// can be undone (rollback) until commit is called.
2626/// CGP does not check if instructions could be speculatively executed when
2627/// moved. Preserving the original location would pessimize the debugging
2628/// experience, as well as negatively impact the quality of sample PGO.
2629class TypePromotionTransaction {
2630 /// This represents the common interface of the individual transaction.
2631 /// Each class implements the logic for doing one specific modification on
2632 /// the IR via the TypePromotionTransaction.
2633 class TypePromotionAction {
2634 protected:
2635 /// The Instruction modified.
2636 Instruction *Inst;
2637
2638 public:
2639 /// Constructor of the action.
2640 /// The constructor performs the related action on the IR.
2641 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2642
2643 virtual ~TypePromotionAction() = default;
2644
2645 /// Undo the modification done by this action.
2646 /// When this method is called, the IR must be in the same state as it was
2647 /// before this action was applied.
2648 /// \pre Undoing the action works if and only if the IR is in the exact same
2649 /// state as it was directly after this action was applied.
2650 virtual void undo() = 0;
2651
2652 /// Advocate every change made by this action.
2653 /// When the results on the IR of the action are to be kept, it is important
2654 /// to call this function, otherwise hidden information may be kept forever.
2655 virtual void commit() {
2656 // Nothing to be done, this action is not doing anything.
2657 }
2658 };
2659
2660 /// Utility to remember the position of an instruction.
2661 class InsertionHandler {
2662 /// Position of an instruction.
2663 /// Either an instruction:
2664 /// - Is the first in a basic block: BB is used.
2665 /// - Has a previous instruction: PrevInst is used.
2666 union {
2667 Instruction *PrevInst;
2668 BasicBlock *BB;
2669 } Point;
2670
2671 /// Remember whether or not the instruction had a previous instruction.
2672 bool HasPrevInstruction;
2673
2674 public:
2675 /// Record the position of \p Inst.
2676 InsertionHandler(Instruction *Inst) {
2677 BasicBlock::iterator It = Inst->getIterator();
2678 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2679 if (HasPrevInstruction)
2680 Point.PrevInst = &*--It;
2681 else
2682 Point.BB = Inst->getParent();
2683 }
2684
2685 /// Insert \p Inst at the recorded position.
2686 void insert(Instruction *Inst) {
2687 if (HasPrevInstruction) {
2688 if (Inst->getParent())
2689 Inst->removeFromParent();
2690 Inst->insertAfter(Point.PrevInst);
2691 } else {
2692 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2693 if (Inst->getParent())
2694 Inst->moveBefore(Position);
2695 else
2696 Inst->insertBefore(Position);
2697 }
2698 }
2699 };
2700
2701 /// Move an instruction before another.
2702 class InstructionMoveBefore : public TypePromotionAction {
2703 /// Original position of the instruction.
2704 InsertionHandler Position;
2705
2706 public:
2707 /// Move \p Inst before \p Before.
2708 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2709 : TypePromotionAction(Inst), Position(Inst) {
2710 LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Beforedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: move: " << *
Inst << "\nbefore: " << *Before << "\n"; } }
while (false)
2711 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: move: " << *
Inst << "\nbefore: " << *Before << "\n"; } }
while (false)
;
2712 Inst->moveBefore(Before);
2713 }
2714
2715 /// Move the instruction back to its original position.
2716 void undo() override {
2717 LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: moveBefore: " <<
*Inst << "\n"; } } while (false)
;
2718 Position.insert(Inst);
2719 }
2720 };
2721
2722 /// Set the operand of an instruction with a new value.
2723 class OperandSetter : public TypePromotionAction {
2724 /// Original operand of the instruction.
2725 Value *Origin;
2726
2727 /// Index of the modified instruction.
2728 unsigned Idx;
2729
2730 public:
2731 /// Set \p Idx operand of \p Inst with \p NewVal.
2732 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2733 : TypePromotionAction(Inst), Idx(Idx) {
2734 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)
2735 << "for:" << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
2736 << "with:" << *NewVal << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: setOperand: " <<
Idx << "\n" << "for:" << *Inst << "\n"
<< "with:" << *NewVal << "\n"; } } while (
false)
;
2737 Origin = Inst->getOperand(Idx);
2738 Inst->setOperand(Idx, NewVal);
2739 }
2740
2741 /// Restore the original value of the instruction.
2742 void undo() override {
2743 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)
2744 << "for: " << *Inst << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
2745 << "with: " << *Origin << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: setOperand:" <<
Idx << "\n" << "for: " << *Inst << "\n"
<< "with: " << *Origin << "\n"; } } while (
false)
;
2746 Inst->setOperand(Idx, Origin);
2747 }
2748 };
2749
2750 /// Hide the operands of an instruction.
2751 /// Do as if this instruction was not using any of its operands.
2752 class OperandsHider : public TypePromotionAction {
2753 /// The list of original operands.
2754 SmallVector<Value *, 4> OriginalValues;
2755
2756 public:
2757 /// Remove \p Inst from the uses of the operands of \p Inst.
2758 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2759 LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: OperandsHider: " <<
*Inst << "\n"; } } while (false)
;
2760 unsigned NumOpnds = Inst->getNumOperands();
2761 OriginalValues.reserve(NumOpnds);
2762 for (unsigned It = 0; It < NumOpnds; ++It) {
2763 // Save the current operand.
2764 Value *Val = Inst->getOperand(It);
2765 OriginalValues.push_back(Val);
2766 // Set a dummy one.
2767 // We could use OperandSetter here, but that would imply an overhead
2768 // that we are not willing to pay.
2769 Inst->setOperand(It, UndefValue::get(Val->getType()));
2770 }
2771 }
2772
2773 /// Restore the original list of uses.
2774 void undo() override {
2775 LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: OperandsHider: "
<< *Inst << "\n"; } } while (false)
;
2776 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2777 Inst->setOperand(It, OriginalValues[It]);
2778 }
2779 };
2780
2781 /// Build a truncate instruction.
2782 class TruncBuilder : public TypePromotionAction {
2783 Value *Val;
2784
2785 public:
2786 /// Build a truncate instruction of \p Opnd producing a \p Ty
2787 /// result.
2788 /// trunc Opnd to Ty.
2789 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2790 IRBuilder<> Builder(Opnd);
2791 Builder.SetCurrentDebugLocation(DebugLoc());
2792 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2793 LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2794 }
2795
2796 /// Get the built value.
2797 Value *getBuiltValue() { return Val; }
2798
2799 /// Remove the built instruction.
2800 void undo() override {
2801 LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: TruncBuilder: " <<
*Val << "\n"; } } while (false)
;
2802 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2803 IVal->eraseFromParent();
2804 }
2805 };
2806
2807 /// Build a sign extension instruction.
2808 class SExtBuilder : public TypePromotionAction {
2809 Value *Val;
2810
2811 public:
2812 /// Build a sign extension instruction of \p Opnd producing a \p Ty
2813 /// result.
2814 /// sext Opnd to Ty.
2815 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2816 : TypePromotionAction(InsertPt) {
2817 IRBuilder<> Builder(InsertPt);
2818 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2819 LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2820 }
2821
2822 /// Get the built value.
2823 Value *getBuiltValue() { return Val; }
2824
2825 /// Remove the built instruction.
2826 void undo() override {
2827 LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: SExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2828 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2829 IVal->eraseFromParent();
2830 }
2831 };
2832
2833 /// Build a zero extension instruction.
2834 class ZExtBuilder : public TypePromotionAction {
2835 Value *Val;
2836
2837 public:
2838 /// Build a zero extension instruction of \p Opnd producing a \p Ty
2839 /// result.
2840 /// zext Opnd to Ty.
2841 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2842 : TypePromotionAction(InsertPt) {
2843 IRBuilder<> Builder(InsertPt);
2844 Builder.SetCurrentDebugLocation(DebugLoc());
2845 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2846 LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2847 }
2848
2849 /// Get the built value.
2850 Value *getBuiltValue() { return Val; }
2851
2852 /// Remove the built instruction.
2853 void undo() override {
2854 LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: ZExtBuilder: " <<
*Val << "\n"; } } while (false)
;
2855 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2856 IVal->eraseFromParent();
2857 }
2858 };
2859
2860 /// Mutate an instruction to another type.
2861 class TypeMutator : public TypePromotionAction {
2862 /// Record the original type.
2863 Type *OrigTy;
2864
2865 public:
2866 /// Mutate the type of \p Inst into \p NewTy.
2867 TypeMutator(Instruction *Inst, Type *NewTy)
2868 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2869 LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
2870 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: MutateType: " <<
*Inst << " with " << *NewTy << "\n"; } } while
(false)
;
2871 Inst->mutateType(NewTy);
2872 }
2873
2874 /// Mutate the instruction back to its original type.
2875 void undo() override {
2876 LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
2877 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: MutateType: " <<
*Inst << " with " << *OrigTy << "\n"; } } while
(false)
;
2878 Inst->mutateType(OrigTy);
2879 }
2880 };
2881
2882 /// Replace the uses of an instruction by another instruction.
2883 class UsesReplacer : public TypePromotionAction {
2884 /// Helper structure to keep track of the replaced uses.
2885 struct InstructionAndIdx {
2886 /// The instruction using the instruction.
2887 Instruction *Inst;
2888
2889 /// The index where this instruction is used for Inst.
2890 unsigned Idx;
2891
2892 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2893 : Inst(Inst), Idx(Idx) {}
2894 };
2895
2896 /// Keep track of the original uses (pair Instruction, Index).
2897 SmallVector<InstructionAndIdx, 4> OriginalUses;
2898 /// Keep track of the debug users.
2899 SmallVector<DbgValueInst *, 1> DbgValues;
2900
2901 /// Keep track of the new value so that we can undo it by replacing
2902 /// instances of the new value with the original value.
2903 Value *New;
2904
2905 using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2906
2907 public:
2908 /// Replace all the use of \p Inst by \p New.
2909 UsesReplacer(Instruction *Inst, Value *New)
2910 : TypePromotionAction(Inst), New(New) {
2911 LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
2912 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: UsersReplacer: " <<
*Inst << " with " << *New << "\n"; } } while
(false)
;
2913 // Record the original uses.
2914 for (Use &U : Inst->uses()) {
2915 Instruction *UserI = cast<Instruction>(U.getUser());
2916 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2917 }
2918 // Record the debug uses separately. They are not in the instruction's
2919 // use list, but they are replaced by RAUW.
2920 findDbgValues(DbgValues, Inst);
2921
2922 // Now, we can replace the uses.
2923 Inst->replaceAllUsesWith(New);
2924 }
2925
2926 /// Reassign the original uses of Inst to Inst.
2927 void undo() override {
2928 LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: UsersReplacer: "
<< *Inst << "\n"; } } while (false)
;
2929 for (InstructionAndIdx &Use : OriginalUses)
2930 Use.Inst->setOperand(Use.Idx, Inst);
2931 // RAUW has replaced all original uses with references to the new value,
2932 // including the debug uses. Since we are undoing the replacements,
2933 // the original debug uses must also be reinstated to maintain the
2934 // correctness and utility of debug value instructions.
2935 for (auto *DVI : DbgValues)
2936 DVI->replaceVariableLocationOp(New, Inst);
2937 }
2938 };
2939
2940 /// Remove an instruction from the IR.
2941 class InstructionRemover : public TypePromotionAction {
2942 /// Original position of the instruction.
2943 InsertionHandler Inserter;
2944
2945 /// Helper structure to hide all the link to the instruction. In other
2946 /// words, this helps to do as if the instruction was removed.
2947 OperandsHider Hider;
2948
2949 /// Keep track of the uses replaced, if any.
2950 UsesReplacer *Replacer = nullptr;
2951
2952 /// Keep track of instructions removed.
2953 SetOfInstrs &RemovedInsts;
2954
2955 public:
2956 /// Remove all reference of \p Inst and optionally replace all its
2957 /// uses with New.
2958 /// \p RemovedInsts Keep track of the instructions removed by this Action.
2959 /// \pre If !Inst->use_empty(), then New != nullptr
2960 InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2961 Value *New = nullptr)
2962 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2963 RemovedInsts(RemovedInsts) {
2964 if (New)
2965 Replacer = new UsesReplacer(Inst, New);
2966 LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Do: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2967 RemovedInsts.insert(Inst);
2968 /// The instructions removed here will be freed after completing
2969 /// optimizeBlock() for all blocks as we need to keep track of the
2970 /// removed instructions during promotion.
2971 Inst->removeFromParent();
2972 }
2973
2974 ~InstructionRemover() override { delete Replacer; }
2975
2976 /// Resurrect the instruction and reassign it to the proper uses if
2977 /// new value was provided when build this action.
2978 void undo() override {
2979 LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Undo: InstructionRemover: "
<< *Inst << "\n"; } } while (false)
;
2980 Inserter.insert(Inst);
2981 if (Replacer)
2982 Replacer->undo();
2983 Hider.undo();
2984 RemovedInsts.erase(Inst);
2985 }
2986 };
2987
2988public:
2989 /// Restoration point.
2990 /// The restoration point is a pointer to an action instead of an iterator
2991 /// because the iterator may be invalidated but not the pointer.
2992 using ConstRestorationPt = const TypePromotionAction *;
2993
2994 TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2995 : RemovedInsts(RemovedInsts) {}
2996
2997 /// Advocate every changes made in that transaction. Return true if any change
2998 /// happen.
2999 bool commit();
3000
3001 /// Undo all the changes made after the given point.
3002 void rollback(ConstRestorationPt Point);
3003
3004 /// Get the current restoration point.
3005 ConstRestorationPt getRestorationPoint() const;
3006
3007 /// \name API for IR modification with state keeping to support rollback.
3008 /// @{
3009 /// Same as Instruction::setOperand.
3010 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3011
3012 /// Same as Instruction::eraseFromParent.
3013 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3014
3015 /// Same as Value::replaceAllUsesWith.
3016 void replaceAllUsesWith(Instruction *Inst, Value *New);
3017
3018 /// Same as Value::mutateType.
3019 void mutateType(Instruction *Inst, Type *NewTy);
3020
3021 /// Same as IRBuilder::createTrunc.
3022 Value *createTrunc(Instruction *Opnd, Type *Ty);
3023
3024 /// Same as IRBuilder::createSExt.
3025 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3026
3027 /// Same as IRBuilder::createZExt.
3028 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3029
3030 /// Same as Instruction::moveBefore.
3031 void moveBefore(Instruction *Inst, Instruction *Before);
3032 /// @}
3033
3034private:
3035 /// The ordered list of actions made so far.
3036 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3037
3038 using CommitPt =
3039 SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
3040
3041 SetOfInstrs &RemovedInsts;
3042};
3043
3044} // end anonymous namespace
3045
3046void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3047 Value *NewVal) {
3048 Actions.push_back(std::make_unique<TypePromotionTransaction::OperandSetter>(
3049 Inst, Idx, NewVal));
3050}
3051
3052void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3053 Value *NewVal) {
3054 Actions.push_back(
3055 std::make_unique<TypePromotionTransaction::InstructionRemover>(
3056 Inst, RemovedInsts, NewVal));
3057}
3058
3059void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3060 Value *New) {
3061 Actions.push_back(
3062 std::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3063}
3064
3065void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3066 Actions.push_back(
3067 std::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3068}
3069
3070Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, Type *Ty) {
3071 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3072 Value *Val = Ptr->getBuiltValue();
3073 Actions.push_back(std::move(Ptr));
3074 return Val;
3075}
3076
3077Value *TypePromotionTransaction::createSExt(Instruction *Inst, Value *Opnd,
3078 Type *Ty) {
3079 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3080 Value *Val = Ptr->getBuiltValue();
3081 Actions.push_back(std::move(Ptr));
3082 return Val;
3083}
3084
3085Value *TypePromotionTransaction::createZExt(Instruction *Inst, Value *Opnd,
3086 Type *Ty) {
3087 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3088 Value *Val = Ptr->getBuiltValue();
3089 Actions.push_back(std::move(Ptr));
3090 return Val;
3091}
3092
3093void TypePromotionTransaction::moveBefore(Instruction *Inst,
3094 Instruction *Before) {
3095 Actions.push_back(
3096 std::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
3097 Inst, Before));
3098}
3099
3100TypePromotionTransaction::ConstRestorationPt
3101TypePromotionTransaction::getRestorationPoint() const {
3102 return !Actions.empty() ? Actions.back().get() : nullptr;
3103}
3104
3105bool TypePromotionTransaction::commit() {
3106 for (std::unique_ptr<TypePromotionAction> &Action : Actions)
3107 Action->commit();
3108 bool Modified = !Actions.empty();
3109 Actions.clear();
3110 return Modified;
3111}
3112
3113void TypePromotionTransaction::rollback(
3114 TypePromotionTransaction::ConstRestorationPt Point) {
3115 while (!Actions.empty() && Point != Actions.back().get()) {
3116 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3117 Curr->undo();
3118 }
3119}
3120
3121namespace {
3122
3123/// A helper class for matching addressing modes.
3124///
3125/// This encapsulates the logic for matching the target-legal addressing modes.
3126class AddressingModeMatcher {
3127 SmallVectorImpl<Instruction *> &AddrModeInsts;
3128 const TargetLowering &TLI;
3129 const TargetRegisterInfo &TRI;
3130 const DataLayout &DL;
3131 const LoopInfo &LI;
3132 const std::function<const DominatorTree &()> getDTFn;
3133
3134 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3135 /// the memory instruction that we're computing this address for.
3136 Type *AccessTy;
3137 unsigned AddrSpace;
3138 Instruction *MemoryInst;
3139
3140 /// This is the addressing mode that we're building up. This is
3141 /// part of the return value of this addressing mode matching stuff.
3142 ExtAddrMode &AddrMode;
3143
3144 /// The instructions inserted by other CodeGenPrepare optimizations.
3145 const SetOfInstrs &InsertedInsts;
3146
3147 /// A map from the instructions to their type before promotion.
3148 InstrToOrigTy &PromotedInsts;
3149
3150 /// The ongoing transaction where every action should be registered.
3151 TypePromotionTransaction &TPT;
3152
3153 // A GEP which has too large offset to be folded into the addressing mode.
3154 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
3155
3156 /// This is set to true when we should not do profitability checks.
3157 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3158 bool IgnoreProfitability;
3159
3160 /// True if we are optimizing for size.
3161 bool OptSize;
3162
3163 ProfileSummaryInfo *PSI;
3164 BlockFrequencyInfo *BFI;
3165
3166 AddressingModeMatcher(
3167 SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI,
3168 const TargetRegisterInfo &TRI, const LoopInfo &LI,
3169 const std::function<const DominatorTree &()> getDTFn, Type *AT,
3170 unsigned AS, Instruction *MI, ExtAddrMode &AM,
3171 const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
3172 TypePromotionTransaction &TPT,
3173 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3174 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI)
3175 : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
3176 DL(MI->getModule()->getDataLayout()), LI(LI), getDTFn(getDTFn),
3177 AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM),
3178 InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT),
3179 LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) {
3180 IgnoreProfitability = false;
3181 }
3182
3183public:
3184 /// Find the maximal addressing mode that a load/store of V can fold,
3185 /// give an access type of AccessTy. This returns a list of involved
3186 /// instructions in AddrModeInsts.
3187 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3188 /// optimizations.
3189 /// \p PromotedInsts maps the instructions to their type before promotion.
3190 /// \p The ongoing transaction where every action should be registered.
3191 static ExtAddrMode
3192 Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
3193 SmallVectorImpl<Instruction *> &AddrModeInsts,
3194 const TargetLowering &TLI, const LoopInfo &LI,
3195 const std::function<const DominatorTree &()> getDTFn,
3196 const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts,
3197 InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
3198 std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP,
3199 bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) {
3200 ExtAddrMode Result;
3201
3202 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, LI, getDTFn,
3203 AccessTy, AS, MemoryInst, Result,
3204 InsertedInsts, PromotedInsts, TPT,
3205 LargeOffsetGEP, OptSize, PSI, BFI)
3206 .matchAddr(V, 0);
3207 (void)Success;
3208 assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?"
) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3208, __extension__ __PRETTY_FUNCTION__
))
;
3209 return Result;
3210 }
3211
3212private:
3213 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3214 bool matchAddr(Value *Addr, unsigned Depth);
3215 bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
3216 bool *MovedAway = nullptr);
3217 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3218 ExtAddrMode &AMBefore,
3219 ExtAddrMode &AMAfter);
3220 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3221 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3222 Value *PromotedOperand) const;
3223};
3224
3225class PhiNodeSet;
3226
3227/// An iterator for PhiNodeSet.
3228class PhiNodeSetIterator {
3229 PhiNodeSet *const Set;
3230 size_t CurrentIndex = 0;
3231
3232public:
3233 /// The constructor. Start should point to either a valid element, or be equal
3234 /// to the size of the underlying SmallVector of the PhiNodeSet.
3235 PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start);
3236 PHINode *operator*() const;
3237 PhiNodeSetIterator &operator++();
3238 bool operator==(const PhiNodeSetIterator &RHS) const;
3239 bool operator!=(const PhiNodeSetIterator &RHS) const;
3240};
3241
3242/// Keeps a set of PHINodes.
3243///
3244/// This is a minimal set implementation for a specific use case:
3245/// It is very fast when there are very few elements, but also provides good
3246/// performance when there are many. It is similar to SmallPtrSet, but also
3247/// provides iteration by insertion order, which is deterministic and stable
3248/// across runs. It is also similar to SmallSetVector, but provides removing
3249/// elements in O(1) time. This is achieved by not actually removing the element
3250/// from the underlying vector, so comes at the cost of using more memory, but
3251/// that is fine, since PhiNodeSets are used as short lived objects.
3252class PhiNodeSet {
3253 friend class PhiNodeSetIterator;
3254
3255 using MapType = SmallDenseMap<PHINode *, size_t, 32>;
3256 using iterator = PhiNodeSetIterator;
3257
3258 /// Keeps the elements in the order of their insertion in the underlying
3259 /// vector. To achieve constant time removal, it never deletes any element.
3260 SmallVector<PHINode *, 32> NodeList;
3261
3262 /// Keeps the elements in the underlying set implementation. This (and not the
3263 /// NodeList defined above) is the source of truth on whether an element
3264 /// is actually in the collection.
3265 MapType NodeMap;
3266
3267 /// Points to the first valid (not deleted) element when the set is not empty
3268 /// and the value is not zero. Equals to the size of the underlying vector
3269 /// when the set is empty. When the value is 0, as in the beginning, the
3270 /// first element may or may not be valid.
3271 size_t FirstValidElement = 0;
3272
3273public:
3274 /// Inserts a new element to the collection.
3275 /// \returns true if the element is actually added, i.e. was not in the
3276 /// collection before the operation.
3277 bool insert(PHINode *Ptr) {
3278 if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
3279 NodeList.push_back(Ptr);
3280 return true;
3281 }
3282 return false;
3283 }
3284
3285 /// Removes the element from the collection.
3286 /// \returns whether the element is actually removed, i.e. was in the
3287 /// collection before the operation.
3288 bool erase(PHINode *Ptr) {
3289 if (NodeMap.erase(Ptr)) {
3290 SkipRemovedElements(FirstValidElement);
3291 return true;
3292 }
3293 return false;
3294 }
3295
3296 /// Removes all elements and clears the collection.
3297 void clear() {
3298 NodeMap.clear();
3299 NodeList.clear();
3300 FirstValidElement = 0;
3301 }
3302
3303 /// \returns an iterator that will iterate the elements in the order of
3304 /// insertion.
3305 iterator begin() {
3306 if (FirstValidElement == 0)
3307 SkipRemovedElements(FirstValidElement);
3308 return PhiNodeSetIterator(this, FirstValidElement);
3309 }
3310
3311 /// \returns an iterator that points to the end of the collection.
3312 iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
3313
3314 /// Returns the number of elements in the collection.
3315 size_t size() const { return NodeMap.size(); }
3316
3317 /// \returns 1 if the given element is in the collection, and 0 if otherwise.
3318 size_t count(PHINode *Ptr) const { return NodeMap.count(Ptr); }
3319
3320private:
3321 /// Updates the CurrentIndex so that it will point to a valid element.
3322 ///
3323 /// If the element of NodeList at CurrentIndex is valid, it does not
3324 /// change it. If there are no more valid elements, it updates CurrentIndex
3325 /// to point to the end of the NodeList.
3326 void SkipRemovedElements(size_t &CurrentIndex) {
3327 while (CurrentIndex < NodeList.size()) {
3328 auto it = NodeMap.find(NodeList[CurrentIndex]);
3329 // If the element has been deleted and added again later, NodeMap will
3330 // point to a different index, so CurrentIndex will still be invalid.
3331 if (it != NodeMap.end() && it->second == CurrentIndex)
3332 break;
3333 ++CurrentIndex;
3334 }
3335 }
3336};
3337
3338PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
3339 : Set(Set), CurrentIndex(Start) {}
3340
3341PHINode *PhiNodeSetIterator::operator*() const {
3342 assert(CurrentIndex < Set->NodeList.size() &&(static_cast <bool> (CurrentIndex < Set->NodeList
.size() && "PhiNodeSet access out of range") ? void (
0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3343, __extension__ __PRETTY_FUNCTION__
))
3343 "PhiNodeSet access out of range")(static_cast <bool> (CurrentIndex < Set->NodeList
.size() && "PhiNodeSet access out of range") ? void (
0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3343, __extension__ __PRETTY_FUNCTION__
))
;
3344 return Set->NodeList[CurrentIndex];
3345}
3346
3347PhiNodeSetIterator &PhiNodeSetIterator::operator++() {
3348 assert(CurrentIndex < Set->NodeList.size() &&(static_cast <bool> (CurrentIndex < Set->NodeList
.size() && "PhiNodeSet access out of range") ? void (
0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3349, __extension__ __PRETTY_FUNCTION__
))
3349 "PhiNodeSet access out of range")(static_cast <bool> (CurrentIndex < Set->NodeList
.size() && "PhiNodeSet access out of range") ? void (
0) : __assert_fail ("CurrentIndex < Set->NodeList.size() && \"PhiNodeSet access out of range\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3349, __extension__ __PRETTY_FUNCTION__
))
;
3350 ++CurrentIndex;
3351 Set->SkipRemovedElements(CurrentIndex);
3352 return *this;
3353}
3354
3355bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
3356 return CurrentIndex == RHS.CurrentIndex;
3357}
3358
3359bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
3360 return !((*this) == RHS);
3361}
3362
3363/// Keep track of simplification of Phi nodes.
3364/// Accept the set of all phi nodes and erase phi node from this set
3365/// if it is simplified.
3366class SimplificationTracker {
3367 DenseMap<Value *, Value *> Storage;
3368 const SimplifyQuery &SQ;
3369 // Tracks newly created Phi nodes. The elements are iterated by insertion
3370 // order.
3371 PhiNodeSet AllPhiNodes;
3372 // Tracks newly created Select nodes.
3373 SmallPtrSet<SelectInst *, 32> AllSelectNodes;
3374
3375public:
3376 SimplificationTracker(const SimplifyQuery &sq) : SQ(sq) {}
3377
3378 Value *Get(Value *V) {
3379 do {
3380 auto SV = Storage.find(V);
3381 if (SV == Storage.end())
3382 return V;
3383 V = SV->second;
3384 } while (true);
3385 }
3386
3387 Value *Simplify(Value *Val) {
3388 SmallVector<Value *, 32> WorkList;
3389 SmallPtrSet<Value *, 32> Visited;
3390 WorkList.push_back(Val);
3391 while (!WorkList.empty()) {
3392 auto *P = WorkList.pop_back_val();
3393 if (!Visited.insert(P).second)
3394 continue;
3395 if (auto *PI = dyn_cast<Instruction>(P))
3396 if (Value *V = simplifyInstruction(cast<Instruction>(PI), SQ)) {
3397 for (auto *U : PI->users())
3398 WorkList.push_back(cast<Value>(U));
3399 Put(PI, V);
3400 PI->replaceAllUsesWith(V);
3401 if (auto *PHI = dyn_cast<PHINode>(PI))
3402 AllPhiNodes.erase(PHI);
3403 if (auto *Select = dyn_cast<SelectInst>(PI))
3404 AllSelectNodes.erase(Select);
3405 PI->eraseFromParent();
3406 }
3407 }
3408 return Get(Val);
3409 }
3410
3411 void Put(Value *From, Value *To) { Storage.insert({From, To}); }
3412
3413 void ReplacePhi(PHINode *From, PHINode *To) {
3414 Value *OldReplacement = Get(From);
3415 while (OldReplacement != From) {
3416 From = To;
3417 To = dyn_cast<PHINode>(OldReplacement);
3418 OldReplacement = Get(From);
3419 }
3420 assert(To && Get(To) == To && "Replacement PHI node is already replaced.")(static_cast <bool> (To && Get(To) == To &&
"Replacement PHI node is already replaced.") ? void (0) : __assert_fail
("To && Get(To) == To && \"Replacement PHI node is already replaced.\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3420, __extension__ __PRETTY_FUNCTION__
))
;
3421 Put(From, To);
3422 From->replaceAllUsesWith(To);
3423 AllPhiNodes.erase(From);
3424 From->eraseFromParent();
3425 }
3426
3427 PhiNodeSet &newPhiNodes() { return AllPhiNodes; }
3428
3429 void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
3430
3431 void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
3432
3433 unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
3434
3435 unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
3436
3437 void destroyNewNodes(Type *CommonType) {
3438 // For safe erasing, replace the uses with dummy value first.
3439 auto *Dummy = PoisonValue::get(CommonType);
3440 for (auto *I : AllPhiNodes) {
3441 I->replaceAllUsesWith(Dummy);
3442 I->eraseFromParent();
3443 }
3444 AllPhiNodes.clear();
3445 for (auto *I : AllSelectNodes) {
3446 I->replaceAllUsesWith(Dummy);
3447 I->eraseFromParent();
3448 }
3449 AllSelectNodes.clear();
3450 }
3451};
3452
3453/// A helper class for combining addressing modes.
3454class AddressingModeCombiner {
3455 typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
3456 typedef std::pair<PHINode *, PHINode *> PHIPair;
3457
3458private:
3459 /// The addressing modes we've collected.
3460 SmallVector<ExtAddrMode, 16> AddrModes;
3461
3462 /// The field in which the AddrModes differ, when we have more than one.
3463 ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
3464
3465 /// Are the AddrModes that we have all just equal to their original values?
3466 bool AllAddrModesTrivial = true;
3467
3468 /// Common Type for all different fields in addressing modes.
3469 Type *CommonType = nullptr;
3470
3471 /// SimplifyQuery for simplifyInstruction utility.
3472 const SimplifyQuery &SQ;
3473
3474 /// Original Address.
3475 Value *Original;
3476
3477public:
3478 AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
3479 : SQ(_SQ), Original(OriginalValue) {}
3480
3481 /// Get the combined AddrMode
3482 const ExtAddrMode &getAddrMode() const { return AddrModes[0]; }
3483
3484 /// Add a new AddrMode if it's compatible with the AddrModes we already
3485 /// have.
3486 /// \return True iff we succeeded in doing so.
3487 bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
3488 // Take note of if we have any non-trivial AddrModes, as we need to detect
3489 // when all AddrModes are trivial as then we would introduce a phi or select
3490 // which just duplicates what's already there.
3491 AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
3492
3493 // If this is the first addrmode then everything is fine.
3494 if (AddrModes.empty()) {
3495 AddrModes.emplace_back(NewAddrMode);
3496 return true;
3497 }
3498
3499 // Figure out how different this is from the other address modes, which we
3500 // can do just by comparing against the first one given that we only care
3501 // about the cumulative difference.
3502 ExtAddrMode::FieldName ThisDifferentField =
3503 AddrModes[0].compare(NewAddrMode);
3504 if (DifferentField == ExtAddrMode::NoField)
3505 DifferentField = ThisDifferentField;
3506 else if (DifferentField != ThisDifferentField)
3507 DifferentField = ExtAddrMode::MultipleFields;
3508
3509 // If NewAddrMode differs in more than one dimension we cannot handle it.
3510 bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
3511
3512 // If Scale Field is different then we reject.
3513 CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
3514
3515 // We also must reject the case when base offset is different and
3516 // scale reg is not null, we cannot handle this case due to merge of
3517 // different offsets will be used as ScaleReg.
3518 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
3519 !NewAddrMode.ScaledReg);
3520
3521 // We also must reject the case when GV is different and BaseReg installed
3522 // due to we want to use base reg as a merge of GV values.
3523 CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
3524 !NewAddrMode.HasBaseReg);
3525
3526 // Even if NewAddMode is the same we still need to collect it due to
3527 // original value is different. And later we will need all original values
3528 // as anchors during finding the common Phi node.
3529 if (CanHandle)
3530 AddrModes.emplace_back(NewAddrMode);
3531 else
3532 AddrModes.clear();
3533
3534 return CanHandle;
3535 }
3536
3537 /// Combine the addressing modes we've collected into a single
3538 /// addressing mode.
3539 /// \return True iff we successfully combined them or we only had one so
3540 /// didn't need to combine them anyway.
3541 bool combineAddrModes() {
3542 // If we have no AddrModes then they can't be combined.
3543 if (AddrModes.size() == 0)
3544 return false;
3545
3546 // A single AddrMode can trivially be combined.
3547 if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3548 return true;
3549
3550 // If the AddrModes we collected are all just equal to the value they are
3551 // derived from then combining them wouldn't do anything useful.
3552 if (AllAddrModesTrivial)
3553 return false;
3554
3555 if (!addrModeCombiningAllowed())
3556 return false;
3557
3558 // Build a map between <original value, basic block where we saw it> to
3559 // value of base register.
3560 // Bail out if there is no common type.
3561 FoldAddrToValueMapping Map;
3562 if (!initializeMap(Map))
3563 return false;
3564
3565 Value *CommonValue = findCommon(Map);
3566 if (CommonValue)
3567 AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3568 return CommonValue != nullptr;
3569 }
3570
3571private:
3572 /// Initialize Map with anchor values. For address seen
3573 /// we set the value of different field saw in this address.
3574 /// At the same time we find a common type for different field we will
3575 /// use to create new Phi/Select nodes. Keep it in CommonType field.
3576 /// Return false if there is no common type found.
3577 bool initializeMap(FoldAddrToValueMapping &Map) {
3578 // Keep track of keys where the value is null. We will need to replace it
3579 // with constant null when we know the common type.
3580 SmallVector<Value *, 2> NullValue;
3581 Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3582 for (auto &AM : AddrModes) {
3583 Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3584 if (DV) {
3585 auto *Type = DV->getType();
3586 if (CommonType && CommonType != Type)
3587 return false;
3588 CommonType = Type;
3589 Map[AM.OriginalValue] = DV;
3590 } else {
3591 NullValue.push_back(AM.OriginalValue);
3592 }
3593 }
3594 assert(CommonType && "At least one non-null value must be!")(static_cast <bool> (CommonType && "At least one non-null value must be!"
) ? void (0) : __assert_fail ("CommonType && \"At least one non-null value must be!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3594, __extension__ __PRETTY_FUNCTION__
))
;
3595 for (auto *V : NullValue)
3596 Map[V] = Constant::getNullValue(CommonType);
3597 return true;
3598 }
3599
3600 /// We have mapping between value A and other value B where B was a field in
3601 /// addressing mode represented by A. Also we have an original value C
3602 /// representing an address we start with. Traversing from C through phi and
3603 /// selects we ended up with A's in a map. This utility function tries to find
3604 /// a value V which is a field in addressing mode C and traversing through phi
3605 /// nodes and selects we will end up in corresponded values B in a map.
3606 /// The utility will create a new Phi/Selects if needed.
3607 // The simple example looks as follows:
3608 // BB1:
3609 // p1 = b1 + 40
3610 // br cond BB2, BB3
3611 // BB2:
3612 // p2 = b2 + 40
3613 // br BB3
3614 // BB3:
3615 // p = phi [p1, BB1], [p2, BB2]
3616 // v = load p
3617 // Map is
3618 // p1 -> b1
3619 // p2 -> b2
3620 // Request is
3621 // p -> ?
3622 // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
3623 Value *findCommon(FoldAddrToValueMapping &Map) {
3624 // Tracks the simplification of newly created phi nodes. The reason we use
3625 // this mapping is because we will add new created Phi nodes in AddrToBase.
3626 // Simplification of Phi nodes is recursive, so some Phi node may
3627 // be simplified after we added it to AddrToBase. In reality this
3628 // simplification is possible only if original phi/selects were not
3629 // simplified yet.
3630 // Using this mapping we can find the current value in AddrToBase.
3631 SimplificationTracker ST(SQ);
3632
3633 // First step, DFS to create PHI nodes for all intermediate blocks.
3634 // Also fill traverse order for the second step.
3635 SmallVector<Value *, 32> TraverseOrder;
3636 InsertPlaceholders(Map, TraverseOrder, ST);
3637
3638 // Second Step, fill new nodes by merged values and simplify if possible.
3639 FillPlaceholders(Map, TraverseOrder, ST);
3640
3641 if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3642 ST.destroyNewNodes(CommonType);
3643 return nullptr;
3644 }
3645
3646 // Now we'd like to match New Phi nodes to existed ones.
3647 unsigned PhiNotMatchedCount = 0;
3648 if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3649 ST.destroyNewNodes(CommonType);
3650 return nullptr;
3651 }
3652
3653 auto *Result = ST.Get(Map.find(Original)->second);
3654 if (Result) {
3655 NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3656 NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3657 }
3658 return Result;
3659 }
3660
3661 /// Try to match PHI node to Candidate.
3662 /// Matcher tracks the matched Phi nodes.
3663 bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3664 SmallSetVector<PHIPair, 8> &Matcher,
3665 PhiNodeSet &PhiNodesToMatch) {
3666 SmallVector<PHIPair, 8> WorkList;
3667 Matcher.insert({PHI, Candidate});
3668 SmallSet<PHINode *, 8> MatchedPHIs;
3669 MatchedPHIs.insert(PHI);
3670 WorkList.push_back({PHI, Candidate});
3671 SmallSet<PHIPair, 8> Visited;
3672 while (!WorkList.empty()) {
3673 auto Item = WorkList.pop_back_val();
3674 if (!Visited.insert(Item).second)
3675 continue;
3676 // We iterate over all incoming values to Phi to compare them.
3677 // If values are different and both of them Phi and the first one is a
3678 // Phi we added (subject to match) and both of them is in the same basic
3679 // block then we can match our pair if values match. So we state that
3680 // these values match and add it to work list to verify that.
3681 for (auto *B : Item.first->blocks()) {
3682 Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3683 Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3684 if (FirstValue == SecondValue)
3685 continue;
3686
3687 PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3688 PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3689
3690 // One of them is not Phi or
3691 // The first one is not Phi node from the set we'd like to match or
3692 // Phi nodes from different basic blocks then
3693 // we will not be able to match.
3694 if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3695 FirstPhi->getParent() != SecondPhi->getParent())
3696 return false;
3697
3698 // If we already matched them then continue.
3699 if (Matcher.count({FirstPhi, SecondPhi}))
3700 continue;
3701 // So the values are different and does not match. So we need them to
3702 // match. (But we register no more than one match per PHI node, so that
3703 // we won't later try to replace them twice.)
3704 if (MatchedPHIs.insert(FirstPhi).second)
3705 Matcher.insert({FirstPhi, SecondPhi});
3706 // But me must check it.
3707 WorkList.push_back({FirstPhi, SecondPhi});
3708 }
3709 }
3710 return true;
3711 }
3712
3713 /// For the given set of PHI nodes (in the SimplificationTracker) try
3714 /// to find their equivalents.
3715 /// Returns false if this matching fails and creation of new Phi is disabled.
3716 bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3717 unsigned &PhiNotMatchedCount) {
3718 // Matched and PhiNodesToMatch iterate their elements in a deterministic
3719 // order, so the replacements (ReplacePhi) are also done in a deterministic
3720 // order.
3721 SmallSetVector<PHIPair, 8> Matched;
3722 SmallPtrSet<PHINode *, 8> WillNotMatch;
3723 PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3724 while (PhiNodesToMatch.size()) {
3725 PHINode *PHI = *PhiNodesToMatch.begin();
3726
3727 // Add us, if no Phi nodes in the basic block we do not match.
3728 WillNotMatch.clear();
3729 WillNotMatch.insert(PHI);
3730
3731 // Traverse all Phis until we found equivalent or fail to do that.
3732 bool IsMatched = false;
3733 for (auto &P : PHI->getParent()->phis()) {
3734 // Skip new Phi nodes.
3735 if (PhiNodesToMatch.count(&P))
3736 continue;
3737 if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3738 break;
3739 // If it does not match, collect all Phi nodes from matcher.
3740 // if we end up with no match, them all these Phi nodes will not match
3741 // later.
3742 for (auto M : Matched)
3743 WillNotMatch.insert(M.first);
3744 Matched.clear();
3745 }
3746 if (IsMatched) {
3747 // Replace all matched values and erase them.
3748 for (auto MV : Matched)
3749 ST.ReplacePhi(MV.first, MV.second);
3750 Matched.clear();
3751 continue;
3752 }
3753 // If we are not allowed to create new nodes then bail out.
3754 if (!AllowNewPhiNodes)
3755 return false;
3756 // Just remove all seen values in matcher. They will not match anything.
3757 PhiNotMatchedCount += WillNotMatch.size();
3758 for (auto *P : WillNotMatch)
3759 PhiNodesToMatch.erase(P);
3760 }
3761 return true;
3762 }
3763 /// Fill the placeholders with values from predecessors and simplify them.
3764 void FillPlaceholders(FoldAddrToValueMapping &Map,
3765 SmallVectorImpl<Value *> &TraverseOrder,
3766 SimplificationTracker &ST) {
3767 while (!TraverseOrder.empty()) {
3768 Value *Current = TraverseOrder.pop_back_val();
3769 assert(Map.find(Current) != Map.end() && "No node to fill!!!")(static_cast <bool> (Map.find(Current) != Map.end() &&
"No node to fill!!!") ? void (0) : __assert_fail ("Map.find(Current) != Map.end() && \"No node to fill!!!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3769, __extension__ __PRETTY_FUNCTION__
))
;
3770 Value *V = Map[Current];
3771
3772 if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3773 // CurrentValue also must be Select.
3774 auto *CurrentSelect = cast<SelectInst>(Current);
3775 auto *TrueValue = CurrentSelect->getTrueValue();
3776 assert(Map.find(TrueValue) != Map.end() && "No True Value!")(static_cast <bool> (Map.find(TrueValue) != Map.end() &&
"No True Value!") ? void (0) : __assert_fail ("Map.find(TrueValue) != Map.end() && \"No True Value!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3776, __extension__ __PRETTY_FUNCTION__
))
;
3777 Select->setTrueValue(ST.Get(Map[TrueValue]));
3778 auto *FalseValue = CurrentSelect->getFalseValue();
3779 assert(Map.find(FalseValue) != Map.end() && "No False Value!")(static_cast <bool> (Map.find(FalseValue) != Map.end() &&
"No False Value!") ? void (0) : __assert_fail ("Map.find(FalseValue) != Map.end() && \"No False Value!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3779, __extension__ __PRETTY_FUNCTION__
))
;
3780 Select->setFalseValue(ST.Get(Map[FalseValue]));
3781 } else {
3782 // Must be a Phi node then.
3783 auto *PHI = cast<PHINode>(V);
3784 // Fill the Phi node with values from predecessors.
3785 for (auto *B : predecessors(PHI->getParent())) {
3786 Value *PV = cast<PHINode>(Current)->getIncomingValueForBlock(B);
3787 assert(Map.find(PV) != Map.end() && "No predecessor Value!")(static_cast <bool> (Map.find(PV) != Map.end() &&
"No predecessor Value!") ? void (0) : __assert_fail ("Map.find(PV) != Map.end() && \"No predecessor Value!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3787, __extension__ __PRETTY_FUNCTION__
))
;
3788 PHI->addIncoming(ST.Get(Map[PV]), B);
3789 }
3790 }
3791 Map[Current] = ST.Simplify(V);
3792 }
3793 }
3794
3795 /// Starting from original value recursively iterates over def-use chain up to
3796 /// known ending values represented in a map. For each traversed phi/select
3797 /// inserts a placeholder Phi or Select.
3798 /// Reports all new created Phi/Select nodes by adding them to set.
3799 /// Also reports and order in what values have been traversed.
3800 void InsertPlaceholders(FoldAddrToValueMapping &Map,
3801 SmallVectorImpl<Value *> &TraverseOrder,
3802 SimplificationTracker &ST) {
3803 SmallVector<Value *, 32> Worklist;
3804 assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&(static_cast <bool> ((isa<PHINode>(Original) || isa
<SelectInst>(Original)) && "Address must be a Phi or Select node"
) ? void (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3805, __extension__ __PRETTY_FUNCTION__
))
3805 "Address must be a Phi or Select node")(static_cast <bool> ((isa<PHINode>(Original) || isa
<SelectInst>(Original)) && "Address must be a Phi or Select node"
) ? void (0) : __assert_fail ("(isa<PHINode>(Original) || isa<SelectInst>(Original)) && \"Address must be a Phi or Select node\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3805, __extension__ __PRETTY_FUNCTION__
))
;
3806 auto *Dummy = PoisonValue::get(CommonType);
3807 Worklist.push_back(Original);
3808 while (!Worklist.empty()) {
3809 Value *Current = Worklist.pop_back_val();
3810 // if it is already visited or it is an ending value then skip it.
3811 if (Map.find(Current) != Map.end())
3812 continue;
3813 TraverseOrder.push_back(Current);
3814
3815 // CurrentValue must be a Phi node or select. All others must be covered
3816 // by anchors.
3817 if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3818 // Is it OK to get metadata from OrigSelect?!
3819 // Create a Select placeholder with dummy value.
3820 SelectInst *Select = SelectInst::Create(
3821 CurrentSelect->getCondition(), Dummy, Dummy,
3822 CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3823 Map[Current] = Select;
3824 ST.insertNewSelect(Select);
3825 // We are interested in True and False values.
3826 Worklist.push_back(CurrentSelect->getTrueValue());
3827 Worklist.push_back(CurrentSelect->getFalseValue());
3828 } else {
3829 // It must be a Phi node then.
3830 PHINode *CurrentPhi = cast<PHINode>(Current);
3831 unsigned PredCount = CurrentPhi->getNumIncomingValues();
3832 PHINode *PHI =
3833 PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3834 Map[Current] = PHI;
3835 ST.insertNewPhi(PHI);
3836 append_range(Worklist, CurrentPhi->incoming_values());
3837 }
3838 }
3839 }
3840
3841 bool addrModeCombiningAllowed() {
3842 if (DisableComplexAddrModes)
3843 return false;
3844 switch (DifferentField) {
3845 default:
3846 return false;
3847 case ExtAddrMode::BaseRegField:
3848 return AddrSinkCombineBaseReg;
3849 case ExtAddrMode::BaseGVField:
3850 return AddrSinkCombineBaseGV;
3851 case ExtAddrMode::BaseOffsField:
3852 return AddrSinkCombineBaseOffs;
3853 case ExtAddrMode::ScaledRegField:
3854 return AddrSinkCombineScaledReg;
3855 }
3856 }
3857};
3858} // end anonymous namespace
3859
3860/// Try adding ScaleReg*Scale to the current addressing mode.
3861/// Return true and update AddrMode if this addr mode is legal for the target,
3862/// false if not.
3863bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3864 unsigned Depth) {
3865 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3866 // mode. Just process that directly.
3867 if (Scale == 1)
3868 return matchAddr(ScaleReg, Depth);
3869
3870 // If the scale is 0, it takes nothing to add this.
3871 if (Scale == 0)
3872 return true;
3873
3874 // If we already have a scale of this value, we can add to it, otherwise, we
3875 // need an available scale field.
3876 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3877 return false;
3878
3879 ExtAddrMode TestAddrMode = AddrMode;
3880
3881 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3882 // [A+B + A*7] -> [B+A*8].
3883 TestAddrMode.Scale += Scale;
3884 TestAddrMode.ScaledReg = ScaleReg;
3885
3886 // If the new address isn't legal, bail out.
3887 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3888 return false;
3889
3890 // It was legal, so commit it.
3891 AddrMode = TestAddrMode;
3892
3893 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3894 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3895 // X*Scale + C*Scale to addr mode. If we found available IV increment, do not
3896 // go any further: we can reuse it and cannot eliminate it.
3897 ConstantInt *CI = nullptr;
3898 Value *AddLHS = nullptr;
3899 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3900 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI))) &&
3901 !isIVIncrement(ScaleReg, &LI) && CI->getValue().isSignedIntN(64)) {
3902 TestAddrMode.InBounds = false;
3903 TestAddrMode.ScaledReg = AddLHS;
3904 TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale;
3905
3906 // If this addressing mode is legal, commit it and remember that we folded
3907 // this instruction.
3908 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3909 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3910 AddrMode = TestAddrMode;
3911 return true;
3912 }
3913 // Restore status quo.
3914 TestAddrMode = AddrMode;
3915 }
3916
3917 // If this is an add recurrence with a constant step, return the increment
3918 // instruction and the canonicalized step.
3919 auto GetConstantStep =
3920 [this](const Value *V) -> Optional<std::pair<Instruction *, APInt>> {
3921 auto *PN = dyn_cast<PHINode>(V);
3922 if (!PN)
3923 return None;
3924 auto IVInc = getIVIncrement(PN, &LI);
3925 if (!IVInc)
3926 return None;
3927 // TODO: The result of the intrinsics above is two-compliment. However when
3928 // IV inc is expressed as add or sub, iv.next is potentially a poison value.
3929 // If it has nuw or nsw flags, we need to make sure that these flags are
3930 // inferrable at the point of memory instruction. Otherwise we are replacing
3931 // well-defined two-compliment computation with poison. Currently, to avoid
3932 // potentially complex analysis needed to prove this, we reject such cases.
3933 if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(IVInc->first))
3934 if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap())
3935 return None;
3936 if (auto *ConstantStep = dyn_cast<ConstantInt>(IVInc->second))
3937 return std::make_pair(IVInc->first, ConstantStep->getValue());
3938 return None;
3939 };
3940
3941 // Try to account for the following special case:
3942 // 1. ScaleReg is an inductive variable;
3943 // 2. We use it with non-zero offset;
3944 // 3. IV's increment is available at the point of memory instruction.
3945 //
3946 // In this case, we may reuse the IV increment instead of the IV Phi to
3947 // achieve the following advantages:
3948 // 1. If IV step matches the offset, we will have no need in the offset;
3949 // 2. Even if they don't match, we will reduce the overlap of living IV
3950 // and IV increment, that will potentially lead to better register
3951 // assignment.
3952 if (AddrMode.BaseOffs) {
3953 if (auto IVStep = GetConstantStep(ScaleReg)) {
3954 Instruction *IVInc = IVStep->first;
3955 // The following assert is important to ensure a lack of infinite loops.
3956 // This transforms is (intentionally) the inverse of the one just above.
3957 // If they don't agree on the definition of an increment, we'd alternate
3958 // back and forth indefinitely.
3959 assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep")(static_cast <bool> (isIVIncrement(IVInc, &LI) &&
"implied by GetConstantStep") ? void (0) : __assert_fail ("isIVIncrement(IVInc, &LI) && \"implied by GetConstantStep\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 3959, __extension__ __PRETTY_FUNCTION__
))
;
3960 APInt Step = IVStep->second;
3961 APInt Offset = Step * AddrMode.Scale;
3962 if (Offset.isSignedIntN(64)) {
3963 TestAddrMode.InBounds = false;
3964 TestAddrMode.ScaledReg = IVInc;
3965 TestAddrMode.BaseOffs -= Offset.getLimitedValue();
3966 // If this addressing mode is legal, commit it..
3967 // (Note that we defer the (expensive) domtree base legality check
3968 // to the very last possible point.)
3969 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace) &&
3970 getDTFn().dominates(IVInc, MemoryInst)) {
3971 AddrModeInsts.push_back(cast<Instruction>(IVInc));
3972 AddrMode = TestAddrMode;
3973 return true;
3974 }
3975 // Restore status quo.
3976 TestAddrMode = AddrMode;
3977 }
3978 }
3979 }
3980
3981 // Otherwise, just return what we have.
3982 return true;
3983}
3984
3985/// This is a little filter, which returns true if an addressing computation
3986/// involving I might be folded into a load/store accessing it.
3987/// This doesn't need to be perfect, but needs to accept at least
3988/// the set of instructions that MatchOperationAddr can.
3989static bool MightBeFoldableInst(Instruction *I) {
3990 switch (I->getOpcode()) {
3991 case Instruction::BitCast:
3992 case Instruction::AddrSpaceCast:
3993 // Don't touch identity bitcasts.
3994 if (I->getType() == I->getOperand(0)->getType())
3995 return false;
3996 return I->getType()->isIntOrPtrTy();
3997 case Instruction::PtrToInt:
3998 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3999 return true;
4000 case Instruction::IntToPtr:
4001 // We know the input is intptr_t, so this is foldable.
4002 return true;
4003 case Instruction::Add:
4004 return true;
4005 case Instruction::Mul:
4006 case Instruction::Shl:
4007 // Can only handle X*C and X << C.
4008 return isa<ConstantInt>(I->getOperand(1));
4009 case Instruction::GetElementPtr:
4010 return true;
4011 default:
4012 return false;
4013 }
4014}
4015
4016/// Check whether or not \p Val is a legal instruction for \p TLI.
4017/// \note \p Val is assumed to be the product of some type promotion.
4018/// Therefore if \p Val has an undefined state in \p TLI, this is assumed
4019/// to be legal, as the non-promoted value would have had the same state.
4020static bool isPromotedInstructionLegal(const TargetLowering &TLI,
4021 const DataLayout &DL, Value *Val) {
4022 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
4023 if (!PromotedInst)
4024 return false;
4025 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
4026 // If the ISDOpcode is undefined, it was undefined before the promotion.
4027 if (!ISDOpcode)
4028 return true;
4029 // Otherwise, check if the promoted instruction is legal or not.
4030 return TLI.isOperationLegalOrCustom(
4031 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
4032}
4033
4034namespace {
4035
4036/// Hepler class to perform type promotion.
4037class TypePromotionHelper {
4038 /// Utility function to add a promoted instruction \p ExtOpnd to
4039 /// \p PromotedInsts and record the type of extension we have seen.
4040 static void addPromotedInst(InstrToOrigTy &PromotedInsts,
4041 Instruction *ExtOpnd, bool IsSExt) {
4042 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4043 InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
4044 if (It != PromotedInsts.end()) {
4045 // If the new extension is same as original, the information in
4046 // PromotedInsts[ExtOpnd] is still correct.
4047 if (It->second.getInt() == ExtTy)
4048 return;
4049
4050 // Now the new extension is different from old extension, we make
4051 // the type information invalid by setting extension type to
4052 // BothExtension.
4053 ExtTy = BothExtension;
4054 }
4055 PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
4056 }
4057
4058 /// Utility function to query the original type of instruction \p Opnd
4059 /// with a matched extension type. If the extension doesn't match, we
4060 /// cannot use the information we had on the original type.
4061 /// BothExtension doesn't match any extension type.
4062 static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
4063 Instruction *Opnd, bool IsSExt) {
4064 ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
4065 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
4066 if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
4067 return It->second.getPointer();
4068 return nullptr;
4069 }
4070
4071 /// Utility function to check whether or not a sign or zero extension
4072 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
4073 /// either using the operands of \p Inst or promoting \p Inst.
4074 /// The type of the extension is defined by \p IsSExt.
4075 /// In other words, check if:
4076 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
4077 /// #1 Promotion applies:
4078 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
4079 /// #2 Operand reuses:
4080 /// ext opnd1 to ConsideredExtType.
4081 /// \p PromotedInsts maps the instructions to their type before promotion.
4082 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
4083 const InstrToOrigTy &PromotedInsts, bool IsSExt);
4084
4085 /// Utility function to determine if \p OpIdx should be promoted when
4086 /// promoting \p Inst.
4087 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
4088 return !(isa<SelectInst>(Inst) && OpIdx == 0);
4089 }
4090
4091 /// Utility function to promote the operand of \p Ext when this
4092 /// operand is a promotable trunc or sext or zext.
4093 /// \p PromotedInsts maps the instructions to their type before promotion.
4094 /// \p CreatedInstsCost[out] contains the cost of all instructions
4095 /// created to promote the operand of Ext.
4096 /// Newly added extensions are inserted in \p Exts.
4097 /// Newly added truncates are inserted in \p Truncs.
4098 /// Should never be called directly.
4099 /// \return The promoted value which is used instead of Ext.
4100 static Value *promoteOperandForTruncAndAnyExt(
4101 Instruction *Ext, TypePromotionTransaction &TPT,
4102 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4103 SmallVectorImpl<Instruction *> *Exts,
4104 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
4105
4106 /// Utility function to promote the operand of \p Ext when this
4107 /// operand is promotable and is not a supported trunc or sext.
4108 /// \p PromotedInsts maps the instructions to their type before promotion.
4109 /// \p CreatedInstsCost[out] contains the cost of all the instructions
4110 /// created to promote the operand of Ext.
4111 /// Newly added extensions are inserted in \p Exts.
4112 /// Newly added truncates are inserted in \p Truncs.
4113 /// Should never be called directly.
4114 /// \return The promoted value which is used instead of Ext.
4115 static Value *promoteOperandForOther(Instruction *Ext,
4116 TypePromotionTransaction &TPT,
4117 InstrToOrigTy &PromotedInsts,
4118 unsigned &CreatedInstsCost,
4119 SmallVectorImpl<Instruction *> *Exts,
4120 SmallVectorImpl<Instruction *> *Truncs,
4121 const TargetLowering &TLI, bool IsSExt);
4122
4123 /// \see promoteOperandForOther.
4124 static Value *signExtendOperandForOther(
4125 Instruction *Ext, TypePromotionTransaction &TPT,
4126 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4127 SmallVectorImpl<Instruction *> *Exts,
4128 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4129 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4130 Exts, Truncs, TLI, true);
4131 }
4132
4133 /// \see promoteOperandForOther.
4134 static Value *zeroExtendOperandForOther(
4135 Instruction *Ext, TypePromotionTransaction &TPT,
4136 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4137 SmallVectorImpl<Instruction *> *Exts,
4138 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4139 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
4140 Exts, Truncs, TLI, false);
4141 }
4142
4143public:
4144 /// Type for the utility function that promotes the operand of Ext.
4145 using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
4146 InstrToOrigTy &PromotedInsts,
4147 unsigned &CreatedInstsCost,
4148 SmallVectorImpl<Instruction *> *Exts,
4149 SmallVectorImpl<Instruction *> *Truncs,
4150 const TargetLowering &TLI);
4151
4152 /// Given a sign/zero extend instruction \p Ext, return the appropriate
4153 /// action to promote the operand of \p Ext instead of using Ext.
4154 /// \return NULL if no promotable action is possible with the current
4155 /// sign extension.
4156 /// \p InsertedInsts keeps track of all the instructions inserted by the
4157 /// other CodeGenPrepare optimizations. This information is important
4158 /// because we do not want to promote these instructions as CodeGenPrepare
4159 /// will reinsert them later. Thus creating an infinite loop: create/remove.
4160 /// \p PromotedInsts maps the instructions to their type before promotion.
4161 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
4162 const TargetLowering &TLI,
4163 const InstrToOrigTy &PromotedInsts);
4164};
4165
4166} // end anonymous namespace
4167
4168bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
4169 Type *ConsideredExtType,
4170 const InstrToOrigTy &PromotedInsts,
4171 bool IsSExt) {
4172 // The promotion helper does not know how to deal with vector types yet.
4173 // To be able to fix that, we would need to fix the places where we
4174 // statically extend, e.g., constants and such.
4175 if (Inst->getType()->isVectorTy())
4176 return false;
4177
4178 // We can always get through zext.
4179 if (isa<ZExtInst>(Inst))
4180 return true;
4181
4182 // sext(sext) is ok too.
4183 if (IsSExt && isa<SExtInst>(Inst))
4184 return true;
4185
4186 // We can get through binary operator, if it is legal. In other words, the
4187 // binary operator must have a nuw or nsw flag.
4188 if (const auto *BinOp = dyn_cast<BinaryOperator>(Inst))
4189 if (isa<OverflowingBinaryOperator>(BinOp) &&
4190 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
4191 (IsSExt && BinOp->hasNoSignedWrap())))
4192 return true;
4193
4194 // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
4195 if ((Inst->getOpcode() == Instruction::And ||
4196 Inst->getOpcode() == Instruction::Or))
4197 return true;
4198
4199 // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
4200 if (Inst->getOpcode() == Instruction::Xor) {
4201 // Make sure it is not a NOT.
4202 if (const auto *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1)))
4203 if (!Cst->getValue().isAllOnes())
4204 return true;
4205 }
4206
4207 // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
4208 // It may change a poisoned value into a regular value, like
4209 // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
4210 // poisoned value regular value
4211 // It should be OK since undef covers valid value.
4212 if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
4213 return true;
4214
4215 // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
4216 // It may change a poisoned value into a regular value, like
4217 // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
4218 // poisoned value regular value
4219 // It should be OK since undef covers valid value.
4220 if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
4221 const auto *ExtInst = cast<const Instruction>(*Inst->user_begin());
4222 if (ExtInst->hasOneUse()) {
4223 const auto *AndInst = dyn_cast<const Instruction>(*ExtInst->user_begin());
4224 if (AndInst && AndInst->getOpcode() == Instruction::And) {
4225 const auto *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
4226 if (Cst &&
4227 Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
4228 return true;
4229 }
4230 }
4231 }
4232
4233 // Check if we can do the following simplification.
4234 // ext(trunc(opnd)) --> ext(opnd)
4235 if (!isa<TruncInst>(Inst))
4236 return false;
4237
4238 Value *OpndVal = Inst->getOperand(0);
4239 // Check if we can use this operand in the extension.
4240 // If the type is larger than the result type of the extension, we cannot.
4241 if (!OpndVal->getType()->isIntegerTy() ||
4242 OpndVal->getType()->getIntegerBitWidth() >
4243 ConsideredExtType->getIntegerBitWidth())
4244 return false;
4245
4246 // If the operand of the truncate is not an instruction, we will not have
4247 // any information on the dropped bits.
4248 // (Actually we could for constant but it is not worth the extra logic).
4249 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
4250 if (!Opnd)
4251 return false;
4252
4253 // Check if the source of the type is narrow enough.
4254 // I.e., check that trunc just drops extended bits of the same kind of
4255 // the extension.
4256 // #1 get the type of the operand and check the kind of the extended bits.
4257 const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
4258 if (OpndType)
4259 ;
4260 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
4261 OpndType = Opnd->getOperand(0)->getType();
4262 else
4263 return false;
4264
4265 // #2 check that the truncate just drops extended bits.
4266 return Inst->getType()->getIntegerBitWidth() >=
4267 OpndType->getIntegerBitWidth();
4268}
4269
4270TypePromotionHelper::Action TypePromotionHelper::getAction(
4271 Instruction *Ext, const SetOfInstrs &InsertedInsts,
4272 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
4273 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&(static_cast <bool> ((isa<SExtInst>(Ext) || isa<
ZExtInst>(Ext)) && "Unexpected instruction type") ?
void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4274, __extension__ __PRETTY_FUNCTION__
))
4274 "Unexpected instruction type")(static_cast <bool> ((isa<SExtInst>(Ext) || isa<
ZExtInst>(Ext)) && "Unexpected instruction type") ?
void (0) : __assert_fail ("(isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && \"Unexpected instruction type\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4274, __extension__ __PRETTY_FUNCTION__
))
;
4275 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
4276 Type *ExtTy = Ext->getType();
4277 bool IsSExt = isa<SExtInst>(Ext);
4278 // If the operand of the extension is not an instruction, we cannot
4279 // get through.
4280 // If it, check we can get through.
4281 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
4282 return nullptr;
4283
4284 // Do not promote if the operand has been added by codegenprepare.
4285 // Otherwise, it means we are undoing an optimization that is likely to be
4286 // redone, thus causing potential infinite loop.
4287 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
4288 return nullptr;
4289
4290 // SExt or Trunc instructions.
4291 // Return the related handler.
4292 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
4293 isa<ZExtInst>(ExtOpnd))
4294 return promoteOperandForTruncAndAnyExt;
4295
4296 // Regular instruction.
4297 // Abort early if we will have to insert non-free instructions.
4298 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
4299 return nullptr;
4300 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
4301}
4302
4303Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
4304 Instruction *SExt, TypePromotionTransaction &TPT,
4305 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4306 SmallVectorImpl<Instruction *> *Exts,
4307 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
4308 // By construction, the operand of SExt is an instruction. Otherwise we cannot
4309 // get through it and this method should not be called.
4310 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
4311 Value *ExtVal = SExt;
4312 bool HasMergedNonFreeExt = false;
4313 if (isa<ZExtInst>(SExtOpnd)) {
4314 // Replace s|zext(zext(opnd))
4315 // => zext(opnd).
4316 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
4317 Value *ZExt =
4318 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
4319 TPT.replaceAllUsesWith(SExt, ZExt);
4320 TPT.eraseInstruction(SExt);
4321 ExtVal = ZExt;
4322 } else {
4323 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
4324 // => z|sext(opnd).
4325 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
4326 }
4327 CreatedInstsCost = 0;
4328
4329 // Remove dead code.
4330 if (SExtOpnd->use_empty())
4331 TPT.eraseInstruction(SExtOpnd);
4332
4333 // Check if the extension is still needed.
4334 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
4335 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
4336 if (ExtInst) {
4337 if (Exts)
4338 Exts->push_back(ExtInst);
4339 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
4340 }
4341 return ExtVal;
4342 }
4343
4344 // At this point we have: ext ty opnd to ty.
4345 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
4346 Value *NextVal = ExtInst->getOperand(0);
4347 TPT.eraseInstruction(ExtInst, NextVal);
4348 return NextVal;
4349}
4350
4351Value *TypePromotionHelper::promoteOperandForOther(
4352 Instruction *Ext, TypePromotionTransaction &TPT,
4353 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
4354 SmallVectorImpl<Instruction *> *Exts,
4355 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
4356 bool IsSExt) {
4357 // By construction, the operand of Ext is an instruction. Otherwise we cannot
4358 // get through it and this method should not be called.
4359 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
4360 CreatedInstsCost = 0;
4361 if (!ExtOpnd->hasOneUse()) {
4362 // ExtOpnd will be promoted.
4363 // All its uses, but Ext, will need to use a truncated value of the
4364 // promoted version.
4365 // Create the truncate now.
4366 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
4367 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
4368 // Insert it just after the definition.
4369 ITrunc->moveAfter(ExtOpnd);
4370 if (Truncs)
4371 Truncs->push_back(ITrunc);
4372 }
4373
4374 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
4375 // Restore the operand of Ext (which has been replaced by the previous call
4376 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
4377 TPT.setOperand(Ext, 0, ExtOpnd);
4378 }
4379
4380 // Get through the Instruction:
4381 // 1. Update its type.
4382 // 2. Replace the uses of Ext by Inst.
4383 // 3. Extend each operand that needs to be extended.
4384
4385 // Remember the original type of the instruction before promotion.
4386 // This is useful to know that the high bits are sign extended bits.
4387 addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
4388 // Step #1.
4389 TPT.mutateType(ExtOpnd, Ext->getType());
4390 // Step #2.
4391 TPT.replaceAllUsesWith(Ext, ExtOpnd);
4392 // Step #3.
4393 Instruction *ExtForOpnd = Ext;
4394
4395 LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Propagate Ext to operands\n"
; } } while (false)
;
4396 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
4397 ++OpIdx) {
4398 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
)
;
4399 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
4400 !shouldExtOperand(ExtOpnd, OpIdx)) {
4401 LLVM_DEBUG(dbgs() << "No need to propagate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "No need to propagate\n"
; } } while (false)
;
4402 continue;
4403 }
4404 // Check if we can statically extend the operand.
4405 Value *Opnd = ExtOpnd->getOperand(OpIdx);
4406 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
4407 LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
4408 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
4409 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
4410 : Cst->getValue().zext(BitWidth);
4411 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
4412 continue;
4413 }
4414 // UndefValue are typed, so we have to statically sign extend them.
4415 if (isa<UndefValue>(Opnd)) {
4416 LLVM_DEBUG(dbgs() << "Statically extend\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Statically extend\n"; }
} while (false)
;
4417 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
4418 continue;
4419 }
4420
4421 // Otherwise we have to explicitly sign extend the operand.
4422 // Check if Ext was reused to extend an operand.
4423 if (!ExtForOpnd) {
4424 // If yes, create a new one.
4425 LLVM_DEBUG(dbgs() << "More operands to ext\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "More operands to ext\n"
; } } while (false)
;
4426 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
4427 : TPT.createZExt(Ext, Opnd, Ext->getType());
4428 if (!isa<Instruction>(ValForExtOpnd)) {
4429 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
4430 continue;
4431 }
4432 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4433 }
4434 if (Exts)
4435 Exts->push_back(ExtForOpnd);
4436 TPT.setOperand(ExtForOpnd, 0, Opnd);
4437
4438 // Move the sign extension before the insertion point.
4439 TPT.moveBefore(ExtForOpnd, ExtOpnd);
4440 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4441 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4442 // If more sext are required, new instructions will have to be created.
4443 ExtForOpnd = nullptr;
4444 }
4445 if (ExtForOpnd == Ext) {
4446 LLVM_DEBUG(dbgs() << "Extension is useless now\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Extension is useless now\n"
; } } while (false)
;
4447 TPT.eraseInstruction(Ext);
4448 }
4449 return ExtOpnd;
4450}
4451
4452/// Check whether or not promoting an instruction to a wider type is profitable.
4453/// \p NewCost gives the cost of extension instructions created by the
4454/// promotion.
4455/// \p OldCost gives the cost of extension instructions before the promotion
4456/// plus the number of instructions that have been
4457/// matched in the addressing mode the promotion.
4458/// \p PromotedOperand is the value that has been promoted.
4459/// \return True if the promotion is profitable, false otherwise.
4460bool AddressingModeMatcher::isPromotionProfitable(
4461 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4462 LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "OldCost: " << OldCost
<< "\tNewCost: " << NewCost << '\n'; } } while
(false)
4463 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "OldCost: " << OldCost
<< "\tNewCost: " << NewCost << '\n'; } } while
(false)
;
4464 // The cost of the new extensions is greater than the cost of the
4465 // old extension plus what we folded.
4466 // This is not profitable.
4467 if (NewCost > OldCost)
4468 return false;
4469 if (NewCost < OldCost)
4470 return true;
4471 // The promotion is neutral but it may help folding the sign extension in
4472 // loads for instance.
4473 // Check that we did not create an illegal instruction.
4474 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4475}
4476
4477/// Given an instruction or constant expr, see if we can fold the operation
4478/// into the addressing mode. If so, update the addressing mode and return
4479/// true, otherwise return false without modifying AddrMode.
4480/// If \p MovedAway is not NULL, it contains the information of whether or
4481/// not AddrInst has to be folded into the addressing mode on success.
4482/// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4483/// because it has been moved away.
4484/// Thus AddrInst must not be added in the matched instructions.
4485/// This state can happen when AddrInst is a sext, since it may be moved away.
4486/// Therefore, AddrInst may not be valid when MovedAway is true and it must
4487/// not be referenced anymore.
4488bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4489 unsigned Depth,
4490 bool *MovedAway) {
4491 // Avoid exponential behavior on extremely deep expression trees.
4492 if (Depth >= 5)
4493 return false;
4494
4495 // By default, all matched instructions stay in place.
4496 if (MovedAway)
4497 *MovedAway = false;
4498
4499 switch (Opcode) {
4500 case Instruction::PtrToInt:
4501 // PtrToInt is always a noop, as we know that the int type is pointer sized.
4502 return matchAddr(AddrInst->getOperand(0), Depth);
4503 case Instruction::IntToPtr: {
4504 auto AS = AddrInst->getType()->getPointerAddressSpace();
4505 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4506 // This inttoptr is a no-op if the integer type is pointer sized.
4507 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4508 return matchAddr(AddrInst->getOperand(0), Depth);
4509 return false;
4510 }
4511 case Instruction::BitCast:
4512 // BitCast is always a noop, and we can handle it as long as it is
4513 // int->int or pointer->pointer (we don't want int<->fp or something).
4514 if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
4515 // Don't touch identity bitcasts. These were probably put here by LSR,
4516 // and we don't want to mess around with them. Assume it knows what it
4517 // is doing.
4518 AddrInst->getOperand(0)->getType() != AddrInst->getType())
4519 return matchAddr(AddrInst->getOperand(0), Depth);
4520 return false;
4521 case Instruction::AddrSpaceCast: {
4522 unsigned SrcAS =
4523 AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4524 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4525 if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
4526 return matchAddr(AddrInst->getOperand(0), Depth);
4527 return false;
4528 }
4529 case Instruction::Add: {
4530 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4531 ExtAddrMode BackupAddrMode = AddrMode;
4532 unsigned OldSize = AddrModeInsts.size();
4533 // Start a transaction at this point.
4534 // The LHS may match but not the RHS.
4535 // Therefore, we need a higher level restoration point to undo partially
4536 // matched operation.
4537 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4538 TPT.getRestorationPoint();
4539
4540 AddrMode.InBounds = false;
4541 if (matchAddr(AddrInst->getOperand(1), Depth + 1) &&
4542 matchAddr(AddrInst->getOperand(0), Depth + 1))
4543 return true;
4544
4545 // Restore the old addr mode info.
4546 AddrMode = BackupAddrMode;
4547 AddrModeInsts.resize(OldSize);
4548 TPT.rollback(LastKnownGood);
4549
4550 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4551 if (matchAddr(AddrInst->getOperand(0), Depth + 1) &&
4552 matchAddr(AddrInst->getOperand(1), Depth + 1))
4553 return true;
4554
4555 // Otherwise we definitely can't merge the ADD in.
4556 AddrMode = BackupAddrMode;
4557 AddrModeInsts.resize(OldSize);
4558 TPT.rollback(LastKnownGood);
4559 break;
4560 }
4561 // case Instruction::Or:
4562 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4563 // break;
4564 case Instruction::Mul:
4565 case Instruction::Shl: {
4566 // Can only handle X*C and X << C.
4567 AddrMode.InBounds = false;
4568 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4569 if (!RHS || RHS->getBitWidth() > 64)
4570 return false;
4571 int64_t Scale = Opcode == Instruction::Shl
4572 ? 1LL << RHS->getLimitedValue(RHS->getBitWidth() - 1)
4573 : RHS->getSExtValue();
4574
4575 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4576 }
4577 case Instruction::GetElementPtr: {
4578 // Scan the GEP. We check it if it contains constant offsets and at most
4579 // one variable offset.
4580 int VariableOperand = -1;
4581 unsigned VariableScale = 0;
4582
4583 int64_t ConstantOffset = 0;
4584 gep_type_iterator GTI = gep_type_begin(AddrInst);
4585 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4586 if (StructType *STy = GTI.getStructTypeOrNull()) {
4587 const StructLayout *SL = DL.getStructLayout(STy);
4588 unsigned Idx =
4589 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4590 ConstantOffset += SL->getElementOffset(Idx);
4591 } else {
4592 TypeSize TS = DL.getTypeAllocSize(GTI.getIndexedType());
4593 if (TS.isNonZero()) {
4594 // The optimisations below currently only work for fixed offsets.
4595 if (TS.isScalable())
4596 return false;
4597 int64_t TypeSize = TS.getFixedSize();
4598 if (ConstantInt *CI =
4599 dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4600 const APInt &CVal = CI->getValue();
4601 if (CVal.getMinSignedBits() <= 64) {
4602 ConstantOffset += CVal.getSExtValue() * TypeSize;
4603 continue;
4604 }
4605 }
4606 // We only allow one variable index at the moment.
4607 if (VariableOperand != -1)
4608 return false;
4609
4610 // Remember the variable index.
4611 VariableOperand = i;
4612 VariableScale = TypeSize;
4613 }
4614 }
4615 }
4616
4617 // A common case is for the GEP to only do a constant offset. In this case,
4618 // just add it to the disp field and check validity.
4619 if (VariableOperand == -1) {
4620 AddrMode.BaseOffs += ConstantOffset;
4621 if (ConstantOffset == 0 ||
4622 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4623 // Check to see if we can fold the base pointer in too.
4624 if (matchAddr(AddrInst->getOperand(0), Depth + 1)) {
4625 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4626 AddrMode.InBounds = false;
4627 return true;
4628 }
4629 } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4630 TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4631 ConstantOffset > 0) {
4632 // Record GEPs with non-zero offsets as candidates for splitting in the
4633 // event that the offset cannot fit into the r+i addressing mode.
4634 // Simple and common case that only one GEP is used in calculating the
4635 // address for the memory access.
4636 Value *Base = AddrInst->getOperand(0);
4637 auto *BaseI = dyn_cast<Instruction>(Base);
4638 auto *GEP = cast<GetElementPtrInst>(AddrInst);
4639 if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4640 (BaseI && !isa<CastInst>(BaseI) &&
4641 !isa<GetElementPtrInst>(BaseI))) {
4642 // Make sure the parent block allows inserting non-PHI instructions
4643 // before the terminator.
4644 BasicBlock *Parent =
4645 BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4646 if (!Parent->getTerminator()->isEHPad())
4647 LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4648 }
4649 }
4650 AddrMode.BaseOffs -= ConstantOffset;
4651 return false;
4652 }
4653
4654 // Save the valid addressing mode in case we can't match.
4655 ExtAddrMode BackupAddrMode = AddrMode;
4656 unsigned OldSize = AddrModeInsts.size();
4657
4658 // See if the scale and offset amount is valid for this target.
4659 AddrMode.BaseOffs += ConstantOffset;
4660 if (!cast<GEPOperator>(AddrInst)->isInBounds())
4661 AddrMode.InBounds = false;
4662
4663 // Match the base operand of the GEP.
4664 if (!matchAddr(AddrInst->getOperand(0), Depth + 1)) {
4665 // If it couldn't be matched, just stuff the value in a register.
4666 if (AddrMode.HasBaseReg) {
4667 AddrMode = BackupAddrMode;
4668 AddrModeInsts.resize(OldSize);
4669 return false;
4670 }
4671 AddrMode.HasBaseReg = true;
4672 AddrMode.BaseReg = AddrInst->getOperand(0);
4673 }
4674
4675 // Match the remaining variable portion of the GEP.
4676 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4677 Depth)) {
4678 // If it couldn't be matched, try stuffing the base into a register
4679 // instead of matching it, and retrying the match of the scale.
4680 AddrMode = BackupAddrMode;
4681 AddrModeInsts.resize(OldSize);
4682 if (AddrMode.HasBaseReg)
4683 return false;
4684 AddrMode.HasBaseReg = true;
4685 AddrMode.BaseReg = AddrInst->getOperand(0);
4686 AddrMode.BaseOffs += ConstantOffset;
4687 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4688 VariableScale, Depth)) {
4689 // If even that didn't work, bail.
4690 AddrMode = BackupAddrMode;
4691 AddrModeInsts.resize(OldSize);
4692 return false;
4693 }
4694 }
4695
4696 return true;
4697 }
4698 case Instruction::SExt:
4699 case Instruction::ZExt: {
4700 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4701 if (!Ext)
4702 return false;
4703
4704 // Try to move this ext out of the way of the addressing mode.
4705 // Ask for a method for doing so.
4706 TypePromotionHelper::Action TPH =
4707 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4708 if (!TPH)
4709 return false;
4710
4711 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4712 TPT.getRestorationPoint();
4713 unsigned CreatedInstsCost = 0;
4714 unsigned ExtCost = !TLI.isExtFree(Ext);
4715 Value *PromotedOperand =
4716 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4717 // SExt has been moved away.
4718 // Thus either it will be rematched later in the recursive calls or it is
4719 // gone. Anyway, we must not fold it into the addressing mode at this point.
4720 // E.g.,
4721 // op = add opnd, 1
4722 // idx = ext op
4723 // addr = gep base, idx
4724 // is now:
4725 // promotedOpnd = ext opnd <- no match here
4726 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4727 // addr = gep base, op <- match
4728 if (MovedAway)
4729 *MovedAway = true;
4730
4731 assert(PromotedOperand &&(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4732, __extension__ __PRETTY_FUNCTION__
))
4732 "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedOperand && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedOperand && \"TypePromotionHelper should have filtered out those cases\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 4732, __extension__ __PRETTY_FUNCTION__
))
;
4733
4734 ExtAddrMode BackupAddrMode = AddrMode;
4735 unsigned OldSize = AddrModeInsts.size();
4736
4737 if (!matchAddr(PromotedOperand, Depth) ||
4738 // The total of the new cost is equal to the cost of the created
4739 // instructions.
4740 // The total of the old cost is equal to the cost of the extension plus
4741 // what we have saved in the addressing mode.
4742 !isPromotionProfitable(CreatedInstsCost,
4743 ExtCost + (AddrModeInsts.size() - OldSize),
4744 PromotedOperand)) {
4745 AddrMode = BackupAddrMode;
4746 AddrModeInsts.resize(OldSize);
4747 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)
;
4748 TPT.rollback(LastKnownGood);
4749 return false;
4750 }
4751 return true;
4752 }
4753 }
4754 return false;
4755}
4756
4757/// If we can, try to add the value of 'Addr' into the current addressing mode.
4758/// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4759/// unmodified. This assumes that Addr is either a pointer type or intptr_t
4760/// for the target.
4761///
4762bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4763 // Start a transaction at this point that we will rollback if the matching
4764 // fails.
4765 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4766 TPT.getRestorationPoint();
4767 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4768 if (CI->getValue().isSignedIntN(64)) {
4769 // Fold in immediates if legal for the target.
4770 AddrMode.BaseOffs += CI->getSExtValue();
4771 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4772 return true;
4773 AddrMode.BaseOffs -= CI->getSExtValue();
4774 }
4775 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4776 // If this is a global variable, try to fold it into the addressing mode.
4777 if (!AddrMode.BaseGV) {
4778 AddrMode.BaseGV = GV;
4779 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4780 return true;
4781 AddrMode.BaseGV = nullptr;
4782 }
4783 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4784 ExtAddrMode BackupAddrMode = AddrMode;
4785 unsigned OldSize = AddrModeInsts.size();
4786
4787 // Check to see if it is possible to fold this operation.
4788 bool MovedAway = false;
4789 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4790 // This instruction may have been moved away. If so, there is nothing
4791 // to check here.
4792 if (MovedAway)
4793 return true;
4794 // Okay, it's possible to fold this. Check to see if it is actually
4795 // *profitable* to do so. We use a simple cost model to avoid increasing
4796 // register pressure too much.
4797 if (I->hasOneUse() ||
4798 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4799 AddrModeInsts.push_back(I);
4800 return true;
4801 }
4802
4803 // It isn't profitable to do this, roll back.
4804 AddrMode = BackupAddrMode;
4805 AddrModeInsts.resize(OldSize);
4806 TPT.rollback(LastKnownGood);
4807 }
4808 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4809 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4810 return true;
4811 TPT.rollback(LastKnownGood);
4812 } else if (isa<ConstantPointerNull>(Addr)) {
4813 // Null pointer gets folded without affecting the addressing mode.
4814 return true;
4815 }
4816
4817 // Worse case, the target should support [reg] addressing modes. :)
4818 if (!AddrMode.HasBaseReg) {
4819 AddrMode.HasBaseReg = true;
4820 AddrMode.BaseReg = Addr;
4821 // Still check for legality in case the target supports [imm] but not [i+r].
4822 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4823 return true;
4824 AddrMode.HasBaseReg = false;
4825 AddrMode.BaseReg = nullptr;
4826 }
4827
4828 // If the base register is already taken, see if we can do [r+r].
4829 if (AddrMode.Scale == 0) {
4830 AddrMode.Scale = 1;
4831 AddrMode.ScaledReg = Addr;
4832 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4833 return true;
4834 AddrMode.Scale = 0;
4835 AddrMode.ScaledReg = nullptr;
4836 }
4837 // Couldn't match.
4838 TPT.rollback(LastKnownGood);
4839 return false;
4840}
4841
4842/// Check to see if all uses of OpVal by the specified inline asm call are due
4843/// to memory operands. If so, return true, otherwise return false.
4844static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4845 const TargetLowering &TLI,
4846 const TargetRegisterInfo &TRI) {
4847 const Function *F = CI->getFunction();
4848 TargetLowering::AsmOperandInfoVector TargetConstraints =
4849 TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, *CI);
4850
4851 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
4852 // Compute the constraint code and ConstraintType to use.
4853 TLI.ComputeConstraintToUse(OpInfo, SDValue());
4854
4855 // If this asm operand is our Value*, and if it isn't an indirect memory
4856 // operand, we can't fold it! TODO: Also handle C_Address?
4857 if (OpInfo.CallOperandVal == OpVal &&
4858 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4859 !OpInfo.isIndirect))
4860 return false;
4861 }
4862
4863 return true;
4864}
4865
4866// Max number of memory uses to look at before aborting the search to conserve
4867// compile time.
4868static constexpr int MaxMemoryUsesToScan = 20;
4869
4870/// Recursively walk all the uses of I until we find a memory use.
4871/// If we find an obviously non-foldable instruction, return true.
4872/// Add accessed addresses and types to MemoryUses.
4873static bool FindAllMemoryUses(
4874 Instruction *I, SmallVectorImpl<std::pair<Value *, Type *>> &MemoryUses,
4875 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4876 const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI,
4877 BlockFrequencyInfo *BFI, int SeenInsts = 0) {
4878 // If we already considered this instruction, we're done.
4879 if (!ConsideredInsts.insert(I).second)
4880 return false;
4881
4882 // If this is an obviously unfoldable instruction, bail out.
4883 if (!MightBeFoldableInst(I))
4884 return true;
4885
4886 // Loop over all the uses, recursively processing them.
4887 for (Use &U : I->uses()) {
4888 // Conservatively return true if we're seeing a large number or a deep chain
4889 // of users. This avoids excessive compilation times in pathological cases.
4890 if (SeenInsts++ >= MaxMemoryUsesToScan)
4891 return true;
4892
4893 Instruction *UserI = cast<Instruction>(U.getUser());
4894 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4895 MemoryUses.push_back({U.get(), LI->getType()});
4896 continue;
4897 }
4898
4899 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4900 if (U.getOperandNo() != StoreInst::getPointerOperandIndex())
4901 return true; // Storing addr, not into addr.
4902 MemoryUses.push_back({U.get(), SI->getValueOperand()->getType()});
4903 continue;
4904 }
4905
4906 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4907 if (U.getOperandNo() != AtomicRMWInst::getPointerOperandIndex())
4908 return true; // Storing addr, not into addr.
4909 MemoryUses.push_back({U.get(), RMW->getValOperand()->getType()});
4910 continue;
4911 }
4912
4913 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4914 if (U.getOperandNo() != AtomicCmpXchgInst::getPointerOperandIndex())
4915 return true; // Storing addr, not into addr.
4916 MemoryUses.push_back({U.get(), CmpX->getCompareOperand()->getType()});
4917 continue;
4918 }
4919
4920 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4921 if (CI->hasFnAttr(Attribute::Cold)) {
4922 // If this is a cold call, we can sink the addressing calculation into
4923 // the cold path. See optimizeCallInst
4924 bool OptForSize =
4925 OptSize || llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI);
4926 if (!OptForSize)
4927 continue;
4928 }
4929
4930 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand());
4931 if (!IA)
4932 return true;
4933
4934 // If this is a memory operand, we're cool, otherwise bail out.
4935 if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4936 return true;
4937 continue;
4938 }
4939
4940 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize,
4941 PSI, BFI, SeenInsts))
4942 return true;
4943 }
4944
4945 return false;
4946}
4947
4948/// Return true if Val is already known to be live at the use site that we're
4949/// folding it into. If so, there is no cost to include it in the addressing
4950/// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4951/// instruction already.
4952bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,
4953 Value *KnownLive1,
4954 Value *KnownLive2) {
4955 // If Val is either of the known-live values, we know it is live!
4956 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4957 return true;
4958
4959 // All values other than instructions and arguments (e.g. constants) are live.
4960 if (!isa<Instruction>(Val) && !isa<Argument>(Val))
4961 return true;
4962
4963 // If Val is a constant sized alloca in the entry block, it is live, this is
4964 // true because it is just a reference to the stack/frame pointer, which is
4965 // live for the whole function.
4966 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4967 if (AI->isStaticAlloca())
4968 return true;
4969
4970 // Check to see if this value is already used in the memory instruction's
4971 // block. If so, it's already live into the block at the very least, so we
4972 // can reasonably fold it.
4973 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4974}
4975
4976/// It is possible for the addressing mode of the machine to fold the specified
4977/// instruction into a load or store that ultimately uses it.
4978/// However, the specified instruction has multiple uses.
4979/// Given this, it may actually increase register pressure to fold it
4980/// into the load. For example, consider this code:
4981///
4982/// X = ...
4983/// Y = X+1
4984/// use(Y) -> nonload/store
4985/// Z = Y+1
4986/// load Z
4987///
4988/// In this case, Y has multiple uses, and can be folded into the load of Z
4989/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4990/// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4991/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4992/// number of computations either.
4993///
4994/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4995/// X was live across 'load Z' for other reasons, we actually *would* want to
4996/// fold the addressing mode in the Z case. This would make Y die earlier.
4997bool AddressingModeMatcher::isProfitableToFoldIntoAddressingMode(
4998 Instruction *I, ExtAddrMode &AMBefore, ExtAddrMode &AMAfter) {
4999 if (IgnoreProfitability)
5000 return true;
5001
5002 // AMBefore is the addressing mode before this instruction was folded into it,
5003 // and AMAfter is the addressing mode after the instruction was folded. Get
5004 // the set of registers referenced by AMAfter and subtract out those
5005 // referenced by AMBefore: this is the set of values which folding in this
5006 // address extends the lifetime of.
5007 //
5008 // Note that there are only two potential values being referenced here,
5009 // BaseReg and ScaleReg (global addresses are always available, as are any
5010 // folded immediates).
5011 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
5012
5013 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
5014 // lifetime wasn't extended by adding this instruction.
5015 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
5016 BaseReg = nullptr;
5017 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
5018 ScaledReg = nullptr;
5019
5020 // If folding this instruction (and it's subexprs) didn't extend any live
5021 // ranges, we're ok with it.
5022 if (!BaseReg && !ScaledReg)
5023 return true;
5024
5025 // If all uses of this instruction can have the address mode sunk into them,
5026 // we can remove the addressing mode and effectively trade one live register
5027 // for another (at worst.) In this context, folding an addressing mode into
5028 // the use is just a particularly nice way of sinking it.
5029 SmallVector<std::pair<Value *, Type *>, 16> MemoryUses;
5030 SmallPtrSet<Instruction *, 16> ConsideredInsts;
5031 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize, PSI,
5032 BFI))
5033 return false; // Has a non-memory, non-foldable use!
5034
5035 // Now that we know that all uses of this instruction are part of a chain of
5036 // computation involving only operations that could theoretically be folded
5037 // into a memory use, loop over each of these memory operation uses and see
5038 // if they could *actually* fold the instruction. The assumption is that
5039 // addressing modes are cheap and that duplicating the computation involved
5040 // many times is worthwhile, even on a fastpath. For sinking candidates
5041 // (i.e. cold call sites), this serves as a way to prevent excessive code
5042 // growth since most architectures have some reasonable small and fast way to
5043 // compute an effective address. (i.e LEA on x86)
5044 SmallVector<Instruction *, 32> MatchedAddrModeInsts;
5045 for (const std::pair<Value *, Type *> &Pair : MemoryUses) {
5046 Value *Address = Pair.first;
5047 Type *AddressAccessTy = Pair.second;
5048 unsigned AS = Address->getType()->getPointerAddressSpace();
5049
5050 // Do a match against the root of this address, ignoring profitability. This
5051 // will tell us if the addressing mode for the memory operation will
5052 // *actually* cover the shared instruction.
5053 ExtAddrMode Result;
5054 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5055 0);
5056 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5057 TPT.getRestorationPoint();
5058 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn,
5059 AddressAccessTy, AS, MemoryInst, Result,
5060 InsertedInsts, PromotedInsts, TPT,
5061 LargeOffsetGEP, OptSize, PSI, BFI);
5062 Matcher.IgnoreProfitability = true;
5063 bool Success = Matcher.matchAddr(Address, 0);
5064 (void)Success;
5065 assert(Success && "Couldn't select *anything*?")(static_cast <bool> (Success && "Couldn't select *anything*?"
) ? void (0) : __assert_fail ("Success && \"Couldn't select *anything*?\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5065, __extension__ __PRETTY_FUNCTION__
))
;
5066
5067 // The match was to check the profitability, the changes made are not
5068 // part of the original matcher. Therefore, they should be dropped
5069 // otherwise the original matcher will not present the right state.
5070 TPT.rollback(LastKnownGood);
5071
5072 // If the match didn't cover I, then it won't be shared by it.
5073 if (!is_contained(MatchedAddrModeInsts, I))
5074 return false;
5075
5076 MatchedAddrModeInsts.clear();
5077 }
5078
5079 return true;
5080}
5081
5082/// Return true if the specified values are defined in a
5083/// different basic block than BB.
5084static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
5085 if (Instruction *I = dyn_cast<Instruction>(V))
5086 return I->getParent() != BB;
5087 return false;
5088}
5089
5090/// Sink addressing mode computation immediate before MemoryInst if doing so
5091/// can be done without increasing register pressure. The need for the
5092/// register pressure constraint means this can end up being an all or nothing
5093/// decision for all uses of the same addressing computation.
5094///
5095/// Load and Store Instructions often have addressing modes that can do
5096/// significant amounts of computation. As such, instruction selection will try
5097/// to get the load or store to do as much computation as possible for the
5098/// program. The problem is that isel can only see within a single block. As
5099/// such, we sink as much legal addressing mode work into the block as possible.
5100///
5101/// This method is used to optimize both load/store and inline asms with memory
5102/// operands. It's also used to sink addressing computations feeding into cold
5103/// call sites into their (cold) basic block.
5104///
5105/// The motivation for handling sinking into cold blocks is that doing so can
5106/// both enable other address mode sinking (by satisfying the register pressure
5107/// constraint above), and reduce register pressure globally (by removing the
5108/// addressing mode computation from the fast path entirely.).
5109bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
5110 Type *AccessTy, unsigned AddrSpace) {
5111 Value *Repl = Addr;
5112
5113 // Try to collapse single-value PHI nodes. This is necessary to undo
5114 // unprofitable PRE transformations.
5115 SmallVector<Value *, 8> worklist;
5116 SmallPtrSet<Value *, 16> Visited;
5117 worklist.push_back(Addr);
5118
5119 // Use a worklist to iteratively look through PHI and select nodes, and
5120 // ensure that the addressing mode obtained from the non-PHI/select roots of
5121 // the graph are compatible.
5122 bool PhiOrSelectSeen = false;
5123 SmallVector<Instruction *, 16> AddrModeInsts;
5124 const SimplifyQuery SQ(*DL, TLInfo);
5125 AddressingModeCombiner AddrModes(SQ, Addr);
5126 TypePromotionTransaction TPT(RemovedInsts);
5127 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5128 TPT.getRestorationPoint();
5129 while (!worklist.empty()) {
5130 Value *V = worklist.pop_back_val();
5131
5132 // We allow traversing cyclic Phi nodes.
5133 // In case of success after this loop we ensure that traversing through
5134 // Phi nodes ends up with all cases to compute address of the form
5135 // BaseGV + Base + Scale * Index + Offset
5136 // where Scale and Offset are constans and BaseGV, Base and Index
5137 // are exactly the same Values in all cases.
5138 // It means that BaseGV, Scale and Offset dominate our memory instruction
5139 // and have the same value as they had in address computation represented
5140 // as Phi. So we can safely sink address computation to memory instruction.
5141 if (!Visited.insert(V).second)
5142 continue;
5143
5144 // For a PHI node, push all of its incoming values.
5145 if (PHINode *P = dyn_cast<PHINode>(V)) {
5146 append_range(worklist, P->incoming_values());
5147 PhiOrSelectSeen = true;
5148 continue;
5149 }
5150 // Similar for select.
5151 if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
5152 worklist.push_back(SI->getFalseValue());
5153 worklist.push_back(SI->getTrueValue());
5154 PhiOrSelectSeen = true;
5155 continue;
5156 }
5157
5158 // For non-PHIs, determine the addressing mode being computed. Note that
5159 // the result may differ depending on what other uses our candidate
5160 // addressing instructions might have.
5161 AddrModeInsts.clear();
5162 std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
5163 0);
5164 // Defer the query (and possible computation of) the dom tree to point of
5165 // actual use. It's expected that most address matches don't actually need
5166 // the domtree.
5167 auto getDTFn = [MemoryInst, this]() -> const DominatorTree & {
5168 Function *F = MemoryInst->getParent()->getParent();
5169 return this->getDT(*F);
5170 };
5171 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
5172 V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *LI, getDTFn,
5173 *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI,
5174 BFI.get());
5175
5176 GetElementPtrInst *GEP = LargeOffsetGEP.first;
5177 if (GEP && !NewGEPBases.count(GEP)) {
5178 // If splitting the underlying data structure can reduce the offset of a
5179 // GEP, collect the GEP. Skip the GEPs that are the new bases of
5180 // previously split data structures.
5181 LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
5182 LargeOffsetGEPID.insert(std::make_pair(GEP, LargeOffsetGEPID.size()));
5183 }
5184
5185 NewAddrMode.OriginalValue = V;
5186 if (!AddrModes.addNewAddrMode(NewAddrMode))
5187 break;
5188 }
5189
5190 // Try to combine the AddrModes we've collected. If we couldn't collect any,
5191 // or we have multiple but either couldn't combine them or combining them
5192 // wouldn't do anything useful, bail out now.
5193 if (!AddrModes.combineAddrModes()) {
5194 TPT.rollback(LastKnownGood);
5195 return false;
5196 }
5197 bool Modified = TPT.commit();
5198
5199 // Get the combined AddrMode (or the only AddrMode, if we only had one).
5200 ExtAddrMode AddrMode = AddrModes.getAddrMode();
5201
5202 // If all the instructions matched are already in this BB, don't do anything.
5203 // If we saw a Phi node then it is not local definitely, and if we saw a
5204 // select then we want to push the address calculation past it even if it's
5205 // already in this BB.
5206 if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
5207 return IsNonLocalValue(V, MemoryInst->getParent());
5208 })) {
5209 LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrModedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Found local addrmode: "
<< AddrMode << "\n"; } } while (false)
5210 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Found local addrmode: "
<< AddrMode << "\n"; } } while (false)
;
5211 return Modified;
5212 }
5213
5214 // Insert this computation right after this user. Since our caller is
5215 // scanning from the top of the BB to the bottom, reuse of the expr are
5216 // guaranteed to happen later.
5217 IRBuilder<> Builder(MemoryInst);
5218
5219 // Now that we determined the addressing expression we want to use and know
5220 // that we have to sink it into this block. Check to see if we have already
5221 // done this for some other load/store instr in this block. If so, reuse
5222 // the computation. Before attempting reuse, check if the address is valid
5223 // as it may have been erased.
5224
5225 WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
5226
5227 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
5228 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5229 if (SunkAddr) {
5230 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)
5231 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: Reusing nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
5232 if (SunkAddr->getType() != Addr->getType()) {
5233 if (SunkAddr->getType()->getPointerAddressSpace() !=
5234 Addr->getType()->getPointerAddressSpace() &&
5235 !DL->isNonIntegralPointerType(Addr->getType())) {
5236 // There are two reasons the address spaces might not match: a no-op
5237 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
5238 // ptrtoint/inttoptr pair to ensure we match the original semantics.
5239 // TODO: allow bitcast between different address space pointers with the
5240 // same size.
5241 SunkAddr = Builder.CreatePtrToInt(SunkAddr, IntPtrTy, "sunkaddr");
5242 SunkAddr =
5243 Builder.CreateIntToPtr(SunkAddr, Addr->getType(), "sunkaddr");
5244 } else
5245 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5246 }
5247 } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() &&
5248 SubtargetInfo->addrSinkUsingGEPs())) {
5249 // By default, we use the GEP-based method when AA is used later. This
5250 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
5251 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)
5252 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
5253 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
5254
5255 // First, find the pointer.
5256 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
5257 ResultPtr = AddrMode.BaseReg;
5258 AddrMode.BaseReg = nullptr;
5259 }
5260
5261 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
5262 // We can't add more than one pointer together, nor can we scale a
5263 // pointer (both of which seem meaningless).
5264 if (ResultPtr || AddrMode.Scale != 1)
5265 return Modified;
5266
5267 ResultPtr = AddrMode.ScaledReg;
5268 AddrMode.Scale = 0;
5269 }
5270
5271 // It is only safe to sign extend the BaseReg if we know that the math
5272 // required to create it did not overflow before we extend it. Since
5273 // the original IR value was tossed in favor of a constant back when
5274 // the AddrMode was created we need to bail out gracefully if widths
5275 // do not match instead of extending it.
5276 //
5277 // (See below for code to add the scale.)
5278 if (AddrMode.Scale) {
5279 Type *ScaledRegTy = AddrMode.ScaledReg->getType();
5280 if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
5281 cast<IntegerType>(ScaledRegTy)->getBitWidth())
5282 return Modified;
5283 }
5284
5285 if (AddrMode.BaseGV) {
5286 if (ResultPtr)
5287 return Modified;
5288
5289 ResultPtr = AddrMode.BaseGV;
5290 }
5291
5292 // If the real base value actually came from an inttoptr, then the matcher
5293 // will look through it and provide only the integer value. In that case,
5294 // use it here.
5295 if (!DL->isNonIntegralPointerType(Addr->getType())) {
5296 if (!ResultPtr && AddrMode.BaseReg) {
5297 ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
5298 "sunkaddr");
5299 AddrMode.BaseReg = nullptr;
5300 } else if (!ResultPtr && AddrMode.Scale == 1) {
5301 ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
5302 "sunkaddr");
5303 AddrMode.Scale = 0;
5304 }
5305 }
5306
5307 if (!ResultPtr && !AddrMode.BaseReg && !AddrMode.Scale &&
5308 !AddrMode.BaseOffs) {
5309 SunkAddr = Constant::getNullValue(Addr->getType());
5310 } else if (!ResultPtr) {
5311 return Modified;
5312 } else {
5313 Type *I8PtrTy =
5314 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
5315 Type *I8Ty = Builder.getInt8Ty();
5316
5317 // Start with the base register. Do this first so that subsequent address
5318 // matching finds it last, which will prevent it from trying to match it
5319 // as the scaled value in case it happens to be a mul. That would be
5320 // problematic if we've sunk a different mul for the scale, because then
5321 // we'd end up sinking both muls.
5322 if (AddrMode.BaseReg) {
5323 Value *V = AddrMode.BaseReg;
5324 if (V->getType() != IntPtrTy)
5325 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5326
5327 ResultIndex = V;
5328 }
5329
5330 // Add the scale value.
5331 if (AddrMode.Scale) {
5332 Value *V = AddrMode.ScaledReg;
5333 if (V->getType() == IntPtrTy) {
5334 // done.
5335 } else {
5336 assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5338, __extension__ __PRETTY_FUNCTION__
))
5337 cast<IntegerType>(V->getType())->getBitWidth() &&(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5338, __extension__ __PRETTY_FUNCTION__
))
5338 "We can't transform if ScaledReg is too narrow")(static_cast <bool> (cast<IntegerType>(IntPtrTy)->
getBitWidth() < cast<IntegerType>(V->getType())->
getBitWidth() && "We can't transform if ScaledReg is too narrow"
) ? void (0) : __assert_fail ("cast<IntegerType>(IntPtrTy)->getBitWidth() < cast<IntegerType>(V->getType())->getBitWidth() && \"We can't transform if ScaledReg is too narrow\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5338, __extension__ __PRETTY_FUNCTION__
))
;
5339 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5340 }
5341
5342 if (AddrMode.Scale != 1)
5343 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5344 "sunkaddr");
5345 if (ResultIndex)
5346 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
5347 else
5348 ResultIndex = V;
5349 }
5350
5351 // Add in the Base Offset if present.
5352 if (AddrMode.BaseOffs) {
5353 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5354 if (ResultIndex) {
5355 // We need to add this separately from the scale above to help with
5356 // SDAG consecutive load/store merging.
5357 if (ResultPtr->getType() != I8PtrTy)
5358 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5359 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex,
5360 "sunkaddr", AddrMode.InBounds);
5361 }
5362
5363 ResultIndex = V;
5364 }
5365
5366 if (!ResultIndex) {
5367 SunkAddr = ResultPtr;
5368 } else {
5369 if (ResultPtr->getType() != I8PtrTy)
5370 ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
5371 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr",
5372 AddrMode.InBounds);
5373 }
5374
5375 if (SunkAddr->getType() != Addr->getType()) {
5376 if (SunkAddr->getType()->getPointerAddressSpace() !=
5377 Addr->getType()->getPointerAddressSpace() &&
5378 !DL->isNonIntegralPointerType(Addr->getType())) {
5379 // There are two reasons the address spaces might not match: a no-op
5380 // addrspacecast, or a ptrtoint/inttoptr pair. Either way, we emit a
5381 // ptrtoint/inttoptr pair to ensure we match the original semantics.
5382 // TODO: allow bitcast between different address space pointers with
5383 // the same size.
5384 SunkAddr = Builder.CreatePtrToInt(SunkAddr, IntPtrTy, "sunkaddr");
5385 SunkAddr =
5386 Builder.CreateIntToPtr(SunkAddr, Addr->getType(), "sunkaddr");
5387 } else
5388 SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
5389 }
5390 }
5391 } else {
5392 // We'd require a ptrtoint/inttoptr down the line, which we can't do for
5393 // non-integral pointers, so in that case bail out now.
5394 Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
5395 Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
5396 PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
5397 PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
5398 if (DL->isNonIntegralPointerType(Addr->getType()) ||
5399 (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
5400 (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
5401 (AddrMode.BaseGV &&
5402 DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
5403 return Modified;
5404
5405 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)
5406 << " for " << *MemoryInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "CGP: SINKING nonlocal addrmode: "
<< AddrMode << " for " << *MemoryInst <<
"\n"; } } while (false)
;
5407 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
5408 Value *Result = nullptr;
5409
5410 // Start with the base register. Do this first so that subsequent address
5411 // matching finds it last, which will prevent it from trying to match it
5412 // as the scaled value in case it happens to be a mul. That would be
5413 // problematic if we've sunk a different mul for the scale, because then
5414 // we'd end up sinking both muls.
5415 if (AddrMode.BaseReg) {
5416 Value *V = AddrMode.BaseReg;
5417 if (V->getType()->isPointerTy())
5418 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5419 if (V->getType() != IntPtrTy)
5420 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
5421 Result = V;
5422 }
5423
5424 // Add the scale value.
5425 if (AddrMode.Scale) {
5426 Value *V = AddrMode.ScaledReg;
5427 if (V->getType() == IntPtrTy) {
5428 // done.
5429 } else if (V->getType()->isPointerTy()) {
5430 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
5431 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
5432 cast<IntegerType>(V->getType())->getBitWidth()) {
5433 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
5434 } else {
5435 // It is only safe to sign extend the BaseReg if we know that the math
5436 // required to create it did not overflow before we extend it. Since
5437 // the original IR value was tossed in favor of a constant back when
5438 // the AddrMode was created we need to bail out gracefully if widths
5439 // do not match instead of extending it.
5440 Instruction *I = dyn_cast_or_null<Instruction>(Result);
5441 if (I && (Result != AddrMode.BaseReg))
5442 I->eraseFromParent();
5443 return Modified;
5444 }
5445 if (AddrMode.Scale != 1)
5446 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
5447 "sunkaddr");
5448 if (Result)
5449 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5450 else
5451 Result = V;
5452 }
5453
5454 // Add in the BaseGV if present.
5455 if (AddrMode.BaseGV) {
5456 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5457 if (Result)
5458 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5459 else
5460 Result = V;
5461 }
5462
5463 // Add in the Base Offset if present.
5464 if (AddrMode.BaseOffs) {
5465 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5466 if (Result)
5467 Result = Builder.CreateAdd(Result, V, "sunkaddr");
5468 else
5469 Result = V;
5470 }
5471
5472 if (!Result)
5473 SunkAddr = Constant::getNullValue(Addr->getType());
5474 else
5475 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5476 }
5477
5478 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5479 // Store the newly computed address into the cache. In the case we reused a
5480 // value, this should be idempotent.
5481 SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
5482
5483 // If we have no uses, recursively delete the value and all dead instructions
5484 // using it.
5485 if (Repl->use_empty()) {
5486 resetIteratorIfInvalidatedWhileCalling(CurInstIterator->getParent(), [&]() {
5487 RecursivelyDeleteTriviallyDeadInstructions(
5488 Repl, TLInfo, nullptr,
5489 [&](Value *V) { removeAllAssertingVHReferences(V); });
5490 });
5491 }
5492 ++NumMemoryInsts;
5493 return true;
5494}
5495
5496/// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
5497/// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
5498/// only handle a 2 operand GEP in the same basic block or a splat constant
5499/// vector. The 2 operands to the GEP must have a scalar pointer and a vector
5500/// index.
5501///
5502/// If the existing GEP has a vector base pointer that is splat, we can look
5503/// through the splat to find the scalar pointer. If we can't find a scalar
5504/// pointer there's nothing we can do.
5505///
5506/// If we have a GEP with more than 2 indices where the middle indices are all
5507/// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
5508///
5509/// If the final index isn't a vector or is a splat, we can emit a scalar GEP
5510/// followed by a GEP with an all zeroes vector index. This will enable
5511/// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a
5512/// zero index.
5513bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
5514 Value *Ptr) {
5515 Value *NewAddr;
5516
5517 if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
5518 // Don't optimize GEPs that don't have indices.
5519 if (!GEP->hasIndices())
5520 return false;
5521
5522 // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
5523 // FIXME: We should support this by sinking the GEP.
5524 if (MemoryInst->getParent() != GEP->getParent())
5525 return false;
5526
5527 SmallVector<Value *, 2> Ops(GEP->operands());
5528
5529 bool RewriteGEP = false;
5530
5531 if (Ops[0]->getType()->isVectorTy()) {
5532 Ops[0] = getSplatValue(Ops[0]);
5533 if (!Ops[0])
5534 return false;
5535 RewriteGEP = true;
5536 }
5537
5538 unsigned FinalIndex = Ops.size() - 1;
5539
5540 // Ensure all but the last index is 0.
5541 // FIXME: This isn't strictly required. All that's required is that they are
5542 // all scalars or splats.
5543 for (unsigned i = 1; i < FinalIndex; ++i) {
5544 auto *C = dyn_cast<Constant>(Ops[i]);
5545 if (!C)
5546 return false;
5547 if (isa<VectorType>(C->getType()))
5548 C = C->getSplatValue();
5549 auto *CI = dyn_cast_or_null<ConstantInt>(C);
5550 if (!CI || !CI->isZero())
5551 return false;
5552 // Scalarize the index if needed.
5553 Ops[i] = CI;
5554 }
5555
5556 // Try to scalarize the final index.
5557 if (Ops[FinalIndex]->getType()->isVectorTy()) {
5558 if (Value *V = getSplatValue(Ops[FinalIndex])) {
5559 auto *C = dyn_cast<ConstantInt>(V);
5560 // Don't scalarize all zeros vector.
5561 if (!C || !C->isZero()) {
5562 Ops[FinalIndex] = V;
5563 RewriteGEP = true;
5564 }
5565 }
5566 }
5567
5568 // If we made any changes or the we have extra operands, we need to generate
5569 // new instructions.
5570 if (!RewriteGEP && Ops.size() == 2)
5571 return false;
5572
5573 auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5574
5575 IRBuilder<> Builder(MemoryInst);
5576
5577 Type *SourceTy = GEP->getSourceElementType();
5578 Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType());
5579
5580 // If the final index isn't a vector, emit a scalar GEP containing all ops
5581 // and a vector GEP with all zeroes final index.
5582 if (!Ops[FinalIndex]->getType()->isVectorTy()) {
5583 NewAddr =
5584 Builder.CreateGEP(SourceTy, Ops[0], makeArrayRef(Ops).drop_front());
5585 auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5586 auto *SecondTy = GetElementPtrInst::getIndexedType(
5587 SourceTy, makeArrayRef(Ops).drop_front());
5588 NewAddr =
5589 Builder.CreateGEP(SecondTy, NewAddr, Constant::getNullValue(IndexTy));
5590 } else {
5591 Value *Base = Ops[0];
5592 Value *Index = Ops[FinalIndex];
5593
5594 // Create a scalar GEP if there are more than 2 operands.
5595 if (Ops.size() != 2) {
5596 // Replace the last index with 0.
5597 Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy);
5598 Base =
5599 Builder.CreateGEP(SourceTy, Base, makeArrayRef(Ops).drop_front());
5600 SourceTy = GetElementPtrInst::getIndexedType(
5601 SourceTy, makeArrayRef(Ops).drop_front());
5602 }
5603
5604 // Now create the GEP with scalar pointer and vector index.
5605 NewAddr = Builder.CreateGEP(SourceTy, Base, Index);
5606 }
5607 } else if (!isa<Constant>(Ptr)) {
5608 // Not a GEP, maybe its a splat and we can create a GEP to enable
5609 // SelectionDAGBuilder to use it as a uniform base.
5610 Value *V = getSplatValue(Ptr);
5611 if (!V)
5612 return false;
5613
5614 auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount();
5615
5616 IRBuilder<> Builder(MemoryInst);
5617
5618 // Emit a vector GEP with a scalar pointer and all 0s vector index.
5619 Type *ScalarIndexTy = DL->getIndexType(V->getType()->getScalarType());
5620 auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
5621 Type *ScalarTy;
5622 if (cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==
5623 Intrinsic::masked_gather) {
5624 ScalarTy = MemoryInst->getType()->getScalarType();
5625 } else {
5626 assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==(static_cast <bool> (cast<IntrinsicInst>(MemoryInst
)->getIntrinsicID() == Intrinsic::masked_scatter) ? void (
0) : __assert_fail ("cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == Intrinsic::masked_scatter"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5627, __extension__ __PRETTY_FUNCTION__
))
5627 Intrinsic::masked_scatter)(static_cast <bool> (cast<IntrinsicInst>(MemoryInst
)->getIntrinsicID() == Intrinsic::masked_scatter) ? void (
0) : __assert_fail ("cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == Intrinsic::masked_scatter"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5627, __extension__ __PRETTY_FUNCTION__
))
;
5628 ScalarTy = MemoryInst->getOperand(0)->getType()->getScalarType();
5629 }
5630 NewAddr = Builder.CreateGEP(ScalarTy, V, Constant::getNullValue(IndexTy));
5631 } else {
5632 // Constant, SelectionDAGBuilder knows to check if its a splat.
5633 return false;
5634 }
5635
5636 MemoryInst->replaceUsesOfWith(Ptr, NewAddr);
5637
5638 // If we have no uses, recursively delete the value and all dead instructions
5639 // using it.
5640 if (Ptr->use_empty())
5641 RecursivelyDeleteTriviallyDeadInstructions(
5642 Ptr, TLInfo, nullptr,
5643 [&](Value *V) { removeAllAssertingVHReferences(V); });
5644
5645 return true;
5646}
5647
5648/// If there are any memory operands, use OptimizeMemoryInst to sink their
5649/// address computing into the block when possible / profitable.
5650bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5651 bool MadeChange = false;
5652
5653 const TargetRegisterInfo *TRI =
5654 TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
5655 TargetLowering::AsmOperandInfoVector TargetConstraints =
5656 TLI->ParseConstraints(*DL, TRI, *CS);
5657 unsigned ArgNo = 0;
5658 for (TargetLowering::AsmOperandInfo &OpInfo : TargetConstraints) {
5659 // Compute the constraint code and ConstraintType to use.
5660 TLI->ComputeConstraintToUse(OpInfo, SDValue());
5661
5662 // TODO: Also handle C_Address?
5663 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5664 OpInfo.isIndirect) {
5665 Value *OpVal = CS->getArgOperand(ArgNo++);
5666 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5667 } else if (OpInfo.Type == InlineAsm::isInput)
5668 ArgNo++;
5669 }
5670
5671 return MadeChange;
5672}
5673
5674/// Check if all the uses of \p Val are equivalent (or free) zero or
5675/// sign extensions.
5676static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
5677 assert(!Val->use_empty() && "Input must have at least one use")(static_cast <bool> (!Val->use_empty() && "Input must have at least one use"
) ? void (0) : __assert_fail ("!Val->use_empty() && \"Input must have at least one use\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5677, __extension__ __PRETTY_FUNCTION__
))
;
5678 const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
5679 bool IsSExt = isa<SExtInst>(FirstUser);
5680 Type *ExtTy = FirstUser->getType();
5681 for (const User *U : Val->users()) {
5682 const Instruction *UI = cast<Instruction>(U);
5683 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5684 return false;
5685 Type *CurTy = UI->getType();
5686 // Same input and output types: Same instruction after CSE.
5687 if (CurTy == ExtTy)
5688 continue;
5689
5690 // If IsSExt is true, we are in this situation:
5691 // a = Val
5692 // b = sext ty1 a to ty2
5693 // c = sext ty1 a to ty3
5694 // Assuming ty2 is shorter than ty3, this could be turned into:
5695 // a = Val
5696 // b = sext ty1 a to ty2
5697 // c = sext ty2 b to ty3
5698 // However, the last sext is not free.
5699 if (IsSExt)
5700 return false;
5701
5702 // This is a ZExt, maybe this is free to extend from one type to another.
5703 // In that case, we would not account for a different use.
5704 Type *NarrowTy;
5705 Type *LargeTy;
5706 if (ExtTy->getScalarType()->getIntegerBitWidth() >
5707 CurTy->getScalarType()->getIntegerBitWidth()) {
5708 NarrowTy = CurTy;
5709 LargeTy = ExtTy;
5710 } else {
5711 NarrowTy = ExtTy;
5712 LargeTy = CurTy;
5713 }
5714
5715 if (!TLI.isZExtFree(NarrowTy, LargeTy))
5716 return false;
5717 }
5718 // All uses are the same or can be derived from one another for free.
5719 return true;
5720}
5721
5722/// Try to speculatively promote extensions in \p Exts and continue
5723/// promoting through newly promoted operands recursively as far as doing so is
5724/// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
5725/// When some promotion happened, \p TPT contains the proper state to revert
5726/// them.
5727///
5728/// \return true if some promotion happened, false otherwise.
5729bool CodeGenPrepare::tryToPromoteExts(
5730 TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
5731 SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
5732 unsigned CreatedInstsCost) {
5733 bool Promoted = false;
5734
5735 // Iterate over all the extensions to try to promote them.
5736 for (auto *I : Exts) {
5737 // Early check if we directly have ext(load).
5738 if (isa<LoadInst>(I->getOperand(0))) {
5739 ProfitablyMovedExts.push_back(I);
5740 continue;
5741 }
5742
5743 // Check whether or not we want to do any promotion. The reason we have
5744 // this check inside the for loop is to catch the case where an extension
5745 // is directly fed by a load because in such case the extension can be moved
5746 // up without any promotion on its operands.
5747 if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5748 return false;
5749
5750 // Get the action to perform the promotion.
5751 TypePromotionHelper::Action TPH =
5752 TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
5753 // Check if we can promote.
5754 if (!TPH) {
5755 // Save the current extension as we cannot move up through its operand.
5756 ProfitablyMovedExts.push_back(I);
5757 continue;
5758 }
5759
5760 // Save the current state.
5761 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5762 TPT.getRestorationPoint();
5763 SmallVector<Instruction *, 4> NewExts;
5764 unsigned NewCreatedInstsCost = 0;
5765 unsigned ExtCost = !TLI->isExtFree(I);
5766 // Promote.
5767 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5768 &NewExts, nullptr, *TLI);
5769 assert(PromotedVal &&(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5770, __extension__ __PRETTY_FUNCTION__
))
5770 "TypePromotionHelper should have filtered out those cases")(static_cast <bool> (PromotedVal && "TypePromotionHelper should have filtered out those cases"
) ? void (0) : __assert_fail ("PromotedVal && \"TypePromotionHelper should have filtered out those cases\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 5770, __extension__ __PRETTY_FUNCTION__
))
;
5771
5772 // We would be able to merge only one extension in a load.
5773 // Therefore, if we have more than 1 new extension we heuristically
5774 // cut this search path, because it means we degrade the code quality.
5775 // With exactly 2, the transformation is neutral, because we will merge
5776 // one extension but leave one. However, we optimistically keep going,
5777 // because the new extension may be removed too.
5778 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5779 // FIXME: It would be possible to propagate a negative value instead of
5780 // conservatively ceiling it to 0.
5781 TotalCreatedInstsCost =
5782 std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
5783 if (!StressExtLdPromotion &&
5784 (TotalCreatedInstsCost > 1 ||
5785 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5786 // This promotion is not profitable, rollback to the previous state, and
5787 // save the current extension in ProfitablyMovedExts as the latest
5788 // speculative promotion turned out to be unprofitable.
5789 TPT.rollback(LastKnownGood);
5790 ProfitablyMovedExts.push_back(I);
5791 continue;
5792 }
5793 // Continue promoting NewExts as far as doing so is profitable.
5794 SmallVector<Instruction *, 2> NewlyMovedExts;
5795 (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5796 bool NewPromoted = false;
5797 for (auto *ExtInst : NewlyMovedExts) {
5798 Instruction *MovedExt = cast<Instruction>(ExtInst);
5799 Value *ExtOperand = MovedExt->getOperand(0);
5800 // If we have reached to a load, we need this extra profitability check
5801 // as it could potentially be merged into an ext(load).
5802 if (isa<LoadInst>(ExtOperand) &&
5803 !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5804 (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5805 continue;
5806
5807 ProfitablyMovedExts.push_back(MovedExt);
5808 NewPromoted = true;
5809 }
5810
5811 // If none of speculative promotions for NewExts is profitable, rollback
5812 // and save the current extension (I) as the last profitable extension.
5813 if (!NewPromoted) {
5814 TPT.rollback(LastKnownGood);
5815 ProfitablyMovedExts.push_back(I);
5816 continue;
5817 }
5818 // The promotion is profitable.
5819 Promoted = true;
5820 }
5821 return Promoted;
5822}
5823
5824/// Merging redundant sexts when one is dominating the other.
5825bool CodeGenPrepare::mergeSExts(Function &F) {
5826 bool Changed = false;
5827 for (auto &Entry : ValToSExtendedUses) {
5828 SExts &Insts = Entry.second;
5829 SExts CurPts;
5830 for (Instruction *Inst : Insts) {
5831 if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5832 Inst->getOperand(0) != Entry.first)
5833 continue;
5834 bool inserted = false;
5835 for (auto &Pt : CurPts) {
5836 if (getDT(F).dominates(Inst, Pt)) {
5837 Pt->replaceAllUsesWith(Inst);
5838 RemovedInsts.insert(Pt);
5839 Pt->removeFromParent();
5840 Pt = Inst;
5841 inserted = true;
5842 Changed = true;
5843 break;
5844 }
5845 if (!getDT(F).dominates(Pt, Inst))
5846 // Give up if we need to merge in a common dominator as the
5847 // experiments show it is not profitable.
5848 continue;
5849 Inst->replaceAllUsesWith(Pt);
5850 RemovedInsts.insert(Inst);
5851 Inst->removeFromParent();
5852 inserted = true;
5853 Changed = true;
5854 break;
5855 }
5856 if (!inserted)
5857 CurPts.push_back(Inst);
5858 }
5859 }
5860 return Changed;
5861}
5862
5863// Splitting large data structures so that the GEPs accessing them can have
5864// smaller offsets so that they can be sunk to the same blocks as their users.
5865// For example, a large struct starting from %base is split into two parts
5866// where the second part starts from %new_base.
5867//
5868// Before:
5869// BB0:
5870// %base =
5871//
5872// BB1:
5873// %gep0 = gep %base, off0
5874// %gep1 = gep %base, off1
5875// %gep2 = gep %base, off2
5876//
5877// BB2:
5878// %load1 = load %gep0
5879// %load2 = load %gep1
5880// %load3 = load %gep2
5881//
5882// After:
5883// BB0:
5884// %base =
5885// %new_base = gep %base, off0
5886//
5887// BB1:
5888// %new_gep0 = %new_base
5889// %new_gep1 = gep %new_base, off1 - off0
5890// %new_gep2 = gep %new_base, off2 - off0
5891//
5892// BB2:
5893// %load1 = load i32, i32* %new_gep0
5894// %load2 = load i32, i32* %new_gep1
5895// %load3 = load i32, i32* %new_gep2
5896//
5897// %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5898// their offsets are smaller enough to fit into the addressing mode.
5899bool CodeGenPrepare::splitLargeGEPOffsets() {
5900 bool Changed = false;
5901 for (auto &Entry : LargeOffsetGEPMap) {
5902 Value *OldBase = Entry.first;
5903 SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>>
5904 &LargeOffsetGEPs = Entry.second;
5905 auto compareGEPOffset =
5906 [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5907 const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5908 if (LHS.first == RHS.first)
5909 return false;
5910 if (LHS.second != RHS.second)
5911 return LHS.second < RHS.second;
5912 return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5913 };
5914 // Sorting all the GEPs of the same data structures based on the offsets.
5915 llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5916 LargeOffsetGEPs.erase(
5917 std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5918 LargeOffsetGEPs.end());
5919 // Skip if all the GEPs have the same offsets.
5920 if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5921 continue;
5922 GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5923 int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5924 Value *NewBaseGEP = nullptr;
5925
5926 auto *LargeOffsetGEP = LargeOffsetGEPs.begin();
5927 while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5928 GetElementPtrInst *GEP = LargeOffsetGEP->first;
5929 int64_t Offset = LargeOffsetGEP->second;
5930 if (Offset != BaseOffset) {
5931 TargetLowering::AddrMode AddrMode;
5932 AddrMode.BaseOffs = Offset - BaseOffset;
5933 // The result type of the GEP might not be the type of the memory
5934 // access.
5935 if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5936 GEP->getResultElementType(),
5937 GEP->getAddressSpace())) {
5938 // We need to create a new base if the offset to the current base is
5939 // too large to fit into the addressing mode. So, a very large struct
5940 // may be split into several parts.
5941 BaseGEP = GEP;
5942 BaseOffset = Offset;
5943 NewBaseGEP = nullptr;
5944 }
5945 }
5946
5947 // Generate a new GEP to replace the current one.
5948 LLVMContext &Ctx = GEP->getContext();
5949 Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5950 Type *I8PtrTy =
5951 Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace());
5952 Type *I8Ty = Type::getInt8Ty(Ctx);
5953
5954 if (!NewBaseGEP) {
5955 // Create a new base if we don't have one yet. Find the insertion
5956 // pointer for the new base first.
5957 BasicBlock::iterator NewBaseInsertPt;
5958 BasicBlock *NewBaseInsertBB;
5959 if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5960 // If the base of the struct is an instruction, the new base will be
5961 // inserted close to it.
5962 NewBaseInsertBB = BaseI->getParent();
5963 if (isa<PHINode>(BaseI))
5964 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5965 else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5966 NewBaseInsertBB =
5967 SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5968 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5969 } else
5970 NewBaseInsertPt = std::next(BaseI->getIterator());
5971 } else {
5972 // If the current base is an argument or global value, the new base
5973 // will be inserted to the entry block.
5974 NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5975 NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5976 }
5977 IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5978 // Create a new base.
5979 Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5980 NewBaseGEP = OldBase;
5981 if (NewBaseGEP->getType() != I8PtrTy)
5982 NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5983 NewBaseGEP =
5984 NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5985 NewGEPBases.insert(NewBaseGEP);
5986 }
5987
5988 IRBuilder<> Builder(GEP);
5989 Value *NewGEP = NewBaseGEP;
5990 if (Offset == BaseOffset) {
5991 if (GEP->getType() != I8PtrTy)
5992 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5993 } else {
5994 // Calculate the new offset for the new GEP.
5995 Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5996 NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5997
5998 if (GEP->getType() != I8PtrTy)
5999 NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
6000 }
6001 GEP->replaceAllUsesWith(NewGEP);
6002 LargeOffsetGEPID.erase(GEP);
6003 LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
6004 GEP->eraseFromParent();
6005 Changed = true;
6006 }
6007 }
6008 return Changed;
6009}
6010
6011bool CodeGenPrepare::optimizePhiType(
6012 PHINode *I, SmallPtrSetImpl<PHINode *> &Visited,
6013 SmallPtrSetImpl<Instruction *> &DeletedInstrs) {
6014 // We are looking for a collection on interconnected phi nodes that together
6015 // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts
6016 // are of the same type. Convert the whole set of nodes to the type of the
6017 // bitcast.
6018 Type *PhiTy = I->getType();
6019 Type *ConvertTy = nullptr;
6020 if (Visited.count(I) ||
6021 (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy()))
6022 return false;
6023
6024 SmallVector<Instruction *, 4> Worklist;
6025 Worklist.push_back(cast<Instruction>(I));
6026 SmallPtrSet<PHINode *, 4> PhiNodes;
6027 PhiNodes.insert(I);
6028 Visited.insert(I);
6029 SmallPtrSet<Instruction *, 4> Defs;
6030 SmallPtrSet<Instruction *, 4> Uses;
6031 // This works by adding extra bitcasts between load/stores and removing
6032 // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi))
6033 // we can get in the situation where we remove a bitcast in one iteration
6034 // just to add it again in the next. We need to ensure that at least one
6035 // bitcast we remove are anchored to something that will not change back.
6036 bool AnyAnchored = false;
6037
6038 while (!Worklist.empty()) {
6039 Instruction *II = Worklist.pop_back_val();
6040
6041 if (auto *Phi = dyn_cast<PHINode>(II)) {
6042 // Handle Defs, which might also be PHI's
6043 for (Value *V : Phi->incoming_values()) {
6044 if (auto *OpPhi = dyn_cast<PHINode>(V)) {
6045 if (!PhiNodes.count(OpPhi)) {
6046 if (!Visited.insert(OpPhi).second)
6047 return false;
6048 PhiNodes.insert(OpPhi);
6049 Worklist.push_back(OpPhi);
6050 }
6051 } else if (auto *OpLoad = dyn_cast<LoadInst>(V)) {
6052 if (!OpLoad->isSimple())
6053 return false;
6054 if (Defs.insert(OpLoad).second)
6055 Worklist.push_back(OpLoad);
6056 } else if (auto *OpEx = dyn_cast<ExtractElementInst>(V)) {
6057 if (Defs.insert(OpEx).second)
6058 Worklist.push_back(OpEx);
6059 } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
6060 if (!ConvertTy)
6061 ConvertTy = OpBC->getOperand(0)->getType();
6062 if (OpBC->getOperand(0)->getType() != ConvertTy)
6063 return false;
6064 if (Defs.insert(OpBC).second) {
6065 Worklist.push_back(OpBC);
6066 AnyAnchored |= !isa<LoadInst>(OpBC->getOperand(0)) &&
6067 !isa<ExtractElementInst>(OpBC->getOperand(0));
6068 }
6069 } else if (!isa<UndefValue>(V)) {
6070 return false;
6071 }
6072 }
6073 }
6074
6075 // Handle uses which might also be phi's
6076 for (User *V : II->users()) {
6077 if (auto *OpPhi = dyn_cast<PHINode>(V)) {
6078 if (!PhiNodes.count(OpPhi)) {
6079 if (Visited.count(OpPhi))
6080 return false;
6081 PhiNodes.insert(OpPhi);
6082 Visited.insert(OpPhi);
6083 Worklist.push_back(OpPhi);
6084 }
6085 } else if (auto *OpStore = dyn_cast<StoreInst>(V)) {
6086 if (!OpStore->isSimple() || OpStore->getOperand(0) != II)
6087 return false;
6088 Uses.insert(OpStore);
6089 } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) {
6090 if (!ConvertTy)
6091 ConvertTy = OpBC->getType();
6092 if (OpBC->getType() != ConvertTy)
6093 return false;
6094 Uses.insert(OpBC);
6095 AnyAnchored |=
6096 any_of(OpBC->users(), [](User *U) { return !isa<StoreInst>(U); });
6097 } else {
6098 return false;
6099 }
6100 }
6101 }
6102
6103 if (!ConvertTy || !AnyAnchored ||
6104 !TLI->shouldConvertPhiType(PhiTy, ConvertTy))
6105 return false;
6106
6107 LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Converting " << *
I << "\n and connected nodes to " << *ConvertTy <<
"\n"; } } while (false)
6108 << *ConvertTy << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Converting " << *
I << "\n and connected nodes to " << *ConvertTy <<
"\n"; } } while (false)
;
6109
6110 // Create all the new phi nodes of the new type, and bitcast any loads to the
6111 // correct type.
6112 ValueToValueMap ValMap;
6113 ValMap[UndefValue::get(PhiTy)] = UndefValue::get(ConvertTy);
6114 for (Instruction *D : Defs) {
6115 if (isa<BitCastInst>(D)) {
6116 ValMap[D] = D->getOperand(0);
6117 DeletedInstrs.insert(D);
6118 } else {
6119 ValMap[D] =
6120 new BitCastInst(D, ConvertTy, D->getName() + ".bc", D->getNextNode());
6121 }
6122 }
6123 for (PHINode *Phi : PhiNodes)
6124 ValMap[Phi] = PHINode::Create(ConvertTy, Phi->getNumIncomingValues(),
6125 Phi->getName() + ".tc", Phi);
6126 // Pipe together all the PhiNodes.
6127 for (PHINode *Phi : PhiNodes) {
6128 PHINode *NewPhi = cast<PHINode>(ValMap[Phi]);
6129 for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++)
6130 NewPhi->addIncoming(ValMap[Phi->getIncomingValue(i)],
6131 Phi->getIncomingBlock(i));
6132 Visited.insert(NewPhi);
6133 }
6134 // And finally pipe up the stores and bitcasts
6135 for (Instruction *U : Uses) {
6136 if (isa<BitCastInst>(U)) {
6137 DeletedInstrs.insert(U);
6138 U->replaceAllUsesWith(ValMap[U->getOperand(0)]);
6139 } else {
6140 U->setOperand(0,
6141 new BitCastInst(ValMap[U->getOperand(0)], PhiTy, "bc", U));
6142 }
6143 }
6144
6145 // Save the removed phis to be deleted later.
6146 for (PHINode *Phi : PhiNodes)
6147 DeletedInstrs.insert(Phi);
6148 return true;
6149}
6150
6151bool CodeGenPrepare::optimizePhiTypes(Function &F) {
6152 if (!OptimizePhiTypes)
6153 return false;
6154
6155 bool Changed = false;
6156 SmallPtrSet<PHINode *, 4> Visited;
6157 SmallPtrSet<Instruction *, 4> DeletedInstrs;
6158
6159 // Attempt to optimize all the phis in the functions to the correct type.
6160 for (auto &BB : F)
6161 for (auto &Phi : BB.phis())
6162 Changed |= optimizePhiType(&Phi, Visited, DeletedInstrs);
6163
6164 // Remove any old phi's that have been converted.
6165 for (auto *I : DeletedInstrs) {
6166 I->replaceAllUsesWith(PoisonValue::get(I->getType()));
6167 I->eraseFromParent();
6168 }
6169
6170 return Changed;
6171}
6172
6173/// Return true, if an ext(load) can be formed from an extension in
6174/// \p MovedExts.
6175bool CodeGenPrepare::canFormExtLd(
6176 const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
6177 Instruction *&Inst, bool HasPromoted) {
6178 for (auto *MovedExtInst : MovedExts) {
6179 if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
6180 LI = cast<LoadInst>(MovedExtInst->getOperand(0));
6181 Inst = MovedExtInst;
6182 break;
6183 }
6184 }
6185 if (!LI)
6186 return false;
6187
6188 // If they're already in the same block, there's nothing to do.
6189 // Make the cheap checks first if we did not promote.
6190 // If we promoted, we need to check if it is indeed profitable.
6191 if (!HasPromoted && LI->getParent() == Inst->getParent())
6192 return false;
6193
6194 return TLI->isExtLoad(LI, Inst, *DL);
6195}
6196
6197/// Move a zext or sext fed by a load into the same basic block as the load,
6198/// unless conditions are unfavorable. This allows SelectionDAG to fold the
6199/// extend into the load.
6200///
6201/// E.g.,
6202/// \code
6203/// %ld = load i32* %addr
6204/// %add = add nuw i32 %ld, 4
6205/// %zext = zext i32 %add to i64
6206// \endcode
6207/// =>
6208/// \code
6209/// %ld = load i32* %addr
6210/// %zext = zext i32 %ld to i64
6211/// %add = add nuw i64 %zext, 4
6212/// \encode
6213/// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
6214/// allow us to match zext(load i32*) to i64.
6215///
6216/// Also, try to promote the computations used to obtain a sign extended
6217/// value used into memory accesses.
6218/// E.g.,
6219/// \code
6220/// a = add nsw i32 b, 3
6221/// d = sext i32 a to i64
6222/// e = getelementptr ..., i64 d
6223/// \endcode
6224/// =>
6225/// \code
6226/// f = sext i32 b to i64
6227/// a = add nsw i64 f, 3
6228/// e = getelementptr ..., i64 a
6229/// \endcode
6230///
6231/// \p Inst[in/out] the extension may be modified during the process if some
6232/// promotions apply.
6233bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
6234 bool AllowPromotionWithoutCommonHeader = false;
6235 /// See if it is an interesting sext operations for the address type
6236 /// promotion before trying to promote it, e.g., the ones with the right
6237 /// type and used in memory accesses.
6238 bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
6239 *Inst, AllowPromotionWithoutCommonHeader);
6240 TypePromotionTransaction TPT(RemovedInsts);
6241 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
6242 TPT.getRestorationPoint();
6243 SmallVector<Instruction *, 1> Exts;
6244 SmallVector<Instruction *, 2> SpeculativelyMovedExts;
6245 Exts.push_back(Inst);
6246
6247 bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
6248
6249 // Look for a load being extended.
6250 LoadInst *LI = nullptr;
6251 Instruction *ExtFedByLoad;
6252
6253 // Try to promote a chain of computation if it allows to form an extended
6254 // load.
6255 if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
6256 assert(LI && ExtFedByLoad && "Expect a valid load and extension")(static_cast <bool> (LI && ExtFedByLoad &&
"Expect a valid load and extension") ? void (0) : __assert_fail
("LI && ExtFedByLoad && \"Expect a valid load and extension\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6256, __extension__ __PRETTY_FUNCTION__
))
;
6257 TPT.commit();
6258 // Move the extend into the same block as the load.
6259 ExtFedByLoad->moveAfter(LI);
6260 ++NumExtsMoved;
6261 Inst = ExtFedByLoad;
6262 return true;
6263 }
6264
6265 // Continue promoting SExts if known as considerable depending on targets.
6266 if (ATPConsiderable &&
6267 performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
6268 HasPromoted, TPT, SpeculativelyMovedExts))
6269 return true;
6270
6271 TPT.rollback(LastKnownGood);
6272 return false;
6273}
6274
6275// Perform address type promotion if doing so is profitable.
6276// If AllowPromotionWithoutCommonHeader == false, we should find other sext
6277// instructions that sign extended the same initial value. However, if
6278// AllowPromotionWithoutCommonHeader == true, we expect promoting the
6279// extension is just profitable.
6280bool CodeGenPrepare::performAddressTypePromotion(
6281 Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
6282 bool HasPromoted, TypePromotionTransaction &TPT,
6283 SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
6284 bool Promoted = false;
6285 SmallPtrSet<Instruction *, 1> UnhandledExts;
6286 bool AllSeenFirst = true;
6287 for (auto *I : SpeculativelyMovedExts) {
6288 Value *HeadOfChain = I->getOperand(0);
6289 DenseMap<Value *, Instruction *>::iterator AlreadySeen =
6290 SeenChainsForSExt.find(HeadOfChain);
6291 // If there is an unhandled SExt which has the same header, try to promote
6292 // it as well.
6293 if (AlreadySeen != SeenChainsForSExt.end()) {
6294 if (AlreadySeen->second != nullptr)
6295 UnhandledExts.insert(AlreadySeen->second);
6296 AllSeenFirst = false;
6297 }
6298 }
6299
6300 if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
6301 SpeculativelyMovedExts.size() == 1)) {
6302 TPT.commit();
6303 if (HasPromoted)
6304 Promoted = true;
6305 for (auto *I : SpeculativelyMovedExts) {
6306 Value *HeadOfChain = I->getOperand(0);
6307 SeenChainsForSExt[HeadOfChain] = nullptr;
6308 ValToSExtendedUses[HeadOfChain].push_back(I);
6309 }
6310 // Update Inst as promotion happen.
6311 Inst = SpeculativelyMovedExts.pop_back_val();
6312 } else {
6313 // This is the first chain visited from the header, keep the current chain
6314 // as unhandled. Defer to promote this until we encounter another SExt
6315 // chain derived from the same header.
6316 for (auto *I : SpeculativelyMovedExts) {
6317 Value *HeadOfChain = I->getOperand(0);
6318 SeenChainsForSExt[HeadOfChain] = Inst;
6319 }
6320 return false;
6321 }
6322
6323 if (!AllSeenFirst && !UnhandledExts.empty())
6324 for (auto *VisitedSExt : UnhandledExts) {
6325 if (RemovedInsts.count(VisitedSExt))
6326 continue;
6327 TypePromotionTransaction TPT(RemovedInsts);
6328 SmallVector<Instruction *, 1> Exts;
6329 SmallVector<Instruction *, 2> Chains;
6330 Exts.push_back(VisitedSExt);
6331 bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
6332 TPT.commit();
6333 if (HasPromoted)
6334 Promoted = true;
6335 for (auto *I : Chains) {
6336 Value *HeadOfChain = I->getOperand(0);
6337 // Mark this as handled.
6338 SeenChainsForSExt[HeadOfChain] = nullptr;
6339 ValToSExtendedUses[HeadOfChain].push_back(I);
6340 }
6341 }
6342 return Promoted;
6343}
6344
6345bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
6346 BasicBlock *DefBB = I->getParent();
6347
6348 // If the result of a {s|z}ext and its source are both live out, rewrite all
6349 // other uses of the source with result of extension.
6350 Value *Src = I->getOperand(0);
6351 if (Src->hasOneUse())
6352 return false;
6353
6354 // Only do this xform if truncating is free.
6355 if (!TLI->isTruncateFree(I->getType(), Src->getType()))
6356 return false;
6357
6358 // Only safe to perform the optimization if the source is also defined in
6359 // this block.
6360 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
6361 return false;
6362
6363 bool DefIsLiveOut = false;
6364 for (User *U : I->users()) {
6365 Instruction *UI = cast<Instruction>(U);
6366
6367 // Figure out which BB this ext is used in.
6368 BasicBlock *UserBB = UI->getParent();
6369 if (UserBB == DefBB)
6370 continue;
6371 DefIsLiveOut = true;
6372 break;
6373 }
6374 if (!DefIsLiveOut)
6375 return false;
6376
6377 // Make sure none of the uses are PHI nodes.
6378 for (User *U : Src->users()) {
6379 Instruction *UI = cast<Instruction>(U);
6380 BasicBlock *UserBB = UI->getParent();
6381 if (UserBB == DefBB)
6382 continue;
6383 // Be conservative. We don't want this xform to end up introducing
6384 // reloads just before load / store instructions.
6385 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
6386 return false;
6387 }
6388
6389 // InsertedTruncs - Only insert one trunc in each block once.
6390 DenseMap<BasicBlock *, Instruction *> InsertedTruncs;
6391
6392 bool MadeChange = false;
6393 for (Use &U : Src->uses()) {
6394 Instruction *User = cast<Instruction>(U.getUser());
6395
6396 // Figure out which BB this ext is used in.
6397 BasicBlock *UserBB = User->getParent();
6398 if (UserBB == DefBB)
6399 continue;
6400
6401 // Both src and def are live in this block. Rewrite the use.
6402 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
6403
6404 if (!InsertedTrunc) {
6405 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
6406 assert(InsertPt != UserBB->end())(static_cast <bool> (InsertPt != UserBB->end()) ? void
(0) : __assert_fail ("InsertPt != UserBB->end()", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 6406, __extension__ __PRETTY_FUNCTION__))
;
6407 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
6408 InsertedInsts.insert(InsertedTrunc);
6409 }
6410
6411 // Replace a use of the {s|z}ext source with a use of the result.
6412 U = InsertedTrunc;
6413 ++NumExtUses;
6414 MadeChange = true;
6415 }
6416
6417 return MadeChange;
6418}
6419
6420// Find loads whose uses only use some of the loaded value's bits. Add an "and"
6421// just after the load if the target can fold this into one extload instruction,
6422// with the hope of eliminating some of the other later "and" instructions using
6423// the loaded value. "and"s that are made trivially redundant by the insertion
6424// of the new "and" are removed by this function, while others (e.g. those whose
6425// path from the load goes through a phi) are left for isel to potentially
6426// remove.
6427//
6428// For example:
6429//
6430// b0:
6431// x = load i32
6432// ...
6433// b1:
6434// y = and x, 0xff
6435// z = use y
6436//
6437// becomes:
6438//
6439// b0:
6440// x = load i32
6441// x' = and x, 0xff
6442// ...
6443// b1:
6444// z = use x'
6445//
6446// whereas:
6447//
6448// b0:
6449// x1 = load i32
6450// ...
6451// b1:
6452// x2 = load i32
6453// ...
6454// b2:
6455// x = phi x1, x2
6456// y = and x, 0xff
6457//
6458// becomes (after a call to optimizeLoadExt for each load):
6459//
6460// b0:
6461// x1 = load i32
6462// x1' = and x1, 0xff
6463// ...
6464// b1:
6465// x2 = load i32
6466// x2' = and x2, 0xff
6467// ...
6468// b2:
6469// x = phi x1', x2'
6470// y = and x, 0xff
6471bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
6472 if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
6473 return false;
6474
6475 // Skip loads we've already transformed.
6476 if (Load->hasOneUse() &&
6477 InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
6478 return false;
6479
6480 // Look at all uses of Load, looking through phis, to determine how many bits
6481 // of the loaded value are needed.
6482 SmallVector<Instruction *, 8> WorkList;
6483 SmallPtrSet<Instruction *, 16> Visited;
6484 SmallVector<Instruction *, 8> AndsToMaybeRemove;
6485 for (auto *U : Load->users())
6486 WorkList.push_back(cast<Instruction>(U));
6487
6488 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
6489 unsigned BitWidth = LoadResultVT.getSizeInBits();
6490 // If the BitWidth is 0, do not try to optimize the type
6491 if (BitWidth == 0)
6492 return false;
6493
6494 APInt DemandBits(BitWidth, 0);
6495 APInt WidestAndBits(BitWidth, 0);
6496
6497 while (!WorkList.empty()) {
6498 Instruction *I = WorkList.pop_back_val();
6499
6500 // Break use-def graph loops.
6501 if (!Visited.insert(I).second)
6502 continue;
6503
6504 // For a PHI node, push all of its users.
6505 if (auto *Phi = dyn_cast<PHINode>(I)) {
6506 for (auto *U : Phi->users())
6507 WorkList.push_back(cast<Instruction>(U));
6508 continue;
6509 }
6510
6511 switch (I->getOpcode()) {
6512 case Instruction::And: {
6513 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
6514 if (!AndC)
6515 return false;
6516 APInt AndBits = AndC->getValue();
6517 DemandBits |= AndBits;
6518 // Keep track of the widest and mask we see.
6519 if (AndBits.ugt(WidestAndBits))
6520 WidestAndBits = AndBits;
6521 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
6522 AndsToMaybeRemove.push_back(I);
6523 break;
6524 }
6525
6526 case Instruction::Shl: {
6527 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
6528 if (!ShlC)
6529 return false;
6530 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
6531 DemandBits.setLowBits(BitWidth - ShiftAmt);
6532 break;
6533 }
6534
6535 case Instruction::Trunc: {
6536 EVT TruncVT = TLI->getValueType(*DL, I->getType());
6537 unsigned TruncBitWidth = TruncVT.getSizeInBits();
6538 DemandBits.setLowBits(TruncBitWidth);
6539 break;
6540 }
6541
6542 default:
6543 return false;
6544 }
6545 }
6546
6547 uint32_t ActiveBits = DemandBits.getActiveBits();
6548 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
6549 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
6550 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
6551 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
6552 // followed by an AND.
6553 // TODO: Look into removing this restriction by fixing backends to either
6554 // return false for isLoadExtLegal for i1 or have them select this pattern to
6555 // a single instruction.
6556 //
6557 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
6558 // mask, since these are the only ands that will be removed by isel.
6559 if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
6560 WidestAndBits != DemandBits)
6561 return false;
6562
6563 LLVMContext &Ctx = Load->getType()->getContext();
6564 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
6565 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
6566
6567 // Reject cases that won't be matched as extloads.
6568 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
6569 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
6570 return false;
6571
6572 IRBuilder<> Builder(Load->getNextNode());
6573 auto *NewAnd = cast<Instruction>(
6574 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
6575 // Mark this instruction as "inserted by CGP", so that other
6576 // optimizations don't touch it.
6577 InsertedInsts.insert(NewAnd);
6578
6579 // Replace all uses of load with new and (except for the use of load in the
6580 // new and itself).
6581 Load->replaceAllUsesWith(NewAnd);
6582 NewAnd->setOperand(0, Load);
6583
6584 // Remove any and instructions that are now redundant.
6585 for (auto *And : AndsToMaybeRemove)
6586 // Check that the and mask is the same as the one we decided to put on the
6587 // new and.
6588 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
6589 And->replaceAllUsesWith(NewAnd);
6590 if (&*CurInstIterator == And)
6591 CurInstIterator = std::next(And->getIterator());
6592 And->eraseFromParent();
6593 ++NumAndUses;
6594 }
6595
6596 ++NumAndsAdded;
6597 return true;
6598}
6599
6600/// Check if V (an operand of a select instruction) is an expensive instruction
6601/// that is only used once.
6602static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
6603 auto *I = dyn_cast<Instruction>(V);
6604 // If it's safe to speculatively execute, then it should not have side
6605 // effects; therefore, it's safe to sink and possibly *not* execute.
6606 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
6607 TTI->isExpensiveToSpeculativelyExecute(I);
6608}
6609
6610/// Returns true if a SelectInst should be turned into an explicit branch.
6611static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
6612 const TargetLowering *TLI,
6613 SelectInst *SI) {
6614 // If even a predictable select is cheap, then a branch can't be cheaper.
6615 if (!TLI->isPredictableSelectExpensive())
6616 return false;
6617
6618 // FIXME: This should use the same heuristics as IfConversion to determine
6619 // whether a select is better represented as a branch.
6620
6621 // If metadata tells us that the select condition is obviously predictable,
6622 // then we want to replace the select with a branch.
6623 uint64_t TrueWeight, FalseWeight;
6624 if (extractBranchWeights(*SI, TrueWeight, FalseWeight)) {
6625 uint64_t Max = std::max(TrueWeight, FalseWeight);
6626 uint64_t Sum = TrueWeight + FalseWeight;
6627 if (Sum != 0) {
6628 auto Probability = BranchProbability::getBranchProbability(Max, Sum);
6629 if (Probability > TTI->getPredictableBranchThreshold())
6630 return true;
6631 }
6632 }
6633
6634 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
6635
6636 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
6637 // comparison condition. If the compare has more than one use, there's
6638 // probably another cmov or setcc around, so it's not worth emitting a branch.
6639 if (!Cmp || !Cmp->hasOneUse())
6640 return false;
6641
6642 // If either operand of the select is expensive and only needed on one side
6643 // of the select, we should form a branch.
6644 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
6645 sinkSelectOperand(TTI, SI->getFalseValue()))
6646 return true;
6647
6648 return false;
6649}
6650
6651/// If \p isTrue is true, return the true value of \p SI, otherwise return
6652/// false value of \p SI. If the true/false value of \p SI is defined by any
6653/// select instructions in \p Selects, look through the defining select
6654/// instruction until the true/false value is not defined in \p Selects.
6655static Value *
6656getTrueOrFalseValue(SelectInst *SI, bool isTrue,
6657 const SmallPtrSet<const Instruction *, 2> &Selects) {
6658 Value *V = nullptr;
6659
6660 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
6661 DefSI = dyn_cast<SelectInst>(V)) {
6662 assert(DefSI->getCondition() == SI->getCondition() &&(static_cast <bool> (DefSI->getCondition() == SI->
getCondition() && "The condition of DefSI does not match with SI"
) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6663, __extension__ __PRETTY_FUNCTION__
))
6663 "The condition of DefSI does not match with SI")(static_cast <bool> (DefSI->getCondition() == SI->
getCondition() && "The condition of DefSI does not match with SI"
) ? void (0) : __assert_fail ("DefSI->getCondition() == SI->getCondition() && \"The condition of DefSI does not match with SI\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6663, __extension__ __PRETTY_FUNCTION__
))
;
6664 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
6665 }
6666
6667 assert(V && "Failed to get select true/false value")(static_cast <bool> (V && "Failed to get select true/false value"
) ? void (0) : __assert_fail ("V && \"Failed to get select true/false value\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6667, __extension__ __PRETTY_FUNCTION__
))
;
6668 return V;
6669}
6670
6671bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) {
6672 assert(Shift->isShift() && "Expected a shift")(static_cast <bool> (Shift->isShift() && "Expected a shift"
) ? void (0) : __assert_fail ("Shift->isShift() && \"Expected a shift\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6672, __extension__ __PRETTY_FUNCTION__
))
;
6673
6674 // If this is (1) a vector shift, (2) shifts by scalars are cheaper than
6675 // general vector shifts, and (3) the shift amount is a select-of-splatted
6676 // values, hoist the shifts before the select:
6677 // shift Op0, (select Cond, TVal, FVal) -->
6678 // select Cond, (shift Op0, TVal), (shift Op0, FVal)
6679 //
6680 // This is inverting a generic IR transform when we know that the cost of a
6681 // general vector shift is more than the cost of 2 shift-by-scalars.
6682 // We can't do this effectively in SDAG because we may not be able to
6683 // determine if the select operands are splats from within a basic block.
6684 Type *Ty = Shift->getType();
6685 if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6686 return false;
6687 Value *Cond, *TVal, *FVal;
6688 if (!match(Shift->getOperand(1),
6689 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6690 return false;
6691 if (!isSplatValue(TVal) || !isSplatValue(FVal))
6692 return false;
6693
6694 IRBuilder<> Builder(Shift);
6695 BinaryOperator::BinaryOps Opcode = Shift->getOpcode();
6696 Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal);
6697 Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal);
6698 Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6699 Shift->replaceAllUsesWith(NewSel);
6700 Shift->eraseFromParent();
6701 return true;
6702}
6703
6704bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) {
6705 Intrinsic::ID Opcode = Fsh->getIntrinsicID();
6706 assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&(static_cast <bool> ((Opcode == Intrinsic::fshl || Opcode
== Intrinsic::fshr) && "Expected a funnel shift") ? void
(0) : __assert_fail ("(Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && \"Expected a funnel shift\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6707, __extension__ __PRETTY_FUNCTION__
))
6707 "Expected a funnel shift")(static_cast <bool> ((Opcode == Intrinsic::fshl || Opcode
== Intrinsic::fshr) && "Expected a funnel shift") ? void
(0) : __assert_fail ("(Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) && \"Expected a funnel shift\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6707, __extension__ __PRETTY_FUNCTION__
))
;
6708
6709 // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper
6710 // than general vector shifts, and (3) the shift amount is select-of-splatted
6711 // values, hoist the funnel shifts before the select:
6712 // fsh Op0, Op1, (select Cond, TVal, FVal) -->
6713 // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal)
6714 //
6715 // This is inverting a generic IR transform when we know that the cost of a
6716 // general vector shift is more than the cost of 2 shift-by-scalars.
6717 // We can't do this effectively in SDAG because we may not be able to
6718 // determine if the select operands are splats from within a basic block.
6719 Type *Ty = Fsh->getType();
6720 if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty))
6721 return false;
6722 Value *Cond, *TVal, *FVal;
6723 if (!match(Fsh->getOperand(2),
6724 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))
6725 return false;
6726 if (!isSplatValue(TVal) || !isSplatValue(FVal))
6727 return false;
6728
6729 IRBuilder<> Builder(Fsh);
6730 Value *X = Fsh->getOperand(0), *Y = Fsh->getOperand(1);
6731 Value *NewTVal = Builder.CreateIntrinsic(Opcode, Ty, {X, Y, TVal});
6732 Value *NewFVal = Builder.CreateIntrinsic(Opcode, Ty, {X, Y, FVal});
6733 Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal);
6734 Fsh->replaceAllUsesWith(NewSel);
6735 Fsh->eraseFromParent();
6736 return true;
6737}
6738
6739/// If we have a SelectInst that will likely profit from branch prediction,
6740/// turn it into a branch.
6741bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
6742 if (DisableSelectToBranch)
6743 return false;
6744
6745 // If the SelectOptimize pass is enabled, selects have already been optimized.
6746 if (!getCGPassBuilderOption().DisableSelectOptimize)
6747 return false;
6748
6749 // Find all consecutive select instructions that share the same condition.
6750 SmallVector<SelectInst *, 2> ASI;
6751 ASI.push_back(SI);
6752 for (BasicBlock::iterator It = ++BasicBlock::iterator(SI);
6753 It != SI->getParent()->end(); ++It) {
6754 SelectInst *I = dyn_cast<SelectInst>(&*It);
6755 if (I && SI->getCondition() == I->getCondition()) {
6756 ASI.push_back(I);
6757 } else {
6758 break;
6759 }
6760 }
6761
6762 SelectInst *LastSI = ASI.back();
6763 // Increment the current iterator to skip all the rest of select instructions
6764 // because they will be either "not lowered" or "all lowered" to branch.
6765 CurInstIterator = std::next(LastSI->getIterator());
6766
6767 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
6768
6769 // Can we convert the 'select' to CF ?
6770 if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable))
6771 return false;
6772
6773 TargetLowering::SelectSupportKind SelectKind;
6774 if (VectorCond)
6775 SelectKind = TargetLowering::VectorMaskSelect;
6776 else if (SI->getType()->isVectorTy())
6777 SelectKind = TargetLowering::ScalarCondVectorVal;
6778 else
6779 SelectKind = TargetLowering::ScalarValSelect;
6780
6781 if (TLI->isSelectSupported(SelectKind) &&
6782 (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) || OptSize ||
6783 llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI.get())))
6784 return false;
6785
6786 // The DominatorTree needs to be rebuilt by any consumers after this
6787 // transformation. We simply reset here rather than setting the ModifiedDT
6788 // flag to avoid restarting the function walk in runOnFunction for each
6789 // select optimized.
6790 DT.reset();
6791
6792 // Transform a sequence like this:
6793 // start:
6794 // %cmp = cmp uge i32 %a, %b
6795 // %sel = select i1 %cmp, i32 %c, i32 %d
6796 //
6797 // Into:
6798 // start:
6799 // %cmp = cmp uge i32 %a, %b
6800 // %cmp.frozen = freeze %cmp
6801 // br i1 %cmp.frozen, label %select.true, label %select.false
6802 // select.true:
6803 // br label %select.end
6804 // select.false:
6805 // br label %select.end
6806 // select.end:
6807 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
6808 //
6809 // %cmp should be frozen, otherwise it may introduce undefined behavior.
6810 // In addition, we may sink instructions that produce %c or %d from
6811 // the entry block into the destination(s) of the new branch.
6812 // If the true or false blocks do not contain a sunken instruction, that
6813 // block and its branch may be optimized away. In that case, one side of the
6814 // first branch will point directly to select.end, and the corresponding PHI
6815 // predecessor block will be the start block.
6816
6817 // First, we split the block containing the select into 2 blocks.
6818 BasicBlock *StartBlock = SI->getParent();
6819 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
6820 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
6821 BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency());
6822
6823 // Delete the unconditional branch that was just created by the split.
6824 StartBlock->getTerminator()->eraseFromParent();
6825
6826 // These are the new basic blocks for the conditional branch.
6827 // At least one will become an actual new basic block.
6828 BasicBlock *TrueBlock = nullptr;
6829 BasicBlock *FalseBlock = nullptr;
6830 BranchInst *TrueBranch = nullptr;
6831 BranchInst *FalseBranch = nullptr;
6832
6833 // Sink expensive instructions into the conditional blocks to avoid executing
6834 // them speculatively.
6835 for (SelectInst *SI : ASI) {
6836 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
6837 if (TrueBlock == nullptr) {
6838 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
6839 EndBlock->getParent(), EndBlock);
6840 TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
6841 TrueBranch->setDebugLoc(SI->getDebugLoc());
6842 }
6843 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
6844 TrueInst->moveBefore(TrueBranch);
6845 }
6846 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
6847 if (FalseBlock == nullptr) {
6848 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
6849 EndBlock->getParent(), EndBlock);
6850 FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6851 FalseBranch->setDebugLoc(SI->getDebugLoc());
6852 }
6853 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
6854 FalseInst->moveBefore(FalseBranch);
6855 }
6856 }
6857
6858 // If there was nothing to sink, then arbitrarily choose the 'false' side
6859 // for a new input value to the PHI.
6860 if (TrueBlock == FalseBlock) {
6861 assert(TrueBlock == nullptr &&(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6862, __extension__ __PRETTY_FUNCTION__
))
6862 "Unexpected basic block transform while optimizing select")(static_cast <bool> (TrueBlock == nullptr && "Unexpected basic block transform while optimizing select"
) ? void (0) : __assert_fail ("TrueBlock == nullptr && \"Unexpected basic block transform while optimizing select\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6862, __extension__ __PRETTY_FUNCTION__
))
;
6863
6864 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
6865 EndBlock->getParent(), EndBlock);
6866 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
6867 FalseBranch->setDebugLoc(SI->getDebugLoc());
6868 }
6869
6870 // Insert the real conditional branch based on the original condition.
6871 // If we did not create a new block for one of the 'true' or 'false' paths
6872 // of the condition, it means that side of the branch goes to the end block
6873 // directly and the path originates from the start block from the point of
6874 // view of the new PHI.
6875 BasicBlock *TT, *FT;
6876 if (TrueBlock == nullptr) {
6877 TT = EndBlock;
6878 FT = FalseBlock;
6879 TrueBlock = StartBlock;
6880 } else if (FalseBlock == nullptr) {
6881 TT = TrueBlock;
6882 FT = EndBlock;
6883 FalseBlock = StartBlock;
6884 } else {
6885 TT = TrueBlock;
6886 FT = FalseBlock;
6887 }
6888 IRBuilder<> IB(SI);
6889 auto *CondFr = IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen");
6890 IB.CreateCondBr(CondFr, TT, FT, SI);
6891
6892 SmallPtrSet<const Instruction *, 2> INS;
6893 INS.insert(ASI.begin(), ASI.end());
6894 // Use reverse iterator because later select may use the value of the
6895 // earlier select, and we need to propagate value through earlier select
6896 // to get the PHI operand.
6897 for (SelectInst *SI : llvm::reverse(ASI)) {
6898 // The select itself is replaced with a PHI Node.
6899 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
6900 PN->takeName(SI);
6901 PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
6902 PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
6903 PN->setDebugLoc(SI->getDebugLoc());
6904
6905 SI->replaceAllUsesWith(PN);
6906 SI->eraseFromParent();
6907 INS.erase(SI);
6908 ++NumSelectsExpanded;
6909 }
6910
6911 // Instruct OptimizeBlock to skip to the next block.
6912 CurInstIterator = StartBlock->end();
6913 return true;
6914}
6915
6916/// Some targets only accept certain types for splat inputs. For example a VDUP
6917/// in MVE takes a GPR (integer) register, and the instruction that incorporate
6918/// a VDUP (such as a VADD qd, qm, rm) also require a gpr register.
6919bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
6920 // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only
6921 if (!match(SVI, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
6922 m_Undef(), m_ZeroMask())))
6923 return false;
6924 Type *NewType = TLI->shouldConvertSplatType(SVI);
6925 if (!NewType)
6926 return false;
6927
6928 auto *SVIVecType = cast<FixedVectorType>(SVI->getType());
6929 assert(!NewType->isVectorTy() && "Expected a scalar type!")(static_cast <bool> (!NewType->isVectorTy() &&
"Expected a scalar type!") ? void (0) : __assert_fail ("!NewType->isVectorTy() && \"Expected a scalar type!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6929, __extension__ __PRETTY_FUNCTION__
))
;
6930 assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&(static_cast <bool> (NewType->getScalarSizeInBits() ==
SVIVecType->getScalarSizeInBits() && "Expected a type of the same size!"
) ? void (0) : __assert_fail ("NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && \"Expected a type of the same size!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6931, __extension__ __PRETTY_FUNCTION__
))
6931 "Expected a type of the same size!")(static_cast <bool> (NewType->getScalarSizeInBits() ==
SVIVecType->getScalarSizeInBits() && "Expected a type of the same size!"
) ? void (0) : __assert_fail ("NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() && \"Expected a type of the same size!\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 6931, __extension__ __PRETTY_FUNCTION__
))
;
6932 auto *NewVecType =
6933 FixedVectorType::get(NewType, SVIVecType->getNumElements());
6934
6935 // Create a bitcast (shuffle (insert (bitcast(..))))
6936 IRBuilder<> Builder(SVI->getContext());
6937 Builder.SetInsertPoint(SVI);
6938 Value *BC1 = Builder.CreateBitCast(
6939 cast<Instruction>(SVI->getOperand(0))->getOperand(1), NewType);
6940 Value *Shuffle = Builder.CreateVectorSplat(NewVecType->getNumElements(), BC1);
6941 Value *BC2 = Builder.CreateBitCast(Shuffle, SVIVecType);
6942
6943 SVI->replaceAllUsesWith(BC2);
6944 RecursivelyDeleteTriviallyDeadInstructions(
6945 SVI, TLInfo, nullptr,
6946 [&](Value *V) { removeAllAssertingVHReferences(V); });
6947
6948 // Also hoist the bitcast up to its operand if it they are not in the same
6949 // block.
6950 if (auto *BCI = dyn_cast<Instruction>(BC1))
6951 if (auto *Op = dyn_cast<Instruction>(BCI->getOperand(0)))
6952 if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Op) &&
6953 !Op->isTerminator() && !Op->isEHPad())
6954 BCI->moveAfter(Op);
6955
6956 return true;
6957}
6958
6959bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) {
6960 // If the operands of I can be folded into a target instruction together with
6961 // I, duplicate and sink them.
6962 SmallVector<Use *, 4> OpsToSink;
6963 if (!TLI->shouldSinkOperands(I, OpsToSink))
6964 return false;
6965
6966 // OpsToSink can contain multiple uses in a use chain (e.g.
6967 // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating
6968 // uses must come first, so we process the ops in reverse order so as to not
6969 // create invalid IR.
6970 BasicBlock *TargetBB = I->getParent();
6971 bool Changed = false;
6972 SmallVector<Use *, 4> ToReplace;
6973 Instruction *InsertPoint = I;
6974 DenseMap<const Instruction *, unsigned long> InstOrdering;
6975 unsigned long InstNumber = 0;
6976 for (const auto &I : *TargetBB)
6977 InstOrdering[&I] = InstNumber++;
6978
6979 for (Use *U : reverse(OpsToSink)) {
6980 auto *UI = cast<Instruction>(U->get());
6981 if (isa<PHINode>(UI))
6982 continue;
6983 if (UI->getParent() == TargetBB) {
6984 if (InstOrdering[UI] < InstOrdering[InsertPoint])
6985 InsertPoint = UI;
6986 continue;
6987 }
6988 ToReplace.push_back(U);
6989 }
6990
6991 SetVector<Instruction *> MaybeDead;
6992 DenseMap<Instruction *, Instruction *> NewInstructions;
6993 for (Use *U : ToReplace) {
6994 auto *UI = cast<Instruction>(U->get());
6995 Instruction *NI = UI->clone();
6996 NewInstructions[UI] = NI;
6997 MaybeDead.insert(UI);
6998 LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Sinking " << *UI
<< " to user " << *I << "\n"; } } while (false
)
;
6999 NI->insertBefore(InsertPoint);
7000 InsertPoint = NI;
7001 InsertedInsts.insert(NI);
7002
7003 // Update the use for the new instruction, making sure that we update the
7004 // sunk instruction uses, if it is part of a chain that has already been
7005 // sunk.
7006 Instruction *OldI = cast<Instruction>(U->getUser());
7007 if (NewInstructions.count(OldI))
7008 NewInstructions[OldI]->setOperand(U->getOperandNo(), NI);
7009 else
7010 U->set(NI);
7011 Changed = true;
7012 }
7013
7014 // Remove instructions that are dead after sinking.
7015 for (auto *I : MaybeDead) {
7016 if (!I->hasNUsesOrMore(1)) {
7017 LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Removing dead instruction: "
<< *I << "\n"; } } while (false)
;
7018 I->eraseFromParent();
7019 }
7020 }
7021
7022 return Changed;
7023}
7024
7025bool CodeGenPrepare::optimizeSwitchType(SwitchInst *SI) {
7026 Value *Cond = SI->getCondition();
7027 Type *OldType = Cond->getType();
7028 LLVMContext &Context = Cond->getContext();
7029 EVT OldVT = TLI->getValueType(*DL, OldType);
7030 MVT RegType = TLI->getPreferredSwitchConditionType(Context, OldVT);
7031 unsigned RegWidth = RegType.getSizeInBits();
7032
7033 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
7034 return false;
7035
7036 // If the register width is greater than the type width, expand the condition
7037 // of the switch instruction and each case constant to the width of the
7038 // register. By widening the type of the switch condition, subsequent
7039 // comparisons (for case comparisons) will not need to be extended to the
7040 // preferred register width, so we will potentially eliminate N-1 extends,
7041 // where N is the number of cases in the switch.
7042 auto *NewType = Type::getIntNTy(Context, RegWidth);
7043
7044 // Extend the switch condition and case constants using the target preferred
7045 // extend unless the switch condition is a function argument with an extend
7046 // attribute. In that case, we can avoid an unnecessary mask/extension by
7047 // matching the argument extension instead.
7048 Instruction::CastOps ExtType = Instruction::ZExt;
7049 // Some targets prefer SExt over ZExt.
7050 if (TLI->isSExtCheaperThanZExt(OldVT, RegType))
7051 ExtType = Instruction::SExt;
7052
7053 if (auto *Arg = dyn_cast<Argument>(Cond)) {
7054 if (Arg->hasSExtAttr())
7055 ExtType = Instruction::SExt;
7056 if (Arg->hasZExtAttr())
7057 ExtType = Instruction::ZExt;
7058 }
7059
7060 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
7061 ExtInst->insertBefore(SI);
7062 ExtInst->setDebugLoc(SI->getDebugLoc());
7063 SI->setCondition(ExtInst);
7064 for (auto Case : SI->cases()) {
7065 const APInt &NarrowConst = Case.getCaseValue()->getValue();
7066 APInt WideConst = (ExtType == Instruction::ZExt)
7067 ? NarrowConst.zext(RegWidth)
7068 : NarrowConst.sext(RegWidth);
7069 Case.setValue(ConstantInt::get(Context, WideConst));
7070 }
7071
7072 return true;
7073}
7074
7075bool CodeGenPrepare::optimizeSwitchPhiConstants(SwitchInst *SI) {
7076 // The SCCP optimization tends to produce code like this:
7077 // switch(x) { case 42: phi(42, ...) }
7078 // Materializing the constant for the phi-argument needs instructions; So we
7079 // change the code to:
7080 // switch(x) { case 42: phi(x, ...) }
7081
7082 Value *Condition = SI->getCondition();
7083 // Avoid endless loop in degenerate case.
7084 if (isa<ConstantInt>(*Condition))
7085 return false;
7086
7087 bool Changed = false;
7088 BasicBlock *SwitchBB = SI->getParent();
7089 Type *ConditionType = Condition->getType();
7090
7091 for (const SwitchInst::CaseHandle &Case : SI->cases()) {
7092 ConstantInt *CaseValue = Case.getCaseValue();
7093 BasicBlock *CaseBB = Case.getCaseSuccessor();
7094 // Set to true if we previously checked that `CaseBB` is only reached by
7095 // a single case from this switch.
7096 bool CheckedForSinglePred = false;
7097 for (PHINode &PHI : CaseBB->phis()) {
7098 Type *PHIType = PHI.getType();
7099 // If ZExt is free then we can also catch patterns like this:
7100 // switch((i32)x) { case 42: phi((i64)42, ...); }
7101 // and replace `(i64)42` with `zext i32 %x to i64`.
7102 bool TryZExt =
7103 PHIType->isIntegerTy() &&
7104 PHIType->getIntegerBitWidth() > ConditionType->getIntegerBitWidth() &&
7105 TLI->isZExtFree(ConditionType, PHIType);
7106 if (PHIType == ConditionType || TryZExt) {
7107 // Set to true to skip this case because of multiple preds.
7108 bool SkipCase = false;
7109 Value *Replacement = nullptr;
7110 for (unsigned I = 0, E = PHI.getNumIncomingValues(); I != E; I++) {
7111 Value *PHIValue = PHI.getIncomingValue(I);
7112 if (PHIValue != CaseValue) {
7113 if (!TryZExt)
7114 continue;
7115 ConstantInt *PHIValueInt = dyn_cast<ConstantInt>(PHIValue);
7116 if (!PHIValueInt ||
7117 PHIValueInt->getValue() !=
7118 CaseValue->getValue().zext(PHIType->getIntegerBitWidth()))
7119 continue;
7120 }
7121 if (PHI.getIncomingBlock(I) != SwitchBB)
7122 continue;
7123 // We cannot optimize if there are multiple case labels jumping to
7124 // this block. This check may get expensive when there are many
7125 // case labels so we test for it last.
7126 if (!CheckedForSinglePred) {
7127 CheckedForSinglePred = true;
7128 if (SI->findCaseDest(CaseBB) == nullptr) {
7129 SkipCase = true;
7130 break;
7131 }
7132 }
7133
7134 if (Replacement == nullptr) {
7135 if (PHIValue == CaseValue) {
7136 Replacement = Condition;
7137 } else {
7138 IRBuilder<> Builder(SI);
7139 Replacement = Builder.CreateZExt(Condition, PHIType);
7140 }
7141 }
7142 PHI.setIncomingValue(I, Replacement);
7143 Changed = true;
7144 }
7145 if (SkipCase)
7146 break;
7147 }
7148 }
7149 }
7150 return Changed;
7151}
7152
7153bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
7154 bool Changed = optimizeSwitchType(SI);
7155 Changed |= optimizeSwitchPhiConstants(SI);
7156 return Changed;
7157}
7158
7159namespace {
7160
7161/// Helper class to promote a scalar operation to a vector one.
7162/// This class is used to move downward extractelement transition.
7163/// E.g.,
7164/// a = vector_op <2 x i32>
7165/// b = extractelement <2 x i32> a, i32 0
7166/// c = scalar_op b
7167/// store c
7168///
7169/// =>
7170/// a = vector_op <2 x i32>
7171/// c = vector_op a (equivalent to scalar_op on the related lane)
7172/// * d = extractelement <2 x i32> c, i32 0
7173/// * store d
7174/// Assuming both extractelement and store can be combine, we get rid of the
7175/// transition.
7176class VectorPromoteHelper {
7177 /// DataLayout associated with the current module.
7178 const DataLayout &DL;
7179
7180 /// Used to perform some checks on the legality of vector operations.
7181 const TargetLowering &TLI;
7182
7183 /// Used to estimated the cost of the promoted chain.
7184 const TargetTransformInfo &TTI;
7185
7186 /// The transition being moved downwards.
7187 Instruction *Transition;
7188
7189 /// The sequence of instructions to be promoted.
7190 SmallVector<Instruction *, 4> InstsToBePromoted;
7191
7192 /// Cost of combining a store and an extract.
7193 unsigned StoreExtractCombineCost;
7194
7195 /// Instruction that will be combined with the transition.
7196 Instruction *CombineInst = nullptr;
7197
7198 /// The instruction that represents the current end of the transition.
7199 /// Since we are faking the promotion until we reach the end of the chain
7200 /// of computation, we need a way to get the current end of the transition.
7201 Instruction *getEndOfTransition() const {
7202 if (InstsToBePromoted.empty())
7203 return Transition;
7204 return InstsToBePromoted.back();
7205 }
7206
7207 /// Return the index of the original value in the transition.
7208 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
7209 /// c, is at index 0.
7210 unsigned getTransitionOriginalValueIdx() const {
7211 assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7212, __extension__ __PRETTY_FUNCTION__
))
7212 "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7212, __extension__ __PRETTY_FUNCTION__
))
;
7213 return 0;
7214 }
7215
7216 /// Return the index of the index in the transition.
7217 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
7218 /// is at index 1.
7219 unsigned getTransitionIdx() const {
7220 assert(isa<ExtractElementInst>(Transition) &&(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7221, __extension__ __PRETTY_FUNCTION__
))
7221 "Other kind of transitions are not supported yet")(static_cast <bool> (isa<ExtractElementInst>(Transition
) && "Other kind of transitions are not supported yet"
) ? void (0) : __assert_fail ("isa<ExtractElementInst>(Transition) && \"Other kind of transitions are not supported yet\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7221, __extension__ __PRETTY_FUNCTION__
))
;
7222 return 1;
7223 }
7224
7225 /// Get the type of the transition.
7226 /// This is the type of the original value.
7227 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
7228 /// transition is <2 x i32>.
7229 Type *getTransitionType() const {
7230 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
7231 }
7232
7233 /// Promote \p ToBePromoted by moving \p Def downward through.
7234 /// I.e., we have the following sequence:
7235 /// Def = Transition <ty1> a to <ty2>
7236 /// b = ToBePromoted <ty2> Def, ...
7237 /// =>
7238 /// b = ToBePromoted <ty1> a, ...
7239 /// Def = Transition <ty1> ToBePromoted to <ty2>
7240 void promoteImpl(Instruction *ToBePromoted);
7241
7242 /// Check whether or not it is profitable to promote all the
7243 /// instructions enqueued to be promoted.
7244 bool isProfitableToPromote() {
7245 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
7246 unsigned Index = isa<ConstantInt>(ValIdx)
7247 ? cast<ConstantInt>(ValIdx)->getZExtValue()
7248 : -1;
7249 Type *PromotedType = getTransitionType();
7250
7251 StoreInst *ST = cast<StoreInst>(CombineInst);
7252 unsigned AS = ST->getPointerAddressSpace();
7253 // Check if this store is supported.
7254 if (!TLI.allowsMisalignedMemoryAccesses(
7255 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
7256 ST->getAlign())) {
7257 // If this is not supported, there is no way we can combine
7258 // the extract with the store.
7259 return false;
7260 }
7261
7262 // The scalar chain of computation has to pay for the transition
7263 // scalar to vector.
7264 // The vector chain has to account for the combining cost.
7265 InstructionCost ScalarCost =
7266 TTI.getVectorInstrCost(*Transition, PromotedType, Index);
7267 InstructionCost VectorCost = StoreExtractCombineCost;
7268 enum TargetTransformInfo::TargetCostKind CostKind =
7269 TargetTransformInfo::TCK_RecipThroughput;
7270 for (const auto &Inst : InstsToBePromoted) {
7271 // Compute the cost.
7272 // By construction, all instructions being promoted are arithmetic ones.
7273 // Moreover, one argument is a constant that can be viewed as a splat
7274 // constant.
7275 Value *Arg0 = Inst->getOperand(0);
7276 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
7277 isa<ConstantFP>(Arg0);
7278 TargetTransformInfo::OperandValueInfo Arg0Info, Arg1Info;
7279 if (IsArg0Constant)
7280 Arg0Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
7281 else
7282 Arg1Info.Kind = TargetTransformInfo::OK_UniformConstantValue;
7283
7284 ScalarCost += TTI.getArithmeticInstrCost(
7285 Inst->getOpcode(), Inst->getType(), CostKind, Arg0Info, Arg1Info);
7286 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
7287 CostKind, Arg0Info, Arg1Info);
7288 }
7289 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
<< ScalarCost << "\nVector: " << VectorCost
<< '\n'; } } while (false)
7290 dbgs() << "Estimated cost of computation to be promoted:\nScalar: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
<< ScalarCost << "\nVector: " << VectorCost
<< '\n'; } } while (false)
7291 << ScalarCost << "\nVector: " << VectorCost << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
<< ScalarCost << "\nVector: " << VectorCost
<< '\n'; } } while (false)
;
7292 return ScalarCost > VectorCost;
7293 }
7294
7295 /// Generate a constant vector with \p Val with the same
7296 /// number of elements as the transition.
7297 /// \p UseSplat defines whether or not \p Val should be replicated
7298 /// across the whole vector.
7299 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
7300 /// otherwise we generate a vector with as many undef as possible:
7301 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
7302 /// used at the index of the extract.
7303 Value *getConstantVector(Constant *Val, bool UseSplat) const {
7304 unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
7305 if (!UseSplat) {
7306 // If we cannot determine where the constant must be, we have to
7307 // use a splat constant.
7308 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
7309 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
7310 ExtractIdx = CstVal->getSExtValue();
7311 else
7312 UseSplat = true;
7313 }
7314
7315 ElementCount EC = cast<VectorType>(getTransitionType())->getElementCount();
7316 if (UseSplat)
7317 return ConstantVector::getSplat(EC, Val);
7318
7319 if (!EC.isScalable()) {
7320 SmallVector<Constant *, 4> ConstVec;
7321 UndefValue *UndefVal = UndefValue::get(Val->getType());
7322 for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) {
7323 if (Idx == ExtractIdx)
7324 ConstVec.push_back(Val);
7325 else
7326 ConstVec.push_back(UndefVal);
7327 }
7328 return ConstantVector::get(ConstVec);
7329 } else
7330 llvm_unreachable(::llvm::llvm_unreachable_internal("Generate scalable vector for non-splat is unimplemented"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7331)
7331 "Generate scalable vector for non-splat is unimplemented")::llvm::llvm_unreachable_internal("Generate scalable vector for non-splat is unimplemented"
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7331)
;
7332 }
7333
7334 /// Check if promoting to a vector type an operand at \p OperandIdx
7335 /// in \p Use can trigger undefined behavior.
7336 static bool canCauseUndefinedBehavior(const Instruction *Use,
7337 unsigned OperandIdx) {
7338 // This is not safe to introduce undef when the operand is on
7339 // the right hand side of a division-like instruction.
7340 if (OperandIdx != 1)
7341 return false;
7342 switch (Use->getOpcode()) {
7343 default:
7344 return false;
7345 case Instruction::SDiv:
7346 case Instruction::UDiv:
7347 case Instruction::SRem:
7348 case Instruction::URem:
7349 return true;
7350 case Instruction::FDiv:
7351 case Instruction::FRem:
7352 return !Use->hasNoNaNs();
7353 }
7354 llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 7354)
;
7355 }
7356
7357public:
7358 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
7359 const TargetTransformInfo &TTI, Instruction *Transition,
7360 unsigned CombineCost)
7361 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
7362 StoreExtractCombineCost(CombineCost) {
7363 assert(Transition && "Do not know how to promote null")(static_cast <bool> (Transition && "Do not know how to promote null"
) ? void (0) : __assert_fail ("Transition && \"Do not know how to promote null\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7363, __extension__ __PRETTY_FUNCTION__
))
;
7364 }
7365
7366 /// Check if we can promote \p ToBePromoted to \p Type.
7367 bool canPromote(const Instruction *ToBePromoted) const {
7368 // We could support CastInst too.
7369 return isa<BinaryOperator>(ToBePromoted);
7370 }
7371
7372 /// Check if it is profitable to promote \p ToBePromoted
7373 /// by moving downward the transition through.
7374 bool shouldPromote(const Instruction *ToBePromoted) const {
7375 // Promote only if all the operands can be statically expanded.
7376 // Indeed, we do not want to introduce any new kind of transitions.
7377 for (const Use &U : ToBePromoted->operands()) {
7378 const Value *Val = U.get();
7379 if (Val == getEndOfTransition()) {
7380 // If the use is a division and the transition is on the rhs,
7381 // we cannot promote the operation, otherwise we may create a
7382 // division by zero.
7383 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
7384 return false;
7385 continue;
7386 }
7387 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
7388 !isa<ConstantFP>(Val))
7389 return false;
7390 }
7391 // Check that the resulting operation is legal.
7392 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
7393 if (!ISDOpcode)
7394 return false;
7395 return StressStoreExtract ||
7396 TLI.isOperationLegalOrCustom(
7397 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
7398 }
7399
7400 /// Check whether or not \p Use can be combined
7401 /// with the transition.
7402 /// I.e., is it possible to do Use(Transition) => AnotherUse?
7403 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
7404
7405 /// Record \p ToBePromoted as part of the chain to be promoted.
7406 void enqueueForPromotion(Instruction *ToBePromoted) {
7407 InstsToBePromoted.push_back(ToBePromoted);
7408 }
7409
7410 /// Set the instruction that will be combined with the transition.
7411 void recordCombineInstruction(Instruction *ToBeCombined) {
7412 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine")(static_cast <bool> (canCombine(ToBeCombined) &&
"Unsupported instruction to combine") ? void (0) : __assert_fail
("canCombine(ToBeCombined) && \"Unsupported instruction to combine\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7412, __extension__ __PRETTY_FUNCTION__
))
;
7413 CombineInst = ToBeCombined;
7414 }
7415
7416 /// Promote all the instructions enqueued for promotion if it is
7417 /// is profitable.
7418 /// \return True if the promotion happened, false otherwise.
7419 bool promote() {
7420 // Check if there is something to promote.
7421 // Right now, if we do not have anything to combine with,
7422 // we assume the promotion is not profitable.
7423 if (InstsToBePromoted.empty() || !CombineInst)
7424 return false;
7425
7426 // Check cost.
7427 if (!StressStoreExtract && !isProfitableToPromote())
7428 return false;
7429
7430 // Promote.
7431 for (auto &ToBePromoted : InstsToBePromoted)
7432 promoteImpl(ToBePromoted);
7433 InstsToBePromoted.clear();
7434 return true;
7435 }
7436};
7437
7438} // end anonymous namespace
7439
7440void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
7441 // At this point, we know that all the operands of ToBePromoted but Def
7442 // can be statically promoted.
7443 // For Def, we need to use its parameter in ToBePromoted:
7444 // b = ToBePromoted ty1 a
7445 // Def = Transition ty1 b to ty2
7446 // Move the transition down.
7447 // 1. Replace all uses of the promoted operation by the transition.
7448 // = ... b => = ... Def.
7449 assert(ToBePromoted->getType() == Transition->getType() &&(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7451, __extension__ __PRETTY_FUNCTION__
))
7450 "The type of the result of the transition does not match "(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7451, __extension__ __PRETTY_FUNCTION__
))
7451 "the final type")(static_cast <bool> (ToBePromoted->getType() == Transition
->getType() && "The type of the result of the transition does not match "
"the final type") ? void (0) : __assert_fail ("ToBePromoted->getType() == Transition->getType() && \"The type of the result of the transition does not match \" \"the final type\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7451, __extension__ __PRETTY_FUNCTION__
))
;
7452 ToBePromoted->replaceAllUsesWith(Transition);
7453 // 2. Update the type of the uses.
7454 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
7455 Type *TransitionTy = getTransitionType();
7456 ToBePromoted->mutateType(TransitionTy);
7457 // 3. Update all the operands of the promoted operation with promoted
7458 // operands.
7459 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
7460 for (Use &U : ToBePromoted->operands()) {
7461 Value *Val = U.get();
7462 Value *NewVal = nullptr;
7463 if (Val == Transition)
7464 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
7465 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
7466 isa<ConstantFP>(Val)) {
7467 // Use a splat constant if it is not safe to use undef.
7468 NewVal = getConstantVector(
7469 cast<Constant>(Val),
7470 isa<UndefValue>(Val) ||
7471 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
7472 } else
7473 llvm_unreachable("Did you modified shouldPromote and forgot to update "::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update "
"this?", "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7474)
7474 "this?")::llvm::llvm_unreachable_internal("Did you modified shouldPromote and forgot to update "
"this?", "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7474)
;
7475 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
7476 }
7477 Transition->moveAfter(ToBePromoted);
7478 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
7479}
7480
7481/// Some targets can do store(extractelement) with one instruction.
7482/// Try to push the extractelement towards the stores when the target
7483/// has this feature and this is profitable.
7484bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
7485 unsigned CombineCost = std::numeric_limits<unsigned>::max();
7486 if (DisableStoreExtract ||
7487 (!StressStoreExtract &&
7488 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
7489 Inst->getOperand(1), CombineCost)))
7490 return false;
7491
7492 // At this point we know that Inst is a vector to scalar transition.
7493 // Try to move it down the def-use chain, until:
7494 // - We can combine the transition with its single use
7495 // => we got rid of the transition.
7496 // - We escape the current basic block
7497 // => we would need to check that we are moving it at a cheaper place and
7498 // we do not do that for now.
7499 BasicBlock *Parent = Inst->getParent();
7500 LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Found an interesting transition: "
<< *Inst << '\n'; } } while (false)
;
7501 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
7502 // If the transition has more than one use, assume this is not going to be
7503 // beneficial.
7504 while (Inst->hasOneUse()) {
7505 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
7506 LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Use: " << *ToBePromoted
<< '\n'; } } while (false)
;
7507
7508 if (ToBePromoted->getParent() != Parent) {
7509 LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
7510 << ToBePromoted->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
7511 << ") than the transition (" << Parent->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
7512 << ").\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Instruction to promote is in a different block ("
<< ToBePromoted->getParent()->getName() <<
") than the transition (" << Parent->getName() <<
").\n"; } } while (false)
;
7513 return false;
7514 }
7515
7516 if (VPH.canCombine(ToBePromoted)) {
7517 LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Assume " << *Inst
<< '\n' << "will be combined with: " << *ToBePromoted
<< '\n'; } } while (false)
7518 << "will be combined with: " << *ToBePromoted << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Assume " << *Inst
<< '\n' << "will be combined with: " << *ToBePromoted
<< '\n'; } } while (false)
;
7519 VPH.recordCombineInstruction(ToBePromoted);
7520 bool Changed = VPH.promote();
7521 NumStoreExtractExposed += Changed;
7522 return Changed;
7523 }
7524
7525 LLVM_DEBUG(dbgs() << "Try promoting.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Try promoting.\n"; } }
while (false)
;
7526 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
7527 return false;
7528
7529 LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Promoting is possible... Enqueue for promotion!\n"
; } } while (false)
;
7530
7531 VPH.enqueueForPromotion(ToBePromoted);
7532 Inst = ToBePromoted;
7533 }
7534 return false;
7535}
7536
7537/// For the instruction sequence of store below, F and I values
7538/// are bundled together as an i64 value before being stored into memory.
7539/// Sometimes it is more efficient to generate separate stores for F and I,
7540/// which can remove the bitwise instructions or sink them to colder places.
7541///
7542/// (store (or (zext (bitcast F to i32) to i64),
7543/// (shl (zext I to i64), 32)), addr) -->
7544/// (store F, addr) and (store I, addr+4)
7545///
7546/// Similarly, splitting for other merged store can also be beneficial, like:
7547/// For pair of {i32, i32}, i64 store --> two i32 stores.
7548/// For pair of {i32, i16}, i64 store --> two i32 stores.
7549/// For pair of {i16, i16}, i32 store --> two i16 stores.
7550/// For pair of {i16, i8}, i32 store --> two i16 stores.
7551/// For pair of {i8, i8}, i16 store --> two i8 stores.
7552///
7553/// We allow each target to determine specifically which kind of splitting is
7554/// supported.
7555///
7556/// The store patterns are commonly seen from the simple code snippet below
7557/// if only std::make_pair(...) is sroa transformed before inlined into hoo.
7558/// void goo(const std::pair<int, float> &);
7559/// hoo() {
7560/// ...
7561/// goo(std::make_pair(tmp, ftmp));
7562/// ...
7563/// }
7564///
7565/// Although we already have similar splitting in DAG Combine, we duplicate
7566/// it in CodeGenPrepare to catch the case in which pattern is across
7567/// multiple BBs. The logic in DAG Combine is kept to catch case generated
7568/// during code expansion.
7569static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL,
7570 const TargetLowering &TLI) {
7571 // Handle simple but common cases only.
7572 Type *StoreType = SI.getValueOperand()->getType();
7573
7574 // The code below assumes shifting a value by <number of bits>,
7575 // whereas scalable vectors would have to be shifted by
7576 // <2log(vscale) + number of bits> in order to store the
7577 // low/high parts. Bailing out for now.
7578 if (isa<ScalableVectorType>(StoreType))
7579 return false;
7580
7581 if (!DL.typeSizeEqualsStoreSize(StoreType) ||
7582 DL.getTypeSizeInBits(StoreType) == 0)
7583 return false;
7584
7585 unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2;
7586 Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize);
7587 if (!DL.typeSizeEqualsStoreSize(SplitStoreType))
7588 return false;
7589
7590 // Don't split the store if it is volatile.
7591 if (SI.isVolatile())
7592 return false;
7593
7594 // Match the following patterns:
7595 // (store (or (zext LValue to i64),
7596 // (shl (zext HValue to i64), 32)), HalfValBitSize)
7597 // or
7598 // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize)
7599 // (zext LValue to i64),
7600 // Expect both operands of OR and the first operand of SHL have only
7601 // one use.
7602 Value *LValue, *HValue;
7603 if (!match(SI.getValueOperand(),
7604 m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))),
7605 m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))),
7606 m_SpecificInt(HalfValBitSize))))))
7607 return false;
7608
7609 // Check LValue and HValue are int with size less or equal than 32.
7610 if (!LValue->getType()->isIntegerTy() ||
7611 DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize ||
7612 !HValue->getType()->isIntegerTy() ||
7613 DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize)
7614 return false;
7615
7616 // If LValue/HValue is a bitcast instruction, use the EVT before bitcast
7617 // as the input of target query.
7618 auto *LBC = dyn_cast<BitCastInst>(LValue);
7619 auto *HBC = dyn_cast<BitCastInst>(HValue);
7620 EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType())
7621 : EVT::getEVT(LValue->getType());
7622 EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType())
7623 : EVT::getEVT(HValue->getType());
7624 if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy))
7625 return false;
7626
7627 // Start to split store.
7628 IRBuilder<> Builder(SI.getContext());
7629 Builder.SetInsertPoint(&SI);
7630
7631 // If LValue/HValue is a bitcast in another BB, create a new one in current
7632 // BB so it may be merged with the splitted stores by dag combiner.
7633 if (LBC && LBC->getParent() != SI.getParent())
7634 LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType());
7635 if (HBC && HBC->getParent() != SI.getParent())
7636 HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType());
7637
7638 bool IsLE = SI.getModule()->getDataLayout().isLittleEndian();
7639 auto CreateSplitStore = [&](Value *V, bool Upper) {
7640 V = Builder.CreateZExtOrBitCast(V, SplitStoreType);
7641 Value *Addr = Builder.CreateBitCast(
7642 SI.getOperand(1),
7643 SplitStoreType->getPointerTo(SI.getPointerAddressSpace()));
7644 Align Alignment = SI.getAlign();
7645 const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper);
7646 if (IsOffsetStore) {
7647 Addr = Builder.CreateGEP(
7648 SplitStoreType, Addr,
7649 ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1));
7650
7651 // When splitting the store in half, naturally one half will retain the
7652 // alignment of the original wider store, regardless of whether it was
7653 // over-aligned or not, while the other will require adjustment.
7654 Alignment = commonAlignment(Alignment, HalfValBitSize / 8);
7655 }
7656 Builder.CreateAlignedStore(V, Addr, Alignment);
7657 };
7658
7659 CreateSplitStore(LValue, false);
7660 CreateSplitStore(HValue, true);
7661
7662 // Delete the old store.
7663 SI.eraseFromParent();
7664 return true;
7665}
7666
7667// Return true if the GEP has two operands, the first operand is of a sequential
7668// type, and the second operand is a constant.
7669static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) {
7670 gep_type_iterator I = gep_type_begin(*GEP);
7671 return GEP->getNumOperands() == 2 && I.isSequential() &&
7672 isa<ConstantInt>(GEP->getOperand(1));
7673}
7674
7675// Try unmerging GEPs to reduce liveness interference (register pressure) across
7676// IndirectBr edges. Since IndirectBr edges tend to touch on many blocks,
7677// reducing liveness interference across those edges benefits global register
7678// allocation. Currently handles only certain cases.
7679//
7680// For example, unmerge %GEPI and %UGEPI as below.
7681//
7682// ---------- BEFORE ----------
7683// SrcBlock:
7684// ...
7685// %GEPIOp = ...
7686// ...
7687// %GEPI = gep %GEPIOp, Idx
7688// ...
7689// indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ]
7690// (* %GEPI is alive on the indirectbr edges due to other uses ahead)
7691// (* %GEPIOp is alive on the indirectbr edges only because of it's used by
7692// %UGEPI)
7693//
7694// DstB0: ... (there may be a gep similar to %UGEPI to be unmerged)
7695// DstB1: ... (there may be a gep similar to %UGEPI to be unmerged)
7696// ...
7697//
7698// DstBi:
7699// ...
7700// %UGEPI = gep %GEPIOp, UIdx
7701// ...
7702// ---------------------------
7703//
7704// ---------- AFTER ----------
7705// SrcBlock:
7706// ... (same as above)
7707// (* %GEPI is still alive on the indirectbr edges)
7708// (* %GEPIOp is no longer alive on the indirectbr edges as a result of the
7709// unmerging)
7710// ...
7711//
7712// DstBi:
7713// ...
7714// %UGEPI = gep %GEPI, (UIdx-Idx)
7715// ...
7716// ---------------------------
7717//
7718// The register pressure on the IndirectBr edges is reduced because %GEPIOp is
7719// no longer alive on them.
7720//
7721// We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging
7722// of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as
7723// not to disable further simplications and optimizations as a result of GEP
7724// merging.
7725//
7726// Note this unmerging may increase the length of the data flow critical path
7727// (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff
7728// between the register pressure and the length of data-flow critical
7729// path. Restricting this to the uncommon IndirectBr case would minimize the
7730// impact of potentially longer critical path, if any, and the impact on compile
7731// time.
7732static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI,
7733 const TargetTransformInfo *TTI) {
7734 BasicBlock *SrcBlock = GEPI->getParent();
7735 // Check that SrcBlock ends with an IndirectBr. If not, give up. The common
7736 // (non-IndirectBr) cases exit early here.
7737 if (!isa<IndirectBrInst>(SrcBlock->getTerminator()))
7738 return false;
7739 // Check that GEPI is a simple gep with a single constant index.
7740 if (!GEPSequentialConstIndexed(GEPI))
7741 return false;
7742 ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1));
7743 // Check that GEPI is a cheap one.
7744 if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType(),
7745 TargetTransformInfo::TCK_SizeAndLatency) >
7746 TargetTransformInfo::TCC_Basic)
7747 return false;
7748 Value *GEPIOp = GEPI->getOperand(0);
7749 // Check that GEPIOp is an instruction that's also defined in SrcBlock.
7750 if (!isa<Instruction>(GEPIOp))
7751 return false;
7752 auto *GEPIOpI = cast<Instruction>(GEPIOp);
7753 if (GEPIOpI->getParent() != SrcBlock)
7754 return false;
7755 // Check that GEP is used outside the block, meaning it's alive on the
7756 // IndirectBr edge(s).
7757 if (llvm::none_of(GEPI->users(), [&](User *Usr) {
7758 if (auto *I = dyn_cast<Instruction>(Usr)) {
7759 if (I->getParent() != SrcBlock) {
7760 return true;
7761 }
7762 }
7763 return false;
7764 }))
7765 return false;
7766 // The second elements of the GEP chains to be unmerged.
7767 std::vector<GetElementPtrInst *> UGEPIs;
7768 // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive
7769 // on IndirectBr edges.
7770 for (User *Usr : GEPIOp->users()) {
7771 if (Usr == GEPI)
7772 continue;
7773 // Check if Usr is an Instruction. If not, give up.
7774 if (!isa<Instruction>(Usr))
7775 return false;
7776 auto *UI = cast<Instruction>(Usr);
7777 // Check if Usr in the same block as GEPIOp, which is fine, skip.
7778 if (UI->getParent() == SrcBlock)
7779 continue;
7780 // Check if Usr is a GEP. If not, give up.
7781 if (!isa<GetElementPtrInst>(Usr))
7782 return false;
7783 auto *UGEPI = cast<GetElementPtrInst>(Usr);
7784 // Check if UGEPI is a simple gep with a single constant index and GEPIOp is
7785 // the pointer operand to it. If so, record it in the vector. If not, give
7786 // up.
7787 if (!GEPSequentialConstIndexed(UGEPI))
7788 return false;
7789 if (UGEPI->getOperand(0) != GEPIOp)
7790 return false;
7791 if (GEPIIdx->getType() !=
7792 cast<ConstantInt>(UGEPI->getOperand(1))->getType())
7793 return false;
7794 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7795 if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType(),
7796 TargetTransformInfo::TCK_SizeAndLatency) >
7797 TargetTransformInfo::TCC_Basic)
7798 return false;
7799 UGEPIs.push_back(UGEPI);
7800 }
7801 if (UGEPIs.size() == 0)
7802 return false;
7803 // Check the materializing cost of (Uidx-Idx).
7804 for (GetElementPtrInst *UGEPI : UGEPIs) {
7805 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7806 APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue();
7807 InstructionCost ImmCost = TTI->getIntImmCost(
7808 NewIdx, GEPIIdx->getType(), TargetTransformInfo::TCK_SizeAndLatency);
7809 if (ImmCost > TargetTransformInfo::TCC_Basic)
7810 return false;
7811 }
7812 // Now unmerge between GEPI and UGEPIs.
7813 for (GetElementPtrInst *UGEPI : UGEPIs) {
7814 UGEPI->setOperand(0, GEPI);
7815 ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1));
7816 Constant *NewUGEPIIdx = ConstantInt::get(
7817 GEPIIdx->getType(), UGEPIIdx->getValue() - GEPIIdx->getValue());
7818 UGEPI->setOperand(1, NewUGEPIIdx);
7819 // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not
7820 // inbounds to avoid UB.
7821 if (!GEPI->isInBounds()) {
7822 UGEPI->setIsInBounds(false);
7823 }
7824 }
7825 // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not
7826 // alive on IndirectBr edges).
7827 assert(llvm::none_of(GEPIOp->users(),(static_cast <bool> (llvm::none_of(GEPIOp->users(), [
&](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("llvm::none_of(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) && \"GEPIOp is used outside SrcBlock\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7831, __extension__ __PRETTY_FUNCTION__
))
7828 [&](User *Usr) {(static_cast <bool> (llvm::none_of(GEPIOp->users(), [
&](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("llvm::none_of(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) && \"GEPIOp is used outside SrcBlock\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7831, __extension__ __PRETTY_FUNCTION__
))
7829 return cast<Instruction>(Usr)->getParent() != SrcBlock;(static_cast <bool> (llvm::none_of(GEPIOp->users(), [
&](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("llvm::none_of(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) && \"GEPIOp is used outside SrcBlock\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7831, __extension__ __PRETTY_FUNCTION__
))
7830 }) &&(static_cast <bool> (llvm::none_of(GEPIOp->users(), [
&](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("llvm::none_of(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) && \"GEPIOp is used outside SrcBlock\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7831, __extension__ __PRETTY_FUNCTION__
))
7831 "GEPIOp is used outside SrcBlock")(static_cast <bool> (llvm::none_of(GEPIOp->users(), [
&](User *Usr) { return cast<Instruction>(Usr)->getParent
() != SrcBlock; }) && "GEPIOp is used outside SrcBlock"
) ? void (0) : __assert_fail ("llvm::none_of(GEPIOp->users(), [&](User *Usr) { return cast<Instruction>(Usr)->getParent() != SrcBlock; }) && \"GEPIOp is used outside SrcBlock\""
, "llvm/lib/CodeGen/CodeGenPrepare.cpp", 7831, __extension__ __PRETTY_FUNCTION__
))
;
7832 return true;
7833}
7834
7835static bool optimizeBranch(BranchInst *Branch, const TargetLowering &TLI) {
7836 // Try and convert
7837 // %c = icmp ult %x, 8
7838 // br %c, bla, blb
7839 // %tc = lshr %x, 3
7840 // to
7841 // %tc = lshr %x, 3
7842 // %c = icmp eq %tc, 0
7843 // br %c, bla, blb
7844 // Creating the cmp to zero can be better for the backend, especially if the
7845 // lshr produces flags that can be used automatically.
7846 if (!TLI.preferZeroCompareBranch() || !Branch->isConditional())
7847 return false;
7848
7849 ICmpInst *Cmp = dyn_cast<ICmpInst>(Branch->getCondition());
7850 if (!Cmp || !isa<ConstantInt>(Cmp->getOperand(1)) || !Cmp->hasOneUse())
7851 return false;
7852
7853 Value *X = Cmp->getOperand(0);
7854 APInt CmpC = cast<ConstantInt>(Cmp->getOperand(1))->getValue();
7855
7856 for (auto *U : X->users()) {
7857 Instruction *UI = dyn_cast<Instruction>(U);
7858 // A quick dominance check
7859 if (!UI ||
7860 (UI->getParent() != Branch->getParent() &&
7861 UI->getParent() != Branch->getSuccessor(0) &&
7862 UI->getParent() != Branch->getSuccessor(1)) ||
7863 (UI->getParent() != Branch->getParent() &&
7864 !UI->getParent()->getSinglePredecessor()))
7865 continue;
7866
7867 if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT &&
7868 match(UI, m_Shr(m_Specific(X), m_SpecificInt(CmpC.logBase2())))) {
7869 IRBuilder<> Builder(Branch);
7870 if (UI->getParent() != Branch->getParent())
7871 UI->moveBefore(Branch);
7872 Value *NewCmp = Builder.CreateCmp(ICmpInst::ICMP_EQ, UI,
7873 ConstantInt::get(UI->getType(), 0));
7874 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Converting " << *
Cmp << "\n"; } } while (false)
;
7875 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << " to compare on zero: "
<< *NewCmp << "\n"; } } while (false)
;
7876 Cmp->replaceAllUsesWith(NewCmp);
7877 return true;
7878 }
7879 if (Cmp->isEquality() &&
7880 (match(UI, m_Add(m_Specific(X), m_SpecificInt(-CmpC))) ||
7881 match(UI, m_Sub(m_Specific(X), m_SpecificInt(CmpC))))) {
7882 IRBuilder<> Builder(Branch);
7883 if (UI->getParent() != Branch->getParent())
7884 UI->moveBefore(Branch);
7885 Value *NewCmp = Builder.CreateCmp(Cmp->getPredicate(), UI,
7886 ConstantInt::get(UI->getType(), 0));
7887 LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Converting " << *
Cmp << "\n"; } } while (false)
;
7888 LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << " to compare on zero: "
<< *NewCmp << "\n"; } } while (false)
;
7889 Cmp->replaceAllUsesWith(NewCmp);
7890 return true;
7891 }
7892 }
7893 return false;
7894}
7895
7896bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) {
7897 // Bail out if we inserted the instruction to prevent optimizations from
7898 // stepping on each other's toes.
7899 if (InsertedInsts.count(I))
7900 return false;
7901
7902 // TODO: Move into the switch on opcode below here.
7903 if (PHINode *P = dyn_cast<PHINode>(I)) {
7904 // It is possible for very late stage optimizations (such as SimplifyCFG)
7905 // to introduce PHI nodes too late to be cleaned up. If we detect such a
7906 // trivial PHI, go ahead and zap it here.
7907 if (Value *V = simplifyInstruction(P, {*DL, TLInfo})) {
7908 LargeOffsetGEPMap.erase(P);
7909 P->replaceAllUsesWith(V);
7910 P->eraseFromParent();
7911 ++NumPHIsElim;
7912 return true;
7913 }
7914 return false;
7915 }
7916
7917 if (CastInst *CI = dyn_cast<CastInst>(I)) {
7918 // If the source of the cast is a constant, then this should have
7919 // already been constant folded. The only reason NOT to constant fold
7920 // it is if something (e.g. LSR) was careful to place the constant
7921 // evaluation in a block other than then one that uses it (e.g. to hoist
7922 // the address of globals out of a loop). If this is the case, we don't
7923 // want to forward-subst the cast.
7924 if (isa<Constant>(CI->getOperand(0)))
7925 return false;
7926
7927 if (OptimizeNoopCopyExpression(CI, *TLI, *DL))
7928 return true;
7929
7930 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7931 /// Sink a zext or sext into its user blocks if the target type doesn't
7932 /// fit in one register
7933 if (TLI->getTypeAction(CI->getContext(),
7934 TLI->getValueType(*DL, CI->getType())) ==
7935 TargetLowering::TypeExpandInteger) {
7936 return SinkCast(CI);
7937 } else {
7938 bool MadeChange = optimizeExt(I);
7939 return MadeChange | optimizeExtUses(I);
7940 }
7941 }
7942 return false;
7943 }
7944
7945 if (auto *Cmp = dyn_cast<CmpInst>(I))
7946 if (optimizeCmp(Cmp, ModifiedDT))
7947 return true;
7948
7949 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7950 LI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7951 bool Modified = optimizeLoadExt(LI);
7952 unsigned AS = LI->getPointerAddressSpace();
7953 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
7954 return Modified;
7955 }
7956
7957 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
7958 if (splitMergedValStore(*SI, *DL, *TLI))
7959 return true;
7960 SI->setMetadata(LLVMContext::MD_invariant_group, nullptr);
7961 unsigned AS = SI->getPointerAddressSpace();
7962 return optimizeMemoryInst(I, SI->getOperand(1),
7963 SI->getOperand(0)->getType(), AS);
7964 }
7965
7966 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
7967 unsigned AS = RMW->getPointerAddressSpace();
7968 return optimizeMemoryInst(I, RMW->getPointerOperand(), RMW->getType(), AS);
7969 }
7970
7971 if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
7972 unsigned AS = CmpX->getPointerAddressSpace();
7973 return optimizeMemoryInst(I, CmpX->getPointerOperand(),
7974 CmpX->getCompareOperand()->getType(), AS);
7975 }
7976
7977 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
7978
7979 if (BinOp && BinOp->getOpcode() == Instruction::And && EnableAndCmpSinking &&
7980 sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts))
7981 return true;
7982
7983 // TODO: Move this into the switch on opcode - it handles shifts already.
7984 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
7985 BinOp->getOpcode() == Instruction::LShr)) {
7986 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
7987 if (CI && TLI->hasExtractBitsInsn())
7988 if (OptimizeExtractBits(BinOp, CI, *TLI, *DL))
7989 return true;
7990 }
7991
7992 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
7993 if (GEPI->hasAllZeroIndices()) {
7994 /// The GEP operand must be a pointer, so must its result -> BitCast
7995 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
7996 GEPI->getName(), GEPI);
7997 NC->setDebugLoc(GEPI->getDebugLoc());
7998 GEPI->replaceAllUsesWith(NC);
7999 GEPI->eraseFromParent();
8000 ++NumGEPsElim;
8001 optimizeInst(NC, ModifiedDT);
8002 return true;
8003 }
8004 if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) {
8005 return true;
8006 }
8007 return false;
8008 }
8009
8010 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
8011 // freeze(icmp a, const)) -> icmp (freeze a), const
8012 // This helps generate efficient conditional jumps.
8013 Instruction *CmpI = nullptr;
8014 if (ICmpInst *II = dyn_cast<ICmpInst>(FI->getOperand(0)))
8015 CmpI = II;
8016 else if (FCmpInst *F = dyn_cast<FCmpInst>(FI->getOperand(0)))
8017 CmpI = F->getFastMathFlags().none() ? F : nullptr;
8018
8019 if (CmpI && CmpI->hasOneUse()) {
8020 auto Op0 = CmpI->getOperand(0), Op1 = CmpI->getOperand(1);
8021 bool Const0 = isa<ConstantInt>(Op0) || isa<ConstantFP>(Op0) ||
8022 isa<ConstantPointerNull>(Op0);
8023 bool Const1 = isa<ConstantInt>(Op1) || isa<ConstantFP>(Op1) ||
8024 isa<ConstantPointerNull>(Op1);
8025 if (Const0 || Const1) {
8026 if (!Const0 || !Const1) {
8027 auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI);
8028 F->takeName(FI);
8029 CmpI->setOperand(Const0 ? 1 : 0, F);
8030 }
8031 FI->replaceAllUsesWith(CmpI);
8032 FI->eraseFromParent();
8033 return true;
8034 }
8035 }
8036 return false;
8037 }
8038
8039 if (tryToSinkFreeOperands(I))
8040 return true;
8041
8042 switch (I->getOpcode()) {
8043 case Instruction::Shl:
8044 case Instruction::LShr:
8045 case Instruction::AShr:
8046 return optimizeShiftInst(cast<BinaryOperator>(I));
8047 case Instruction::Call:
8048 return optimizeCallInst(cast<CallInst>(I), ModifiedDT);
8049 case Instruction::Select:
8050 return optimizeSelectInst(cast<SelectInst>(I));
8051 case Instruction::ShuffleVector:
8052 return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I));
8053 case Instruction::Switch:
8054 return optimizeSwitchInst(cast<SwitchInst>(I));
8055 case Instruction::ExtractElement:
8056 return optimizeExtractElementInst(cast<ExtractElementInst>(I));
8057 case Instruction::Br:
8058 return optimizeBranch(cast<BranchInst>(I), *TLI);
8059 }
8060
8061 return false;
8062}
8063
8064/// Given an OR instruction, check to see if this is a bitreverse
8065/// idiom. If so, insert the new intrinsic and return true.
8066bool CodeGenPrepare::makeBitReverse(Instruction &I) {
8067 if (!I.getType()->isIntegerTy() ||
8068 !TLI->isOperationLegalOrCustom(ISD::BITREVERSE,
8069 TLI->getValueType(*DL, I.getType(), true)))
8070 return false;
8071
8072 SmallVector<Instruction *, 4> Insts;
8073 if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts))
8074 return false;
8075 Instruction *LastInst = Insts.back();
8076 I.replaceAllUsesWith(LastInst);
8077 RecursivelyDeleteTriviallyDeadInstructions(
8078 &I, TLInfo, nullptr,
8079 [&](Value *V) { removeAllAssertingVHReferences(V); });
8080 return true;
8081}
8082
8083// In this pass we look for GEP and cast instructions that are used
8084// across basic blocks and rewrite them to improve basic-block-at-a-time
8085// selection.
8086bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) {
8087 SunkAddrs.clear();
8088 bool MadeChange = false;
8089
8090 CurInstIterator = BB.begin();
8091 while (CurInstIterator != BB.end()) {
8092 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
8093 if (ModifiedDT)
8094 return true;
8095 }
8096
8097 bool MadeBitReverse = true;
8098 while (MadeBitReverse) {
8099 MadeBitReverse = false;
8100 for (auto &I : reverse(BB)) {
8101 if (makeBitReverse(I)) {
8102 MadeBitReverse = MadeChange = true;
8103 break;
8104 }
8105 }
8106 }
8107 MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT);
8108
8109 return MadeChange;
8110}
8111
8112// Some CGP optimizations may move or alter what's computed in a block. Check
8113// whether a dbg.value intrinsic could be pointed at a more appropriate operand.
8114bool CodeGenPrepare::fixupDbgValue(Instruction *I) {
8115 assert(isa<DbgValueInst>(I))(static_cast <bool> (isa<DbgValueInst>(I)) ? void
(0) : __assert_fail ("isa<DbgValueInst>(I)", "llvm/lib/CodeGen/CodeGenPrepare.cpp"
, 8115, __extension__ __PRETTY_FUNCTION__))
;
8116 DbgValueInst &DVI = *cast<DbgValueInst>(I);
8117
8118 // Does this dbg.value refer to a sunk address calculation?
8119 bool AnyChange = false;
8120 SmallDenseSet<Value *> LocationOps(DVI.location_ops().begin(),
8121 DVI.location_ops().end());
8122 for (Value *Location : LocationOps) {
8123 WeakTrackingVH SunkAddrVH = SunkAddrs[Location];
8124 Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
8125 if (SunkAddr) {
8126 // Point dbg.value at locally computed address, which should give the best
8127 // opportunity to be accurately lowered. This update may change the type
8128 // of pointer being referred to; however this makes no difference to
8129 // debugging information, and we can't generate bitcasts that may affect
8130 // codegen.
8131 DVI.replaceVariableLocationOp(Location, SunkAddr);
8132 AnyChange = true;
8133 }
8134 }
8135 return AnyChange;
8136}
8137
8138// A llvm.dbg.value may be using a value before its definition, due to
8139// optimizations in this pass and others. Scan for such dbg.values, and rescue
8140// them by moving the dbg.value to immediately after the value definition.
8141// FIXME: Ideally this should never be necessary, and this has the potential
8142// to re-order dbg.value intrinsics.
8143bool CodeGenPrepare::placeDbgValues(Function &F) {
8144 bool MadeChange = false;
8145 DominatorTree DT(F);
8146
8147 for (BasicBlock &BB : F) {
8148 for (Instruction &Insn : llvm::make_early_inc_range(BB)) {
8149 DbgValueInst *DVI = dyn_cast<DbgValueInst>(&Insn);
8150 if (!DVI)
8151 continue;
8152
8153 SmallVector<Instruction *, 4> VIs;
8154 for (Value *V : DVI->getValues())
8155 if (Instruction *VI = dyn_cast_or_null<Instruction>(V))
8156 VIs.push_back(VI);
8157
8158 // This DVI may depend on multiple instructions, complicating any
8159 // potential sink. This block takes the defensive approach, opting to
8160 // "undef" the DVI if it has more than one instruction and any of them do
8161 // not dominate DVI.
8162 for (Instruction *VI : VIs) {
8163 if (VI->isTerminator())
8164 continue;
8165
8166 // If VI is a phi in a block with an EHPad terminator, we can't insert
8167 // after it.
8168 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
8169 continue;
8170
8171 // If the defining instruction dominates the dbg.value, we do not need
8172 // to move the dbg.value.
8173 if (DT.dominates(VI, DVI))
8174 continue;
8175
8176 // If we depend on multiple instructions and any of them doesn't
8177 // dominate this DVI, we probably can't salvage it: moving it to
8178 // after any of the instructions could cause us to lose the others.
8179 if (VIs.size() > 1) {
8180 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n"
<< *DVI; } } while (false)
8181 dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n"
<< *DVI; } } while (false)
8182 << "Unable to find valid location for Debug Value, undefing:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n"
<< *DVI; } } while (false)
8183 << *DVI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Unable to find valid location for Debug Value, undefing:\n"
<< *DVI; } } while (false)
;
8184 DVI->setUndef();
8185 break;
8186 }
8187
8188 LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Moving Debug Value before :\n"
<< *DVI << ' ' << *VI; } } while (false)
8189 << *DVI << ' ' << *VI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Moving Debug Value before :\n"
<< *DVI << ' ' << *VI; } } while (false)
;
8190 DVI->removeFromParent();
8191 if (isa<PHINode>(VI))
8192 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
8193 else
8194 DVI->insertAfter(VI);
8195 MadeChange = true;
8196 ++NumDbgValueMoved;
8197 }
8198 }
8199 }
8200 return MadeChange;
8201}
8202
8203// Group scattered pseudo probes in a block to favor SelectionDAG. Scattered
8204// probes can be chained dependencies of other regular DAG nodes and block DAG
8205// combine optimizations.
8206bool CodeGenPrepare::placePseudoProbes(Function &F) {
8207 bool MadeChange = false;
8208 for (auto &Block : F) {
8209 // Move the rest probes to the beginning of the block.
8210 auto FirstInst = Block.getFirstInsertionPt();
8211 while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst())
8212 ++FirstInst;
8213 BasicBlock::iterator I(FirstInst);
8214 I++;
8215 while (I != Block.end()) {
8216 if (auto *II = dyn_cast<PseudoProbeInst>(I++)) {
8217 II->moveBefore(&*FirstInst);
8218 MadeChange = true;
8219 }
8220 }
8221 }
8222 return MadeChange;
8223}
8224
8225/// Scale down both weights to fit into uint32_t.
8226static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
8227 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
8228 uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1;
8229 NewTrue = NewTrue / Scale;
8230 NewFalse = NewFalse / Scale;
8231}
8232
8233/// Some targets prefer to split a conditional branch like:
8234/// \code
8235/// %0 = icmp ne i32 %a, 0
8236/// %1 = icmp ne i32 %b, 0
8237/// %or.cond = or i1 %0, %1
8238/// br i1 %or.cond, label %TrueBB, label %FalseBB
8239/// \endcode
8240/// into multiple branch instructions like:
8241/// \code
8242/// bb1:
8243/// %0 = icmp ne i32 %a, 0
8244/// br i1 %0, label %TrueBB, label %bb2
8245/// bb2:
8246/// %1 = icmp ne i32 %b, 0
8247/// br i1 %1, label %TrueBB, label %FalseBB
8248/// \endcode
8249/// This usually allows instruction selection to do even further optimizations
8250/// and combine the compare with the branch instruction. Currently this is
8251/// applied for targets which have "cheap" jump instructions.
8252///
8253/// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
8254///
8255bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) {
8256 if (!TM->Options.EnableFastISel || TLI->isJumpExpensive())
8257 return false;
8258
8259 bool MadeChange = false;
8260 for (auto &BB : F) {
8261 // Does this BB end with the following?
8262 // %cond1 = icmp|fcmp|binary instruction ...
8263 // %cond2 = icmp|fcmp|binary instruction ...
8264 // %cond.or = or|and i1 %cond1, cond2
8265 // br i1 %cond.or label %dest1, label %dest2"
8266 Instruction *LogicOp;
8267 BasicBlock *TBB, *FBB;
8268 if (!match(BB.getTerminator(),
8269 m_Br(m_OneUse(m_Instruction(LogicOp)), TBB, FBB)))
8270 continue;
8271
8272 auto *Br1 = cast<BranchInst>(BB.getTerminator());
8273 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
8274 continue;
8275
8276 // The merging of mostly empty BB can cause a degenerate branch.
8277 if (TBB == FBB)
8278 continue;
8279
8280 unsigned Opc;
8281 Value *Cond1, *Cond2;
8282 if (match(LogicOp,
8283 m_LogicalAnd(m_OneUse(m_Value(Cond1)), m_OneUse(m_Value(Cond2)))))
8284 Opc = Instruction::And;
8285 else if (match(LogicOp, m_LogicalOr(m_OneUse(m_Value(Cond1)),
8286 m_OneUse(m_Value(Cond2)))))
8287 Opc = Instruction::Or;
8288 else
8289 continue;
8290
8291 auto IsGoodCond = [](Value *Cond) {
8292 return match(
8293 Cond,
8294 m_CombineOr(m_Cmp(), m_CombineOr(m_LogicalAnd(m_Value(), m_Value()),
8295 m_LogicalOr(m_Value(), m_Value()))));
8296 };
8297 if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2))
8298 continue;
8299
8300 LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "Before branch condition splitting\n"
; BB.dump(); } } while (false)
;
8301
8302 // Create a new BB.
8303 auto *TmpBB =
8304 BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split",
8305 BB.getParent(), BB.getNextNode());
8306
8307 // Update original basic block by using the first condition directly by the
8308 // branch instruction and removing the no longer needed and/or instruction.
8309 Br1->setCondition(Cond1);
8310 LogicOp->eraseFromParent();
8311
8312 // Depending on the condition we have to either replace the true or the
8313 // false successor of the original branch instruction.
8314 if (Opc == Instruction::And)
8315 Br1->setSuccessor(0, TmpBB);
8316 else
8317 Br1->setSuccessor(1, TmpBB);
8318
8319 // Fill in the new basic block.
8320 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
8321 if (auto *I = dyn_cast<Instruction>(Cond2)) {
8322 I->removeFromParent();
8323 I->insertBefore(Br2);
8324 }
8325
8326 // Update PHI nodes in both successors. The original BB needs to be
8327 // replaced in one successor's PHI nodes, because the branch comes now from
8328 // the newly generated BB (NewBB). In the other successor we need to add one
8329 // incoming edge to the PHI nodes, because both branch instructions target
8330 // now the same successor. Depending on the original branch condition
8331 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
8332 // we perform the correct update for the PHI nodes.
8333 // This doesn't change the successor order of the just created branch
8334 // instruction (or any other instruction).
8335 if (Opc == Instruction::Or)
8336 std::swap(TBB, FBB);
8337
8338 // Replace the old BB with the new BB.
8339 TBB->replacePhiUsesWith(&BB, TmpBB);
8340
8341 // Add another incoming edge form the new BB.
8342 for (PHINode &PN : FBB->phis()) {
8343 auto *Val = PN.getIncomingValueForBlock(&BB);
8344 PN.addIncoming(Val, TmpBB);
8345 }
8346
8347 // Update the branch weights (from SelectionDAGBuilder::
8348 // FindMergedConditions).
8349 if (Opc == Instruction::Or) {
8350 // Codegen X | Y as:
8351 // BB1:
8352 // jmp_if_X TBB
8353 // jmp TmpBB
8354 // TmpBB:
8355 // jmp_if_Y TBB
8356 // jmp FBB
8357 //
8358
8359 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
8360 // The requirement is that
8361 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
8362 // = TrueProb for original BB.
8363 // Assuming the original weights are A and B, one choice is to set BB1's
8364 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
8365 // assumes that
8366 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
8367 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
8368 // TmpBB, but the math is more complicated.
8369 uint64_t TrueWeight, FalseWeight;
8370 if (extractBranchWeights(*Br1, TrueWeight, FalseWeight)) {
8371 uint64_t NewTrueWeight = TrueWeight;
8372 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
8373 scaleWeights(NewTrueWeight, NewFalseWeight);
8374 Br1->setMetadata(LLVMContext::MD_prof,
8375 MDBuilder(Br1->getContext())
8376 .createBranchWeights(TrueWeight, FalseWeight));
8377
8378 NewTrueWeight = TrueWeight;
8379 NewFalseWeight = 2 * FalseWeight;
8380 scaleWeights(NewTrueWeight, NewFalseWeight);
8381 Br2->setMetadata(LLVMContext::MD_prof,
8382 MDBuilder(Br2->getContext())
8383 .createBranchWeights(TrueWeight, FalseWeight));
8384 }
8385 } else {
8386 // Codegen X & Y as:
8387 // BB1:
8388 // jmp_if_X TmpBB
8389 // jmp FBB
8390 // TmpBB:
8391 // jmp_if_Y TBB
8392 // jmp FBB
8393 //
8394 // This requires creation of TmpBB after CurBB.
8395
8396 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
8397 // The requirement is that
8398 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
8399 // = FalseProb for original BB.
8400 // Assuming the original weights are A and B, one choice is to set BB1's
8401 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
8402 // assumes that
8403 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
8404 uint64_t TrueWeight, FalseWeight;
8405 if (extractBranchWeights(*Br1, TrueWeight, FalseWeight)) {
8406 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
8407 uint64_t NewFalseWeight = FalseWeight;
8408 scaleWeights(NewTrueWeight, NewFalseWeight);
8409 Br1->setMetadata(LLVMContext::MD_prof,
8410 MDBuilder(Br1->getContext())
8411 .createBranchWeights(TrueWeight, FalseWeight));
8412
8413 NewTrueWeight = 2 * TrueWeight;
8414 NewFalseWeight = FalseWeight;
8415 scaleWeights(NewTrueWeight, NewFalseWeight);
8416 Br2->setMetadata(LLVMContext::MD_prof,
8417 MDBuilder(Br2->getContext())
8418 .createBranchWeights(TrueWeight, FalseWeight));
8419 }
8420 }
8421
8422 ModifiedDT = true;
8423 MadeChange = true;
8424
8425 LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "After branch condition splitting\n"
; BB.dump(); TmpBB->dump(); } } while (false)
8426 TmpBB->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("codegenprepare")) { dbgs() << "After branch condition splitting\n"
; BB.dump(); TmpBB->dump(); } } while (false)
;
8427 }
8428 return MadeChange;
8429}

/build/llvm-toolchain-snapshot-16~++20220904122748+c444af1c20b3/llvm/include/llvm/IR/Instructions.h

1//===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file exposes the class definitions of all of the subclasses of the
10// Instruction class. This is meant to be an easy way to get access to all
11// instruction subclasses.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_IR_INSTRUCTIONS_H
16#define LLVM_IR_INSTRUCTIONS_H
17
18#include "llvm/ADT/ArrayRef.h"
19#include "llvm/ADT/Bitfields.h"
20#include "llvm/ADT/MapVector.h"
21#include "llvm/ADT/None.h"
22#include "llvm/ADT/STLExtras.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/Twine.h"
25#include "llvm/ADT/iterator.h"
26#include "llvm/ADT/iterator_range.h"
27#include "llvm/IR/CFG.h"
28#include "llvm/IR/Constant.h"
29#include "llvm/IR/DerivedTypes.h"
30#include "llvm/IR/InstrTypes.h"
31#include "llvm/IR/Instruction.h"
32#include "llvm/IR/OperandTraits.h"
33#include "llvm/IR/Use.h"
34#include "llvm/IR/User.h"
35#include "llvm/Support/AtomicOrdering.h"
36#include "llvm/Support/ErrorHandling.h"
37#include <cassert>
38#include <cstddef>
39#include <cstdint>
40#include <iterator>
41
42namespace llvm {
43
44class APFloat;
45class APInt;
46class BasicBlock;
47class BlockAddress;
48class ConstantInt;
49class DataLayout;
50class StringRef;
51class Type;
52class Value;
53
54//===----------------------------------------------------------------------===//
55// AllocaInst Class
56//===----------------------------------------------------------------------===//
57
58/// an instruction to allocate memory on the stack
59class AllocaInst : public UnaryInstruction {
60 Type *AllocatedType;
61
62 using AlignmentField = AlignmentBitfieldElementT<0>;
63 using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>;
64 using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>;
65 static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField,
66 SwiftErrorField>(),
67 "Bitfields must be contiguous");
68
69protected:
70 // Note: Instruction needs to be a friend here to call cloneImpl.
71 friend class Instruction;
72
73 AllocaInst *cloneImpl() const;
74
75public:
76 explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
77 const Twine &Name, Instruction *InsertBefore);
78 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
79 const Twine &Name, BasicBlock *InsertAtEnd);
80
81 AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name,
82 Instruction *InsertBefore);
83 AllocaInst(Type *Ty, unsigned AddrSpace,
84 const Twine &Name, BasicBlock *InsertAtEnd);
85
86 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
87 const Twine &Name = "", Instruction *InsertBefore = nullptr);
88 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
89 const Twine &Name, BasicBlock *InsertAtEnd);
90
91 /// Return true if there is an allocation size parameter to the allocation
92 /// instruction that is not 1.
93 bool isArrayAllocation() const;
94
95 /// Get the number of elements allocated. For a simple allocation of a single
96 /// element, this will return a constant 1 value.
97 const Value *getArraySize() const { return getOperand(0); }
98 Value *getArraySize() { return getOperand(0); }
99
100 /// Overload to return most specific pointer type.
101 PointerType *getType() const {
102 return cast<PointerType>(Instruction::getType());
103 }
104
105 /// Return the address space for the allocation.
106 unsigned getAddressSpace() const {
107 return getType()->getAddressSpace();
108 }
109
110 /// Get allocation size in bits. Returns None if size can't be determined,
111 /// e.g. in case of a VLA.
112 Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const;
113
114 /// Return the type that is being allocated by the instruction.
115 Type *getAllocatedType() const { return AllocatedType; }
116 /// for use only in special circumstances that need to generically
117 /// transform a whole instruction (eg: IR linking and vectorization).
118 void setAllocatedType(Type *Ty) { AllocatedType = Ty; }
119
120 /// Return the alignment of the memory that is being allocated by the
121 /// instruction.
122 Align getAlign() const {
123 return Align(1ULL << getSubclassData<AlignmentField>());
124 }
125
126 void setAlignment(Align Align) {
127 setSubclassData<AlignmentField>(Log2(Align));
128 }
129
130 /// Return true if this alloca is in the entry block of the function and is a
131 /// constant size. If so, the code generator will fold it into the
132 /// prolog/epilog code, so it is basically free.
133 bool isStaticAlloca() const;
134
135 /// Return true if this alloca is used as an inalloca argument to a call. Such
136 /// allocas are never considered static even if they are in the entry block.
137 bool isUsedWithInAlloca() const {
138 return getSubclassData<UsedWithInAllocaField>();
139 }
140
141 /// Specify whether this alloca is used to represent the arguments to a call.
142 void setUsedWithInAlloca(bool V) {
143 setSubclassData<UsedWithInAllocaField>(V);
144 }
145
146 /// Return true if this alloca is used as a swifterror argument to a call.
147 bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); }
148 /// Specify whether this alloca is used to represent a swifterror.
149 void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); }
150
151 // Methods for support type inquiry through isa, cast, and dyn_cast:
152 static bool classof(const Instruction *I) {
153 return (I->getOpcode() == Instruction::Alloca);
154 }
155 static bool classof(const Value *V) {
156 return isa<Instruction>(V) && classof(cast<Instruction>(V));
157 }
158
159private:
160 // Shadow Instruction::setInstructionSubclassData with a private forwarding
161 // method so that subclasses cannot accidentally use it.
162 template <typename Bitfield>
163 void setSubclassData(typename Bitfield::Type Value) {
164 Instruction::setSubclassData<Bitfield>(Value);
165 }
166};
167
168//===----------------------------------------------------------------------===//
169// LoadInst Class
170//===----------------------------------------------------------------------===//
171
172/// An instruction for reading from memory. This uses the SubclassData field in
173/// Value to store whether or not the load is volatile.
174class LoadInst : public UnaryInstruction {
175 using VolatileField = BoolBitfieldElementT<0>;
176 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
177 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
178 static_assert(
179 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
180 "Bitfields must be contiguous");
181
182 void AssertOK();
183
184protected:
185 // Note: Instruction needs to be a friend here to call cloneImpl.
186 friend class Instruction;
187
188 LoadInst *cloneImpl() const;
189
190public:
191 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr,
192 Instruction *InsertBefore);
193 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd);
194 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
195 Instruction *InsertBefore);
196 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
197 BasicBlock *InsertAtEnd);
198 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
199 Align Align, Instruction *InsertBefore = nullptr);
200 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
201 Align Align, BasicBlock *InsertAtEnd);
202 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
203 Align Align, AtomicOrdering Order,
204 SyncScope::ID SSID = SyncScope::System,
205 Instruction *InsertBefore = nullptr);
206 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
207 Align Align, AtomicOrdering Order, SyncScope::ID SSID,
208 BasicBlock *InsertAtEnd);
209
210 /// Return true if this is a load from a volatile memory location.
211 bool isVolatile() const { return getSubclassData<VolatileField>(); }
212
213 /// Specify whether this is a volatile load or not.
214 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
215
216 /// Return the alignment of the access that is being performed.
217 Align getAlign() const {
218 return Align(1ULL << (getSubclassData<AlignmentField>()));
219 }
220
221 void setAlignment(Align Align) {
222 setSubclassData<AlignmentField>(Log2(Align));
223 }
224
225 /// Returns the ordering constraint of this load instruction.
226 AtomicOrdering getOrdering() const {
227 return getSubclassData<OrderingField>();
228 }
229 /// Sets the ordering constraint of this load instruction. May not be Release
230 /// or AcquireRelease.
231 void setOrdering(AtomicOrdering Ordering) {
232 setSubclassData<OrderingField>(Ordering);
233 }
234
235 /// Returns the synchronization scope ID of this load instruction.
236 SyncScope::ID getSyncScopeID() const {
237 return SSID;
238 }
239
240 /// Sets the synchronization scope ID of this load instruction.
241 void setSyncScopeID(SyncScope::ID SSID) {
242 this->SSID = SSID;
243 }
244
245 /// Sets the ordering constraint and the synchronization scope ID of this load
246 /// instruction.
247 void setAtomic(AtomicOrdering Ordering,
248 SyncScope::ID SSID = SyncScope::System) {
249 setOrdering(Ordering);
250 setSyncScopeID(SSID);
251 }
252
253 bool isSimple() const { return !isAtomic() && !isVolatile(); }
254
255 bool isUnordered() const {
256 return (getOrdering() == AtomicOrdering::NotAtomic ||
257 getOrdering() == AtomicOrdering::Unordered) &&
258 !isVolatile();
259 }
260
261 Value *getPointerOperand() { return getOperand(0); }
262 const Value *getPointerOperand() const { return getOperand(0); }
263 static unsigned getPointerOperandIndex() { return 0U; }
264 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
265
266 /// Returns the address space of the pointer operand.
267 unsigned getPointerAddressSpace() const {
268 return getPointerOperandType()->getPointerAddressSpace();
269 }
270
271 // Methods for support type inquiry through isa, cast, and dyn_cast:
272 static bool classof(const Instruction *I) {
273 return I->getOpcode() == Instruction::Load;
274 }
275 static bool classof(const Value *V) {
276 return isa<Instruction>(V) && classof(cast<Instruction>(V));
277 }
278
279private:
280 // Shadow Instruction::setInstructionSubclassData with a private forwarding
281 // method so that subclasses cannot accidentally use it.
282 template <typename Bitfield>
283 void setSubclassData(typename Bitfield::Type Value) {
284 Instruction::setSubclassData<Bitfield>(Value);
285 }
286
287 /// The synchronization scope ID of this load instruction. Not quite enough
288 /// room in SubClassData for everything, so synchronization scope ID gets its
289 /// own field.
290 SyncScope::ID SSID;
291};
292
293//===----------------------------------------------------------------------===//
294// StoreInst Class
295//===----------------------------------------------------------------------===//
296
297/// An instruction for storing to memory.
298class StoreInst : public Instruction {
299 using VolatileField = BoolBitfieldElementT<0>;
300 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
301 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
302 static_assert(
303 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
304 "Bitfields must be contiguous");
305
306 void AssertOK();
307
308protected:
309 // Note: Instruction needs to be a friend here to call cloneImpl.
310 friend class Instruction;
311
312 StoreInst *cloneImpl() const;
313
314public:
315 StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore);
316 StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd);
317 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore);
318 StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd);
319 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
320 Instruction *InsertBefore = nullptr);
321 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
322 BasicBlock *InsertAtEnd);
323 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
324 AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System,
325 Instruction *InsertBefore = nullptr);
326 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
327 AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd);
328
329 // allocate space for exactly two operands
330 void *operator new(size_t S) { return User::operator new(S, 2); }
331 void operator delete(void *Ptr) { User::operator delete(Ptr); }
332
333 /// Return true if this is a store to a volatile memory location.
334 bool isVolatile() const { return getSubclassData<VolatileField>(); }
335
336 /// Specify whether this is a volatile store or not.
337 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
338
339 /// Transparently provide more efficient getOperand methods.
340 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
341
342 Align getAlign() const {
343 return Align(1ULL << (getSubclassData<AlignmentField>()));
344 }
345
346 void setAlignment(Align Align) {
347 setSubclassData<AlignmentField>(Log2(Align));
348 }
349
350 /// Returns the ordering constraint of this store instruction.
351 AtomicOrdering getOrdering() const {
352 return getSubclassData<OrderingField>();
353 }
354
355 /// Sets the ordering constraint of this store instruction. May not be
356 /// Acquire or AcquireRelease.
357 void setOrdering(AtomicOrdering Ordering) {
358 setSubclassData<OrderingField>(Ordering);
359 }
360
361 /// Returns the synchronization scope ID of this store instruction.
362 SyncScope::ID getSyncScopeID() const {
363 return SSID;
364 }
365
366 /// Sets the synchronization scope ID of this store instruction.
367 void setSyncScopeID(SyncScope::ID SSID) {
368 this->SSID = SSID;
369 }
370
371 /// Sets the ordering constraint and the synchronization scope ID of this
372 /// store instruction.
373 void setAtomic(AtomicOrdering Ordering,
374 SyncScope::ID SSID = SyncScope::System) {
375 setOrdering(Ordering);
376 setSyncScopeID(SSID);
377 }
378
379 bool isSimple() const { return !isAtomic() && !isVolatile(); }
380
381 bool isUnordered() const {
382 return (getOrdering() == AtomicOrdering::NotAtomic ||
383 getOrdering() == AtomicOrdering::Unordered) &&
384 !isVolatile();
385 }
386
387 Value *getValueOperand() { return getOperand(0); }
388 const Value *getValueOperand() const { return getOperand(0); }
389
390 Value *getPointerOperand() { return getOperand(1); }
391 const Value *getPointerOperand() const { return getOperand(1); }
392 static unsigned getPointerOperandIndex() { return 1U; }
393 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
394
395 /// Returns the address space of the pointer operand.
396 unsigned getPointerAddressSpace() const {
397 return getPointerOperandType()->getPointerAddressSpace();
398 }
399
400 // Methods for support type inquiry through isa, cast, and dyn_cast:
401 static bool classof(const Instruction *I) {
402 return I->getOpcode() == Instruction::Store;
403 }
404 static bool classof(const Value *V) {
405 return isa<Instruction>(V) && classof(cast<Instruction>(V));
406 }
407
408private:
409 // Shadow Instruction::setInstructionSubclassData with a private forwarding
410 // method so that subclasses cannot accidentally use it.
411 template <typename Bitfield>
412 void setSubclassData(typename Bitfield::Type Value) {
413 Instruction::setSubclassData<Bitfield>(Value);
414 }
415
416 /// The synchronization scope ID of this store instruction. Not quite enough
417 /// room in SubClassData for everything, so synchronization scope ID gets its
418 /// own field.
419 SyncScope::ID SSID;
420};
421
422template <>
423struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> {
424};
425
426DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits
<StoreInst>::op_begin(this); } StoreInst::const_op_iterator
StoreInst::op_begin() const { return OperandTraits<StoreInst
>::op_begin(const_cast<StoreInst*>(this)); } StoreInst
::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst
>::op_end(this); } StoreInst::const_op_iterator StoreInst::
op_end() const { return OperandTraits<StoreInst>::op_end
(const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand
(unsigned i_nocapture) const { (static_cast <bool> (i_nocapture
< OperandTraits<StoreInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 426, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<StoreInst
>::op_begin(const_cast<StoreInst*>(this))[i_nocapture
].get()); } void StoreInst::setOperand(unsigned i_nocapture, Value
*Val_nocapture) { (static_cast <bool> (i_nocapture <
OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!"
) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 426, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<StoreInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned StoreInst::getNumOperands() const
{ return OperandTraits<StoreInst>::operands(this); } template
<int Idx_nocapture> Use &StoreInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &StoreInst::Op() const { return this->OpFrom
<Idx_nocapture>(this); }
427
428//===----------------------------------------------------------------------===//
429// FenceInst Class
430//===----------------------------------------------------------------------===//
431
432/// An instruction for ordering other memory operations.
433class FenceInst : public Instruction {
434 using OrderingField = AtomicOrderingBitfieldElementT<0>;
435
436 void Init(AtomicOrdering Ordering, SyncScope::ID SSID);
437
438protected:
439 // Note: Instruction needs to be a friend here to call cloneImpl.
440 friend class Instruction;
441
442 FenceInst *cloneImpl() const;
443
444public:
445 // Ordering may only be Acquire, Release, AcquireRelease, or
446 // SequentiallyConsistent.
447 FenceInst(LLVMContext &C, AtomicOrdering Ordering,
448 SyncScope::ID SSID = SyncScope::System,
449 Instruction *InsertBefore = nullptr);
450 FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID,
451 BasicBlock *InsertAtEnd);
452
453 // allocate space for exactly zero operands
454 void *operator new(size_t S) { return User::operator new(S, 0); }
455 void operator delete(void *Ptr) { User::operator delete(Ptr); }
456
457 /// Returns the ordering constraint of this fence instruction.
458 AtomicOrdering getOrdering() const {
459 return getSubclassData<OrderingField>();
460 }
461
462 /// Sets the ordering constraint of this fence instruction. May only be
463 /// Acquire, Release, AcquireRelease, or SequentiallyConsistent.
464 void setOrdering(AtomicOrdering Ordering) {
465 setSubclassData<OrderingField>(Ordering);
466 }
467
468 /// Returns the synchronization scope ID of this fence instruction.
469 SyncScope::ID getSyncScopeID() const {
470 return SSID;
471 }
472
473 /// Sets the synchronization scope ID of this fence instruction.
474 void setSyncScopeID(SyncScope::ID SSID) {
475 this->SSID = SSID;
476 }
477
478 // Methods for support type inquiry through isa, cast, and dyn_cast:
479 static bool classof(const Instruction *I) {
480 return I->getOpcode() == Instruction::Fence;
481 }
482 static bool classof(const Value *V) {
483 return isa<Instruction>(V) && classof(cast<Instruction>(V));
484 }
485
486private:
487 // Shadow Instruction::setInstructionSubclassData with a private forwarding
488 // method so that subclasses cannot accidentally use it.
489 template <typename Bitfield>
490 void setSubclassData(typename Bitfield::Type Value) {
491 Instruction::setSubclassData<Bitfield>(Value);
492 }
493
494 /// The synchronization scope ID of this fence instruction. Not quite enough
495 /// room in SubClassData for everything, so synchronization scope ID gets its
496 /// own field.
497 SyncScope::ID SSID;
498};
499
500//===----------------------------------------------------------------------===//
501// AtomicCmpXchgInst Class
502//===----------------------------------------------------------------------===//
503
504/// An instruction that atomically checks whether a
505/// specified value is in a memory location, and, if it is, stores a new value
506/// there. The value returned by this instruction is a pair containing the
507/// original value as first element, and an i1 indicating success (true) or
508/// failure (false) as second element.
509///
510class AtomicCmpXchgInst : public Instruction {
511 void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align,
512 AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering,
513 SyncScope::ID SSID);
514
515 template <unsigned Offset>
516 using AtomicOrderingBitfieldElement =
517 typename Bitfield::Element<AtomicOrdering, Offset, 3,
518 AtomicOrdering::LAST>;
519
520protected:
521 // Note: Instruction needs to be a friend here to call cloneImpl.
522 friend class Instruction;
523
524 AtomicCmpXchgInst *cloneImpl() const;
525
526public:
527 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
528 AtomicOrdering SuccessOrdering,
529 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
530 Instruction *InsertBefore = nullptr);
531 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
532 AtomicOrdering SuccessOrdering,
533 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
534 BasicBlock *InsertAtEnd);
535
536 // allocate space for exactly three operands
537 void *operator new(size_t S) { return User::operator new(S, 3); }
538 void operator delete(void *Ptr) { User::operator delete(Ptr); }
539
540 using VolatileField = BoolBitfieldElementT<0>;
541 using WeakField = BoolBitfieldElementT<VolatileField::NextBit>;
542 using SuccessOrderingField =
543 AtomicOrderingBitfieldElementT<WeakField::NextBit>;
544 using FailureOrderingField =
545 AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>;
546 using AlignmentField =
547 AlignmentBitfieldElementT<FailureOrderingField::NextBit>;
548 static_assert(
549 Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField,
550 FailureOrderingField, AlignmentField>(),
551 "Bitfields must be contiguous");
552
553 /// Return the alignment of the memory that is being allocated by the
554 /// instruction.
555 Align getAlign() const {
556 return Align(1ULL << getSubclassData<AlignmentField>());
557 }
558
559 void setAlignment(Align Align) {
560 setSubclassData<AlignmentField>(Log2(Align));
561 }
562
563 /// Return true if this is a cmpxchg from a volatile memory
564 /// location.
565 ///
566 bool isVolatile() const { return getSubclassData<VolatileField>(); }
567
568 /// Specify whether this is a volatile cmpxchg.
569 ///
570 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
571
572 /// Return true if this cmpxchg may spuriously fail.
573 bool isWeak() const { return getSubclassData<WeakField>(); }
574
575 void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); }
576
577 /// Transparently provide more efficient getOperand methods.
578 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
579
580 static bool isValidSuccessOrdering(AtomicOrdering Ordering) {
581 return Ordering != AtomicOrdering::NotAtomic &&
582 Ordering != AtomicOrdering::Unordered;
583 }
584
585 static bool isValidFailureOrdering(AtomicOrdering Ordering) {
586 return Ordering != AtomicOrdering::NotAtomic &&
587 Ordering != AtomicOrdering::Unordered &&
588 Ordering != AtomicOrdering::AcquireRelease &&
589 Ordering != AtomicOrdering::Release;
590 }
591
592 /// Returns the success ordering constraint of this cmpxchg instruction.
593 AtomicOrdering getSuccessOrdering() const {
594 return getSubclassData<SuccessOrderingField>();
595 }
596
597 /// Sets the success ordering constraint of this cmpxchg instruction.
598 void setSuccessOrdering(AtomicOrdering Ordering) {
599 assert(isValidSuccessOrdering(Ordering) &&(static_cast <bool> (isValidSuccessOrdering(Ordering) &&
"invalid CmpXchg success ordering") ? void (0) : __assert_fail
("isValidSuccessOrdering(Ordering) && \"invalid CmpXchg success ordering\""
, "llvm/include/llvm/IR/Instructions.h", 600, __extension__ __PRETTY_FUNCTION__
))
600 "invalid CmpXchg success ordering")(static_cast <bool> (isValidSuccessOrdering(Ordering) &&
"invalid CmpXchg success ordering") ? void (0) : __assert_fail
("isValidSuccessOrdering(Ordering) && \"invalid CmpXchg success ordering\""
, "llvm/include/llvm/IR/Instructions.h", 600, __extension__ __PRETTY_FUNCTION__
))
;
601 setSubclassData<SuccessOrderingField>(Ordering);
602 }
603
604 /// Returns the failure ordering constraint of this cmpxchg instruction.
605 AtomicOrdering getFailureOrdering() const {
606 return getSubclassData<FailureOrderingField>();
607 }
608
609 /// Sets the failure ordering constraint of this cmpxchg instruction.
610 void setFailureOrdering(AtomicOrdering Ordering) {
611 assert(isValidFailureOrdering(Ordering) &&(static_cast <bool> (isValidFailureOrdering(Ordering) &&
"invalid CmpXchg failure ordering") ? void (0) : __assert_fail
("isValidFailureOrdering(Ordering) && \"invalid CmpXchg failure ordering\""
, "llvm/include/llvm/IR/Instructions.h", 612, __extension__ __PRETTY_FUNCTION__
))
612 "invalid CmpXchg failure ordering")(static_cast <bool> (isValidFailureOrdering(Ordering) &&
"invalid CmpXchg failure ordering") ? void (0) : __assert_fail
("isValidFailureOrdering(Ordering) && \"invalid CmpXchg failure ordering\""
, "llvm/include/llvm/IR/Instructions.h", 612, __extension__ __PRETTY_FUNCTION__
))
;
613 setSubclassData<FailureOrderingField>(Ordering);
614 }
615
616 /// Returns a single ordering which is at least as strong as both the
617 /// success and failure orderings for this cmpxchg.
618 AtomicOrdering getMergedOrdering() const {
619 if (getFailureOrdering() == AtomicOrdering::SequentiallyConsistent)
620 return AtomicOrdering::SequentiallyConsistent;
621 if (getFailureOrdering() == AtomicOrdering::Acquire) {
622 if (getSuccessOrdering() == AtomicOrdering::Monotonic)
623 return AtomicOrdering::Acquire;
624 if (getSuccessOrdering() == AtomicOrdering::Release)
625 return AtomicOrdering::AcquireRelease;
626 }
627 return getSuccessOrdering();
628 }
629
630 /// Returns the synchronization scope ID of this cmpxchg instruction.
631 SyncScope::ID getSyncScopeID() const {
632 return SSID;
633 }
634
635 /// Sets the synchronization scope ID of this cmpxchg instruction.
636 void setSyncScopeID(SyncScope::ID SSID) {
637 this->SSID = SSID;
638 }
639
640 Value *getPointerOperand() { return getOperand(0); }
641 const Value *getPointerOperand() const { return getOperand(0); }
642 static unsigned getPointerOperandIndex() { return 0U; }
643
644 Value *getCompareOperand() { return getOperand(1); }
645 const Value *getCompareOperand() const { return getOperand(1); }
646
647 Value *getNewValOperand() { return getOperand(2); }
648 const Value *getNewValOperand() const { return getOperand(2); }
649
650 /// Returns the address space of the pointer operand.
651 unsigned getPointerAddressSpace() const {
652 return getPointerOperand()->getType()->getPointerAddressSpace();
653 }
654
655 /// Returns the strongest permitted ordering on failure, given the
656 /// desired ordering on success.
657 ///
658 /// If the comparison in a cmpxchg operation fails, there is no atomic store
659 /// so release semantics cannot be provided. So this function drops explicit
660 /// Release requests from the AtomicOrdering. A SequentiallyConsistent
661 /// operation would remain SequentiallyConsistent.
662 static AtomicOrdering
663 getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) {
664 switch (SuccessOrdering) {
665 default:
666 llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering"
, "llvm/include/llvm/IR/Instructions.h", 666)
;
667 case AtomicOrdering::Release:
668 case AtomicOrdering::Monotonic:
669 return AtomicOrdering::Monotonic;
670 case AtomicOrdering::AcquireRelease:
671 case AtomicOrdering::Acquire:
672 return AtomicOrdering::Acquire;
673 case AtomicOrdering::SequentiallyConsistent:
674 return AtomicOrdering::SequentiallyConsistent;
675 }
676 }
677
678 // Methods for support type inquiry through isa, cast, and dyn_cast:
679 static bool classof(const Instruction *I) {
680 return I->getOpcode() == Instruction::AtomicCmpXchg;
681 }
682 static bool classof(const Value *V) {
683 return isa<Instruction>(V) && classof(cast<Instruction>(V));
684 }
685
686private:
687 // Shadow Instruction::setInstructionSubclassData with a private forwarding
688 // method so that subclasses cannot accidentally use it.
689 template <typename Bitfield>
690 void setSubclassData(typename Bitfield::Type Value) {
691 Instruction::setSubclassData<Bitfield>(Value);
692 }
693
694 /// The synchronization scope ID of this cmpxchg instruction. Not quite
695 /// enough room in SubClassData for everything, so synchronization scope ID
696 /// gets its own field.
697 SyncScope::ID SSID;
698};
699
700template <>
701struct OperandTraits<AtomicCmpXchgInst> :
702 public FixedNumOperandTraits<AtomicCmpXchgInst, 3> {
703};
704
705DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() {
return OperandTraits<AtomicCmpXchgInst>::op_begin(this
); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst::
op_begin() const { return OperandTraits<AtomicCmpXchgInst>
::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst
::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits
<AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst::
const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits
<AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst
*>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned
i_nocapture) const { (static_cast <bool> (i_nocapture <
OperandTraits<AtomicCmpXchgInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 705, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<AtomicCmpXchgInst
>::op_begin(const_cast<AtomicCmpXchgInst*>(this))[i_nocapture
].get()); } void AtomicCmpXchgInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<AtomicCmpXchgInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 705, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<AtomicCmpXchgInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned AtomicCmpXchgInst::getNumOperands
() const { return OperandTraits<AtomicCmpXchgInst>::operands
(this); } template <int Idx_nocapture> Use &AtomicCmpXchgInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &AtomicCmpXchgInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
706
707//===----------------------------------------------------------------------===//
708// AtomicRMWInst Class
709//===----------------------------------------------------------------------===//
710
711/// an instruction that atomically reads a memory location,
712/// combines it with another value, and then stores the result back. Returns
713/// the old value.
714///
715class AtomicRMWInst : public Instruction {
716protected:
717 // Note: Instruction needs to be a friend here to call cloneImpl.
718 friend class Instruction;
719
720 AtomicRMWInst *cloneImpl() const;
721
722public:
723 /// This enumeration lists the possible modifications atomicrmw can make. In
724 /// the descriptions, 'p' is the pointer to the instruction's memory location,
725 /// 'old' is the initial value of *p, and 'v' is the other value passed to the
726 /// instruction. These instructions always return 'old'.
727 enum BinOp : unsigned {
728 /// *p = v
729 Xchg,
730 /// *p = old + v
731 Add,
732 /// *p = old - v
733 Sub,
734 /// *p = old & v
735 And,
736 /// *p = ~(old & v)
737 Nand,
738 /// *p = old | v
739 Or,
740 /// *p = old ^ v
741 Xor,
742 /// *p = old >signed v ? old : v
743 Max,
744 /// *p = old <signed v ? old : v
745 Min,
746 /// *p = old >unsigned v ? old : v
747 UMax,
748 /// *p = old <unsigned v ? old : v
749 UMin,
750
751 /// *p = old + v
752 FAdd,
753
754 /// *p = old - v
755 FSub,
756
757 /// *p = maxnum(old, v)
758 /// \p maxnum matches the behavior of \p llvm.maxnum.*.
759 FMax,
760
761 /// *p = minnum(old, v)
762 /// \p minnum matches the behavior of \p llvm.minnum.*.
763 FMin,
764
765 FIRST_BINOP = Xchg,
766 LAST_BINOP = FMin,
767 BAD_BINOP
768 };
769
770private:
771 template <unsigned Offset>
772 using AtomicOrderingBitfieldElement =
773 typename Bitfield::Element<AtomicOrdering, Offset, 3,
774 AtomicOrdering::LAST>;
775
776 template <unsigned Offset>
777 using BinOpBitfieldElement =
778 typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>;
779
780public:
781 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
782 AtomicOrdering Ordering, SyncScope::ID SSID,
783 Instruction *InsertBefore = nullptr);
784 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
785 AtomicOrdering Ordering, SyncScope::ID SSID,
786 BasicBlock *InsertAtEnd);
787
788 // allocate space for exactly two operands
789 void *operator new(size_t S) { return User::operator new(S, 2); }
790 void operator delete(void *Ptr) { User::operator delete(Ptr); }
791
792 using VolatileField = BoolBitfieldElementT<0>;
793 using AtomicOrderingField =
794 AtomicOrderingBitfieldElementT<VolatileField::NextBit>;
795 using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>;
796 using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>;
797 static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField,
798 OperationField, AlignmentField>(),
799 "Bitfields must be contiguous");
800
801 BinOp getOperation() const { return getSubclassData<OperationField>(); }
802
803 static StringRef getOperationName(BinOp Op);
804
805 static bool isFPOperation(BinOp Op) {
806 switch (Op) {
807 case AtomicRMWInst::FAdd:
808 case AtomicRMWInst::FSub:
809 case AtomicRMWInst::FMax:
810 case AtomicRMWInst::FMin:
811 return true;
812 default:
813 return false;
814 }
815 }
816
817 void setOperation(BinOp Operation) {
818 setSubclassData<OperationField>(Operation);
819 }
820
821 /// Return the alignment of the memory that is being allocated by the
822 /// instruction.
823 Align getAlign() const {
824 return Align(1ULL << getSubclassData<AlignmentField>());
825 }
826
827 void setAlignment(Align Align) {
828 setSubclassData<AlignmentField>(Log2(Align));
829 }
830
831 /// Return true if this is a RMW on a volatile memory location.
832 ///
833 bool isVolatile() const { return getSubclassData<VolatileField>(); }
834
835 /// Specify whether this is a volatile RMW or not.
836 ///
837 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
838
839 /// Transparently provide more efficient getOperand methods.
840 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
841
842 /// Returns the ordering constraint of this rmw instruction.
843 AtomicOrdering getOrdering() const {
844 return getSubclassData<AtomicOrderingField>();
845 }
846
847 /// Sets the ordering constraint of this rmw instruction.
848 void setOrdering(AtomicOrdering Ordering) {
849 assert(Ordering != AtomicOrdering::NotAtomic &&(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic
&& "atomicrmw instructions can only be atomic.") ? void
(0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "llvm/include/llvm/IR/Instructions.h", 850, __extension__ __PRETTY_FUNCTION__
))
850 "atomicrmw instructions can only be atomic.")(static_cast <bool> (Ordering != AtomicOrdering::NotAtomic
&& "atomicrmw instructions can only be atomic.") ? void
(0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "llvm/include/llvm/IR/Instructions.h", 850, __extension__ __PRETTY_FUNCTION__
))
;
851 assert(Ordering != AtomicOrdering::Unordered &&(static_cast <bool> (Ordering != AtomicOrdering::Unordered
&& "atomicrmw instructions cannot be unordered.") ? void
(0) : __assert_fail ("Ordering != AtomicOrdering::Unordered && \"atomicrmw instructions cannot be unordered.\""
, "llvm/include/llvm/IR/Instructions.h", 852, __extension__ __PRETTY_FUNCTION__
))
852 "atomicrmw instructions cannot be unordered.")(static_cast <bool> (Ordering != AtomicOrdering::Unordered
&& "atomicrmw instructions cannot be unordered.") ? void
(0) : __assert_fail ("Ordering != AtomicOrdering::Unordered && \"atomicrmw instructions cannot be unordered.\""
, "llvm/include/llvm/IR/Instructions.h", 852, __extension__ __PRETTY_FUNCTION__
))
;
853 setSubclassData<AtomicOrderingField>(Ordering);
854 }
855
856 /// Returns the synchronization scope ID of this rmw instruction.
857 SyncScope::ID getSyncScopeID() const {
858 return SSID;
859 }
860
861 /// Sets the synchronization scope ID of this rmw instruction.
862 void setSyncScopeID(SyncScope::ID SSID) {
863 this->SSID = SSID;
864 }
865
866 Value *getPointerOperand() { return getOperand(0); }
867 const Value *getPointerOperand() const { return getOperand(0); }
868 static unsigned getPointerOperandIndex() { return 0U; }
869
870 Value *getValOperand() { return getOperand(1); }
871 const Value *getValOperand() const { return getOperand(1); }
872
873 /// Returns the address space of the pointer operand.
874 unsigned getPointerAddressSpace() const {
875 return getPointerOperand()->getType()->getPointerAddressSpace();
876 }
877
878 bool isFloatingPointOperation() const {
879 return isFPOperation(getOperation());
880 }
881
882 // Methods for support type inquiry through isa, cast, and dyn_cast:
883 static bool classof(const Instruction *I) {
884 return I->getOpcode() == Instruction::AtomicRMW;
885 }
886 static bool classof(const Value *V) {
887 return isa<Instruction>(V) && classof(cast<Instruction>(V));
888 }
889
890private:
891 void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align,
892 AtomicOrdering Ordering, SyncScope::ID SSID);
893
894 // Shadow Instruction::setInstructionSubclassData with a private forwarding
895 // method so that subclasses cannot accidentally use it.
896 template <typename Bitfield>
897 void setSubclassData(typename Bitfield::Type Value) {
898 Instruction::setSubclassData<Bitfield>(Value);
899 }
900
901 /// The synchronization scope ID of this rmw instruction. Not quite enough
902 /// room in SubClassData for everything, so synchronization scope ID gets its
903 /// own field.
904 SyncScope::ID SSID;
905};
906
907template <>
908struct OperandTraits<AtomicRMWInst>
909 : public FixedNumOperandTraits<AtomicRMWInst,2> {
910};
911
912DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return
OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst
::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits
<AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*>
(this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end()
{ return OperandTraits<AtomicRMWInst>::op_end(this); }
AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const
{ return OperandTraits<AtomicRMWInst>::op_end(const_cast
<AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand
(unsigned i_nocapture) const { (static_cast <bool> (i_nocapture
< OperandTraits<AtomicRMWInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 912, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<AtomicRMWInst
>::op_begin(const_cast<AtomicRMWInst*>(this))[i_nocapture
].get()); } void AtomicRMWInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<AtomicRMWInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 912, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<AtomicRMWInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned AtomicRMWInst::getNumOperands()
const { return OperandTraits<AtomicRMWInst>::operands(
this); } template <int Idx_nocapture> Use &AtomicRMWInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &AtomicRMWInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
913
914//===----------------------------------------------------------------------===//
915// GetElementPtrInst Class
916//===----------------------------------------------------------------------===//
917
918// checkGEPType - Simple wrapper function to give a better assertion failure
919// message on bad indexes for a gep instruction.
920//
921inline Type *checkGEPType(Type *Ty) {
922 assert(Ty && "Invalid GetElementPtrInst indices for type!")(static_cast <bool> (Ty && "Invalid GetElementPtrInst indices for type!"
) ? void (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\""
, "llvm/include/llvm/IR/Instructions.h", 922, __extension__ __PRETTY_FUNCTION__
))
;
923 return Ty;
924}
925
926/// an instruction for type-safe pointer arithmetic to
927/// access elements of arrays and structs
928///
929class GetElementPtrInst : public Instruction {
930 Type *SourceElementType;
931 Type *ResultElementType;
932
933 GetElementPtrInst(const GetElementPtrInst &GEPI);
934
935 /// Constructors - Create a getelementptr instruction with a base pointer an
936 /// list of indices. The first ctor can optionally insert before an existing
937 /// instruction, the second appends the new instruction to the specified
938 /// BasicBlock.
939 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
940 ArrayRef<Value *> IdxList, unsigned Values,
941 const Twine &NameStr, Instruction *InsertBefore);
942 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
943 ArrayRef<Value *> IdxList, unsigned Values,
944 const Twine &NameStr, BasicBlock *InsertAtEnd);
945
946 void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr);
947
948protected:
949 // Note: Instruction needs to be a friend here to call cloneImpl.
950 friend class Instruction;
951
952 GetElementPtrInst *cloneImpl() const;
953
954public:
955 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
956 ArrayRef<Value *> IdxList,
957 const Twine &NameStr = "",
958 Instruction *InsertBefore = nullptr) {
959 unsigned Values = 1 + unsigned(IdxList.size());
960 assert(PointeeType && "Must specify element type")(static_cast <bool> (PointeeType && "Must specify element type"
) ? void (0) : __assert_fail ("PointeeType && \"Must specify element type\""
, "llvm/include/llvm/IR/Instructions.h", 960, __extension__ __PRETTY_FUNCTION__
))
;
961 assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast <bool> (cast<PointerType>(Ptr->getType
()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType
)) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)"
, "llvm/include/llvm/IR/Instructions.h", 962, __extension__ __PRETTY_FUNCTION__
))
962 ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast <bool> (cast<PointerType>(Ptr->getType
()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType
)) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)"
, "llvm/include/llvm/IR/Instructions.h", 962, __extension__ __PRETTY_FUNCTION__
))
;
963 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
964 NameStr, InsertBefore);
965 }
966
967 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
968 ArrayRef<Value *> IdxList,
969 const Twine &NameStr,
970 BasicBlock *InsertAtEnd) {
971 unsigned Values = 1 + unsigned(IdxList.size());
972 assert(PointeeType && "Must specify element type")(static_cast <bool> (PointeeType && "Must specify element type"
) ? void (0) : __assert_fail ("PointeeType && \"Must specify element type\""
, "llvm/include/llvm/IR/Instructions.h", 972, __extension__ __PRETTY_FUNCTION__
))
;
973 assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast <bool> (cast<PointerType>(Ptr->getType
()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType
)) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)"
, "llvm/include/llvm/IR/Instructions.h", 974, __extension__ __PRETTY_FUNCTION__
))
974 ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast <bool> (cast<PointerType>(Ptr->getType
()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType
)) ? void (0) : __assert_fail ("cast<PointerType>(Ptr->getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(PointeeType)"
, "llvm/include/llvm/IR/Instructions.h", 974, __extension__ __PRETTY_FUNCTION__
))
;
975 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
976 NameStr, InsertAtEnd);
977 }
978
979 /// Create an "inbounds" getelementptr. See the documentation for the
980 /// "inbounds" flag in LangRef.html for details.
981 static GetElementPtrInst *
982 CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList,
983 const Twine &NameStr = "",
984 Instruction *InsertBefore = nullptr) {
985 GetElementPtrInst *GEP =
986 Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore);
987 GEP->setIsInBounds(true);
988 return GEP;
989 }
990
991 static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr,
992 ArrayRef<Value *> IdxList,
993 const Twine &NameStr,
994 BasicBlock *InsertAtEnd) {
995 GetElementPtrInst *GEP =
996 Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd);
997 GEP->setIsInBounds(true);
998 return GEP;
999 }
1000
1001 /// Transparently provide more efficient getOperand methods.
1002 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1003
1004 Type *getSourceElementType() const { return SourceElementType; }
1005
1006 void setSourceElementType(Type *Ty) { SourceElementType = Ty; }
1007 void setResultElementType(Type *Ty) { ResultElementType = Ty; }
1008
1009 Type *getResultElementType() const {
1010 assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1011, __extension__ __PRETTY_FUNCTION__
))
1011 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1011, __extension__ __PRETTY_FUNCTION__
))
;
1012 return ResultElementType;
1013 }
1014
1015 /// Returns the address space of this instruction's pointer type.
1016 unsigned getAddressSpace() const {
1017 // Note that this is always the same as the pointer operand's address space
1018 // and that is cheaper to compute, so cheat here.
1019 return getPointerAddressSpace();
1020 }
1021
1022 /// Returns the result type of a getelementptr with the given source
1023 /// element type and indexes.
1024 ///
1025 /// Null is returned if the indices are invalid for the specified
1026 /// source element type.
1027 static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList);
1028 static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList);
1029 static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList);
1030
1031 /// Return the type of the element at the given index of an indexable
1032 /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})".
1033 ///
1034 /// Returns null if the type can't be indexed, or the given index is not
1035 /// legal for the given type.
1036 static Type *getTypeAtIndex(Type *Ty, Value *Idx);
1037 static Type *getTypeAtIndex(Type *Ty, uint64_t Idx);
1038
1039 inline op_iterator idx_begin() { return op_begin()+1; }
1040 inline const_op_iterator idx_begin() const { return op_begin()+1; }
1041 inline op_iterator idx_end() { return op_end(); }
1042 inline const_op_iterator idx_end() const { return op_end(); }
1043
1044 inline iterator_range<op_iterator> indices() {
1045 return make_range(idx_begin(), idx_end());
1046 }
1047
1048 inline iterator_range<const_op_iterator> indices() const {
1049 return make_range(idx_begin(), idx_end());
1050 }
1051
1052 Value *getPointerOperand() {
1053 return getOperand(0);
1054 }
1055 const Value *getPointerOperand() const {
1056 return getOperand(0);
1057 }
1058 static unsigned getPointerOperandIndex() {
1059 return 0U; // get index for modifying correct operand.
1060 }
1061
1062 /// Method to return the pointer operand as a
1063 /// PointerType.
1064 Type *getPointerOperandType() const {
1065 return getPointerOperand()->getType();
1066 }
1067
1068 /// Returns the address space of the pointer operand.
1069 unsigned getPointerAddressSpace() const {
1070 return getPointerOperandType()->getPointerAddressSpace();
1071 }
1072
1073 /// Returns the pointer type returned by the GEP
1074 /// instruction, which may be a vector of pointers.
1075 static Type *getGEPReturnType(Type *ElTy, Value *Ptr,
1076 ArrayRef<Value *> IdxList) {
1077 PointerType *OrigPtrTy = cast<PointerType>(Ptr->getType()->getScalarType());
1078 unsigned AddrSpace = OrigPtrTy->getAddressSpace();
1079 Type *ResultElemTy = checkGEPType(getIndexedType(ElTy, IdxList));
1080 Type *PtrTy = OrigPtrTy->isOpaque()
1081 ? PointerType::get(OrigPtrTy->getContext(), AddrSpace)
1082 : PointerType::get(ResultElemTy, AddrSpace);
1083 // Vector GEP
1084 if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) {
1085 ElementCount EltCount = PtrVTy->getElementCount();
1086 return VectorType::get(PtrTy, EltCount);
1087 }
1088 for (Value *Index : IdxList)
1089 if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) {
1090 ElementCount EltCount = IndexVTy->getElementCount();
1091 return VectorType::get(PtrTy, EltCount);
1092 }
1093 // Scalar GEP
1094 return PtrTy;
1095 }
1096
1097 unsigned getNumIndices() const { // Note: always non-negative
1098 return getNumOperands() - 1;
1099 }
1100
1101 bool hasIndices() const {
1102 return getNumOperands() > 1;
1103 }
1104
1105 /// Return true if all of the indices of this GEP are
1106 /// zeros. If so, the result pointer and the first operand have the same
1107 /// value, just potentially different types.
1108 bool hasAllZeroIndices() const;
1109
1110 /// Return true if all of the indices of this GEP are
1111 /// constant integers. If so, the result pointer and the first operand have
1112 /// a constant offset between them.
1113 bool hasAllConstantIndices() const;
1114
1115 /// Set or clear the inbounds flag on this GEP instruction.
1116 /// See LangRef.html for the meaning of inbounds on a getelementptr.
1117 void setIsInBounds(bool b = true);
1118
1119 /// Determine whether the GEP has the inbounds flag.
1120 bool isInBounds() const;
1121
1122 /// Accumulate the constant address offset of this GEP if possible.
1123 ///
1124 /// This routine accepts an APInt into which it will accumulate the constant
1125 /// offset of this GEP if the GEP is in fact constant. If the GEP is not
1126 /// all-constant, it returns false and the value of the offset APInt is
1127 /// undefined (it is *not* preserved!). The APInt passed into this routine
1128 /// must be at least as wide as the IntPtr type for the address space of
1129 /// the base GEP pointer.
1130 bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const;
1131 bool collectOffset(const DataLayout &DL, unsigned BitWidth,
1132 MapVector<Value *, APInt> &VariableOffsets,
1133 APInt &ConstantOffset) const;
1134 // Methods for support type inquiry through isa, cast, and dyn_cast:
1135 static bool classof(const Instruction *I) {
1136 return (I->getOpcode() == Instruction::GetElementPtr);
1137 }
1138 static bool classof(const Value *V) {
1139 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1140 }
1141};
1142
1143template <>
1144struct OperandTraits<GetElementPtrInst> :
1145 public VariadicOperandTraits<GetElementPtrInst, 1> {
1146};
1147
1148GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1149 ArrayRef<Value *> IdxList, unsigned Values,
1150 const Twine &NameStr,
1151 Instruction *InsertBefore)
1152 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1153 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1154 Values, InsertBefore),
1155 SourceElementType(PointeeType),
1156 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1157 assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1158, __extension__ __PRETTY_FUNCTION__
))
1158 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1158, __extension__ __PRETTY_FUNCTION__
))
;
1159 init(Ptr, IdxList, NameStr);
1160}
1161
1162GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1163 ArrayRef<Value *> IdxList, unsigned Values,
1164 const Twine &NameStr,
1165 BasicBlock *InsertAtEnd)
1166 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1167 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1168 Values, InsertAtEnd),
1169 SourceElementType(PointeeType),
1170 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1171 assert(cast<PointerType>(getType()->getScalarType())(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1172, __extension__ __PRETTY_FUNCTION__
))
1172 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast <bool> (cast<PointerType>(getType()->
getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType
)) ? void (0) : __assert_fail ("cast<PointerType>(getType()->getScalarType()) ->isOpaqueOrPointeeTypeMatches(ResultElementType)"
, "llvm/include/llvm/IR/Instructions.h", 1172, __extension__ __PRETTY_FUNCTION__
))
;
1173 init(Ptr, IdxList, NameStr);
1174}
1175
1176DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() {
return OperandTraits<GetElementPtrInst>::op_begin(this
); } GetElementPtrInst::const_op_iterator GetElementPtrInst::
op_begin() const { return OperandTraits<GetElementPtrInst>
::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst
::op_iterator GetElementPtrInst::op_end() { return OperandTraits
<GetElementPtrInst>::op_end(this); } GetElementPtrInst::
const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits
<GetElementPtrInst>::op_end(const_cast<GetElementPtrInst
*>(this)); } Value *GetElementPtrInst::getOperand(unsigned
i_nocapture) const { (static_cast <bool> (i_nocapture <
OperandTraits<GetElementPtrInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1176, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<GetElementPtrInst
>::op_begin(const_cast<GetElementPtrInst*>(this))[i_nocapture
].get()); } void GetElementPtrInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<GetElementPtrInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1176, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<GetElementPtrInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned GetElementPtrInst::getNumOperands
() const { return OperandTraits<GetElementPtrInst>::operands
(this); } template <int Idx_nocapture> Use &GetElementPtrInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &GetElementPtrInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
1177
1178//===----------------------------------------------------------------------===//
1179// ICmpInst Class
1180//===----------------------------------------------------------------------===//
1181
1182/// This instruction compares its operands according to the predicate given
1183/// to the constructor. It only operates on integers or pointers. The operands
1184/// must be identical types.
1185/// Represent an integer comparison operator.
1186class ICmpInst: public CmpInst {
1187 void AssertOK() {
1188 assert(isIntPredicate() &&(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value"
) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "llvm/include/llvm/IR/Instructions.h", 1189, __extension__ __PRETTY_FUNCTION__
))
1189 "Invalid ICmp predicate value")(static_cast <bool> (isIntPredicate() && "Invalid ICmp predicate value"
) ? void (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "llvm/include/llvm/IR/Instructions.h", 1189, __extension__ __PRETTY_FUNCTION__
))
;
1190 assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast <bool> (getOperand(0)->getType() == getOperand
(1)->getType() && "Both operands to ICmp instruction are not of the same type!"
) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "llvm/include/llvm/IR/Instructions.h", 1191, __extension__ __PRETTY_FUNCTION__
))
1191 "Both operands to ICmp instruction are not of the same type!")(static_cast <bool> (getOperand(0)->getType() == getOperand
(1)->getType() && "Both operands to ICmp instruction are not of the same type!"
) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "llvm/include/llvm/IR/Instructions.h", 1191, __extension__ __PRETTY_FUNCTION__
))
;
1192 // Check that the operands are the right type
1193 assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy
() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&
"Invalid operand types for ICmp instruction") ? void (0) : __assert_fail
("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "llvm/include/llvm/IR/Instructions.h", 1195, __extension__ __PRETTY_FUNCTION__
))
1194 getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy
() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&
"Invalid operand types for ICmp instruction") ? void (0) : __assert_fail
("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "llvm/include/llvm/IR/Instructions.h", 1195, __extension__ __PRETTY_FUNCTION__
))
1195 "Invalid operand types for ICmp instruction")(static_cast <bool> ((getOperand(0)->getType()->isIntOrIntVectorTy
() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&
"Invalid operand types for ICmp instruction") ? void (0) : __assert_fail
("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "llvm/include/llvm/IR/Instructions.h", 1195, __extension__ __PRETTY_FUNCTION__
))
;
1196 }
1197
1198protected:
1199 // Note: Instruction needs to be a friend here to call cloneImpl.
1200 friend class Instruction;
1201
1202 /// Clone an identical ICmpInst
1203 ICmpInst *cloneImpl() const;
1204
1205public:
1206 /// Constructor with insert-before-instruction semantics.
1207 ICmpInst(
1208 Instruction *InsertBefore, ///< Where to insert
1209 Predicate pred, ///< The predicate to use for the comparison
1210 Value *LHS, ///< The left-hand-side of the expression
1211 Value *RHS, ///< The right-hand-side of the expression
1212 const Twine &NameStr = "" ///< Name of the instruction
1213 ) : CmpInst(makeCmpResultType(LHS->getType()),
1214 Instruction::ICmp, pred, LHS, RHS, NameStr,
1215 InsertBefore) {
1216#ifndef NDEBUG
1217 AssertOK();
1218#endif
1219 }
1220
1221 /// Constructor with insert-at-end semantics.
1222 ICmpInst(
1223 BasicBlock &InsertAtEnd, ///< Block to insert into.
1224 Predicate pred, ///< The predicate to use for the comparison
1225 Value *LHS, ///< The left-hand-side of the expression
1226 Value *RHS, ///< The right-hand-side of the expression
1227 const Twine &NameStr = "" ///< Name of the instruction
1228 ) : CmpInst(makeCmpResultType(LHS->getType()),
1229 Instruction::ICmp, pred, LHS, RHS, NameStr,
1230 &InsertAtEnd) {
1231#ifndef NDEBUG
1232 AssertOK();
1233#endif
1234 }
1235
1236 /// Constructor with no-insertion semantics
1237 ICmpInst(
1238 Predicate pred, ///< The predicate to use for the comparison
1239 Value *LHS, ///< The left-hand-side of the expression
1240 Value *RHS, ///< The right-hand-side of the expression
1241 const Twine &NameStr = "" ///< Name of the instruction
1242 ) : CmpInst(makeCmpResultType(LHS->getType()),
1243 Instruction::ICmp, pred, LHS, RHS, NameStr) {
1244#ifndef NDEBUG
1245 AssertOK();
1246#endif
1247 }
1248
1249 /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc.
1250 /// @returns the predicate that would be the result if the operand were
1251 /// regarded as signed.
1252 /// Return the signed version of the predicate
1253 Predicate getSignedPredicate() const {
1254 return getSignedPredicate(getPredicate());
1255 }
1256
1257 /// This is a static version that you can use without an instruction.
1258 /// Return the signed version of the predicate.
1259 static Predicate getSignedPredicate(Predicate pred);
1260
1261 /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc.
1262 /// @returns the predicate that would be the result if the operand were
1263 /// regarded as unsigned.
1264 /// Return the unsigned version of the predicate
1265 Predicate getUnsignedPredicate() const {
1266 return getUnsignedPredicate(getPredicate());
1267 }
1268
1269 /// This is a static version that you can use without an instruction.
1270 /// Return the unsigned version of the predicate.
1271 static Predicate getUnsignedPredicate(Predicate pred);
1272
1273 /// Return true if this predicate is either EQ or NE. This also
1274 /// tests for commutativity.
1275 static bool isEquality(Predicate P) {
1276 return P == ICMP_EQ || P == ICMP_NE;
1277 }
1278
1279 /// Return true if this predicate is either EQ or NE. This also
1280 /// tests for commutativity.
1281 bool isEquality() const {
1282 return isEquality(getPredicate());
1283 }
1284
1285 /// @returns true if the predicate of this ICmpInst is commutative
1286 /// Determine if this relation is commutative.
1287 bool isCommutative() const { return isEquality(); }
1288
1289 /// Return true if the predicate is relational (not EQ or NE).
1290 ///
1291 bool isRelational() const {
1292 return !isEquality();
1293 }
1294
1295 /// Return true if the predicate is relational (not EQ or NE).
1296 ///
1297 static bool isRelational(Predicate P) {
1298 return !isEquality(P);
1299 }
1300
1301 /// Return true if the predicate is SGT or UGT.
1302 ///
1303 static bool isGT(Predicate P) {
1304 return P == ICMP_SGT || P == ICMP_UGT;
1305 }
1306
1307 /// Return true if the predicate is SLT or ULT.
1308 ///
1309 static bool isLT(Predicate P) {
1310 return P == ICMP_SLT || P == ICMP_ULT;
1311 }
1312
1313 /// Return true if the predicate is SGE or UGE.
1314 ///
1315 static bool isGE(Predicate P) {
1316 return P == ICMP_SGE || P == ICMP_UGE;
1317 }
1318
1319 /// Return true if the predicate is SLE or ULE.
1320 ///
1321 static bool isLE(Predicate P) {
1322 return P == ICMP_SLE || P == ICMP_ULE;
1323 }
1324
1325 /// Returns the sequence of all ICmp predicates.
1326 ///
1327 static auto predicates() { return ICmpPredicates(); }
1328
1329 /// Exchange the two operands to this instruction in such a way that it does
1330 /// not modify the semantics of the instruction. The predicate value may be
1331 /// changed to retain the same result if the predicate is order dependent
1332 /// (e.g. ult).
1333 /// Swap operands and adjust predicate.
1334 void swapOperands() {
1335 setPredicate(getSwappedPredicate());
1336 Op<0>().swap(Op<1>());
1337 }
1338
1339 /// Return result of `LHS Pred RHS` comparison.
1340 static bool compare(const APInt &LHS, const APInt &RHS,
1341 ICmpInst::Predicate Pred);
1342
1343 // Methods for support type inquiry through isa, cast, and dyn_cast:
1344 static bool classof(const Instruction *I) {
1345 return I->getOpcode() == Instruction::ICmp;
1346 }
1347 static bool classof(const Value *V) {
1348 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1349 }
1350};
1351
1352//===----------------------------------------------------------------------===//
1353// FCmpInst Class
1354//===----------------------------------------------------------------------===//
1355
1356/// This instruction compares its operands according to the predicate given
1357/// to the constructor. It only operates on floating point values or packed
1358/// vectors of floating point values. The operands must be identical types.
1359/// Represents a floating point comparison operator.
1360class FCmpInst: public CmpInst {
1361 void AssertOK() {
1362 assert(isFPPredicate() && "Invalid FCmp predicate value")(static_cast <bool> (isFPPredicate() && "Invalid FCmp predicate value"
) ? void (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\""
, "llvm/include/llvm/IR/Instructions.h", 1362, __extension__ __PRETTY_FUNCTION__
))
;
1363 assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast <bool> (getOperand(0)->getType() == getOperand
(1)->getType() && "Both operands to FCmp instruction are not of the same type!"
) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "llvm/include/llvm/IR/Instructions.h", 1364, __extension__ __PRETTY_FUNCTION__
))
1364 "Both operands to FCmp instruction are not of the same type!")(static_cast <bool> (getOperand(0)->getType() == getOperand
(1)->getType() && "Both operands to FCmp instruction are not of the same type!"
) ? void (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "llvm/include/llvm/IR/Instructions.h", 1364, __extension__ __PRETTY_FUNCTION__
))
;
1365 // Check that the operands are the right type
1366 assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&(static_cast <bool> (getOperand(0)->getType()->isFPOrFPVectorTy
() && "Invalid operand types for FCmp instruction") ?
void (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "llvm/include/llvm/IR/Instructions.h", 1367, __extension__ __PRETTY_FUNCTION__
))
1367 "Invalid operand types for FCmp instruction")(static_cast <bool> (getOperand(0)->getType()->isFPOrFPVectorTy
() && "Invalid operand types for FCmp instruction") ?
void (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "llvm/include/llvm/IR/Instructions.h", 1367, __extension__ __PRETTY_FUNCTION__
))
;
1368 }
1369
1370protected:
1371 // Note: Instruction needs to be a friend here to call cloneImpl.
1372 friend class Instruction;
1373
1374 /// Clone an identical FCmpInst
1375 FCmpInst *cloneImpl() const;
1376
1377public:
1378 /// Constructor with insert-before-instruction semantics.
1379 FCmpInst(
1380 Instruction *InsertBefore, ///< Where to insert
1381 Predicate pred, ///< The predicate to use for the comparison
1382 Value *LHS, ///< The left-hand-side of the expression
1383 Value *RHS, ///< The right-hand-side of the expression
1384 const Twine &NameStr = "" ///< Name of the instruction
1385 ) : CmpInst(makeCmpResultType(LHS->getType()),
1386 Instruction::FCmp, pred, LHS, RHS, NameStr,
1387 InsertBefore) {
1388 AssertOK();
1389 }
1390
1391 /// Constructor with insert-at-end semantics.
1392 FCmpInst(
1393 BasicBlock &InsertAtEnd, ///< Block to insert into.
1394 Predicate pred, ///< The predicate to use for the comparison
1395 Value *LHS, ///< The left-hand-side of the expression
1396 Value *RHS, ///< The right-hand-side of the expression
1397 const Twine &NameStr = "" ///< Name of the instruction
1398 ) : CmpInst(makeCmpResultType(LHS->getType()),
1399 Instruction::FCmp, pred, LHS, RHS, NameStr,
1400 &InsertAtEnd) {
1401 AssertOK();
1402 }
1403
1404 /// Constructor with no-insertion semantics
1405 FCmpInst(
1406 Predicate Pred, ///< The predicate to use for the comparison
1407 Value *LHS, ///< The left-hand-side of the expression
1408 Value *RHS, ///< The right-hand-side of the expression
1409 const Twine &NameStr = "", ///< Name of the instruction
1410 Instruction *FlagsSource = nullptr
1411 ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS,
1412 RHS, NameStr, nullptr, FlagsSource) {
1413 AssertOK();
1414 }
1415
1416 /// @returns true if the predicate of this instruction is EQ or NE.
1417 /// Determine if this is an equality predicate.
1418 static bool isEquality(Predicate Pred) {
1419 return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ ||
1420 Pred == FCMP_UNE;
1421 }
1422
1423 /// @returns true if the predicate of this instruction is EQ or NE.
1424 /// Determine if this is an equality predicate.
1425 bool isEquality() const { return isEquality(getPredicate()); }
1426
1427 /// @returns true if the predicate of this instruction is commutative.
1428 /// Determine if this is a commutative predicate.
1429 bool isCommutative() const {
1430 return isEquality() ||
1431 getPredicate() == FCMP_FALSE ||
1432 getPredicate() == FCMP_TRUE ||
1433 getPredicate() == FCMP_ORD ||
1434 getPredicate() == FCMP_UNO;
1435 }
1436
1437 /// @returns true if the predicate is relational (not EQ or NE).
1438 /// Determine if this a relational predicate.
1439 bool isRelational() const { return !isEquality(); }
1440
1441 /// Exchange the two operands to this instruction in such a way that it does
1442 /// not modify the semantics of the instruction. The predicate value may be
1443 /// changed to retain the same result if the predicate is order dependent
1444 /// (e.g. ult).
1445 /// Swap operands and adjust predicate.
1446 void swapOperands() {
1447 setPredicate(getSwappedPredicate());
1448 Op<0>().swap(Op<1>());
1449 }
1450
1451 /// Returns the sequence of all FCmp predicates.
1452 ///
1453 static auto predicates() { return FCmpPredicates(); }
1454
1455 /// Return result of `LHS Pred RHS` comparison.
1456 static bool compare(const APFloat &LHS, const APFloat &RHS,
1457 FCmpInst::Predicate Pred);
1458
1459 /// Methods for support type inquiry through isa, cast, and dyn_cast:
1460 static bool classof(const Instruction *I) {
1461 return I->getOpcode() == Instruction::FCmp;
1462 }
1463 static bool classof(const Value *V) {
1464 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1465 }
1466};
1467
1468//===----------------------------------------------------------------------===//
1469/// This class represents a function call, abstracting a target
1470/// machine's calling convention. This class uses low bit of the SubClassData
1471/// field to indicate whether or not this is a tail call. The rest of the bits
1472/// hold the calling convention of the call.
1473///
1474class CallInst : public CallBase {
1475 CallInst(const CallInst &CI);
1476
1477 /// Construct a CallInst given a range of arguments.
1478 /// Construct a CallInst from a range of arguments
1479 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1480 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1481 Instruction *InsertBefore);
1482
1483 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1484 const Twine &NameStr, Instruction *InsertBefore)
1485 : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {}
1486
1487 /// Construct a CallInst given a range of arguments.
1488 /// Construct a CallInst from a range of arguments
1489 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1490 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1491 BasicBlock *InsertAtEnd);
1492
1493 explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr,
1494 Instruction *InsertBefore);
1495
1496 CallInst(FunctionType *ty, Value *F, const Twine &NameStr,
1497 BasicBlock *InsertAtEnd);
1498
1499 void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args,
1500 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
1501 void init(FunctionType *FTy, Value *Func, const Twine &NameStr);
1502
1503 /// Compute the number of operands to allocate.
1504 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
1505 // We need one operand for the called function, plus the input operand
1506 // counts provided.
1507 return 1 + NumArgs + NumBundleInputs;
1508 }
1509
1510protected:
1511 // Note: Instruction needs to be a friend here to call cloneImpl.
1512 friend class Instruction;
1513
1514 CallInst *cloneImpl() const;
1515
1516public:
1517 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "",
1518 Instruction *InsertBefore = nullptr) {
1519 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore);
1520 }
1521
1522 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1523 const Twine &NameStr,
1524 Instruction *InsertBefore = nullptr) {
1525 return new (ComputeNumOperands(Args.size()))
1526 CallInst(Ty, Func, Args, None, NameStr, InsertBefore);
1527 }
1528
1529 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1530 ArrayRef<OperandBundleDef> Bundles = None,
1531 const Twine &NameStr = "",
1532 Instruction *InsertBefore = nullptr) {
1533 const int NumOperands =
1534 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1535 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1536
1537 return new (NumOperands, DescriptorBytes)
1538 CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore);
1539 }
1540
1541 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr,
1542 BasicBlock *InsertAtEnd) {
1543 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd);
1544 }
1545
1546 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1547 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1548 return new (ComputeNumOperands(Args.size()))
1549 CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd);
1550 }
1551
1552 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1553 ArrayRef<OperandBundleDef> Bundles,
1554 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1555 const int NumOperands =
1556 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1557 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1558
1559 return new (NumOperands, DescriptorBytes)
1560 CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd);
1561 }
1562
1563 static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "",
1564 Instruction *InsertBefore = nullptr) {
1565 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1566 InsertBefore);
1567 }
1568
1569 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1570 ArrayRef<OperandBundleDef> Bundles = None,
1571 const Twine &NameStr = "",
1572 Instruction *InsertBefore = nullptr) {
1573 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1574 NameStr, InsertBefore);
1575 }
1576
1577 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1578 const Twine &NameStr,
1579 Instruction *InsertBefore = nullptr) {
1580 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1581 InsertBefore);
1582 }
1583
1584 static CallInst *Create(FunctionCallee Func, const Twine &NameStr,
1585 BasicBlock *InsertAtEnd) {
1586 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1587 InsertAtEnd);
1588 }
1589
1590 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1591 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1592 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1593 InsertAtEnd);
1594 }
1595
1596 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1597 ArrayRef<OperandBundleDef> Bundles,
1598 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1599 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1600 NameStr, InsertAtEnd);
1601 }
1602
1603 /// Create a clone of \p CI with a different set of operand bundles and
1604 /// insert it before \p InsertPt.
1605 ///
1606 /// The returned call instruction is identical \p CI in every way except that
1607 /// the operand bundles for the new instruction are set to the operand bundles
1608 /// in \p Bundles.
1609 static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles,
1610 Instruction *InsertPt = nullptr);
1611
1612 /// Generate the IR for a call to malloc:
1613 /// 1. Compute the malloc call's argument as the specified type's size,
1614 /// possibly multiplied by the array size if the array size is not
1615 /// constant 1.
1616 /// 2. Call malloc with that argument.
1617 /// 3. Bitcast the result of the malloc call to the specified type.
1618 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1619 Type *AllocTy, Value *AllocSize,
1620 Value *ArraySize = nullptr,
1621 Function *MallocF = nullptr,
1622 const Twine &Name = "");
1623 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1624 Type *AllocTy, Value *AllocSize,
1625 Value *ArraySize = nullptr,
1626 Function *MallocF = nullptr,
1627 const Twine &Name = "");
1628 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1629 Type *AllocTy, Value *AllocSize,
1630 Value *ArraySize = nullptr,
1631 ArrayRef<OperandBundleDef> Bundles = None,
1632 Function *MallocF = nullptr,
1633 const Twine &Name = "");
1634 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1635 Type *AllocTy, Value *AllocSize,
1636 Value *ArraySize = nullptr,
1637 ArrayRef<OperandBundleDef> Bundles = None,
1638 Function *MallocF = nullptr,
1639 const Twine &Name = "");
1640 /// Generate the IR for a call to the builtin free function.
1641 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1642 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1643 static Instruction *CreateFree(Value *Source,
1644 ArrayRef<OperandBundleDef> Bundles,
1645 Instruction *InsertBefore);
1646 static Instruction *CreateFree(Value *Source,
1647 ArrayRef<OperandBundleDef> Bundles,
1648 BasicBlock *InsertAtEnd);
1649
1650 // Note that 'musttail' implies 'tail'.
1651 enum TailCallKind : unsigned {
1652 TCK_None = 0,
1653 TCK_Tail = 1,
1654 TCK_MustTail = 2,
1655 TCK_NoTail = 3,
1656 TCK_LAST = TCK_NoTail
1657 };
1658
1659 using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>;
1660 static_assert(
1661 Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(),
1662 "Bitfields must be contiguous");
1663
1664 TailCallKind getTailCallKind() const {
1665 return getSubclassData<TailCallKindField>();
1666 }
1667
1668 bool isTailCall() const {
1669 TailCallKind Kind = getTailCallKind();
1670 return Kind == TCK_Tail || Kind == TCK_MustTail;
1671 }
1672
1673 bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; }
1674
1675 bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; }
1676
1677 void setTailCallKind(TailCallKind TCK) {
1678 setSubclassData<TailCallKindField>(TCK);
1679 }
1680
1681 void setTailCall(bool IsTc = true) {
1682 setTailCallKind(IsTc ? TCK_Tail : TCK_None);
1683 }
1684
1685 /// Return true if the call can return twice
1686 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1687 void setCanReturnTwice() { addFnAttr(Attribute::ReturnsTwice); }
1688
1689 // Methods for support type inquiry through isa, cast, and dyn_cast:
1690 static bool classof(const Instruction *I) {
1691 return I->getOpcode() == Instruction::Call;
1692 }
1693 static bool classof(const Value *V) {
1694 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1695 }
1696
1697 /// Updates profile metadata by scaling it by \p S / \p T.
1698 void updateProfWeight(uint64_t S, uint64_t T);
1699
1700private:
1701 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1702 // method so that subclasses cannot accidentally use it.
1703 template <typename Bitfield>
1704 void setSubclassData(typename Bitfield::Type Value) {
1705 Instruction::setSubclassData<Bitfield>(Value);
1706 }
1707};
1708
1709CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1710 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1711 BasicBlock *InsertAtEnd)
1712 : CallBase(Ty->getReturnType(), Instruction::Call,
1713 OperandTraits<CallBase>::op_end(this) -
1714 (Args.size() + CountBundleInputs(Bundles) + 1),
1715 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1716 InsertAtEnd) {
1717 init(Ty, Func, Args, Bundles, NameStr);
1718}
1719
1720CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1721 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1722 Instruction *InsertBefore)
1723 : CallBase(Ty->getReturnType(), Instruction::Call,
1724 OperandTraits<CallBase>::op_end(this) -
1725 (Args.size() + CountBundleInputs(Bundles) + 1),
1726 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1727 InsertBefore) {
1728 init(Ty, Func, Args, Bundles, NameStr);
1729}
1730
1731//===----------------------------------------------------------------------===//
1732// SelectInst Class
1733//===----------------------------------------------------------------------===//
1734
1735/// This class represents the LLVM 'select' instruction.
1736///
1737class SelectInst : public Instruction {
1738 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1739 Instruction *InsertBefore)
1740 : Instruction(S1->getType(), Instruction::Select,
1741 &Op<0>(), 3, InsertBefore) {
1742 init(C, S1, S2);
1743 setName(NameStr);
1744 }
1745
1746 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1747 BasicBlock *InsertAtEnd)
1748 : Instruction(S1->getType(), Instruction::Select,
1749 &Op<0>(), 3, InsertAtEnd) {
1750 init(C, S1, S2);
1751 setName(NameStr);
1752 }
1753
1754 void init(Value *C, Value *S1, Value *S2) {
1755 assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")(static_cast <bool> (!areInvalidOperands(C, S1, S2) &&
"Invalid operands for select") ? void (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\""
, "llvm/include/llvm/IR/Instructions.h", 1755, __extension__ __PRETTY_FUNCTION__
))
;
1756 Op<0>() = C;
1757 Op<1>() = S1;
1758 Op<2>() = S2;
1759 }
1760
1761protected:
1762 // Note: Instruction needs to be a friend here to call cloneImpl.
1763 friend class Instruction;
1764
1765 SelectInst *cloneImpl() const;
1766
1767public:
1768 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1769 const Twine &NameStr = "",
1770 Instruction *InsertBefore = nullptr,
1771 Instruction *MDFrom = nullptr) {
1772 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1773 if (MDFrom)
1774 Sel->copyMetadata(*MDFrom);
1775 return Sel;
1776 }
1777
1778 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1779 const Twine &NameStr,
1780 BasicBlock *InsertAtEnd) {
1781 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1782 }
1783
1784 const Value *getCondition() const { return Op<0>(); }
1785 const Value *getTrueValue() const { return Op<1>(); }
1786 const Value *getFalseValue() const { return Op<2>(); }
1787 Value *getCondition() { return Op<0>(); }
1788 Value *getTrueValue() { return Op<1>(); }
1789 Value *getFalseValue() { return Op<2>(); }
1790
1791 void setCondition(Value *V) { Op<0>() = V; }
1792 void setTrueValue(Value *V) { Op<1>() = V; }
1793 void setFalseValue(Value *V) { Op<2>() = V; }
1794
1795 /// Swap the true and false values of the select instruction.
1796 /// This doesn't swap prof metadata.
1797 void swapValues() { Op<1>().swap(Op<2>()); }
1798
1799 /// Return a string if the specified operands are invalid
1800 /// for a select operation, otherwise return null.
1801 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1802
1803 /// Transparently provide more efficient getOperand methods.
1804 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1805
1806 OtherOps getOpcode() const {
1807 return static_cast<OtherOps>(Instruction::getOpcode());
1808 }
1809
1810 // Methods for support type inquiry through isa, cast, and dyn_cast:
1811 static bool classof(const Instruction *I) {
1812 return I->getOpcode() == Instruction::Select;
1813 }
1814 static bool classof(const Value *V) {
1815 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1816 }
1817};
1818
1819template <>
1820struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1821};
1822
1823DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits
<SelectInst>::op_begin(this); } SelectInst::const_op_iterator
SelectInst::op_begin() const { return OperandTraits<SelectInst
>::op_begin(const_cast<SelectInst*>(this)); } SelectInst
::op_iterator SelectInst::op_end() { return OperandTraits<
SelectInst>::op_end(this); } SelectInst::const_op_iterator
SelectInst::op_end() const { return OperandTraits<SelectInst
>::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<SelectInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1823, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<SelectInst
>::op_begin(const_cast<SelectInst*>(this))[i_nocapture
].get()); } void SelectInst::setOperand(unsigned i_nocapture,
Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<SelectInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1823, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<SelectInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned SelectInst::getNumOperands() const
{ return OperandTraits<SelectInst>::operands(this); } template
<int Idx_nocapture> Use &SelectInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &SelectInst::Op() const { return
this->OpFrom<Idx_nocapture>(this); }
1824
1825//===----------------------------------------------------------------------===//
1826// VAArgInst Class
1827//===----------------------------------------------------------------------===//
1828
1829/// This class represents the va_arg llvm instruction, which returns
1830/// an argument of the specified type given a va_list and increments that list
1831///
1832class VAArgInst : public UnaryInstruction {
1833protected:
1834 // Note: Instruction needs to be a friend here to call cloneImpl.
1835 friend class Instruction;
1836
1837 VAArgInst *cloneImpl() const;
1838
1839public:
1840 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1841 Instruction *InsertBefore = nullptr)
1842 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1843 setName(NameStr);
1844 }
1845
1846 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1847 BasicBlock *InsertAtEnd)
1848 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1849 setName(NameStr);
1850 }
1851
1852 Value *getPointerOperand() { return getOperand(0); }
1853 const Value *getPointerOperand() const { return getOperand(0); }
1854 static unsigned getPointerOperandIndex() { return 0U; }
1855
1856 // Methods for support type inquiry through isa, cast, and dyn_cast:
1857 static bool classof(const Instruction *I) {
1858 return I->getOpcode() == VAArg;
1859 }
1860 static bool classof(const Value *V) {
1861 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1862 }
1863};
1864
1865//===----------------------------------------------------------------------===//
1866// ExtractElementInst Class
1867//===----------------------------------------------------------------------===//
1868
1869/// This instruction extracts a single (scalar)
1870/// element from a VectorType value
1871///
1872class ExtractElementInst : public Instruction {
1873 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1874 Instruction *InsertBefore = nullptr);
1875 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1876 BasicBlock *InsertAtEnd);
1877
1878protected:
1879 // Note: Instruction needs to be a friend here to call cloneImpl.
1880 friend class Instruction;
1881
1882 ExtractElementInst *cloneImpl() const;
1883
1884public:
1885 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1886 const Twine &NameStr = "",
1887 Instruction *InsertBefore = nullptr) {
1888 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1889 }
1890
1891 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1892 const Twine &NameStr,
1893 BasicBlock *InsertAtEnd) {
1894 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1895 }
1896
1897 /// Return true if an extractelement instruction can be
1898 /// formed with the specified operands.
1899 static bool isValidOperands(const Value *Vec, const Value *Idx);
1900
1901 Value *getVectorOperand() { return Op<0>(); }
1902 Value *getIndexOperand() { return Op<1>(); }
1903 const Value *getVectorOperand() const { return Op<0>(); }
1904 const Value *getIndexOperand() const { return Op<1>(); }
1905
1906 VectorType *getVectorOperandType() const {
1907 return cast<VectorType>(getVectorOperand()->getType());
1908 }
1909
1910 /// Transparently provide more efficient getOperand methods.
1911 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1912
1913 // Methods for support type inquiry through isa, cast, and dyn_cast:
1914 static bool classof(const Instruction *I) {
1915 return I->getOpcode() == Instruction::ExtractElement;
1916 }
1917 static bool classof(const Value *V) {
1918 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1919 }
1920};
1921
1922template <>
1923struct OperandTraits<ExtractElementInst> :
1924 public FixedNumOperandTraits<ExtractElementInst, 2> {
1925};
1926
1927DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin(
) { return OperandTraits<ExtractElementInst>::op_begin(
this); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_begin() const { return OperandTraits<ExtractElementInst
>::op_begin(const_cast<ExtractElementInst*>(this)); }
ExtractElementInst::op_iterator ExtractElementInst::op_end()
{ return OperandTraits<ExtractElementInst>::op_end(this
); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_end() const { return OperandTraits<ExtractElementInst
>::op_end(const_cast<ExtractElementInst*>(this)); } Value
*ExtractElementInst::getOperand(unsigned i_nocapture) const {
(static_cast <bool> (i_nocapture < OperandTraits<
ExtractElementInst>::operands(this) && "getOperand() out of range!"
) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1927, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<ExtractElementInst
>::op_begin(const_cast<ExtractElementInst*>(this))[i_nocapture
].get()); } void ExtractElementInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<ExtractElementInst>::operands(this)
&& "setOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1927, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<ExtractElementInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned ExtractElementInst::getNumOperands
() const { return OperandTraits<ExtractElementInst>::operands
(this); } template <int Idx_nocapture> Use &ExtractElementInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &ExtractElementInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
1928
1929//===----------------------------------------------------------------------===//
1930// InsertElementInst Class
1931//===----------------------------------------------------------------------===//
1932
1933/// This instruction inserts a single (scalar)
1934/// element into a VectorType value
1935///
1936class InsertElementInst : public Instruction {
1937 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1938 const Twine &NameStr = "",
1939 Instruction *InsertBefore = nullptr);
1940 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1941 BasicBlock *InsertAtEnd);
1942
1943protected:
1944 // Note: Instruction needs to be a friend here to call cloneImpl.
1945 friend class Instruction;
1946
1947 InsertElementInst *cloneImpl() const;
1948
1949public:
1950 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1951 const Twine &NameStr = "",
1952 Instruction *InsertBefore = nullptr) {
1953 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1954 }
1955
1956 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1957 const Twine &NameStr,
1958 BasicBlock *InsertAtEnd) {
1959 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1960 }
1961
1962 /// Return true if an insertelement instruction can be
1963 /// formed with the specified operands.
1964 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1965 const Value *Idx);
1966
1967 /// Overload to return most specific vector type.
1968 ///
1969 VectorType *getType() const {
1970 return cast<VectorType>(Instruction::getType());
1971 }
1972
1973 /// Transparently provide more efficient getOperand methods.
1974 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1975
1976 // Methods for support type inquiry through isa, cast, and dyn_cast:
1977 static bool classof(const Instruction *I) {
1978 return I->getOpcode() == Instruction::InsertElement;
1979 }
1980 static bool classof(const Value *V) {
1981 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1982 }
1983};
1984
1985template <>
1986struct OperandTraits<InsertElementInst> :
1987 public FixedNumOperandTraits<InsertElementInst, 3> {
1988};
1989
1990DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() {
return OperandTraits<InsertElementInst>::op_begin(this
); } InsertElementInst::const_op_iterator InsertElementInst::
op_begin() const { return OperandTraits<InsertElementInst>
::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst
::op_iterator InsertElementInst::op_end() { return OperandTraits
<InsertElementInst>::op_end(this); } InsertElementInst::
const_op_iterator InsertElementInst::op_end() const { return OperandTraits
<InsertElementInst>::op_end(const_cast<InsertElementInst
*>(this)); } Value *InsertElementInst::getOperand(unsigned
i_nocapture) const { (static_cast <bool> (i_nocapture <
OperandTraits<InsertElementInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1990, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<InsertElementInst
>::op_begin(const_cast<InsertElementInst*>(this))[i_nocapture
].get()); } void InsertElementInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<InsertElementInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 1990, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<InsertElementInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned InsertElementInst::getNumOperands
() const { return OperandTraits<InsertElementInst>::operands
(this); } template <int Idx_nocapture> Use &InsertElementInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &InsertElementInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
1991
1992//===----------------------------------------------------------------------===//
1993// ShuffleVectorInst Class
1994//===----------------------------------------------------------------------===//
1995
1996constexpr int UndefMaskElem = -1;
1997
1998/// This instruction constructs a fixed permutation of two
1999/// input vectors.
2000///
2001/// For each element of the result vector, the shuffle mask selects an element
2002/// from one of the input vectors to copy to the result. Non-negative elements
2003/// in the mask represent an index into the concatenated pair of input vectors.
2004/// UndefMaskElem (-1) specifies that the result element is undefined.
2005///
2006/// For scalable vectors, all the elements of the mask must be 0 or -1. This
2007/// requirement may be relaxed in the future.
2008class ShuffleVectorInst : public Instruction {
2009 SmallVector<int, 4> ShuffleMask;
2010 Constant *ShuffleMaskForBitcode;
2011
2012protected:
2013 // Note: Instruction needs to be a friend here to call cloneImpl.
2014 friend class Instruction;
2015
2016 ShuffleVectorInst *cloneImpl() const;
2017
2018public:
2019 ShuffleVectorInst(Value *V1, Value *Mask, const Twine &NameStr = "",
2020 Instruction *InsertBefore = nullptr);
2021 ShuffleVectorInst(Value *V1, Value *Mask, const Twine &NameStr,
2022 BasicBlock *InsertAtEnd);
2023 ShuffleVectorInst(Value *V1, ArrayRef<int> Mask, const Twine &NameStr = "",
2024 Instruction *InsertBefore = nullptr);
2025 ShuffleVectorInst(Value *V1, ArrayRef<int> Mask, const Twine &NameStr,
2026 BasicBlock *InsertAtEnd);
2027 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2028 const Twine &NameStr = "",
2029 Instruction *InsertBefor = nullptr);
2030 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2031 const Twine &NameStr, BasicBlock *InsertAtEnd);
2032 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2033 const Twine &NameStr = "",
2034 Instruction *InsertBefor = nullptr);
2035 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2036 const Twine &NameStr, BasicBlock *InsertAtEnd);
2037
2038 void *operator new(size_t S) { return User::operator new(S, 2); }
2039 void operator delete(void *Ptr) { return User::operator delete(Ptr); }
2040
2041 /// Swap the operands and adjust the mask to preserve the semantics
2042 /// of the instruction.
2043 void commute();
2044
2045 /// Return true if a shufflevector instruction can be
2046 /// formed with the specified operands.
2047 static bool isValidOperands(const Value *V1, const Value *V2,
2048 const Value *Mask);
2049 static bool isValidOperands(const Value *V1, const Value *V2,
2050 ArrayRef<int> Mask);
2051
2052 /// Overload to return most specific vector type.
2053 ///
2054 VectorType *getType() const {
2055 return cast<VectorType>(Instruction::getType());
2056 }
2057
2058 /// Transparently provide more efficient getOperand methods.
2059 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2060
2061 /// Return the shuffle mask value of this instruction for the given element
2062 /// index. Return UndefMaskElem if the element is undef.
2063 int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; }
2064
2065 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2066 /// elements of the mask are returned as UndefMaskElem.
2067 static void getShuffleMask(const Constant *Mask,
2068 SmallVectorImpl<int> &Result);
2069
2070 /// Return the mask for this instruction as a vector of integers. Undefined
2071 /// elements of the mask are returned as UndefMaskElem.
2072 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2073 Result.assign(ShuffleMask.begin(), ShuffleMask.end());
2074 }
2075
2076 /// Return the mask for this instruction, for use in bitcode.
2077 ///
2078 /// TODO: This is temporary until we decide a new bitcode encoding for
2079 /// shufflevector.
2080 Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; }
2081
2082 static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask,
2083 Type *ResultTy);
2084
2085 void setShuffleMask(ArrayRef<int> Mask);
2086
2087 ArrayRef<int> getShuffleMask() const { return ShuffleMask; }
2088
2089 /// Return true if this shuffle returns a vector with a different number of
2090 /// elements than its source vectors.
2091 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2092 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2093 bool changesLength() const {
2094 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2095 ->getElementCount()
2096 .getKnownMinValue();
2097 unsigned NumMaskElts = ShuffleMask.size();
2098 return NumSourceElts != NumMaskElts;
2099 }
2100
2101 /// Return true if this shuffle returns a vector with a greater number of
2102 /// elements than its source vectors.
2103 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2104 bool increasesLength() const {
2105 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2106 ->getElementCount()
2107 .getKnownMinValue();
2108 unsigned NumMaskElts = ShuffleMask.size();
2109 return NumSourceElts < NumMaskElts;
2110 }
2111
2112 /// Return true if this shuffle mask chooses elements from exactly one source
2113 /// vector.
2114 /// Example: <7,5,undef,7>
2115 /// This assumes that vector operands are the same length as the mask.
2116 static bool isSingleSourceMask(ArrayRef<int> Mask);
2117 static bool isSingleSourceMask(const Constant *Mask) {
2118 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2118, __extension__ __PRETTY_FUNCTION__
))
;
2119 SmallVector<int, 16> MaskAsInts;
2120 getShuffleMask(Mask, MaskAsInts);
2121 return isSingleSourceMask(MaskAsInts);
2122 }
2123
2124 /// Return true if this shuffle chooses elements from exactly one source
2125 /// vector without changing the length of that vector.
2126 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2127 /// TODO: Optionally allow length-changing shuffles.
2128 bool isSingleSource() const {
2129 return !changesLength() && isSingleSourceMask(ShuffleMask);
2130 }
2131
2132 /// Return true if this shuffle mask chooses elements from exactly one source
2133 /// vector without lane crossings. A shuffle using this mask is not
2134 /// necessarily a no-op because it may change the number of elements from its
2135 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2136 /// Example: <undef,undef,2,3>
2137 static bool isIdentityMask(ArrayRef<int> Mask);
2138 static bool isIdentityMask(const Constant *Mask) {
2139 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2139, __extension__ __PRETTY_FUNCTION__
))
;
2140
2141 // Not possible to express a shuffle mask for a scalable vector for this
2142 // case.
2143 if (isa<ScalableVectorType>(Mask->getType()))
2144 return false;
2145
2146 SmallVector<int, 16> MaskAsInts;
2147 getShuffleMask(Mask, MaskAsInts);
2148 return isIdentityMask(MaskAsInts);
2149 }
2150
2151 /// Return true if this shuffle chooses elements from exactly one source
2152 /// vector without lane crossings and does not change the number of elements
2153 /// from its input vectors.
2154 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2155 bool isIdentity() const {
2156 // Not possible to express a shuffle mask for a scalable vector for this
2157 // case.
2158 if (isa<ScalableVectorType>(getType()))
2159 return false;
2160
2161 return !changesLength() && isIdentityMask(ShuffleMask);
2162 }
2163
2164 /// Return true if this shuffle lengthens exactly one source vector with
2165 /// undefs in the high elements.
2166 bool isIdentityWithPadding() const;
2167
2168 /// Return true if this shuffle extracts the first N elements of exactly one
2169 /// source vector.
2170 bool isIdentityWithExtract() const;
2171
2172 /// Return true if this shuffle concatenates its 2 source vectors. This
2173 /// returns false if either input is undefined. In that case, the shuffle is
2174 /// is better classified as an identity with padding operation.
2175 bool isConcat() const;
2176
2177 /// Return true if this shuffle mask chooses elements from its source vectors
2178 /// without lane crossings. A shuffle using this mask would be
2179 /// equivalent to a vector select with a constant condition operand.
2180 /// Example: <4,1,6,undef>
2181 /// This returns false if the mask does not choose from both input vectors.
2182 /// In that case, the shuffle is better classified as an identity shuffle.
2183 /// This assumes that vector operands are the same length as the mask
2184 /// (a length-changing shuffle can never be equivalent to a vector select).
2185 static bool isSelectMask(ArrayRef<int> Mask);
2186 static bool isSelectMask(const Constant *Mask) {
2187 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2187, __extension__ __PRETTY_FUNCTION__
))
;
2188 SmallVector<int, 16> MaskAsInts;
2189 getShuffleMask(Mask, MaskAsInts);
2190 return isSelectMask(MaskAsInts);
2191 }
2192
2193 /// Return true if this shuffle chooses elements from its source vectors
2194 /// without lane crossings and all operands have the same number of elements.
2195 /// In other words, this shuffle is equivalent to a vector select with a
2196 /// constant condition operand.
2197 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2198 /// This returns false if the mask does not choose from both input vectors.
2199 /// In that case, the shuffle is better classified as an identity shuffle.
2200 /// TODO: Optionally allow length-changing shuffles.
2201 bool isSelect() const {
2202 return !changesLength() && isSelectMask(ShuffleMask);
2203 }
2204
2205 /// Return true if this shuffle mask swaps the order of elements from exactly
2206 /// one source vector.
2207 /// Example: <7,6,undef,4>
2208 /// This assumes that vector operands are the same length as the mask.
2209 static bool isReverseMask(ArrayRef<int> Mask);
2210 static bool isReverseMask(const Constant *Mask) {
2211 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2211, __extension__ __PRETTY_FUNCTION__
))
;
2212 SmallVector<int, 16> MaskAsInts;
2213 getShuffleMask(Mask, MaskAsInts);
2214 return isReverseMask(MaskAsInts);
2215 }
2216
2217 /// Return true if this shuffle swaps the order of elements from exactly
2218 /// one source vector.
2219 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2220 /// TODO: Optionally allow length-changing shuffles.
2221 bool isReverse() const {
2222 return !changesLength() && isReverseMask(ShuffleMask);
2223 }
2224
2225 /// Return true if this shuffle mask chooses all elements with the same value
2226 /// as the first element of exactly one source vector.
2227 /// Example: <4,undef,undef,4>
2228 /// This assumes that vector operands are the same length as the mask.
2229 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2230 static bool isZeroEltSplatMask(const Constant *Mask) {
2231 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2231, __extension__ __PRETTY_FUNCTION__
))
;
2232 SmallVector<int, 16> MaskAsInts;
2233 getShuffleMask(Mask, MaskAsInts);
2234 return isZeroEltSplatMask(MaskAsInts);
2235 }
2236
2237 /// Return true if all elements of this shuffle are the same value as the
2238 /// first element of exactly one source vector without changing the length
2239 /// of that vector.
2240 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2241 /// TODO: Optionally allow length-changing shuffles.
2242 /// TODO: Optionally allow splats from other elements.
2243 bool isZeroEltSplat() const {
2244 return !changesLength() && isZeroEltSplatMask(ShuffleMask);
2245 }
2246
2247 /// Return true if this shuffle mask is a transpose mask.
2248 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2249 /// even- or odd-numbered vector elements from two n-dimensional source
2250 /// vectors and write each result into consecutive elements of an
2251 /// n-dimensional destination vector. Two shuffles are necessary to complete
2252 /// the transpose, one for the even elements and another for the odd elements.
2253 /// This description closely follows how the TRN1 and TRN2 AArch64
2254 /// instructions operate.
2255 ///
2256 /// For example, a simple 2x2 matrix can be transposed with:
2257 ///
2258 /// ; Original matrix
2259 /// m0 = < a, b >
2260 /// m1 = < c, d >
2261 ///
2262 /// ; Transposed matrix
2263 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2264 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2265 ///
2266 /// For matrices having greater than n columns, the resulting nx2 transposed
2267 /// matrix is stored in two result vectors such that one vector contains
2268 /// interleaved elements from all the even-numbered rows and the other vector
2269 /// contains interleaved elements from all the odd-numbered rows. For example,
2270 /// a 2x4 matrix can be transposed with:
2271 ///
2272 /// ; Original matrix
2273 /// m0 = < a, b, c, d >
2274 /// m1 = < e, f, g, h >
2275 ///
2276 /// ; Transposed matrix
2277 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2278 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2279 static bool isTransposeMask(ArrayRef<int> Mask);
2280 static bool isTransposeMask(const Constant *Mask) {
2281 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2281, __extension__ __PRETTY_FUNCTION__
))
;
2282 SmallVector<int, 16> MaskAsInts;
2283 getShuffleMask(Mask, MaskAsInts);
2284 return isTransposeMask(MaskAsInts);
2285 }
2286
2287 /// Return true if this shuffle transposes the elements of its inputs without
2288 /// changing the length of the vectors. This operation may also be known as a
2289 /// merge or interleave. See the description for isTransposeMask() for the
2290 /// exact specification.
2291 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2292 bool isTranspose() const {
2293 return !changesLength() && isTransposeMask(ShuffleMask);
2294 }
2295
2296 /// Return true if this shuffle mask is a splice mask, concatenating the two
2297 /// inputs together and then extracts an original width vector starting from
2298 /// the splice index.
2299 /// Example: shufflevector <4 x n> A, <4 x n> B, <1,2,3,4>
2300 static bool isSpliceMask(ArrayRef<int> Mask, int &Index);
2301 static bool isSpliceMask(const Constant *Mask, int &Index) {
2302 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2302, __extension__ __PRETTY_FUNCTION__
))
;
2303 SmallVector<int, 16> MaskAsInts;
2304 getShuffleMask(Mask, MaskAsInts);
2305 return isSpliceMask(MaskAsInts, Index);
2306 }
2307
2308 /// Return true if this shuffle splices two inputs without changing the length
2309 /// of the vectors. This operation concatenates the two inputs together and
2310 /// then extracts an original width vector starting from the splice index.
2311 /// Example: shufflevector <4 x n> A, <4 x n> B, <1,2,3,4>
2312 bool isSplice(int &Index) const {
2313 return !changesLength() && isSpliceMask(ShuffleMask, Index);
2314 }
2315
2316 /// Return true if this shuffle mask is an extract subvector mask.
2317 /// A valid extract subvector mask returns a smaller vector from a single
2318 /// source operand. The base extraction index is returned as well.
2319 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2320 int &Index);
2321 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2322 int &Index) {
2323 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2323, __extension__ __PRETTY_FUNCTION__
))
;
2324 // Not possible to express a shuffle mask for a scalable vector for this
2325 // case.
2326 if (isa<ScalableVectorType>(Mask->getType()))
2327 return false;
2328 SmallVector<int, 16> MaskAsInts;
2329 getShuffleMask(Mask, MaskAsInts);
2330 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2331 }
2332
2333 /// Return true if this shuffle mask is an extract subvector mask.
2334 bool isExtractSubvectorMask(int &Index) const {
2335 // Not possible to express a shuffle mask for a scalable vector for this
2336 // case.
2337 if (isa<ScalableVectorType>(getType()))
2338 return false;
2339
2340 int NumSrcElts =
2341 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2342 return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index);
2343 }
2344
2345 /// Return true if this shuffle mask is an insert subvector mask.
2346 /// A valid insert subvector mask inserts the lowest elements of a second
2347 /// source operand into an in-place first source operand operand.
2348 /// Both the sub vector width and the insertion index is returned.
2349 static bool isInsertSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2350 int &NumSubElts, int &Index);
2351 static bool isInsertSubvectorMask(const Constant *Mask, int NumSrcElts,
2352 int &NumSubElts, int &Index) {
2353 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2353, __extension__ __PRETTY_FUNCTION__
))
;
2354 // Not possible to express a shuffle mask for a scalable vector for this
2355 // case.
2356 if (isa<ScalableVectorType>(Mask->getType()))
2357 return false;
2358 SmallVector<int, 16> MaskAsInts;
2359 getShuffleMask(Mask, MaskAsInts);
2360 return isInsertSubvectorMask(MaskAsInts, NumSrcElts, NumSubElts, Index);
2361 }
2362
2363 /// Return true if this shuffle mask is an insert subvector mask.
2364 bool isInsertSubvectorMask(int &NumSubElts, int &Index) const {
2365 // Not possible to express a shuffle mask for a scalable vector for this
2366 // case.
2367 if (isa<ScalableVectorType>(getType()))
2368 return false;
2369
2370 int NumSrcElts =
2371 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2372 return isInsertSubvectorMask(ShuffleMask, NumSrcElts, NumSubElts, Index);
2373 }
2374
2375 /// Return true if this shuffle mask replicates each of the \p VF elements
2376 /// in a vector \p ReplicationFactor times.
2377 /// For example, the mask for \p ReplicationFactor=3 and \p VF=4 is:
2378 /// <0,0,0,1,1,1,2,2,2,3,3,3>
2379 static bool isReplicationMask(ArrayRef<int> Mask, int &ReplicationFactor,
2380 int &VF);
2381 static bool isReplicationMask(const Constant *Mask, int &ReplicationFactor,
2382 int &VF) {
2383 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast <bool> (Mask->getType()->isVectorTy(
) && "Shuffle needs vector constant.") ? void (0) : __assert_fail
("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "llvm/include/llvm/IR/Instructions.h", 2383, __extension__ __PRETTY_FUNCTION__
))
;
2384 // Not possible to express a shuffle mask for a scalable vector for this
2385 // case.
2386 if (isa<ScalableVectorType>(Mask->getType()))
2387 return false;
2388 SmallVector<int, 16> MaskAsInts;
2389 getShuffleMask(Mask, MaskAsInts);
2390 return isReplicationMask(MaskAsInts, ReplicationFactor, VF);
2391 }
2392
2393 /// Return true if this shuffle mask is a replication mask.
2394 bool isReplicationMask(int &ReplicationFactor, int &VF) const;
2395
2396 /// Change values in a shuffle permute mask assuming the two vector operands
2397 /// of length InVecNumElts have swapped position.
2398 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2399 unsigned InVecNumElts) {
2400 for (int &Idx : Mask) {
2401 if (Idx == -1)
2402 continue;
2403 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2404 assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&(static_cast <bool> (Idx >= 0 && Idx < (int
)InVecNumElts * 2 && "shufflevector mask index out of range"
) ? void (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "llvm/include/llvm/IR/Instructions.h", 2405, __extension__ __PRETTY_FUNCTION__
))
2405 "shufflevector mask index out of range")(static_cast <bool> (Idx >= 0 && Idx < (int
)InVecNumElts * 2 && "shufflevector mask index out of range"
) ? void (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "llvm/include/llvm/IR/Instructions.h", 2405, __extension__ __PRETTY_FUNCTION__
))
;
2406 }
2407 }
2408
2409 // Methods for support type inquiry through isa, cast, and dyn_cast:
2410 static bool classof(const Instruction *I) {
2411 return I->getOpcode() == Instruction::ShuffleVector;
2412 }
2413 static bool classof(const Value *V) {
2414 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2415 }
2416};
2417
2418template <>
2419struct OperandTraits<ShuffleVectorInst>
2420 : public FixedNumOperandTraits<ShuffleVectorInst, 2> {};
2421
2422DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() {
return OperandTraits<ShuffleVectorInst>::op_begin(this
); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst::
op_begin() const { return OperandTraits<ShuffleVectorInst>
::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst
::op_iterator ShuffleVectorInst::op_end() { return OperandTraits
<ShuffleVectorInst>::op_end(this); } ShuffleVectorInst::
const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits
<ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst
*>(this)); } Value *ShuffleVectorInst::getOperand(unsigned
i_nocapture) const { (static_cast <bool> (i_nocapture <
OperandTraits<ShuffleVectorInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2422, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<ShuffleVectorInst
>::op_begin(const_cast<ShuffleVectorInst*>(this))[i_nocapture
].get()); } void ShuffleVectorInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<ShuffleVectorInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2422, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<ShuffleVectorInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned ShuffleVectorInst::getNumOperands
() const { return OperandTraits<ShuffleVectorInst>::operands
(this); } template <int Idx_nocapture> Use &ShuffleVectorInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &ShuffleVectorInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
2423
2424//===----------------------------------------------------------------------===//
2425// ExtractValueInst Class
2426//===----------------------------------------------------------------------===//
2427
2428/// This instruction extracts a struct member or array
2429/// element value from an aggregate value.
2430///
2431class ExtractValueInst : public UnaryInstruction {
2432 SmallVector<unsigned, 4> Indices;
2433
2434 ExtractValueInst(const ExtractValueInst &EVI);
2435
2436 /// Constructors - Create a extractvalue instruction with a base aggregate
2437 /// value and a list of indices. The first ctor can optionally insert before
2438 /// an existing instruction, the second appends the new instruction to the
2439 /// specified BasicBlock.
2440 inline ExtractValueInst(Value *Agg,
2441 ArrayRef<unsigned> Idxs,
2442 const Twine &NameStr,
2443 Instruction *InsertBefore);
2444 inline ExtractValueInst(Value *Agg,
2445 ArrayRef<unsigned> Idxs,
2446 const Twine &NameStr, BasicBlock *InsertAtEnd);
2447
2448 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2449
2450protected:
2451 // Note: Instruction needs to be a friend here to call cloneImpl.
2452 friend class Instruction;
2453
2454 ExtractValueInst *cloneImpl() const;
2455
2456public:
2457 static ExtractValueInst *Create(Value *Agg,
2458 ArrayRef<unsigned> Idxs,
2459 const Twine &NameStr = "",
2460 Instruction *InsertBefore = nullptr) {
2461 return new
2462 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2463 }
2464
2465 static ExtractValueInst *Create(Value *Agg,
2466 ArrayRef<unsigned> Idxs,
2467 const Twine &NameStr,
2468 BasicBlock *InsertAtEnd) {
2469 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2470 }
2471
2472 /// Returns the type of the element that would be extracted
2473 /// with an extractvalue instruction with the specified parameters.
2474 ///
2475 /// Null is returned if the indices are invalid for the specified type.
2476 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2477
2478 using idx_iterator = const unsigned*;
2479
2480 inline idx_iterator idx_begin() const { return Indices.begin(); }
2481 inline idx_iterator idx_end() const { return Indices.end(); }
2482 inline iterator_range<idx_iterator> indices() const {
2483 return make_range(idx_begin(), idx_end());
2484 }
2485
2486 Value *getAggregateOperand() {
2487 return getOperand(0);
2488 }
2489 const Value *getAggregateOperand() const {
2490 return getOperand(0);
2491 }
2492 static unsigned getAggregateOperandIndex() {
2493 return 0U; // get index for modifying correct operand
2494 }
2495
2496 ArrayRef<unsigned> getIndices() const {
2497 return Indices;
2498 }
2499
2500 unsigned getNumIndices() const {
2501 return (unsigned)Indices.size();
2502 }
2503
2504 bool hasIndices() const {
2505 return true;
2506 }
2507
2508 // Methods for support type inquiry through isa, cast, and dyn_cast:
2509 static bool classof(const Instruction *I) {
2510 return I->getOpcode() == Instruction::ExtractValue;
2511 }
2512 static bool classof(const Value *V) {
2513 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2514 }
2515};
2516
2517ExtractValueInst::ExtractValueInst(Value *Agg,
2518 ArrayRef<unsigned> Idxs,
2519 const Twine &NameStr,
2520 Instruction *InsertBefore)
2521 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2522 ExtractValue, Agg, InsertBefore) {
2523 init(Idxs, NameStr);
2524}
2525
2526ExtractValueInst::ExtractValueInst(Value *Agg,
2527 ArrayRef<unsigned> Idxs,
2528 const Twine &NameStr,
2529 BasicBlock *InsertAtEnd)
2530 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2531 ExtractValue, Agg, InsertAtEnd) {
2532 init(Idxs, NameStr);
2533}
2534
2535//===----------------------------------------------------------------------===//
2536// InsertValueInst Class
2537//===----------------------------------------------------------------------===//
2538
2539/// This instruction inserts a struct field of array element
2540/// value into an aggregate value.
2541///
2542class InsertValueInst : public Instruction {
2543 SmallVector<unsigned, 4> Indices;
2544
2545 InsertValueInst(const InsertValueInst &IVI);
2546
2547 /// Constructors - Create a insertvalue instruction with a base aggregate
2548 /// value, a value to insert, and a list of indices. The first ctor can
2549 /// optionally insert before an existing instruction, the second appends
2550 /// the new instruction to the specified BasicBlock.
2551 inline InsertValueInst(Value *Agg, Value *Val,
2552 ArrayRef<unsigned> Idxs,
2553 const Twine &NameStr,
2554 Instruction *InsertBefore);
2555 inline InsertValueInst(Value *Agg, Value *Val,
2556 ArrayRef<unsigned> Idxs,
2557 const Twine &NameStr, BasicBlock *InsertAtEnd);
2558
2559 /// Constructors - These two constructors are convenience methods because one
2560 /// and two index insertvalue instructions are so common.
2561 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2562 const Twine &NameStr = "",
2563 Instruction *InsertBefore = nullptr);
2564 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2565 BasicBlock *InsertAtEnd);
2566
2567 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2568 const Twine &NameStr);
2569
2570protected:
2571 // Note: Instruction needs to be a friend here to call cloneImpl.
2572 friend class Instruction;
2573
2574 InsertValueInst *cloneImpl() const;
2575
2576public:
2577 // allocate space for exactly two operands
2578 void *operator new(size_t S) { return User::operator new(S, 2); }
2579 void operator delete(void *Ptr) { User::operator delete(Ptr); }
2580
2581 static InsertValueInst *Create(Value *Agg, Value *Val,
2582 ArrayRef<unsigned> Idxs,
2583 const Twine &NameStr = "",
2584 Instruction *InsertBefore = nullptr) {
2585 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore);
2586 }
2587
2588 static InsertValueInst *Create(Value *Agg, Value *Val,
2589 ArrayRef<unsigned> Idxs,
2590 const Twine &NameStr,
2591 BasicBlock *InsertAtEnd) {
2592 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd);
2593 }
2594
2595 /// Transparently provide more efficient getOperand methods.
2596 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2597
2598 using idx_iterator = const unsigned*;
2599
2600 inline idx_iterator idx_begin() const { return Indices.begin(); }
2601 inline idx_iterator idx_end() const { return Indices.end(); }
2602 inline iterator_range<idx_iterator> indices() const {
2603 return make_range(idx_begin(), idx_end());
2604 }
2605
2606 Value *getAggregateOperand() {
2607 return getOperand(0);
2608 }
2609 const Value *getAggregateOperand() const {
2610 return getOperand(0);
2611 }
2612 static unsigned getAggregateOperandIndex() {
2613 return 0U; // get index for modifying correct operand
2614 }
2615
2616 Value *getInsertedValueOperand() {
2617 return getOperand(1);
2618 }
2619 const Value *getInsertedValueOperand() const {
2620 return getOperand(1);
2621 }
2622 static unsigned getInsertedValueOperandIndex() {
2623 return 1U; // get index for modifying correct operand
2624 }
2625
2626 ArrayRef<unsigned> getIndices() const {
2627 return Indices;
2628 }
2629
2630 unsigned getNumIndices() const {
2631 return (unsigned)Indices.size();
2632 }
2633
2634 bool hasIndices() const {
2635 return true;
2636 }
2637
2638 // Methods for support type inquiry through isa, cast, and dyn_cast:
2639 static bool classof(const Instruction *I) {
2640 return I->getOpcode() == Instruction::InsertValue;
2641 }
2642 static bool classof(const Value *V) {
2643 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2644 }
2645};
2646
2647template <>
2648struct OperandTraits<InsertValueInst> :
2649 public FixedNumOperandTraits<InsertValueInst, 2> {
2650};
2651
2652InsertValueInst::InsertValueInst(Value *Agg,
2653 Value *Val,
2654 ArrayRef<unsigned> Idxs,
2655 const Twine &NameStr,
2656 Instruction *InsertBefore)
2657 : Instruction(Agg->getType(), InsertValue,
2658 OperandTraits<InsertValueInst>::op_begin(this),
2659 2, InsertBefore) {
2660 init(Agg, Val, Idxs, NameStr);
2661}
2662
2663InsertValueInst::InsertValueInst(Value *Agg,
2664 Value *Val,
2665 ArrayRef<unsigned> Idxs,
2666 const Twine &NameStr,
2667 BasicBlock *InsertAtEnd)
2668 : Instruction(Agg->getType(), InsertValue,
2669 OperandTraits<InsertValueInst>::op_begin(this),
2670 2, InsertAtEnd) {
2671 init(Agg, Val, Idxs, NameStr);
2672}
2673
2674DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertValueInst, Value)InsertValueInst::op_iterator InsertValueInst::op_begin() { return
OperandTraits<InsertValueInst>::op_begin(this); } InsertValueInst
::const_op_iterator InsertValueInst::op_begin() const { return
OperandTraits<InsertValueInst>::op_begin(const_cast<
InsertValueInst*>(this)); } InsertValueInst::op_iterator InsertValueInst
::op_end() { return OperandTraits<InsertValueInst>::op_end
(this); } InsertValueInst::const_op_iterator InsertValueInst::
op_end() const { return OperandTraits<InsertValueInst>::
op_end(const_cast<InsertValueInst*>(this)); } Value *InsertValueInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<InsertValueInst>::
operands(this) && "getOperand() out of range!") ? void
(0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2674, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<InsertValueInst
>::op_begin(const_cast<InsertValueInst*>(this))[i_nocapture
].get()); } void InsertValueInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<InsertValueInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2674, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<InsertValueInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned InsertValueInst::getNumOperands
() const { return OperandTraits<InsertValueInst>::operands
(this); } template <int Idx_nocapture> Use &InsertValueInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &InsertValueInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
2675
2676//===----------------------------------------------------------------------===//
2677// PHINode Class
2678//===----------------------------------------------------------------------===//
2679
2680// PHINode - The PHINode class is used to represent the magical mystical PHI
2681// node, that can not exist in nature, but can be synthesized in a computer
2682// scientist's overactive imagination.
2683//
2684class PHINode : public Instruction {
2685 /// The number of operands actually allocated. NumOperands is
2686 /// the number actually in use.
2687 unsigned ReservedSpace;
2688
2689 PHINode(const PHINode &PN);
2690
2691 explicit PHINode(Type *Ty, unsigned NumReservedValues,
2692 const Twine &NameStr = "",
2693 Instruction *InsertBefore = nullptr)
2694 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore),
2695 ReservedSpace(NumReservedValues) {
2696 assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast <bool> (!Ty->isTokenTy() && "PHI nodes cannot have token type!"
) ? void (0) : __assert_fail ("!Ty->isTokenTy() && \"PHI nodes cannot have token type!\""
, "llvm/include/llvm/IR/Instructions.h", 2696, __extension__ __PRETTY_FUNCTION__
))
;
2697 setName(NameStr);
2698 allocHungoffUses(ReservedSpace);
2699 }
2700
2701 PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr,
2702 BasicBlock *InsertAtEnd)
2703 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd),
2704 ReservedSpace(NumReservedValues) {
2705 assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast <bool> (!Ty->isTokenTy() && "PHI nodes cannot have token type!"
) ? void (0) : __assert_fail ("!Ty->isTokenTy() && \"PHI nodes cannot have token type!\""
, "llvm/include/llvm/IR/Instructions.h", 2705, __extension__ __PRETTY_FUNCTION__
))
;
2706 setName(NameStr);
2707 allocHungoffUses(ReservedSpace);
2708 }
2709
2710protected:
2711 // Note: Instruction needs to be a friend here to call cloneImpl.
2712 friend class Instruction;
2713
2714 PHINode *cloneImpl() const;
2715
2716 // allocHungoffUses - this is more complicated than the generic
2717 // User::allocHungoffUses, because we have to allocate Uses for the incoming
2718 // values and pointers to the incoming blocks, all in one allocation.
2719 void allocHungoffUses(unsigned N) {
2720 User::allocHungoffUses(N, /* IsPhi */ true);
2721 }
2722
2723public:
2724 /// Constructors - NumReservedValues is a hint for the number of incoming
2725 /// edges that this phi node will have (use 0 if you really have no idea).
2726 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2727 const Twine &NameStr = "",
2728 Instruction *InsertBefore = nullptr) {
2729 return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore);
2730 }
2731
2732 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2733 const Twine &NameStr, BasicBlock *InsertAtEnd) {
2734 return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd);
2735 }
2736
2737 /// Provide fast operand accessors
2738 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2739
2740 // Block iterator interface. This provides access to the list of incoming
2741 // basic blocks, which parallels the list of incoming values.
2742
2743 using block_iterator = BasicBlock **;
2744 using const_block_iterator = BasicBlock * const *;
2745
2746 block_iterator block_begin() {
2747 return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace);
2748 }
2749
2750 const_block_iterator block_begin() const {
2751 return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace);
2752 }
2753
2754 block_iterator block_end() {
2755 return block_begin() + getNumOperands();
2756 }
2757
2758 const_block_iterator block_end() const {
2759 return block_begin() + getNumOperands();
2760 }
2761
2762 iterator_range<block_iterator> blocks() {
2763 return make_range(block_begin(), block_end());
2764 }
2765
2766 iterator_range<const_block_iterator> blocks() const {
2767 return make_range(block_begin(), block_end());
2768 }
2769
2770 op_range incoming_values() { return operands(); }
2771
2772 const_op_range incoming_values() const { return operands(); }
2773
2774 /// Return the number of incoming edges
2775 ///
2776 unsigned getNumIncomingValues() const { return getNumOperands(); }
2777
2778 /// Return incoming value number x
2779 ///
2780 Value *getIncomingValue(unsigned i) const {
2781 return getOperand(i);
2782 }
2783 void setIncomingValue(unsigned i, Value *V) {
2784 assert(V && "PHI node got a null value!")(static_cast <bool> (V && "PHI node got a null value!"
) ? void (0) : __assert_fail ("V && \"PHI node got a null value!\""
, "llvm/include/llvm/IR/Instructions.h", 2784, __extension__ __PRETTY_FUNCTION__
))
;
2785 assert(getType() == V->getType() &&(static_cast <bool> (getType() == V->getType() &&
"All operands to PHI node must be the same type as the PHI node!"
) ? void (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\""
, "llvm/include/llvm/IR/Instructions.h", 2786, __extension__ __PRETTY_FUNCTION__
))
2786 "All operands to PHI node must be the same type as the PHI node!")(static_cast <bool> (getType() == V->getType() &&
"All operands to PHI node must be the same type as the PHI node!"
) ? void (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\""
, "llvm/include/llvm/IR/Instructions.h", 2786, __extension__ __PRETTY_FUNCTION__
))
;
2787 setOperand(i, V);
2788 }
2789
2790 static unsigned getOperandNumForIncomingValue(unsigned i) {
2791 return i;
2792 }
2793
2794 static unsigned getIncomingValueNumForOperand(unsigned i) {
2795 return i;
2796 }
2797
2798 /// Return incoming basic block number @p i.
2799 ///
2800 BasicBlock *getIncomingBlock(unsigned i) const {
2801 return block_begin()[i];
2802 }
2803
2804 /// Return incoming basic block corresponding
2805 /// to an operand of the PHI.
2806 ///
2807 BasicBlock *getIncomingBlock(const Use &U) const {
2808 assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")(static_cast <bool> (this == U.getUser() && "Iterator doesn't point to PHI's Uses?"
) ? void (0) : __assert_fail ("this == U.getUser() && \"Iterator doesn't point to PHI's Uses?\""
, "llvm/include/llvm/IR/Instructions.h", 2808, __extension__ __PRETTY_FUNCTION__
))
;
2809 return getIncomingBlock(unsigned(&U - op_begin()));
2810 }
2811
2812 /// Return incoming basic block corresponding
2813 /// to value use iterator.
2814 ///
2815 BasicBlock *getIncomingBlock(Value::const_user_iterator I) const {
2816 return getIncomingBlock(I.getUse());
2817 }
2818
2819 void setIncomingBlock(unsigned i, BasicBlock *BB) {
2820 assert(BB && "PHI node got a null basic block!")(static_cast <bool> (BB && "PHI node got a null basic block!"
) ? void (0) : __assert_fail ("BB && \"PHI node got a null basic block!\""
, "llvm/include/llvm/IR/Instructions.h", 2820, __extension__ __PRETTY_FUNCTION__
))
;
2821 block_begin()[i] = BB;
2822 }
2823
2824 /// Replace every incoming basic block \p Old to basic block \p New.
2825 void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) {
2826 assert(New && Old && "PHI node got a null basic block!")(static_cast <bool> (New && Old && "PHI node got a null basic block!"
) ? void (0) : __assert_fail ("New && Old && \"PHI node got a null basic block!\""
, "llvm/include/llvm/IR/Instructions.h", 2826, __extension__ __PRETTY_FUNCTION__
))
;
2827 for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op)
2828 if (getIncomingBlock(Op) == Old)
2829 setIncomingBlock(Op, New);
2830 }
2831
2832 /// Add an incoming value to the end of the PHI list
2833 ///
2834 void addIncoming(Value *V, BasicBlock *BB) {
2835 if (getNumOperands() == ReservedSpace)
2836 growOperands(); // Get more space!
2837 // Initialize some new operands.
2838 setNumHungOffUseOperands(getNumOperands() + 1);
2839 setIncomingValue(getNumOperands() - 1, V);
2840 setIncomingBlock(getNumOperands() - 1, BB);
2841 }
2842
2843 /// Remove an incoming value. This is useful if a
2844 /// predecessor basic block is deleted. The value removed is returned.
2845 ///
2846 /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty
2847 /// is true), the PHI node is destroyed and any uses of it are replaced with
2848 /// dummy values. The only time there should be zero incoming values to a PHI
2849 /// node is when the block is dead, so this strategy is sound.
2850 ///
2851 Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true);
2852
2853 Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) {
2854 int Idx = getBasicBlockIndex(BB);
2855 assert(Idx >= 0 && "Invalid basic block argument to remove!")(static_cast <bool> (Idx >= 0 && "Invalid basic block argument to remove!"
) ? void (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument to remove!\""
, "llvm/include/llvm/IR/Instructions.h", 2855, __extension__ __PRETTY_FUNCTION__
))
;
2856 return removeIncomingValue(Idx, DeletePHIIfEmpty);
2857 }
2858
2859 /// Return the first index of the specified basic
2860 /// block in the value list for this PHI. Returns -1 if no instance.
2861 ///
2862 int getBasicBlockIndex(const BasicBlock *BB) const {
2863 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2864 if (block_begin()[i] == BB)
2865 return i;
2866 return -1;
2867 }
2868
2869 Value *getIncomingValueForBlock(const BasicBlock *BB) const {
2870 int Idx = getBasicBlockIndex(BB);
2871 assert(Idx >= 0 && "Invalid basic block argument!")(static_cast <bool> (Idx >= 0 && "Invalid basic block argument!"
) ? void (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument!\""
, "llvm/include/llvm/IR/Instructions.h", 2871, __extension__ __PRETTY_FUNCTION__
))
;
2872 return getIncomingValue(Idx);
2873 }
2874
2875 /// Set every incoming value(s) for block \p BB to \p V.
2876 void setIncomingValueForBlock(const BasicBlock *BB, Value *V) {
2877 assert(BB && "PHI node got a null basic block!")(static_cast <bool> (BB && "PHI node got a null basic block!"
) ? void (0) : __assert_fail ("BB && \"PHI node got a null basic block!\""
, "llvm/include/llvm/IR/Instructions.h", 2877, __extension__ __PRETTY_FUNCTION__
))
;
2878 bool Found = false;
2879 for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op)
2880 if (getIncomingBlock(Op) == BB) {
2881 Found = true;
2882 setIncomingValue(Op, V);
2883 }
2884 (void)Found;
2885 assert(Found && "Invalid basic block argument to set!")(static_cast <bool> (Found && "Invalid basic block argument to set!"
) ? void (0) : __assert_fail ("Found && \"Invalid basic block argument to set!\""
, "llvm/include/llvm/IR/Instructions.h", 2885, __extension__ __PRETTY_FUNCTION__
))
;
2886 }
2887
2888 /// If the specified PHI node always merges together the
2889 /// same value, return the value, otherwise return null.
2890 Value *hasConstantValue() const;
2891
2892 /// Whether the specified PHI node always merges
2893 /// together the same value, assuming undefs are equal to a unique
2894 /// non-undef value.
2895 bool hasConstantOrUndefValue() const;
2896
2897 /// If the PHI node is complete which means all of its parent's predecessors
2898 /// have incoming value in this PHI, return true, otherwise return false.
2899 bool isComplete() const {
2900 return llvm::all_of(predecessors(getParent()),
2901 [this](const BasicBlock *Pred) {
2902 return getBasicBlockIndex(Pred) >= 0;
2903 });
2904 }
2905
2906 /// Methods for support type inquiry through isa, cast, and dyn_cast:
2907 static bool classof(const Instruction *I) {
2908 return I->getOpcode() == Instruction::PHI;
2909 }
2910 static bool classof(const Value *V) {
2911 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2912 }
2913
2914private:
2915 void growOperands();
2916};
2917
2918template <>
2919struct OperandTraits<PHINode> : public HungoffOperandTraits<2> {
2920};
2921
2922DEFINE_TRANSPARENT_OPERAND_ACCESSORS(PHINode, Value)PHINode::op_iterator PHINode::op_begin() { return OperandTraits
<PHINode>::op_begin(this); } PHINode::const_op_iterator
PHINode::op_begin() const { return OperandTraits<PHINode>
::op_begin(const_cast<PHINode*>(this)); } PHINode::op_iterator
PHINode::op_end() { return OperandTraits<PHINode>::op_end
(this); } PHINode::const_op_iterator PHINode::op_end() const {
return OperandTraits<PHINode>::op_end(const_cast<PHINode
*>(this)); } Value *PHINode::getOperand(unsigned i_nocapture
) const { (static_cast <bool> (i_nocapture < OperandTraits
<PHINode>::operands(this) && "getOperand() out of range!"
) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2922, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<PHINode
>::op_begin(const_cast<PHINode*>(this))[i_nocapture]
.get()); } void PHINode::setOperand(unsigned i_nocapture, Value
*Val_nocapture) { (static_cast <bool> (i_nocapture <
OperandTraits<PHINode>::operands(this) && "setOperand() out of range!"
) ? void (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 2922, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<PHINode>::op_begin(this)[i_nocapture]
= Val_nocapture; } unsigned PHINode::getNumOperands() const {
return OperandTraits<PHINode>::operands(this); } template
<int Idx_nocapture> Use &PHINode::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &PHINode::Op() const { return this->OpFrom
<Idx_nocapture>(this); }
2923
2924//===----------------------------------------------------------------------===//
2925// LandingPadInst Class
2926//===----------------------------------------------------------------------===//
2927
2928//===---------------------------------------------------------------------------
2929/// The landingpad instruction holds all of the information
2930/// necessary to generate correct exception handling. The landingpad instruction
2931/// cannot be moved from the top of a landing pad block, which itself is
2932/// accessible only from the 'unwind' edge of an invoke. This uses the
2933/// SubclassData field in Value to store whether or not the landingpad is a
2934/// cleanup.
2935///
2936class LandingPadInst : public Instruction {
2937 using CleanupField = BoolBitfieldElementT<0>;
2938
2939 /// The number of operands actually allocated. NumOperands is
2940 /// the number actually in use.
2941 unsigned ReservedSpace;
2942
2943 LandingPadInst(const LandingPadInst &LP);
2944
2945public:
2946 enum ClauseType { Catch, Filter };
2947
2948private:
2949 explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues,
2950 const Twine &NameStr, Instruction *InsertBefore);
2951 explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues,
2952 const Twine &NameStr, BasicBlock *InsertAtEnd);
2953
2954 // Allocate space for exactly zero operands.
2955 void *operator new(size_t S) { return User::operator new(S); }
2956
2957 void growOperands(unsigned Size);
2958 void init(unsigned NumReservedValues, const Twine &NameStr);
2959
2960protected:
2961 // Note: Instruction needs to be a friend here to call cloneImpl.
2962 friend class Instruction;
2963
2964 LandingPadInst *cloneImpl() const;
2965
2966public:
2967 void operator delete(void *Ptr) { User::operator delete(Ptr); }
2968
2969 /// Constructors - NumReservedClauses is a hint for the number of incoming
2970 /// clauses that this landingpad will have (use 0 if you really have no idea).
2971 static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses,
2972 const Twine &NameStr = "",
2973 Instruction *InsertBefore = nullptr);
2974 static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses,
2975 const Twine &NameStr, BasicBlock *InsertAtEnd);
2976
2977 /// Provide fast operand accessors
2978 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2979
2980 /// Return 'true' if this landingpad instruction is a
2981 /// cleanup. I.e., it should be run when unwinding even if its landing pad
2982 /// doesn't catch the exception.
2983 bool isCleanup() const { return getSubclassData<CleanupField>(); }
2984
2985 /// Indicate that this landingpad instruction is a cleanup.
2986 void setCleanup(bool V) { setSubclassData<CleanupField>(V); }
2987
2988 /// Add a catch or filter clause to the landing pad.
2989 void addClause(Constant *ClauseVal);
2990
2991 /// Get the value of the clause at index Idx. Use isCatch/isFilter to
2992 /// determine what type of clause this is.
2993 Constant *getClause(unsigned Idx) const {
2994 return cast<Constant>(getOperandList()[Idx]);
2995 }
2996
2997 /// Return 'true' if the clause and index Idx is a catch clause.
2998 bool isCatch(unsigned Idx) const {
2999 return !isa<ArrayType>(getOperandList()[Idx]->getType());
3000 }
3001
3002 /// Return 'true' if the clause and index Idx is a filter clause.
3003 bool isFilter(unsigned Idx) const {
3004 return isa<ArrayType>(getOperandList()[Idx]->getType());
3005 }
3006
3007 /// Get the number of clauses for this landing pad.
3008 unsigned getNumClauses() const { return getNumOperands(); }
3009
3010 /// Grow the size of the operand list to accommodate the new
3011 /// number of clauses.
3012 void reserveClauses(unsigned Size) { growOperands(Size); }
3013
3014 // Methods for support type inquiry through isa, cast, and dyn_cast:
3015 static bool classof(const Instruction *I) {
3016 return I->getOpcode() == Instruction::LandingPad;
3017 }
3018 static bool classof(const Value *V) {
3019 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3020 }
3021};
3022
3023template <>
3024struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> {
3025};
3026
3027DEFINE_TRANSPARENT_OPERAND_ACCESSORS(LandingPadInst, Value)LandingPadInst::op_iterator LandingPadInst::op_begin() { return
OperandTraits<LandingPadInst>::op_begin(this); } LandingPadInst
::const_op_iterator LandingPadInst::op_begin() const { return
OperandTraits<LandingPadInst>::op_begin(const_cast<
LandingPadInst*>(this)); } LandingPadInst::op_iterator LandingPadInst
::op_end() { return OperandTraits<LandingPadInst>::op_end
(this); } LandingPadInst::const_op_iterator LandingPadInst::op_end
() const { return OperandTraits<LandingPadInst>::op_end
(const_cast<LandingPadInst*>(this)); } Value *LandingPadInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<LandingPadInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3027, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<LandingPadInst
>::op_begin(const_cast<LandingPadInst*>(this))[i_nocapture
].get()); } void LandingPadInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<LandingPadInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3027, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<LandingPadInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned LandingPadInst::getNumOperands(
) const { return OperandTraits<LandingPadInst>::operands
(this); } template <int Idx_nocapture> Use &LandingPadInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &LandingPadInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
3028
3029//===----------------------------------------------------------------------===//
3030// ReturnInst Class
3031//===----------------------------------------------------------------------===//
3032
3033//===---------------------------------------------------------------------------
3034/// Return a value (possibly void), from a function. Execution
3035/// does not continue in this function any longer.
3036///
3037class ReturnInst : public Instruction {
3038 ReturnInst(const ReturnInst &RI);
3039
3040private:
3041 // ReturnInst constructors:
3042 // ReturnInst() - 'ret void' instruction
3043 // ReturnInst( null) - 'ret void' instruction
3044 // ReturnInst(Value* X) - 'ret X' instruction
3045 // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I
3046 // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I
3047 // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B
3048 // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B
3049 //
3050 // NOTE: If the Value* passed is of type void then the constructor behaves as
3051 // if it was passed NULL.
3052 explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr,
3053 Instruction *InsertBefore = nullptr);
3054 ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd);
3055 explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd);
3056
3057protected:
3058 // Note: Instruction needs to be a friend here to call cloneImpl.
3059 friend class Instruction;
3060
3061 ReturnInst *cloneImpl() const;
3062
3063public:
3064 static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr,
3065 Instruction *InsertBefore = nullptr) {
3066 return new(!!retVal) ReturnInst(C, retVal, InsertBefore);
3067 }
3068
3069 static ReturnInst* Create(LLVMContext &C, Value *retVal,
3070 BasicBlock *InsertAtEnd) {
3071 return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd);
3072 }
3073
3074 static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) {
3075 return new(0) ReturnInst(C, InsertAtEnd);
3076 }
3077
3078 /// Provide fast operand accessors
3079 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
3080
3081 /// Convenience accessor. Returns null if there is no return value.
3082 Value *getReturnValue() const {
3083 return getNumOperands() != 0 ? getOperand(0) : nullptr;
4
Assuming the condition is false
5
'?' condition is false
6
Returning null pointer, which participates in a condition later
3084 }
3085
3086 unsigned getNumSuccessors() const { return 0; }
3087
3088 // Methods for support type inquiry through isa, cast, and dyn_cast:
3089 static bool classof(const Instruction *I) {
3090 return (I->getOpcode() == Instruction::Ret);
3091 }
3092 static bool classof(const Value *V) {
3093 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3094 }
3095
3096private:
3097 BasicBlock *getSuccessor(unsigned idx) const {
3098 llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 3098)
;
3099 }
3100
3101 void setSuccessor(unsigned idx, BasicBlock *B) {
3102 llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 3102)
;
3103 }
3104};
3105
3106template <>
3107struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> {
3108};
3109
3110DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ReturnInst, Value)ReturnInst::op_iterator ReturnInst::op_begin() { return OperandTraits
<ReturnInst>::op_begin(this); } ReturnInst::const_op_iterator
ReturnInst::op_begin() const { return OperandTraits<ReturnInst
>::op_begin(const_cast<ReturnInst*>(this)); } ReturnInst
::op_iterator ReturnInst::op_end() { return OperandTraits<
ReturnInst>::op_end(this); } ReturnInst::const_op_iterator
ReturnInst::op_end() const { return OperandTraits<ReturnInst
>::op_end(const_cast<ReturnInst*>(this)); } Value *ReturnInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<ReturnInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3110, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<ReturnInst
>::op_begin(const_cast<ReturnInst*>(this))[i_nocapture
].get()); } void ReturnInst::setOperand(unsigned i_nocapture,
Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<ReturnInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3110, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<ReturnInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned ReturnInst::getNumOperands() const
{ return OperandTraits<ReturnInst>::operands(this); } template
<int Idx_nocapture> Use &ReturnInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &ReturnInst::Op() const { return
this->OpFrom<Idx_nocapture>(this); }
3111
3112//===----------------------------------------------------------------------===//
3113// BranchInst Class
3114//===----------------------------------------------------------------------===//
3115
3116//===---------------------------------------------------------------------------
3117/// Conditional or Unconditional Branch instruction.
3118///
3119class BranchInst : public Instruction {
3120 /// Ops list - Branches are strange. The operands are ordered:
3121 /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because
3122 /// they don't have to check for cond/uncond branchness. These are mostly
3123 /// accessed relative from op_end().
3124 BranchInst(const BranchInst &BI);
3125 // BranchInst constructors (where {B, T, F} are blocks, and C is a condition):
3126 // BranchInst(BB *B) - 'br B'
3127 // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F'
3128 // BranchInst(BB* B, Inst *I) - 'br B' insert before I
3129 // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I
3130 // BranchInst(BB* B, BB *I) - 'br B' insert at end
3131 // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end
3132 explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr);
3133 BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond,
3134 Instruction *InsertBefore = nullptr);
3135 BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd);
3136 BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond,
3137 BasicBlock *InsertAtEnd);
3138
3139 void AssertOK();
3140
3141protected:
3142 // Note: Instruction needs to be a friend here to call cloneImpl.
3143 friend class Instruction;
3144
3145 BranchInst *cloneImpl() const;
3146
3147public:
3148 /// Iterator type that casts an operand to a basic block.
3149 ///
3150 /// This only makes sense because the successors are stored as adjacent
3151 /// operands for branch instructions.
3152 struct succ_op_iterator
3153 : iterator_adaptor_base<succ_op_iterator, value_op_iterator,
3154 std::random_access_iterator_tag, BasicBlock *,
3155 ptrdiff_t, BasicBlock *, BasicBlock *> {
3156 explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {}
3157
3158 BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3159 BasicBlock *operator->() const { return operator*(); }
3160 };
3161
3162 /// The const version of `succ_op_iterator`.
3163 struct const_succ_op_iterator
3164 : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator,
3165 std::random_access_iterator_tag,
3166 const BasicBlock *, ptrdiff_t, const BasicBlock *,
3167 const BasicBlock *> {
3168 explicit const_succ_op_iterator(const_value_op_iterator I)
3169 : iterator_adaptor_base(I) {}
3170
3171 const BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3172 const BasicBlock *operator->() const { return operator*(); }
3173 };
3174
3175 static BranchInst *Create(BasicBlock *IfTrue,
3176 Instruction *InsertBefore = nullptr) {
3177 return new(1) BranchInst(IfTrue, InsertBefore);
3178 }
3179
3180 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse,
3181 Value *Cond, Instruction *InsertBefore = nullptr) {
3182 return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore);
3183 }
3184
3185 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) {
3186 return new(1) BranchInst(IfTrue, InsertAtEnd);
3187 }
3188
3189 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse,
3190 Value *Cond, BasicBlock *InsertAtEnd) {
3191 return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd);
3192 }
3193
3194 /// Transparently provide more efficient getOperand methods.
3195 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
3196
3197 bool isUnconditional() const { return getNumOperands() == 1; }
3198 bool isConditional() const { return getNumOperands() == 3; }
3199
3200 Value *getCondition() const {
3201 assert(isConditional() && "Cannot get condition of an uncond branch!")(static_cast <bool> (isConditional() && "Cannot get condition of an uncond branch!"
) ? void (0) : __assert_fail ("isConditional() && \"Cannot get condition of an uncond branch!\""
, "llvm/include/llvm/IR/Instructions.h", 3201, __extension__ __PRETTY_FUNCTION__
))
;
3202 return Op<-3>();
3203 }
3204
3205 void setCondition(Value *V) {
3206 assert(isConditional() && "Cannot set condition of unconditional branch!")(static_cast <bool> (isConditional() && "Cannot set condition of unconditional branch!"
) ? void (0) : __assert_fail ("isConditional() && \"Cannot set condition of unconditional branch!\""
, "llvm/include/llvm/IR/Instructions.h", 3206, __extension__ __PRETTY_FUNCTION__
))
;
3207 Op<-3>() = V;
3208 }
3209
3210 unsigned getNumSuccessors() const { return 1+isConditional(); }
3211
3212 BasicBlock *getSuccessor(unsigned i) const {
3213 assert(i < getNumSuccessors() && "Successor # out of range for Branch!")(static_cast <bool> (i < getNumSuccessors() &&
"Successor # out of range for Branch!") ? void (0) : __assert_fail
("i < getNumSuccessors() && \"Successor # out of range for Branch!\""
, "llvm/include/llvm/IR/Instructions.h", 3213, __extension__ __PRETTY_FUNCTION__
))
;
3214 return cast_or_null<BasicBlock>((&Op<-1>() - i)->get());
3215 }
3216
3217 void setSuccessor(unsigned idx, BasicBlock *NewSucc) {
3218 assert(idx < getNumSuccessors() && "Successor # out of range for Branch!")(static_cast <bool> (idx < getNumSuccessors() &&
"Successor # out of range for Branch!") ? void (0) : __assert_fail
("idx < getNumSuccessors() && \"Successor # out of range for Branch!\""
, "llvm/include/llvm/IR/Instructions.h", 3218, __extension__ __PRETTY_FUNCTION__
))
;
3219 *(&Op<-1>() - idx) = NewSucc;
3220 }
3221
3222 /// Swap the successors of this branch instruction.
3223 ///
3224 /// Swaps the successors of the branch instruction. This also swaps any
3225 /// branch weight metadata associated with the instruction so that it
3226 /// continues to map correctly to each operand.
3227 void swapSuccessors();
3228
3229 iterator_range<succ_op_iterator> successors() {
3230 return make_range(
3231 succ_op_iterator(std::next(value_op_begin(), isConditional() ? 1 : 0)),
3232 succ_op_iterator(value_op_end()));
3233 }
3234
3235 iterator_range<const_succ_op_iterator> successors() const {
3236 return make_range(const_succ_op_iterator(
3237 std::next(value_op_begin(), isConditional() ? 1 : 0)),
3238 const_succ_op_iterator(value_op_end()));
3239 }
3240
3241 // Methods for support type inquiry through isa, cast, and dyn_cast:
3242 static bool classof(const Instruction *I) {
3243 return (I->getOpcode() == Instruction::Br);
3244 }
3245 static bool classof(const Value *V) {
3246 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3247 }
3248};
3249
3250template <>
3251struct OperandTraits<BranchInst> : public VariadicOperandTraits<BranchInst, 1> {
3252};
3253
3254DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BranchInst, Value)BranchInst::op_iterator BranchInst::op_begin() { return OperandTraits
<BranchInst>::op_begin(this); } BranchInst::const_op_iterator
BranchInst::op_begin() const { return OperandTraits<BranchInst
>::op_begin(const_cast<BranchInst*>(this)); } BranchInst
::op_iterator BranchInst::op_end() { return OperandTraits<
BranchInst>::op_end(this); } BranchInst::const_op_iterator
BranchInst::op_end() const { return OperandTraits<BranchInst
>::op_end(const_cast<BranchInst*>(this)); } Value *BranchInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<BranchInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<BranchInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3254, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<BranchInst
>::op_begin(const_cast<BranchInst*>(this))[i_nocapture
].get()); } void BranchInst::setOperand(unsigned i_nocapture,
Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<BranchInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<BranchInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3254, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<BranchInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned BranchInst::getNumOperands() const
{ return OperandTraits<BranchInst>::operands(this); } template
<int Idx_nocapture> Use &BranchInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &BranchInst::Op() const { return
this->OpFrom<Idx_nocapture>(this); }
3255
3256//===----------------------------------------------------------------------===//
3257// SwitchInst Class
3258//===----------------------------------------------------------------------===//
3259
3260//===---------------------------------------------------------------------------
3261/// Multiway switch
3262///
3263class SwitchInst : public Instruction {
3264 unsigned ReservedSpace;
3265
3266 // Operand[0] = Value to switch on
3267 // Operand[1] = Default basic block destination
3268 // Operand[2n ] = Value to match
3269 // Operand[2n+1] = BasicBlock to go to on match
3270 SwitchInst(const SwitchInst &SI);
3271
3272 /// Create a new switch instruction, specifying a value to switch on and a
3273 /// default destination. The number of additional cases can be specified here
3274 /// to make memory allocation more efficient. This constructor can also
3275 /// auto-insert before another instruction.
3276 SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases,
3277 Instruction *InsertBefore);
3278
3279 /// Create a new switch instruction, specifying a value to switch on and a
3280 /// default destination. The number of additional cases can be specified here
3281 /// to make memory allocation more efficient. This constructor also
3282 /// auto-inserts at the end of the specified BasicBlock.
3283 SwitchInst(Value *Value, BasicBlock *Default, unsigned NumCases,
3284 BasicBlock *InsertAtEnd);
3285
3286 // allocate space for exactly zero operands
3287 void *operator new(size_t S) { return User::operator new(S); }
3288
3289 void init(Value *Value, BasicBlock *Default, unsigned NumReserved);
3290 void growOperands();
3291
3292protected:
3293 // Note: Instruction needs to be a friend here to call cloneImpl.
3294 friend class Instruction;
3295
3296 SwitchInst *cloneImpl() const;
3297
3298public:
3299 void operator delete(void *Ptr) { User::operator delete(Ptr); }
3300
3301 // -2
3302 static const unsigned DefaultPseudoIndex = static_cast<unsigned>(~0L-1);
3303
3304 template <typename CaseHandleT> class CaseIteratorImpl;
3305
3306 /// A handle to a particular switch case. It exposes a convenient interface
3307 /// to both the case value and the successor block.
3308 ///
3309 /// We define this as a template and instantiate it to form both a const and
3310 /// non-const handle.
3311 template <typename SwitchInstT, typename ConstantIntT, typename BasicBlockT>
3312 class CaseHandleImpl {
3313 // Directly befriend both const and non-const iterators.
3314 friend class SwitchInst::CaseIteratorImpl<
3315 CaseHandleImpl<SwitchInstT, ConstantIntT, BasicBlockT>>;
3316
3317 protected:
3318 // Expose the switch type we're parameterized with to the iterator.
3319 using SwitchInstType = SwitchInstT;
3320
3321 SwitchInstT *SI;
3322 ptrdiff_t Index;
3323
3324 CaseHandleImpl() = default;
3325 CaseHandleImpl(SwitchInstT *SI, ptrdiff_t Index) : SI(SI), Index(Index) {}
3326
3327 public:
3328 /// Resolves case value for current case.
3329 ConstantIntT *getCaseValue() const {
3330 assert((unsigned)Index < SI->getNumCases() &&(static_cast <bool> ((unsigned)Index < SI->getNumCases
() && "Index out the number of cases.") ? void (0) : __assert_fail
("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3331, __extension__ __PRETTY_FUNCTION__
))
3331 "Index out the number of cases.")(static_cast <bool> ((unsigned)Index < SI->getNumCases
() && "Index out the number of cases.") ? void (0) : __assert_fail
("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3331, __extension__ __PRETTY_FUNCTION__
))
;
3332 return reinterpret_cast<ConstantIntT *>(SI->getOperand(2 + Index * 2));
3333 }
3334
3335 /// Resolves successor for current case.
3336 BasicBlockT *getCaseSuccessor() const {
3337 assert(((unsigned)Index < SI->getNumCases() ||(static_cast <bool> (((unsigned)Index < SI->getNumCases
() || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3339, __extension__ __PRETTY_FUNCTION__
))
3338 (unsigned)Index == DefaultPseudoIndex) &&(static_cast <bool> (((unsigned)Index < SI->getNumCases
() || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3339, __extension__ __PRETTY_FUNCTION__
))
3339 "Index out the number of cases.")(static_cast <bool> (((unsigned)Index < SI->getNumCases
() || (unsigned)Index == DefaultPseudoIndex) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index < SI->getNumCases() || (unsigned)Index == DefaultPseudoIndex) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3339, __extension__ __PRETTY_FUNCTION__
))
;
3340 return SI->getSuccessor(getSuccessorIndex());
3341 }
3342
3343 /// Returns number of current case.
3344 unsigned getCaseIndex() const { return Index; }
3345
3346 /// Returns successor index for current case successor.
3347 unsigned getSuccessorIndex() const {
3348 assert(((unsigned)Index == DefaultPseudoIndex ||(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex
|| (unsigned)Index < SI->getNumCases()) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3350, __extension__ __PRETTY_FUNCTION__
))
3349 (unsigned)Index < SI->getNumCases()) &&(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex
|| (unsigned)Index < SI->getNumCases()) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3350, __extension__ __PRETTY_FUNCTION__
))
3350 "Index out the number of cases.")(static_cast <bool> (((unsigned)Index == DefaultPseudoIndex
|| (unsigned)Index < SI->getNumCases()) && "Index out the number of cases."
) ? void (0) : __assert_fail ("((unsigned)Index == DefaultPseudoIndex || (unsigned)Index < SI->getNumCases()) && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3350, __extension__ __PRETTY_FUNCTION__
))
;
3351 return (unsigned)Index != DefaultPseudoIndex ? Index + 1 : 0;
3352 }
3353
3354 bool operator==(const CaseHandleImpl &RHS) const {
3355 assert(SI == RHS.SI && "Incompatible operators.")(static_cast <bool> (SI == RHS.SI && "Incompatible operators."
) ? void (0) : __assert_fail ("SI == RHS.SI && \"Incompatible operators.\""
, "llvm/include/llvm/IR/Instructions.h", 3355, __extension__ __PRETTY_FUNCTION__
))
;
3356 return Index == RHS.Index;
3357 }
3358 };
3359
3360 using ConstCaseHandle =
3361 CaseHandleImpl<const SwitchInst, const ConstantInt, const BasicBlock>;
3362
3363 class CaseHandle
3364 : public CaseHandleImpl<SwitchInst, ConstantInt, BasicBlock> {
3365 friend class SwitchInst::CaseIteratorImpl<CaseHandle>;
3366
3367 public:
3368 CaseHandle(SwitchInst *SI, ptrdiff_t Index) : CaseHandleImpl(SI, Index) {}
3369
3370 /// Sets the new value for current case.
3371 void setValue(ConstantInt *V) const {
3372 assert((unsigned)Index < SI->getNumCases() &&(static_cast <bool> ((unsigned)Index < SI->getNumCases
() && "Index out the number of cases.") ? void (0) : __assert_fail
("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3373, __extension__ __PRETTY_FUNCTION__
))
3373 "Index out the number of cases.")(static_cast <bool> ((unsigned)Index < SI->getNumCases
() && "Index out the number of cases.") ? void (0) : __assert_fail
("(unsigned)Index < SI->getNumCases() && \"Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3373, __extension__ __PRETTY_FUNCTION__
))
;
3374 SI->setOperand(2 + Index*2, reinterpret_cast<Value*>(V));
3375 }
3376
3377 /// Sets the new successor for current case.
3378 void setSuccessor(BasicBlock *S) const {
3379 SI->setSuccessor(getSuccessorIndex(), S);
3380 }
3381 };
3382
3383 template <typename CaseHandleT>
3384 class CaseIteratorImpl
3385 : public iterator_facade_base<CaseIteratorImpl<CaseHandleT>,
3386 std::random_access_iterator_tag,
3387 const CaseHandleT> {
3388 using SwitchInstT = typename CaseHandleT::SwitchInstType;
3389
3390 CaseHandleT Case;
3391
3392 public:
3393 /// Default constructed iterator is in an invalid state until assigned to
3394 /// a case for a particular switch.
3395 CaseIteratorImpl() = default;
3396
3397 /// Initializes case iterator for given SwitchInst and for given
3398 /// case number.
3399 CaseIteratorImpl(SwitchInstT *SI, unsigned CaseNum) : Case(SI, CaseNum) {}
3400
3401 /// Initializes case iterator for given SwitchInst and for given
3402 /// successor index.
3403 static CaseIteratorImpl fromSuccessorIndex(SwitchInstT *SI,
3404 unsigned SuccessorIndex) {
3405 assert(SuccessorIndex < SI->getNumSuccessors() &&(static_cast <bool> (SuccessorIndex < SI->getNumSuccessors
() && "Successor index # out of range!") ? void (0) :
__assert_fail ("SuccessorIndex < SI->getNumSuccessors() && \"Successor index # out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3406, __extension__ __PRETTY_FUNCTION__
))
3406 "Successor index # out of range!")(static_cast <bool> (SuccessorIndex < SI->getNumSuccessors
() && "Successor index # out of range!") ? void (0) :
__assert_fail ("SuccessorIndex < SI->getNumSuccessors() && \"Successor index # out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3406, __extension__ __PRETTY_FUNCTION__
))
;
3407 return SuccessorIndex != 0 ? CaseIteratorImpl(SI, SuccessorIndex - 1)
3408 : CaseIteratorImpl(SI, DefaultPseudoIndex);
3409 }
3410
3411 /// Support converting to the const variant. This will be a no-op for const
3412 /// variant.
3413 operator CaseIteratorImpl<ConstCaseHandle>() const {
3414 return CaseIteratorImpl<ConstCaseHandle>(Case.SI, Case.Index);
3415 }
3416
3417 CaseIteratorImpl &operator+=(ptrdiff_t N) {
3418 // Check index correctness after addition.
3419 // Note: Index == getNumCases() means end().
3420 assert(Case.Index + N >= 0 &&(static_cast <bool> (Case.Index + N >= 0 && (
unsigned)(Case.Index + N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3422, __extension__ __PRETTY_FUNCTION__
))
3421 (unsigned)(Case.Index + N) <= Case.SI->getNumCases() &&(static_cast <bool> (Case.Index + N >= 0 && (
unsigned)(Case.Index + N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3422, __extension__ __PRETTY_FUNCTION__
))
3422 "Case.Index out the number of cases.")(static_cast <bool> (Case.Index + N >= 0 && (
unsigned)(Case.Index + N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index + N >= 0 && (unsigned)(Case.Index + N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3422, __extension__ __PRETTY_FUNCTION__
))
;
3423 Case.Index += N;
3424 return *this;
3425 }
3426 CaseIteratorImpl &operator-=(ptrdiff_t N) {
3427 // Check index correctness after subtraction.
3428 // Note: Case.Index == getNumCases() means end().
3429 assert(Case.Index - N >= 0 &&(static_cast <bool> (Case.Index - N >= 0 && (
unsigned)(Case.Index - N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3431, __extension__ __PRETTY_FUNCTION__
))
3430 (unsigned)(Case.Index - N) <= Case.SI->getNumCases() &&(static_cast <bool> (Case.Index - N >= 0 && (
unsigned)(Case.Index - N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3431, __extension__ __PRETTY_FUNCTION__
))
3431 "Case.Index out the number of cases.")(static_cast <bool> (Case.Index - N >= 0 && (
unsigned)(Case.Index - N) <= Case.SI->getNumCases() &&
"Case.Index out the number of cases.") ? void (0) : __assert_fail
("Case.Index - N >= 0 && (unsigned)(Case.Index - N) <= Case.SI->getNumCases() && \"Case.Index out the number of cases.\""
, "llvm/include/llvm/IR/Instructions.h", 3431, __extension__ __PRETTY_FUNCTION__
))
;
3432 Case.Index -= N;
3433 return *this;
3434 }
3435 ptrdiff_t operator-(const CaseIteratorImpl &RHS) const {
3436 assert(Case.SI == RHS.Case.SI && "Incompatible operators.")(static_cast <bool> (Case.SI == RHS.Case.SI && "Incompatible operators."
) ? void (0) : __assert_fail ("Case.SI == RHS.Case.SI && \"Incompatible operators.\""
, "llvm/include/llvm/IR/Instructions.h", 3436, __extension__ __PRETTY_FUNCTION__
))
;
3437 return Case.Index - RHS.Case.Index;
3438 }
3439 bool operator==(const CaseIteratorImpl &RHS) const {
3440 return Case == RHS.Case;
3441 }
3442 bool operator<(const CaseIteratorImpl &RHS) const {
3443 assert(Case.SI == RHS.Case.SI && "Incompatible operators.")(static_cast <bool> (Case.SI == RHS.Case.SI && "Incompatible operators."
) ? void (0) : __assert_fail ("Case.SI == RHS.Case.SI && \"Incompatible operators.\""
, "llvm/include/llvm/IR/Instructions.h", 3443, __extension__ __PRETTY_FUNCTION__
))
;
3444 return Case.Index < RHS.Case.Index;
3445 }
3446 const CaseHandleT &operator*() const { return Case; }
3447 };
3448
3449 using CaseIt = CaseIteratorImpl<CaseHandle>;
3450 using ConstCaseIt = CaseIteratorImpl<ConstCaseHandle>;
3451
3452 static SwitchInst *Create(Value *Value, BasicBlock *Default,
3453 unsigned NumCases,
3454 Instruction *InsertBefore = nullptr) {
3455 return new SwitchInst(Value, Default, NumCases, InsertBefore);
3456 }
3457
3458 static SwitchInst *Create(Value *Value, BasicBlock *Default,
3459 unsigned NumCases, BasicBlock *InsertAtEnd) {
3460 return new SwitchInst(Value, Default, NumCases, InsertAtEnd);
3461 }
3462
3463 /// Provide fast operand accessors
3464 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
3465
3466 // Accessor Methods for Switch stmt
3467 Value *getCondition() const { return getOperand(0); }
3468 void setCondition(Value *V) { setOperand(0, V); }
3469
3470 BasicBlock *getDefaultDest() const {
3471 return cast<BasicBlock>(getOperand(1));
3472 }
3473
3474 void setDefaultDest(BasicBlock *DefaultCase) {
3475 setOperand(1, reinterpret_cast<Value*>(DefaultCase));
3476 }
3477
3478 /// Return the number of 'cases' in this switch instruction, excluding the
3479 /// default case.
3480 unsigned getNumCases() const {
3481 return getNumOperands()/2 - 1;
3482 }
3483
3484 /// Returns a read/write iterator that points to the first case in the
3485 /// SwitchInst.
3486 CaseIt case_begin() {
3487 return CaseIt(this, 0);
3488 }
3489
3490 /// Returns a read-only iterator that points to the first case in the
3491 /// SwitchInst.
3492 ConstCaseIt case_begin() const {
3493 return ConstCaseIt(this, 0);
3494 }
3495
3496 /// Returns a read/write iterator that points one past the last in the
3497 /// SwitchInst.
3498 CaseIt case_end() {
3499 return CaseIt(this, getNumCases());
3500 }
3501
3502 /// Returns a read-only iterator that points one past the last in the
3503 /// SwitchInst.
3504 ConstCaseIt case_end() const {
3505 return ConstCaseIt(this, getNumCases());
3506 }
3507
3508 /// Iteration adapter for range-for loops.
3509 iterator_range<CaseIt> cases() {
3510 return make_range(case_begin(), case_end());
3511 }
3512
3513 /// Constant iteration adapter for range-for loops.
3514 iterator_range<ConstCaseIt> cases() const {
3515 return make_range(case_begin(), case_end());
3516 }
3517
3518 /// Returns an iterator that points to the default case.
3519 /// Note: this iterator allows to resolve successor only. Attempt
3520 /// to resolve case value causes an assertion.
3521 /// Also note, that increment and decrement also causes an assertion and
3522 /// makes iterator invalid.
3523 CaseIt case_default() {
3524 return CaseIt(this, DefaultPseudoIndex);
3525 }
3526 ConstCaseIt case_default() const {
3527 return ConstCaseIt(this, DefaultPseudoIndex);
3528 }
3529
3530 /// Search all of the case values for the specified constant. If it is
3531 /// explicitly handled, return the case iterator of it, otherwise return
3532 /// default case iterator to indicate that it is handled by the default
3533 /// handler.
3534 CaseIt findCaseValue(const ConstantInt *C) {
3535 return CaseIt(
3536 this,
3537 const_cast<const SwitchInst *>(this)->findCaseValue(C)->getCaseIndex());
3538 }
3539 ConstCaseIt findCaseValue(const ConstantInt *C) const {
3540 ConstCaseIt I = llvm::find_if(cases(), [C](const ConstCaseHandle &Case) {
3541 return Case.getCaseValue() == C;
3542 });
3543 if (I != case_end())
3544 return I;
3545
3546 return case_default();
3547 }
3548
3549 /// Finds the unique case value for a given successor. Returns null if the
3550 /// successor is not found, not unique, or is the default case.
3551 ConstantInt *findCaseDest(BasicBlock *BB) {
3552 if (BB == getDefaultDest())
3553 return nullptr;
3554
3555 ConstantInt *CI = nullptr;
3556 for (auto Case : cases()) {
3557 if (Case.getCaseSuccessor() != BB)
3558 continue;
3559
3560 if (CI)
3561 return nullptr; // Multiple cases lead to BB.
3562
3563 CI = Case.getCaseValue();
3564 }
3565
3566 return CI;
3567 }
3568
3569 /// Add an entry to the switch instruction.
3570 /// Note:
3571 /// This action invalidates case_end(). Old case_end() iterator will
3572 /// point to the added case.
3573 void addCase(ConstantInt *OnVal, BasicBlock *Dest);
3574
3575 /// This method removes the specified case and its successor from the switch
3576 /// instruction. Note that this operation may reorder the remaining cases at
3577 /// index idx and above.
3578 /// Note:
3579 /// This action invalidates iterators for all cases following the one removed,
3580 /// including the case_end() iterator. It returns an iterator for the next
3581 /// case.
3582 CaseIt removeCase(CaseIt I);
3583
3584 unsigned getNumSuccessors() const { return getNumOperands()/2; }
3585 BasicBlock *getSuccessor(unsigned idx) const {
3586 assert(idx < getNumSuccessors() &&"Successor idx out of range for switch!")(static_cast <bool> (idx < getNumSuccessors() &&
"Successor idx out of range for switch!") ? void (0) : __assert_fail
("idx < getNumSuccessors() &&\"Successor idx out of range for switch!\""
, "llvm/include/llvm/IR/Instructions.h", 3586, __extension__ __PRETTY_FUNCTION__
))
;
3587 return cast<BasicBlock>(getOperand(idx*2+1));
3588 }
3589 void setSuccessor(unsigned idx, BasicBlock *NewSucc) {
3590 assert(idx < getNumSuccessors() && "Successor # out of range for switch!")(static_cast <bool> (idx < getNumSuccessors() &&
"Successor # out of range for switch!") ? void (0) : __assert_fail
("idx < getNumSuccessors() && \"Successor # out of range for switch!\""
, "llvm/include/llvm/IR/Instructions.h", 3590, __extension__ __PRETTY_FUNCTION__
))
;
3591 setOperand(idx * 2 + 1, NewSucc);
3592 }
3593
3594 // Methods for support type inquiry through isa, cast, and dyn_cast:
3595 static bool classof(const Instruction *I) {
3596 return I->getOpcode() == Instruction::Switch;
3597 }
3598 static bool classof(const Value *V) {
3599 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3600 }
3601};
3602
3603/// A wrapper class to simplify modification of SwitchInst cases along with
3604/// their prof branch_weights metadata.
3605class SwitchInstProfUpdateWrapper {
3606 SwitchInst &SI;
3607 Optional<SmallVector<uint32_t, 8> > Weights = None;
3608 bool Changed = false;
3609
3610protected:
3611 static MDNode *getProfBranchWeightsMD(const SwitchInst &SI);
3612
3613 MDNode *buildProfBranchWeightsMD();
3614
3615 void init();
3616
3617public:
3618 using CaseWeightOpt = Optional<uint32_t>;
3619 SwitchInst *operator->() { return &SI; }
3620 SwitchInst &operator*() { return SI; }
3621 operator SwitchInst *() { return &SI; }
3622
3623 SwitchInstProfUpdateWrapper(SwitchInst &SI) : SI(SI) { init(); }
3624
3625 ~SwitchInstProfUpdateWrapper() {
3626 if (Changed)
3627 SI.setMetadata(LLVMContext::MD_prof, buildProfBranchWeightsMD());
3628 }
3629
3630 /// Delegate the call to the underlying SwitchInst::removeCase() and remove
3631 /// correspondent branch weight.
3632 SwitchInst::CaseIt removeCase(SwitchInst::CaseIt I);
3633
3634 /// Delegate the call to the underlying SwitchInst::addCase() and set the
3635 /// specified branch weight for the added case.
3636 void addCase(ConstantInt *OnVal, BasicBlock *Dest, CaseWeightOpt W);
3637
3638 /// Delegate the call to the underlying SwitchInst::eraseFromParent() and mark
3639 /// this object to not touch the underlying SwitchInst in destructor.
3640 SymbolTableList<Instruction>::iterator eraseFromParent();
3641
3642 void setSuccessorWeight(unsigned idx, CaseWeightOpt W);
3643 CaseWeightOpt getSuccessorWeight(unsigned idx);
3644
3645 static CaseWeightOpt getSuccessorWeight(const SwitchInst &SI, unsigned idx);
3646};
3647
3648template <>
3649struct OperandTraits<SwitchInst> : public HungoffOperandTraits<2> {
3650};
3651
3652DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SwitchInst, Value)SwitchInst::op_iterator SwitchInst::op_begin() { return OperandTraits
<SwitchInst>::op_begin(this); } SwitchInst::const_op_iterator
SwitchInst::op_begin() const { return OperandTraits<SwitchInst
>::op_begin(const_cast<SwitchInst*>(this)); } SwitchInst
::op_iterator SwitchInst::op_end() { return OperandTraits<
SwitchInst>::op_end(this); } SwitchInst::const_op_iterator
SwitchInst::op_end() const { return OperandTraits<SwitchInst
>::op_end(const_cast<SwitchInst*>(this)); } Value *SwitchInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<SwitchInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<SwitchInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3652, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<SwitchInst
>::op_begin(const_cast<SwitchInst*>(this))[i_nocapture
].get()); } void SwitchInst::setOperand(unsigned i_nocapture,
Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<SwitchInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<SwitchInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3652, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<SwitchInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned SwitchInst::getNumOperands() const
{ return OperandTraits<SwitchInst>::operands(this); } template
<int Idx_nocapture> Use &SwitchInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &SwitchInst::Op() const { return
this->OpFrom<Idx_nocapture>(this); }
3653
3654//===----------------------------------------------------------------------===//
3655// IndirectBrInst Class
3656//===----------------------------------------------------------------------===//
3657
3658//===---------------------------------------------------------------------------
3659/// Indirect Branch Instruction.
3660///
3661class IndirectBrInst : public Instruction {
3662 unsigned ReservedSpace;
3663
3664 // Operand[0] = Address to jump to
3665 // Operand[n+1] = n-th destination
3666 IndirectBrInst(const IndirectBrInst &IBI);
3667
3668 /// Create a new indirectbr instruction, specifying an
3669 /// Address to jump to. The number of expected destinations can be specified
3670 /// here to make memory allocation more efficient. This constructor can also
3671 /// autoinsert before another instruction.
3672 IndirectBrInst(Value *Address, unsigned NumDests, Instruction *InsertBefore);
3673
3674 /// Create a new indirectbr instruction, specifying an
3675 /// Address to jump to. The number of expected destinations can be specified
3676 /// here to make memory allocation more efficient. This constructor also
3677 /// autoinserts at the end of the specified BasicBlock.
3678 IndirectBrInst(Value *Address, unsigned NumDests, BasicBlock *InsertAtEnd);
3679
3680 // allocate space for exactly zero operands
3681 void *operator new(size_t S) { return User::operator new(S); }
3682
3683 void init(Value *Address, unsigned NumDests);
3684 void growOperands();
3685
3686protected:
3687 // Note: Instruction needs to be a friend here to call cloneImpl.
3688 friend class Instruction;
3689
3690 IndirectBrInst *cloneImpl() const;
3691
3692public:
3693 void operator delete(void *Ptr) { User::operator delete(Ptr); }
3694
3695 /// Iterator type that casts an operand to a basic block.
3696 ///
3697 /// This only makes sense because the successors are stored as adjacent
3698 /// operands for indirectbr instructions.
3699 struct succ_op_iterator
3700 : iterator_adaptor_base<succ_op_iterator, value_op_iterator,
3701 std::random_access_iterator_tag, BasicBlock *,
3702 ptrdiff_t, BasicBlock *, BasicBlock *> {
3703 explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {}
3704
3705 BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3706 BasicBlock *operator->() const { return operator*(); }
3707 };
3708
3709 /// The const version of `succ_op_iterator`.
3710 struct const_succ_op_iterator
3711 : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator,
3712 std::random_access_iterator_tag,
3713 const BasicBlock *, ptrdiff_t, const BasicBlock *,
3714 const BasicBlock *> {
3715 explicit const_succ_op_iterator(const_value_op_iterator I)
3716 : iterator_adaptor_base(I) {}
3717
3718 const BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3719 const BasicBlock *operator->() const { return operator*(); }
3720 };
3721
3722 static IndirectBrInst *Create(Value *Address, unsigned NumDests,
3723 Instruction *InsertBefore = nullptr) {
3724 return new IndirectBrInst(Address, NumDests, InsertBefore);
3725 }
3726
3727 static IndirectBrInst *Create(Value *Address, unsigned NumDests,
3728 BasicBlock *InsertAtEnd) {
3729 return new IndirectBrInst(Address, NumDests, InsertAtEnd);
3730 }
3731
3732 /// Provide fast operand accessors.
3733 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
3734
3735 // Accessor Methods for IndirectBrInst instruction.
3736 Value *getAddress() { return getOperand(0); }
3737 const Value *getAddress() const { return getOperand(0); }
3738 void setAddress(Value *V) { setOperand(0, V); }
3739
3740 /// return the number of possible destinations in this
3741 /// indirectbr instruction.
3742 unsigned getNumDestinations() const { return getNumOperands()-1; }
3743
3744 /// Return the specified destination.
3745 BasicBlock *getDestination(unsigned i) { return getSuccessor(i); }
3746 const BasicBlock *getDestination(unsigned i) const { return getSuccessor(i); }
3747
3748 /// Add a destination.
3749 ///
3750 void addDestination(BasicBlock *Dest);
3751
3752 /// This method removes the specified successor from the
3753 /// indirectbr instruction.
3754 void removeDestination(unsigned i);
3755
3756 unsigned getNumSuccessors() const { return getNumOperands()-1; }
3757 BasicBlock *getSuccessor(unsigned i) const {
3758 return cast<BasicBlock>(getOperand(i+1));
3759 }
3760 void setSuccessor(unsigned i, BasicBlock *NewSucc) {
3761 setOperand(i + 1, NewSucc);
3762 }
3763
3764 iterator_range<succ_op_iterator> successors() {
3765 return make_range(succ_op_iterator(std::next(value_op_begin())),
3766 succ_op_iterator(value_op_end()));
3767 }
3768
3769 iterator_range<const_succ_op_iterator> successors() const {
3770 return make_range(const_succ_op_iterator(std::next(value_op_begin())),
3771 const_succ_op_iterator(value_op_end()));
3772 }
3773
3774 // Methods for support type inquiry through isa, cast, and dyn_cast:
3775 static bool classof(const Instruction *I) {
3776 return I->getOpcode() == Instruction::IndirectBr;
3777 }
3778 static bool classof(const Value *V) {
3779 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3780 }
3781};
3782
3783template <>
3784struct OperandTraits<IndirectBrInst> : public HungoffOperandTraits<1> {
3785};
3786
3787DEFINE_TRANSPARENT_OPERAND_ACCESSORS(IndirectBrInst, Value)IndirectBrInst::op_iterator IndirectBrInst::op_begin() { return
OperandTraits<IndirectBrInst>::op_begin(this); } IndirectBrInst
::const_op_iterator IndirectBrInst::op_begin() const { return
OperandTraits<IndirectBrInst>::op_begin(const_cast<
IndirectBrInst*>(this)); } IndirectBrInst::op_iterator IndirectBrInst
::op_end() { return OperandTraits<IndirectBrInst>::op_end
(this); } IndirectBrInst::const_op_iterator IndirectBrInst::op_end
() const { return OperandTraits<IndirectBrInst>::op_end
(const_cast<IndirectBrInst*>(this)); } Value *IndirectBrInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<IndirectBrInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<IndirectBrInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3787, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<IndirectBrInst
>::op_begin(const_cast<IndirectBrInst*>(this))[i_nocapture
].get()); } void IndirectBrInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<IndirectBrInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<IndirectBrInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 3787, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<IndirectBrInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned IndirectBrInst::getNumOperands(
) const { return OperandTraits<IndirectBrInst>::operands
(this); } template <int Idx_nocapture> Use &IndirectBrInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &IndirectBrInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
3788
3789//===----------------------------------------------------------------------===//
3790// InvokeInst Class
3791//===----------------------------------------------------------------------===//
3792
3793/// Invoke instruction. The SubclassData field is used to hold the
3794/// calling convention of the call.
3795///
3796class InvokeInst : public CallBase {
3797 /// The number of operands for this call beyond the called function,
3798 /// arguments, and operand bundles.
3799 static constexpr int NumExtraOperands = 2;
3800
3801 /// The index from the end of the operand array to the normal destination.
3802 static constexpr int NormalDestOpEndIdx = -3;
3803
3804 /// The index from the end of the operand array to the unwind destination.
3805 static constexpr int UnwindDestOpEndIdx = -2;
3806
3807 InvokeInst(const InvokeInst &BI);
3808
3809 /// Construct an InvokeInst given a range of arguments.
3810 ///
3811 /// Construct an InvokeInst from a range of arguments
3812 inline InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3813 BasicBlock *IfException, ArrayRef<Value *> Args,
3814 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
3815 const Twine &NameStr, Instruction *InsertBefore);
3816
3817 inline InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3818 BasicBlock *IfException, ArrayRef<Value *> Args,
3819 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
3820 const Twine &NameStr, BasicBlock *InsertAtEnd);
3821
3822 void init(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3823 BasicBlock *IfException, ArrayRef<Value *> Args,
3824 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
3825
3826 /// Compute the number of operands to allocate.
3827 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
3828 // We need one operand for the called function, plus our extra operands and
3829 // the input operand counts provided.
3830 return 1 + NumExtraOperands + NumArgs + NumBundleInputs;
3831 }
3832
3833protected:
3834 // Note: Instruction needs to be a friend here to call cloneImpl.
3835 friend class Instruction;
3836
3837 InvokeInst *cloneImpl() const;
3838
3839public:
3840 static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3841 BasicBlock *IfException, ArrayRef<Value *> Args,
3842 const Twine &NameStr,
3843 Instruction *InsertBefore = nullptr) {
3844 int NumOperands = ComputeNumOperands(Args.size());
3845 return new (NumOperands)
3846 InvokeInst(Ty, Func, IfNormal, IfException, Args, None, NumOperands,
3847 NameStr, InsertBefore);
3848 }
3849
3850 static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3851 BasicBlock *IfException, ArrayRef<Value *> Args,
3852 ArrayRef<OperandBundleDef> Bundles = None,
3853 const Twine &NameStr = "",
3854 Instruction *InsertBefore = nullptr) {
3855 int NumOperands =
3856 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
3857 unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
3858
3859 return new (NumOperands, DescriptorBytes)
3860 InvokeInst(Ty, Func, IfNormal, IfException, Args, Bundles, NumOperands,
3861 NameStr, InsertBefore);
3862 }
3863
3864 static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3865 BasicBlock *IfException, ArrayRef<Value *> Args,
3866 const Twine &NameStr, BasicBlock *InsertAtEnd) {
3867 int NumOperands = ComputeNumOperands(Args.size());
3868 return new (NumOperands)
3869 InvokeInst(Ty, Func, IfNormal, IfException, Args, None, NumOperands,
3870 NameStr, InsertAtEnd);
3871 }
3872
3873 static InvokeInst *Create(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3874 BasicBlock *IfException, ArrayRef<Value *> Args,
3875 ArrayRef<OperandBundleDef> Bundles,
3876 const Twine &NameStr, BasicBlock *InsertAtEnd) {
3877 int NumOperands =
3878 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
3879 unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
3880
3881 return new (NumOperands, DescriptorBytes)
3882 InvokeInst(Ty, Func, IfNormal, IfException, Args, Bundles, NumOperands,
3883 NameStr, InsertAtEnd);
3884 }
3885
3886 static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal,
3887 BasicBlock *IfException, ArrayRef<Value *> Args,
3888 const Twine &NameStr,
3889 Instruction *InsertBefore = nullptr) {
3890 return Create(Func.getFunctionType(), Func.getCallee(), IfNormal,
3891 IfException, Args, None, NameStr, InsertBefore);
3892 }
3893
3894 static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal,
3895 BasicBlock *IfException, ArrayRef<Value *> Args,
3896 ArrayRef<OperandBundleDef> Bundles = None,
3897 const Twine &NameStr = "",
3898 Instruction *InsertBefore = nullptr) {
3899 return Create(Func.getFunctionType(), Func.getCallee(), IfNormal,
3900 IfException, Args, Bundles, NameStr, InsertBefore);
3901 }
3902
3903 static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal,
3904 BasicBlock *IfException, ArrayRef<Value *> Args,
3905 const Twine &NameStr, BasicBlock *InsertAtEnd) {
3906 return Create(Func.getFunctionType(), Func.getCallee(), IfNormal,
3907 IfException, Args, NameStr, InsertAtEnd);
3908 }
3909
3910 static InvokeInst *Create(FunctionCallee Func, BasicBlock *IfNormal,
3911 BasicBlock *IfException, ArrayRef<Value *> Args,
3912 ArrayRef<OperandBundleDef> Bundles,
3913 const Twine &NameStr, BasicBlock *InsertAtEnd) {
3914 return Create(Func.getFunctionType(), Func.getCallee(), IfNormal,
3915 IfException, Args, Bundles, NameStr, InsertAtEnd);
3916 }
3917
3918 /// Create a clone of \p II with a different set of operand bundles and
3919 /// insert it before \p InsertPt.
3920 ///
3921 /// The returned invoke instruction is identical to \p II in every way except
3922 /// that the operand bundles for the new instruction are set to the operand
3923 /// bundles in \p Bundles.
3924 static InvokeInst *Create(InvokeInst *II, ArrayRef<OperandBundleDef> Bundles,
3925 Instruction *InsertPt = nullptr);
3926
3927 // get*Dest - Return the destination basic blocks...
3928 BasicBlock *getNormalDest() const {
3929 return cast<BasicBlock>(Op<NormalDestOpEndIdx>());
3930 }
3931 BasicBlock *getUnwindDest() const {
3932 return cast<BasicBlock>(Op<UnwindDestOpEndIdx>());
3933 }
3934 void setNormalDest(BasicBlock *B) {
3935 Op<NormalDestOpEndIdx>() = reinterpret_cast<Value *>(B);
3936 }
3937 void setUnwindDest(BasicBlock *B) {
3938 Op<UnwindDestOpEndIdx>() = reinterpret_cast<Value *>(B);
3939 }
3940
3941 /// Get the landingpad instruction from the landing pad
3942 /// block (the unwind destination).
3943 LandingPadInst *getLandingPadInst() const;
3944
3945 BasicBlock *getSuccessor(unsigned i) const {
3946 assert(i < 2 && "Successor # out of range for invoke!")(static_cast <bool> (i < 2 && "Successor # out of range for invoke!"
) ? void (0) : __assert_fail ("i < 2 && \"Successor # out of range for invoke!\""
, "llvm/include/llvm/IR/Instructions.h", 3946, __extension__ __PRETTY_FUNCTION__
))
;
3947 return i == 0 ? getNormalDest() : getUnwindDest();
3948 }
3949
3950 void setSuccessor(unsigned i, BasicBlock *NewSucc) {
3951 assert(i < 2 && "Successor # out of range for invoke!")(static_cast <bool> (i < 2 && "Successor # out of range for invoke!"
) ? void (0) : __assert_fail ("i < 2 && \"Successor # out of range for invoke!\""
, "llvm/include/llvm/IR/Instructions.h", 3951, __extension__ __PRETTY_FUNCTION__
))
;
3952 if (i == 0)
3953 setNormalDest(NewSucc);
3954 else
3955 setUnwindDest(NewSucc);
3956 }
3957
3958 unsigned getNumSuccessors() const { return 2; }
3959
3960 // Methods for support type inquiry through isa, cast, and dyn_cast:
3961 static bool classof(const Instruction *I) {
3962 return (I->getOpcode() == Instruction::Invoke);
3963 }
3964 static bool classof(const Value *V) {
3965 return isa<Instruction>(V) && classof(cast<Instruction>(V));
3966 }
3967
3968private:
3969 // Shadow Instruction::setInstructionSubclassData with a private forwarding
3970 // method so that subclasses cannot accidentally use it.
3971 template <typename Bitfield>
3972 void setSubclassData(typename Bitfield::Type Value) {
3973 Instruction::setSubclassData<Bitfield>(Value);
3974 }
3975};
3976
3977InvokeInst::InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3978 BasicBlock *IfException, ArrayRef<Value *> Args,
3979 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
3980 const Twine &NameStr, Instruction *InsertBefore)
3981 : CallBase(Ty->getReturnType(), Instruction::Invoke,
3982 OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands,
3983 InsertBefore) {
3984 init(Ty, Func, IfNormal, IfException, Args, Bundles, NameStr);
3985}
3986
3987InvokeInst::InvokeInst(FunctionType *Ty, Value *Func, BasicBlock *IfNormal,
3988 BasicBlock *IfException, ArrayRef<Value *> Args,
3989 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
3990 const Twine &NameStr, BasicBlock *InsertAtEnd)
3991 : CallBase(Ty->getReturnType(), Instruction::Invoke,
3992 OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands,
3993 InsertAtEnd) {
3994 init(Ty, Func, IfNormal, IfException, Args, Bundles, NameStr);
3995}
3996
3997//===----------------------------------------------------------------------===//
3998// CallBrInst Class
3999//===----------------------------------------------------------------------===//
4000
4001/// CallBr instruction, tracking function calls that may not return control but
4002/// instead transfer it to a third location. The SubclassData field is used to
4003/// hold the calling convention of the call.
4004///
4005class CallBrInst : public CallBase {
4006
4007 unsigned NumIndirectDests;
4008
4009 CallBrInst(const CallBrInst &BI);
4010
4011 /// Construct a CallBrInst given a range of arguments.
4012 ///
4013 /// Construct a CallBrInst from a range of arguments
4014 inline CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest,
4015 ArrayRef<BasicBlock *> IndirectDests,
4016 ArrayRef<Value *> Args,
4017 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
4018 const Twine &NameStr, Instruction *InsertBefore);
4019
4020 inline CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest,
4021 ArrayRef<BasicBlock *> IndirectDests,
4022 ArrayRef<Value *> Args,
4023 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
4024 const Twine &NameStr, BasicBlock *InsertAtEnd);
4025
4026 void init(FunctionType *FTy, Value *Func, BasicBlock *DefaultDest,
4027 ArrayRef<BasicBlock *> IndirectDests, ArrayRef<Value *> Args,
4028 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
4029
4030 /// Compute the number of operands to allocate.
4031 static int ComputeNumOperands(int NumArgs, int NumIndirectDests,
4032 int NumBundleInputs = 0) {
4033 // We need one operand for the called function, plus our extra operands and
4034 // the input operand counts provided.
4035 return 2 + NumIndirectDests + NumArgs + NumBundleInputs;
4036 }
4037
4038protected:
4039 // Note: Instruction needs to be a friend here to call cloneImpl.
4040 friend class Instruction;
4041
4042 CallBrInst *cloneImpl() const;
4043
4044public:
4045 static CallBrInst *Create(FunctionType *Ty, Value *Func,
4046 BasicBlock *DefaultDest,
4047 ArrayRef<BasicBlock *> IndirectDests,
4048 ArrayRef<Value *> Args, const Twine &NameStr,
4049 Instruction *InsertBefore = nullptr) {
4050 int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size());
4051 return new (NumOperands)
4052 CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, None,
4053 NumOperands, NameStr, InsertBefore);
4054 }
4055
4056 static CallBrInst *Create(FunctionType *Ty, Value *Func,
4057 BasicBlock *DefaultDest,
4058 ArrayRef<BasicBlock *> IndirectDests,
4059 ArrayRef<Value *> Args,
4060 ArrayRef<OperandBundleDef> Bundles = None,
4061 const Twine &NameStr = "",
4062 Instruction *InsertBefore = nullptr) {
4063 int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size(),
4064 CountBundleInputs(Bundles));
4065 unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
4066
4067 return new (NumOperands, DescriptorBytes)
4068 CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, Bundles,
4069 NumOperands, NameStr, InsertBefore);
4070 }
4071
4072 static CallBrInst *Create(FunctionType *Ty, Value *Func,
4073 BasicBlock *DefaultDest,
4074 ArrayRef<BasicBlock *> IndirectDests,
4075 ArrayRef<Value *> Args, const Twine &NameStr,
4076 BasicBlock *InsertAtEnd) {
4077 int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size());
4078 return new (NumOperands)
4079 CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, None,
4080 NumOperands, NameStr, InsertAtEnd);
4081 }
4082
4083 static CallBrInst *Create(FunctionType *Ty, Value *Func,
4084 BasicBlock *DefaultDest,
4085 ArrayRef<BasicBlock *> IndirectDests,
4086 ArrayRef<Value *> Args,
4087 ArrayRef<OperandBundleDef> Bundles,
4088 const Twine &NameStr, BasicBlock *InsertAtEnd) {
4089 int NumOperands = ComputeNumOperands(Args.size(), IndirectDests.size(),
4090 CountBundleInputs(Bundles));
4091 unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
4092
4093 return new (NumOperands, DescriptorBytes)
4094 CallBrInst(Ty, Func, DefaultDest, IndirectDests, Args, Bundles,
4095 NumOperands, NameStr, InsertAtEnd);
4096 }
4097
4098 static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest,
4099 ArrayRef<BasicBlock *> IndirectDests,
4100 ArrayRef<Value *> Args, const Twine &NameStr,
4101 Instruction *InsertBefore = nullptr) {
4102 return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest,
4103 IndirectDests, Args, NameStr, InsertBefore);
4104 }
4105
4106 static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest,
4107 ArrayRef<BasicBlock *> IndirectDests,
4108 ArrayRef<Value *> Args,
4109 ArrayRef<OperandBundleDef> Bundles = None,
4110 const Twine &NameStr = "",
4111 Instruction *InsertBefore = nullptr) {
4112 return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest,
4113 IndirectDests, Args, Bundles, NameStr, InsertBefore);
4114 }
4115
4116 static CallBrInst *Create(FunctionCallee Func, BasicBlock *DefaultDest,
4117 ArrayRef<BasicBlock *> IndirectDests,
4118 ArrayRef<Value *> Args, const Twine &NameStr,
4119 BasicBlock *InsertAtEnd) {
4120 return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest,
4121 IndirectDests, Args, NameStr, InsertAtEnd);
4122 }
4123
4124 static CallBrInst *Create(FunctionCallee Func,
4125 BasicBlock *DefaultDest,
4126 ArrayRef<BasicBlock *> IndirectDests,
4127 ArrayRef<Value *> Args,
4128 ArrayRef<OperandBundleDef> Bundles,
4129 const Twine &NameStr, BasicBlock *InsertAtEnd) {
4130 return Create(Func.getFunctionType(), Func.getCallee(), DefaultDest,
4131 IndirectDests, Args, Bundles, NameStr, InsertAtEnd);
4132 }
4133
4134 /// Create a clone of \p CBI with a different set of operand bundles and
4135 /// insert it before \p InsertPt.
4136 ///
4137 /// The returned callbr instruction is identical to \p CBI in every way
4138 /// except that the operand bundles for the new instruction are set to the
4139 /// operand bundles in \p Bundles.
4140 static CallBrInst *Create(CallBrInst *CBI,
4141 ArrayRef<OperandBundleDef> Bundles,
4142 Instruction *InsertPt = nullptr);
4143
4144 /// Return the number of callbr indirect dest labels.
4145 ///
4146 unsigned getNumIndirectDests() const { return NumIndirectDests; }
4147
4148 /// getIndirectDestLabel - Return the i-th indirect dest label.
4149 ///
4150 Value *getIndirectDestLabel(unsigned i) const {
4151 assert(i < getNumIndirectDests() && "Out of bounds!")(static_cast <bool> (i < getNumIndirectDests() &&
"Out of bounds!") ? void (0) : __assert_fail ("i < getNumIndirectDests() && \"Out of bounds!\""
, "llvm/include/llvm/IR/Instructions.h", 4151, __extension__ __PRETTY_FUNCTION__
))
;
4152 return getOperand(i + arg_size() + getNumTotalBundleOperands() + 1);
4153 }
4154
4155 Value *getIndirectDestLabelUse(unsigned i) const {
4156 assert(i < getNumIndirectDests() && "Out of bounds!")(static_cast <bool> (i < getNumIndirectDests() &&
"Out of bounds!") ? void (0) : __assert_fail ("i < getNumIndirectDests() && \"Out of bounds!\""
, "llvm/include/llvm/IR/Instructions.h", 4156, __extension__ __PRETTY_FUNCTION__
))
;
4157 return getOperandUse(i + arg_size() + getNumTotalBundleOperands() + 1);
4158 }
4159
4160 // Return the destination basic blocks...
4161 BasicBlock *getDefaultDest() const {
4162 return cast<BasicBlock>(*(&Op<-1>() - getNumIndirectDests() - 1));
4163 }
4164 BasicBlock *getIndirectDest(unsigned i) const {
4165 return cast_or_null<BasicBlock>(*(&Op<-1>() - getNumIndirectDests() + i));
4166 }
4167 SmallVector<BasicBlock *, 16> getIndirectDests() const {
4168 SmallVector<BasicBlock *, 16> IndirectDests;
4169 for (unsigned i = 0, e = getNumIndirectDests(); i < e; ++i)
4170 IndirectDests.push_back(getIndirectDest(i));
4171 return IndirectDests;
4172 }
4173 void setDefaultDest(BasicBlock *B) {
4174 *(&Op<-1>() - getNumIndirectDests() - 1) = reinterpret_cast<Value *>(B);
4175 }
4176 void setIndirectDest(unsigned i, BasicBlock *B) {
4177 *(&Op<-1>() - getNumIndirectDests() + i) = reinterpret_cast<Value *>(B);
4178 }
4179
4180 BasicBlock *getSuccessor(unsigned i) const {
4181 assert(i < getNumSuccessors() + 1 &&(static_cast <bool> (i < getNumSuccessors() + 1 &&
"Successor # out of range for callbr!") ? void (0) : __assert_fail
("i < getNumSuccessors() + 1 && \"Successor # out of range for callbr!\""
, "llvm/include/llvm/IR/Instructions.h", 4182, __extension__ __PRETTY_FUNCTION__
))
4182 "Successor # out of range for callbr!")(static_cast <bool> (i < getNumSuccessors() + 1 &&
"Successor # out of range for callbr!") ? void (0) : __assert_fail
("i < getNumSuccessors() + 1 && \"Successor # out of range for callbr!\""
, "llvm/include/llvm/IR/Instructions.h", 4182, __extension__ __PRETTY_FUNCTION__
))
;
4183 return i == 0 ? getDefaultDest() : getIndirectDest(i - 1);
4184 }
4185
4186 void setSuccessor(unsigned i, BasicBlock *NewSucc) {
4187 assert(i < getNumIndirectDests() + 1 &&(static_cast <bool> (i < getNumIndirectDests() + 1 &&
"Successor # out of range for callbr!") ? void (0) : __assert_fail
("i < getNumIndirectDests() + 1 && \"Successor # out of range for callbr!\""
, "llvm/include/llvm/IR/Instructions.h", 4188, __extension__ __PRETTY_FUNCTION__
))
4188 "Successor # out of range for callbr!")(static_cast <bool> (i < getNumIndirectDests() + 1 &&
"Successor # out of range for callbr!") ? void (0) : __assert_fail
("i < getNumIndirectDests() + 1 && \"Successor # out of range for callbr!\""
, "llvm/include/llvm/IR/Instructions.h", 4188, __extension__ __PRETTY_FUNCTION__
))
;
4189 return i == 0 ? setDefaultDest(NewSucc) : setIndirectDest(i - 1, NewSucc);
4190 }
4191
4192 unsigned getNumSuccessors() const { return getNumIndirectDests() + 1; }
4193
4194 // Methods for support type inquiry through isa, cast, and dyn_cast:
4195 static bool classof(const Instruction *I) {
4196 return (I->getOpcode() == Instruction::CallBr);
4197 }
4198 static bool classof(const Value *V) {
4199 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4200 }
4201
4202private:
4203 // Shadow Instruction::setInstructionSubclassData with a private forwarding
4204 // method so that subclasses cannot accidentally use it.
4205 template <typename Bitfield>
4206 void setSubclassData(typename Bitfield::Type Value) {
4207 Instruction::setSubclassData<Bitfield>(Value);
4208 }
4209};
4210
4211CallBrInst::CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest,
4212 ArrayRef<BasicBlock *> IndirectDests,
4213 ArrayRef<Value *> Args,
4214 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
4215 const Twine &NameStr, Instruction *InsertBefore)
4216 : CallBase(Ty->getReturnType(), Instruction::CallBr,
4217 OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands,
4218 InsertBefore) {
4219 init(Ty, Func, DefaultDest, IndirectDests, Args, Bundles, NameStr);
4220}
4221
4222CallBrInst::CallBrInst(FunctionType *Ty, Value *Func, BasicBlock *DefaultDest,
4223 ArrayRef<BasicBlock *> IndirectDests,
4224 ArrayRef<Value *> Args,
4225 ArrayRef<OperandBundleDef> Bundles, int NumOperands,
4226 const Twine &NameStr, BasicBlock *InsertAtEnd)
4227 : CallBase(Ty->getReturnType(), Instruction::CallBr,
4228 OperandTraits<CallBase>::op_end(this) - NumOperands, NumOperands,
4229 InsertAtEnd) {
4230 init(Ty, Func, DefaultDest, IndirectDests, Args, Bundles, NameStr);
4231}
4232
4233//===----------------------------------------------------------------------===//
4234// ResumeInst Class
4235//===----------------------------------------------------------------------===//
4236
4237//===---------------------------------------------------------------------------
4238/// Resume the propagation of an exception.
4239///
4240class ResumeInst : public Instruction {
4241 ResumeInst(const ResumeInst &RI);
4242
4243 explicit ResumeInst(Value *Exn, Instruction *InsertBefore=nullptr);
4244 ResumeInst(Value *Exn, BasicBlock *InsertAtEnd);
4245
4246protected:
4247 // Note: Instruction needs to be a friend here to call cloneImpl.
4248 friend class Instruction;
4249
4250 ResumeInst *cloneImpl() const;
4251
4252public:
4253 static ResumeInst *Create(Value *Exn, Instruction *InsertBefore = nullptr) {
4254 return new(1) ResumeInst(Exn, InsertBefore);
4255 }
4256
4257 static ResumeInst *Create(Value *Exn, BasicBlock *InsertAtEnd) {
4258 return new(1) ResumeInst(Exn, InsertAtEnd);
4259 }
4260
4261 /// Provide fast operand accessors
4262 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
4263
4264 /// Convenience accessor.
4265 Value *getValue() const { return Op<0>(); }
4266
4267 unsigned getNumSuccessors() const { return 0; }
4268
4269 // Methods for support type inquiry through isa, cast, and dyn_cast:
4270 static bool classof(const Instruction *I) {
4271 return I->getOpcode() == Instruction::Resume;
4272 }
4273 static bool classof(const Value *V) {
4274 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4275 }
4276
4277private:
4278 BasicBlock *getSuccessor(unsigned idx) const {
4279 llvm_unreachable("ResumeInst has no successors!")::llvm::llvm_unreachable_internal("ResumeInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 4279)
;
4280 }
4281
4282 void setSuccessor(unsigned idx, BasicBlock *NewSucc) {
4283 llvm_unreachable("ResumeInst has no successors!")::llvm::llvm_unreachable_internal("ResumeInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 4283)
;
4284 }
4285};
4286
4287template <>
4288struct OperandTraits<ResumeInst> :
4289 public FixedNumOperandTraits<ResumeInst, 1> {
4290};
4291
4292DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ResumeInst, Value)ResumeInst::op_iterator ResumeInst::op_begin() { return OperandTraits
<ResumeInst>::op_begin(this); } ResumeInst::const_op_iterator
ResumeInst::op_begin() const { return OperandTraits<ResumeInst
>::op_begin(const_cast<ResumeInst*>(this)); } ResumeInst
::op_iterator ResumeInst::op_end() { return OperandTraits<
ResumeInst>::op_end(this); } ResumeInst::const_op_iterator
ResumeInst::op_end() const { return OperandTraits<ResumeInst
>::op_end(const_cast<ResumeInst*>(this)); } Value *ResumeInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<ResumeInst>::operands
(this) && "getOperand() out of range!") ? void (0) : __assert_fail
("i_nocapture < OperandTraits<ResumeInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4292, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<ResumeInst
>::op_begin(const_cast<ResumeInst*>(this))[i_nocapture
].get()); } void ResumeInst::setOperand(unsigned i_nocapture,
Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<ResumeInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<ResumeInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4292, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<ResumeInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned ResumeInst::getNumOperands() const
{ return OperandTraits<ResumeInst>::operands(this); } template
<int Idx_nocapture> Use &ResumeInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &ResumeInst::Op() const { return
this->OpFrom<Idx_nocapture>(this); }
4293
4294//===----------------------------------------------------------------------===//
4295// CatchSwitchInst Class
4296//===----------------------------------------------------------------------===//
4297class CatchSwitchInst : public Instruction {
4298 using UnwindDestField = BoolBitfieldElementT<0>;
4299
4300 /// The number of operands actually allocated. NumOperands is
4301 /// the number actually in use.
4302 unsigned ReservedSpace;
4303
4304 // Operand[0] = Outer scope
4305 // Operand[1] = Unwind block destination
4306 // Operand[n] = BasicBlock to go to on match
4307 CatchSwitchInst(const CatchSwitchInst &CSI);
4308
4309 /// Create a new switch instruction, specifying a
4310 /// default destination. The number of additional handlers can be specified
4311 /// here to make memory allocation more efficient.
4312 /// This constructor can also autoinsert before another instruction.
4313 CatchSwitchInst(Value *ParentPad, BasicBlock *UnwindDest,
4314 unsigned NumHandlers, const Twine &NameStr,
4315 Instruction *InsertBefore);
4316
4317 /// Create a new switch instruction, specifying a
4318 /// default destination. The number of additional handlers can be specified
4319 /// here to make memory allocation more efficient.
4320 /// This constructor also autoinserts at the end of the specified BasicBlock.
4321 CatchSwitchInst(Value *ParentPad, BasicBlock *UnwindDest,
4322 unsigned NumHandlers, const Twine &NameStr,
4323 BasicBlock *InsertAtEnd);
4324
4325 // allocate space for exactly zero operands
4326 void *operator new(size_t S) { return User::operator new(S); }
4327
4328 void init(Value *ParentPad, BasicBlock *UnwindDest, unsigned NumReserved);
4329 void growOperands(unsigned Size);
4330
4331protected:
4332 // Note: Instruction needs to be a friend here to call cloneImpl.
4333 friend class Instruction;
4334
4335 CatchSwitchInst *cloneImpl() const;
4336
4337public:
4338 void operator delete(void *Ptr) { return User::operator delete(Ptr); }
4339
4340 static CatchSwitchInst *Create(Value *ParentPad, BasicBlock *UnwindDest,
4341 unsigned NumHandlers,
4342 const Twine &NameStr = "",
4343 Instruction *InsertBefore = nullptr) {
4344 return new CatchSwitchInst(ParentPad, UnwindDest, NumHandlers, NameStr,
4345 InsertBefore);
4346 }
4347
4348 static CatchSwitchInst *Create(Value *ParentPad, BasicBlock *UnwindDest,
4349 unsigned NumHandlers, const Twine &NameStr,
4350 BasicBlock *InsertAtEnd) {
4351 return new CatchSwitchInst(ParentPad, UnwindDest, NumHandlers, NameStr,
4352 InsertAtEnd);
4353 }
4354
4355 /// Provide fast operand accessors
4356 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
4357
4358 // Accessor Methods for CatchSwitch stmt
4359 Value *getParentPad() const { return getOperand(0); }
4360 void setParentPad(Value *ParentPad) { setOperand(0, ParentPad); }
4361
4362 // Accessor Methods for CatchSwitch stmt
4363 bool hasUnwindDest() const { return getSubclassData<UnwindDestField>(); }
4364 bool unwindsToCaller() const { return !hasUnwindDest(); }
4365 BasicBlock *getUnwindDest() const {
4366 if (hasUnwindDest())
4367 return cast<BasicBlock>(getOperand(1));
4368 return nullptr;
4369 }
4370 void setUnwindDest(BasicBlock *UnwindDest) {
4371 assert(UnwindDest)(static_cast <bool> (UnwindDest) ? void (0) : __assert_fail
("UnwindDest", "llvm/include/llvm/IR/Instructions.h", 4371, __extension__
__PRETTY_FUNCTION__))
;
4372 assert(hasUnwindDest())(static_cast <bool> (hasUnwindDest()) ? void (0) : __assert_fail
("hasUnwindDest()", "llvm/include/llvm/IR/Instructions.h", 4372
, __extension__ __PRETTY_FUNCTION__))
;
4373 setOperand(1, UnwindDest);
4374 }
4375
4376 /// return the number of 'handlers' in this catchswitch
4377 /// instruction, except the default handler
4378 unsigned getNumHandlers() const {
4379 if (hasUnwindDest())
4380 return getNumOperands() - 2;
4381 return getNumOperands() - 1;
4382 }
4383
4384private:
4385 static BasicBlock *handler_helper(Value *V) { return cast<BasicBlock>(V); }
4386 static const BasicBlock *handler_helper(const Value *V) {
4387 return cast<BasicBlock>(V);
4388 }
4389
4390public:
4391 using DerefFnTy = BasicBlock *(*)(Value *);
4392 using handler_iterator = mapped_iterator<op_iterator, DerefFnTy>;
4393 using handler_range = iterator_range<handler_iterator>;
4394 using ConstDerefFnTy = const BasicBlock *(*)(const Value *);
4395 using const_handler_iterator =
4396 mapped_iterator<const_op_iterator, ConstDerefFnTy>;
4397 using const_handler_range = iterator_range<const_handler_iterator>;
4398
4399 /// Returns an iterator that points to the first handler in CatchSwitchInst.
4400 handler_iterator handler_begin() {
4401 op_iterator It = op_begin() + 1;
4402 if (hasUnwindDest())
4403 ++It;
4404 return handler_iterator(It, DerefFnTy(handler_helper));
4405 }
4406
4407 /// Returns an iterator that points to the first handler in the
4408 /// CatchSwitchInst.
4409 const_handler_iterator handler_begin() const {
4410 const_op_iterator It = op_begin() + 1;
4411 if (hasUnwindDest())
4412 ++It;
4413 return const_handler_iterator(It, ConstDerefFnTy(handler_helper));
4414 }
4415
4416 /// Returns a read-only iterator that points one past the last
4417 /// handler in the CatchSwitchInst.
4418 handler_iterator handler_end() {
4419 return handler_iterator(op_end(), DerefFnTy(handler_helper));
4420 }
4421
4422 /// Returns an iterator that points one past the last handler in the
4423 /// CatchSwitchInst.
4424 const_handler_iterator handler_end() const {
4425 return const_handler_iterator(op_end(), ConstDerefFnTy(handler_helper));
4426 }
4427
4428 /// iteration adapter for range-for loops.
4429 handler_range handlers() {
4430 return make_range(handler_begin(), handler_end());
4431 }
4432
4433 /// iteration adapter for range-for loops.
4434 const_handler_range handlers() const {
4435 return make_range(handler_begin(), handler_end());
4436 }
4437
4438 /// Add an entry to the switch instruction...
4439 /// Note:
4440 /// This action invalidates handler_end(). Old handler_end() iterator will
4441 /// point to the added handler.
4442 void addHandler(BasicBlock *Dest);
4443
4444 void removeHandler(handler_iterator HI);
4445
4446 unsigned getNumSuccessors() const { return getNumOperands() - 1; }
4447 BasicBlock *getSuccessor(unsigned Idx) const {
4448 assert(Idx < getNumSuccessors() &&(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchswitch!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\""
, "llvm/include/llvm/IR/Instructions.h", 4449, __extension__ __PRETTY_FUNCTION__
))
4449 "Successor # out of range for catchswitch!")(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchswitch!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\""
, "llvm/include/llvm/IR/Instructions.h", 4449, __extension__ __PRETTY_FUNCTION__
))
;
4450 return cast<BasicBlock>(getOperand(Idx + 1));
4451 }
4452 void setSuccessor(unsigned Idx, BasicBlock *NewSucc) {
4453 assert(Idx < getNumSuccessors() &&(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchswitch!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\""
, "llvm/include/llvm/IR/Instructions.h", 4454, __extension__ __PRETTY_FUNCTION__
))
4454 "Successor # out of range for catchswitch!")(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchswitch!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchswitch!\""
, "llvm/include/llvm/IR/Instructions.h", 4454, __extension__ __PRETTY_FUNCTION__
))
;
4455 setOperand(Idx + 1, NewSucc);
4456 }
4457
4458 // Methods for support type inquiry through isa, cast, and dyn_cast:
4459 static bool classof(const Instruction *I) {
4460 return I->getOpcode() == Instruction::CatchSwitch;
4461 }
4462 static bool classof(const Value *V) {
4463 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4464 }
4465};
4466
4467template <>
4468struct OperandTraits<CatchSwitchInst> : public HungoffOperandTraits<2> {};
4469
4470DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CatchSwitchInst, Value)CatchSwitchInst::op_iterator CatchSwitchInst::op_begin() { return
OperandTraits<CatchSwitchInst>::op_begin(this); } CatchSwitchInst
::const_op_iterator CatchSwitchInst::op_begin() const { return
OperandTraits<CatchSwitchInst>::op_begin(const_cast<
CatchSwitchInst*>(this)); } CatchSwitchInst::op_iterator CatchSwitchInst
::op_end() { return OperandTraits<CatchSwitchInst>::op_end
(this); } CatchSwitchInst::const_op_iterator CatchSwitchInst::
op_end() const { return OperandTraits<CatchSwitchInst>::
op_end(const_cast<CatchSwitchInst*>(this)); } Value *CatchSwitchInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<CatchSwitchInst>::
operands(this) && "getOperand() out of range!") ? void
(0) : __assert_fail ("i_nocapture < OperandTraits<CatchSwitchInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4470, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<CatchSwitchInst
>::op_begin(const_cast<CatchSwitchInst*>(this))[i_nocapture
].get()); } void CatchSwitchInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<CatchSwitchInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchSwitchInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4470, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<CatchSwitchInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned CatchSwitchInst::getNumOperands
() const { return OperandTraits<CatchSwitchInst>::operands
(this); } template <int Idx_nocapture> Use &CatchSwitchInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &CatchSwitchInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
4471
4472//===----------------------------------------------------------------------===//
4473// CleanupPadInst Class
4474//===----------------------------------------------------------------------===//
4475class CleanupPadInst : public FuncletPadInst {
4476private:
4477 explicit CleanupPadInst(Value *ParentPad, ArrayRef<Value *> Args,
4478 unsigned Values, const Twine &NameStr,
4479 Instruction *InsertBefore)
4480 : FuncletPadInst(Instruction::CleanupPad, ParentPad, Args, Values,
4481 NameStr, InsertBefore) {}
4482 explicit CleanupPadInst(Value *ParentPad, ArrayRef<Value *> Args,
4483 unsigned Values, const Twine &NameStr,
4484 BasicBlock *InsertAtEnd)
4485 : FuncletPadInst(Instruction::CleanupPad, ParentPad, Args, Values,
4486 NameStr, InsertAtEnd) {}
4487
4488public:
4489 static CleanupPadInst *Create(Value *ParentPad, ArrayRef<Value *> Args = None,
4490 const Twine &NameStr = "",
4491 Instruction *InsertBefore = nullptr) {
4492 unsigned Values = 1 + Args.size();
4493 return new (Values)
4494 CleanupPadInst(ParentPad, Args, Values, NameStr, InsertBefore);
4495 }
4496
4497 static CleanupPadInst *Create(Value *ParentPad, ArrayRef<Value *> Args,
4498 const Twine &NameStr, BasicBlock *InsertAtEnd) {
4499 unsigned Values = 1 + Args.size();
4500 return new (Values)
4501 CleanupPadInst(ParentPad, Args, Values, NameStr, InsertAtEnd);
4502 }
4503
4504 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4505 static bool classof(const Instruction *I) {
4506 return I->getOpcode() == Instruction::CleanupPad;
4507 }
4508 static bool classof(const Value *V) {
4509 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4510 }
4511};
4512
4513//===----------------------------------------------------------------------===//
4514// CatchPadInst Class
4515//===----------------------------------------------------------------------===//
4516class CatchPadInst : public FuncletPadInst {
4517private:
4518 explicit CatchPadInst(Value *CatchSwitch, ArrayRef<Value *> Args,
4519 unsigned Values, const Twine &NameStr,
4520 Instruction *InsertBefore)
4521 : FuncletPadInst(Instruction::CatchPad, CatchSwitch, Args, Values,
4522 NameStr, InsertBefore) {}
4523 explicit CatchPadInst(Value *CatchSwitch, ArrayRef<Value *> Args,
4524 unsigned Values, const Twine &NameStr,
4525 BasicBlock *InsertAtEnd)
4526 : FuncletPadInst(Instruction::CatchPad, CatchSwitch, Args, Values,
4527 NameStr, InsertAtEnd) {}
4528
4529public:
4530 static CatchPadInst *Create(Value *CatchSwitch, ArrayRef<Value *> Args,
4531 const Twine &NameStr = "",
4532 Instruction *InsertBefore = nullptr) {
4533 unsigned Values = 1 + Args.size();
4534 return new (Values)
4535 CatchPadInst(CatchSwitch, Args, Values, NameStr, InsertBefore);
4536 }
4537
4538 static CatchPadInst *Create(Value *CatchSwitch, ArrayRef<Value *> Args,
4539 const Twine &NameStr, BasicBlock *InsertAtEnd) {
4540 unsigned Values = 1 + Args.size();
4541 return new (Values)
4542 CatchPadInst(CatchSwitch, Args, Values, NameStr, InsertAtEnd);
4543 }
4544
4545 /// Convenience accessors
4546 CatchSwitchInst *getCatchSwitch() const {
4547 return cast<CatchSwitchInst>(Op<-1>());
4548 }
4549 void setCatchSwitch(Value *CatchSwitch) {
4550 assert(CatchSwitch)(static_cast <bool> (CatchSwitch) ? void (0) : __assert_fail
("CatchSwitch", "llvm/include/llvm/IR/Instructions.h", 4550,
__extension__ __PRETTY_FUNCTION__))
;
4551 Op<-1>() = CatchSwitch;
4552 }
4553
4554 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4555 static bool classof(const Instruction *I) {
4556 return I->getOpcode() == Instruction::CatchPad;
4557 }
4558 static bool classof(const Value *V) {
4559 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4560 }
4561};
4562
4563//===----------------------------------------------------------------------===//
4564// CatchReturnInst Class
4565//===----------------------------------------------------------------------===//
4566
4567class CatchReturnInst : public Instruction {
4568 CatchReturnInst(const CatchReturnInst &RI);
4569 CatchReturnInst(Value *CatchPad, BasicBlock *BB, Instruction *InsertBefore);
4570 CatchReturnInst(Value *CatchPad, BasicBlock *BB, BasicBlock *InsertAtEnd);
4571
4572 void init(Value *CatchPad, BasicBlock *BB);
4573
4574protected:
4575 // Note: Instruction needs to be a friend here to call cloneImpl.
4576 friend class Instruction;
4577
4578 CatchReturnInst *cloneImpl() const;
4579
4580public:
4581 static CatchReturnInst *Create(Value *CatchPad, BasicBlock *BB,
4582 Instruction *InsertBefore = nullptr) {
4583 assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail
("CatchPad", "llvm/include/llvm/IR/Instructions.h", 4583, __extension__
__PRETTY_FUNCTION__))
;
4584 assert(BB)(static_cast <bool> (BB) ? void (0) : __assert_fail ("BB"
, "llvm/include/llvm/IR/Instructions.h", 4584, __extension__ __PRETTY_FUNCTION__
))
;
4585 return new (2) CatchReturnInst(CatchPad, BB, InsertBefore);
4586 }
4587
4588 static CatchReturnInst *Create(Value *CatchPad, BasicBlock *BB,
4589 BasicBlock *InsertAtEnd) {
4590 assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail
("CatchPad", "llvm/include/llvm/IR/Instructions.h", 4590, __extension__
__PRETTY_FUNCTION__))
;
4591 assert(BB)(static_cast <bool> (BB) ? void (0) : __assert_fail ("BB"
, "llvm/include/llvm/IR/Instructions.h", 4591, __extension__ __PRETTY_FUNCTION__
))
;
4592 return new (2) CatchReturnInst(CatchPad, BB, InsertAtEnd);
4593 }
4594
4595 /// Provide fast operand accessors
4596 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
4597
4598 /// Convenience accessors.
4599 CatchPadInst *getCatchPad() const { return cast<CatchPadInst>(Op<0>()); }
4600 void setCatchPad(CatchPadInst *CatchPad) {
4601 assert(CatchPad)(static_cast <bool> (CatchPad) ? void (0) : __assert_fail
("CatchPad", "llvm/include/llvm/IR/Instructions.h", 4601, __extension__
__PRETTY_FUNCTION__))
;
4602 Op<0>() = CatchPad;
4603 }
4604
4605 BasicBlock *getSuccessor() const { return cast<BasicBlock>(Op<1>()); }
4606 void setSuccessor(BasicBlock *NewSucc) {
4607 assert(NewSucc)(static_cast <bool> (NewSucc) ? void (0) : __assert_fail
("NewSucc", "llvm/include/llvm/IR/Instructions.h", 4607, __extension__
__PRETTY_FUNCTION__))
;
4608 Op<1>() = NewSucc;
4609 }
4610 unsigned getNumSuccessors() const { return 1; }
4611
4612 /// Get the parentPad of this catchret's catchpad's catchswitch.
4613 /// The successor block is implicitly a member of this funclet.
4614 Value *getCatchSwitchParentPad() const {
4615 return getCatchPad()->getCatchSwitch()->getParentPad();
4616 }
4617
4618 // Methods for support type inquiry through isa, cast, and dyn_cast:
4619 static bool classof(const Instruction *I) {
4620 return (I->getOpcode() == Instruction::CatchRet);
4621 }
4622 static bool classof(const Value *V) {
4623 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4624 }
4625
4626private:
4627 BasicBlock *getSuccessor(unsigned Idx) const {
4628 assert(Idx < getNumSuccessors() && "Successor # out of range for catchret!")(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchret!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchret!\""
, "llvm/include/llvm/IR/Instructions.h", 4628, __extension__ __PRETTY_FUNCTION__
))
;
4629 return getSuccessor();
4630 }
4631
4632 void setSuccessor(unsigned Idx, BasicBlock *B) {
4633 assert(Idx < getNumSuccessors() && "Successor # out of range for catchret!")(static_cast <bool> (Idx < getNumSuccessors() &&
"Successor # out of range for catchret!") ? void (0) : __assert_fail
("Idx < getNumSuccessors() && \"Successor # out of range for catchret!\""
, "llvm/include/llvm/IR/Instructions.h", 4633, __extension__ __PRETTY_FUNCTION__
))
;
4634 setSuccessor(B);
4635 }
4636};
4637
4638template <>
4639struct OperandTraits<CatchReturnInst>
4640 : public FixedNumOperandTraits<CatchReturnInst, 2> {};
4641
4642DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CatchReturnInst, Value)CatchReturnInst::op_iterator CatchReturnInst::op_begin() { return
OperandTraits<CatchReturnInst>::op_begin(this); } CatchReturnInst
::const_op_iterator CatchReturnInst::op_begin() const { return
OperandTraits<CatchReturnInst>::op_begin(const_cast<
CatchReturnInst*>(this)); } CatchReturnInst::op_iterator CatchReturnInst
::op_end() { return OperandTraits<CatchReturnInst>::op_end
(this); } CatchReturnInst::const_op_iterator CatchReturnInst::
op_end() const { return OperandTraits<CatchReturnInst>::
op_end(const_cast<CatchReturnInst*>(this)); } Value *CatchReturnInst
::getOperand(unsigned i_nocapture) const { (static_cast <bool
> (i_nocapture < OperandTraits<CatchReturnInst>::
operands(this) && "getOperand() out of range!") ? void
(0) : __assert_fail ("i_nocapture < OperandTraits<CatchReturnInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4642, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<CatchReturnInst
>::op_begin(const_cast<CatchReturnInst*>(this))[i_nocapture
].get()); } void CatchReturnInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<CatchReturnInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CatchReturnInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4642, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<CatchReturnInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned CatchReturnInst::getNumOperands
() const { return OperandTraits<CatchReturnInst>::operands
(this); } template <int Idx_nocapture> Use &CatchReturnInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &CatchReturnInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
4643
4644//===----------------------------------------------------------------------===//
4645// CleanupReturnInst Class
4646//===----------------------------------------------------------------------===//
4647
4648class CleanupReturnInst : public Instruction {
4649 using UnwindDestField = BoolBitfieldElementT<0>;
4650
4651private:
4652 CleanupReturnInst(const CleanupReturnInst &RI);
4653 CleanupReturnInst(Value *CleanupPad, BasicBlock *UnwindBB, unsigned Values,
4654 Instruction *InsertBefore = nullptr);
4655 CleanupReturnInst(Value *CleanupPad, BasicBlock *UnwindBB, unsigned Values,
4656 BasicBlock *InsertAtEnd);
4657
4658 void init(Value *CleanupPad, BasicBlock *UnwindBB);
4659
4660protected:
4661 // Note: Instruction needs to be a friend here to call cloneImpl.
4662 friend class Instruction;
4663
4664 CleanupReturnInst *cloneImpl() const;
4665
4666public:
4667 static CleanupReturnInst *Create(Value *CleanupPad,
4668 BasicBlock *UnwindBB = nullptr,
4669 Instruction *InsertBefore = nullptr) {
4670 assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail
("CleanupPad", "llvm/include/llvm/IR/Instructions.h", 4670, __extension__
__PRETTY_FUNCTION__))
;
4671 unsigned Values = 1;
4672 if (UnwindBB)
4673 ++Values;
4674 return new (Values)
4675 CleanupReturnInst(CleanupPad, UnwindBB, Values, InsertBefore);
4676 }
4677
4678 static CleanupReturnInst *Create(Value *CleanupPad, BasicBlock *UnwindBB,
4679 BasicBlock *InsertAtEnd) {
4680 assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail
("CleanupPad", "llvm/include/llvm/IR/Instructions.h", 4680, __extension__
__PRETTY_FUNCTION__))
;
4681 unsigned Values = 1;
4682 if (UnwindBB)
4683 ++Values;
4684 return new (Values)
4685 CleanupReturnInst(CleanupPad, UnwindBB, Values, InsertAtEnd);
4686 }
4687
4688 /// Provide fast operand accessors
4689 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
4690
4691 bool hasUnwindDest() const { return getSubclassData<UnwindDestField>(); }
4692 bool unwindsToCaller() const { return !hasUnwindDest(); }
4693
4694 /// Convenience accessor.
4695 CleanupPadInst *getCleanupPad() const {
4696 return cast<CleanupPadInst>(Op<0>());
4697 }
4698 void setCleanupPad(CleanupPadInst *CleanupPad) {
4699 assert(CleanupPad)(static_cast <bool> (CleanupPad) ? void (0) : __assert_fail
("CleanupPad", "llvm/include/llvm/IR/Instructions.h", 4699, __extension__
__PRETTY_FUNCTION__))
;
4700 Op<0>() = CleanupPad;
4701 }
4702
4703 unsigned getNumSuccessors() const { return hasUnwindDest() ? 1 : 0; }
4704
4705 BasicBlock *getUnwindDest() const {
4706 return hasUnwindDest() ? cast<BasicBlock>(Op<1>()) : nullptr;
4707 }
4708 void setUnwindDest(BasicBlock *NewDest) {
4709 assert(NewDest)(static_cast <bool> (NewDest) ? void (0) : __assert_fail
("NewDest", "llvm/include/llvm/IR/Instructions.h", 4709, __extension__
__PRETTY_FUNCTION__))
;
4710 assert(hasUnwindDest())(static_cast <bool> (hasUnwindDest()) ? void (0) : __assert_fail
("hasUnwindDest()", "llvm/include/llvm/IR/Instructions.h", 4710
, __extension__ __PRETTY_FUNCTION__))
;
4711 Op<1>() = NewDest;
4712 }
4713
4714 // Methods for support type inquiry through isa, cast, and dyn_cast:
4715 static bool classof(const Instruction *I) {
4716 return (I->getOpcode() == Instruction::CleanupRet);
4717 }
4718 static bool classof(const Value *V) {
4719 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4720 }
4721
4722private:
4723 BasicBlock *getSuccessor(unsigned Idx) const {
4724 assert(Idx == 0)(static_cast <bool> (Idx == 0) ? void (0) : __assert_fail
("Idx == 0", "llvm/include/llvm/IR/Instructions.h", 4724, __extension__
__PRETTY_FUNCTION__))
;
4725 return getUnwindDest();
4726 }
4727
4728 void setSuccessor(unsigned Idx, BasicBlock *B) {
4729 assert(Idx == 0)(static_cast <bool> (Idx == 0) ? void (0) : __assert_fail
("Idx == 0", "llvm/include/llvm/IR/Instructions.h", 4729, __extension__
__PRETTY_FUNCTION__))
;
4730 setUnwindDest(B);
4731 }
4732
4733 // Shadow Instruction::setInstructionSubclassData with a private forwarding
4734 // method so that subclasses cannot accidentally use it.
4735 template <typename Bitfield>
4736 void setSubclassData(typename Bitfield::Type Value) {
4737 Instruction::setSubclassData<Bitfield>(Value);
4738 }
4739};
4740
4741template <>
4742struct OperandTraits<CleanupReturnInst>
4743 : public VariadicOperandTraits<CleanupReturnInst, /*MINARITY=*/1> {};
4744
4745DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CleanupReturnInst, Value)CleanupReturnInst::op_iterator CleanupReturnInst::op_begin() {
return OperandTraits<CleanupReturnInst>::op_begin(this
); } CleanupReturnInst::const_op_iterator CleanupReturnInst::
op_begin() const { return OperandTraits<CleanupReturnInst>
::op_begin(const_cast<CleanupReturnInst*>(this)); } CleanupReturnInst
::op_iterator CleanupReturnInst::op_end() { return OperandTraits
<CleanupReturnInst>::op_end(this); } CleanupReturnInst::
const_op_iterator CleanupReturnInst::op_end() const { return OperandTraits
<CleanupReturnInst>::op_end(const_cast<CleanupReturnInst
*>(this)); } Value *CleanupReturnInst::getOperand(unsigned
i_nocapture) const { (static_cast <bool> (i_nocapture <
OperandTraits<CleanupReturnInst>::operands(this) &&
"getOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && \"getOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4745, __extension__ __PRETTY_FUNCTION__
)); return cast_or_null<Value>( OperandTraits<CleanupReturnInst
>::op_begin(const_cast<CleanupReturnInst*>(this))[i_nocapture
].get()); } void CleanupReturnInst::setOperand(unsigned i_nocapture
, Value *Val_nocapture) { (static_cast <bool> (i_nocapture
< OperandTraits<CleanupReturnInst>::operands(this) &&
"setOperand() out of range!") ? void (0) : __assert_fail ("i_nocapture < OperandTraits<CleanupReturnInst>::operands(this) && \"setOperand() out of range!\""
, "llvm/include/llvm/IR/Instructions.h", 4745, __extension__ __PRETTY_FUNCTION__
)); OperandTraits<CleanupReturnInst>::op_begin(this)[i_nocapture
] = Val_nocapture; } unsigned CleanupReturnInst::getNumOperands
() const { return OperandTraits<CleanupReturnInst>::operands
(this); } template <int Idx_nocapture> Use &CleanupReturnInst
::Op() { return this->OpFrom<Idx_nocapture>(this); }
template <int Idx_nocapture> const Use &CleanupReturnInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
4746
4747//===----------------------------------------------------------------------===//
4748// UnreachableInst Class
4749//===----------------------------------------------------------------------===//
4750
4751//===---------------------------------------------------------------------------
4752/// This function has undefined behavior. In particular, the
4753/// presence of this instruction indicates some higher level knowledge that the
4754/// end of the block cannot be reached.
4755///
4756class UnreachableInst : public Instruction {
4757protected:
4758 // Note: Instruction needs to be a friend here to call cloneImpl.
4759 friend class Instruction;
4760
4761 UnreachableInst *cloneImpl() const;
4762
4763public:
4764 explicit UnreachableInst(LLVMContext &C, Instruction *InsertBefore = nullptr);
4765 explicit UnreachableInst(LLVMContext &C, BasicBlock *InsertAtEnd);
4766
4767 // allocate space for exactly zero operands
4768 void *operator new(size_t S) { return User::operator new(S, 0); }
4769 void operator delete(void *Ptr) { User::operator delete(Ptr); }
4770
4771 unsigned getNumSuccessors() const { return 0; }
4772
4773 // Methods for support type inquiry through isa, cast, and dyn_cast:
4774 static bool classof(const Instruction *I) {
4775 return I->getOpcode() == Instruction::Unreachable;
4776 }
4777 static bool classof(const Value *V) {
4778 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4779 }
4780
4781private:
4782 BasicBlock *getSuccessor(unsigned idx) const {
4783 llvm_unreachable("UnreachableInst has no successors!")::llvm::llvm_unreachable_internal("UnreachableInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 4783)
;
4784 }
4785
4786 void setSuccessor(unsigned idx, BasicBlock *B) {
4787 llvm_unreachable("UnreachableInst has no successors!")::llvm::llvm_unreachable_internal("UnreachableInst has no successors!"
, "llvm/include/llvm/IR/Instructions.h", 4787)
;
4788 }
4789};
4790
4791//===----------------------------------------------------------------------===//
4792// TruncInst Class
4793//===----------------------------------------------------------------------===//
4794
4795/// This class represents a truncation of integer types.
4796class TruncInst : public CastInst {
4797protected:
4798 // Note: Instruction needs to be a friend here to call cloneImpl.
4799 friend class Instruction;
4800
4801 /// Clone an identical TruncInst
4802 TruncInst *cloneImpl() const;
4803
4804public:
4805 /// Constructor with insert-before-instruction semantics
4806 TruncInst(
4807 Value *S, ///< The value to be truncated
4808 Type *Ty, ///< The (smaller) type to truncate to
4809 const Twine &NameStr = "", ///< A name for the new instruction
4810 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
4811 );
4812
4813 /// Constructor with insert-at-end-of-block semantics
4814 TruncInst(
4815 Value *S, ///< The value to be truncated
4816 Type *Ty, ///< The (smaller) type to truncate to
4817 const Twine &NameStr, ///< A name for the new instruction
4818 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
4819 );
4820
4821 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4822 static bool classof(const Instruction *I) {
4823 return I->getOpcode() == Trunc;
4824 }
4825 static bool classof(const Value *V) {
4826 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4827 }
4828};
4829
4830//===----------------------------------------------------------------------===//
4831// ZExtInst Class
4832//===----------------------------------------------------------------------===//
4833
4834/// This class represents zero extension of integer types.
4835class ZExtInst : public CastInst {
4836protected:
4837 // Note: Instruction needs to be a friend here to call cloneImpl.
4838 friend class Instruction;
4839
4840 /// Clone an identical ZExtInst
4841 ZExtInst *cloneImpl() const;
4842
4843public:
4844 /// Constructor with insert-before-instruction semantics
4845 ZExtInst(
4846 Value *S, ///< The value to be zero extended
4847 Type *Ty, ///< The type to zero extend to
4848 const Twine &NameStr = "", ///< A name for the new instruction
4849 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
4850 );
4851
4852 /// Constructor with insert-at-end semantics.
4853 ZExtInst(
4854 Value *S, ///< The value to be zero extended
4855 Type *Ty, ///< The type to zero extend to
4856 const Twine &NameStr, ///< A name for the new instruction
4857 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
4858 );
4859
4860 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4861 static bool classof(const Instruction *I) {
4862 return I->getOpcode() == ZExt;
4863 }
4864 static bool classof(const Value *V) {
4865 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4866 }
4867};
4868
4869//===----------------------------------------------------------------------===//
4870// SExtInst Class
4871//===----------------------------------------------------------------------===//
4872
4873/// This class represents a sign extension of integer types.
4874class SExtInst : public CastInst {
4875protected:
4876 // Note: Instruction needs to be a friend here to call cloneImpl.
4877 friend class Instruction;
4878
4879 /// Clone an identical SExtInst
4880 SExtInst *cloneImpl() const;
4881
4882public:
4883 /// Constructor with insert-before-instruction semantics
4884 SExtInst(
4885 Value *S, ///< The value to be sign extended
4886 Type *Ty, ///< The type to sign extend to
4887 const Twine &NameStr = "", ///< A name for the new instruction
4888 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
4889 );
4890
4891 /// Constructor with insert-at-end-of-block semantics
4892 SExtInst(
4893 Value *S, ///< The value to be sign extended
4894 Type *Ty, ///< The type to sign extend to
4895 const Twine &NameStr, ///< A name for the new instruction
4896 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
4897 );
4898
4899 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4900 static bool classof(const Instruction *I) {
4901 return I->getOpcode() == SExt;
4902 }
4903 static bool classof(const Value *V) {
4904 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4905 }
4906};
4907
4908//===----------------------------------------------------------------------===//
4909// FPTruncInst Class
4910//===----------------------------------------------------------------------===//
4911
4912/// This class represents a truncation of floating point types.
4913class FPTruncInst : public CastInst {
4914protected:
4915 // Note: Instruction needs to be a friend here to call cloneImpl.
4916 friend class Instruction;
4917
4918 /// Clone an identical FPTruncInst
4919 FPTruncInst *cloneImpl() const;
4920
4921public:
4922 /// Constructor with insert-before-instruction semantics
4923 FPTruncInst(
4924 Value *S, ///< The value to be truncated
4925 Type *Ty, ///< The type to truncate to
4926 const Twine &NameStr = "", ///< A name for the new instruction
4927 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
4928 );
4929
4930 /// Constructor with insert-before-instruction semantics
4931 FPTruncInst(
4932 Value *S, ///< The value to be truncated
4933 Type *Ty, ///< The type to truncate to
4934 const Twine &NameStr, ///< A name for the new instruction
4935 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
4936 );
4937
4938 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4939 static bool classof(const Instruction *I) {
4940 return I->getOpcode() == FPTrunc;
4941 }
4942 static bool classof(const Value *V) {
4943 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4944 }
4945};
4946
4947//===----------------------------------------------------------------------===//
4948// FPExtInst Class
4949//===----------------------------------------------------------------------===//
4950
4951/// This class represents an extension of floating point types.
4952class FPExtInst : public CastInst {
4953protected:
4954 // Note: Instruction needs to be a friend here to call cloneImpl.
4955 friend class Instruction;
4956
4957 /// Clone an identical FPExtInst
4958 FPExtInst *cloneImpl() const;
4959
4960public:
4961 /// Constructor with insert-before-instruction semantics
4962 FPExtInst(
4963 Value *S, ///< The value to be extended
4964 Type *Ty, ///< The type to extend to
4965 const Twine &NameStr = "", ///< A name for the new instruction
4966 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
4967 );
4968
4969 /// Constructor with insert-at-end-of-block semantics
4970 FPExtInst(
4971 Value *S, ///< The value to be extended
4972 Type *Ty, ///< The type to extend to
4973 const Twine &NameStr, ///< A name for the new instruction
4974 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
4975 );
4976
4977 /// Methods for support type inquiry through isa, cast, and dyn_cast:
4978 static bool classof(const Instruction *I) {
4979 return I->getOpcode() == FPExt;
4980 }
4981 static bool classof(const Value *V) {
4982 return isa<Instruction>(V) && classof(cast<Instruction>(V));
4983 }
4984};
4985
4986//===----------------------------------------------------------------------===//
4987// UIToFPInst Class
4988//===----------------------------------------------------------------------===//
4989
4990/// This class represents a cast unsigned integer to floating point.
4991class UIToFPInst : public CastInst {
4992protected:
4993 // Note: Instruction needs to be a friend here to call cloneImpl.
4994 friend class Instruction;
4995
4996 /// Clone an identical UIToFPInst
4997 UIToFPInst *cloneImpl() const;
4998
4999public:
5000 /// Constructor with insert-before-instruction semantics
5001 UIToFPInst(
5002 Value *S, ///< The value to be converted
5003 Type *Ty, ///< The type to convert to
5004 const Twine &NameStr = "", ///< A name for the new instruction
5005 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5006 );
5007
5008 /// Constructor with insert-at-end-of-block semantics
5009 UIToFPInst(
5010 Value *S, ///< The value to be converted
5011 Type *Ty, ///< The type to convert to
5012 const Twine &NameStr, ///< A name for the new instruction
5013 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5014 );
5015
5016 /// Methods for support type inquiry through isa, cast, and dyn_cast:
5017 static bool classof(const Instruction *I) {
5018 return I->getOpcode() == UIToFP;
5019 }
5020 static bool classof(const Value *V) {
5021 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5022 }
5023};
5024
5025//===----------------------------------------------------------------------===//
5026// SIToFPInst Class
5027//===----------------------------------------------------------------------===//
5028
5029/// This class represents a cast from signed integer to floating point.
5030class SIToFPInst : public CastInst {
5031protected:
5032 // Note: Instruction needs to be a friend here to call cloneImpl.
5033 friend class Instruction;
5034
5035 /// Clone an identical SIToFPInst
5036 SIToFPInst *cloneImpl() const;
5037
5038public:
5039 /// Constructor with insert-before-instruction semantics
5040 SIToFPInst(
5041 Value *S, ///< The value to be converted
5042 Type *Ty, ///< The type to convert to
5043 const Twine &NameStr = "", ///< A name for the new instruction
5044 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5045 );
5046
5047 /// Constructor with insert-at-end-of-block semantics
5048 SIToFPInst(
5049 Value *S, ///< The value to be converted
5050 Type *Ty, ///< The type to convert to
5051 const Twine &NameStr, ///< A name for the new instruction
5052 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5053 );
5054
5055 /// Methods for support type inquiry through isa, cast, and dyn_cast:
5056 static bool classof(const Instruction *I) {
5057 return I->getOpcode() == SIToFP;
5058 }
5059 static bool classof(const Value *V) {
5060 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5061 }
5062};
5063
5064//===----------------------------------------------------------------------===//
5065// FPToUIInst Class
5066//===----------------------------------------------------------------------===//
5067
5068/// This class represents a cast from floating point to unsigned integer
5069class FPToUIInst : public CastInst {
5070protected:
5071 // Note: Instruction needs to be a friend here to call cloneImpl.
5072 friend class Instruction;
5073
5074 /// Clone an identical FPToUIInst
5075 FPToUIInst *cloneImpl() const;
5076
5077public:
5078 /// Constructor with insert-before-instruction semantics
5079 FPToUIInst(
5080 Value *S, ///< The value to be converted
5081 Type *Ty, ///< The type to convert to
5082 const Twine &NameStr = "", ///< A name for the new instruction
5083 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5084 );
5085
5086 /// Constructor with insert-at-end-of-block semantics
5087 FPToUIInst(
5088 Value *S, ///< The value to be converted
5089 Type *Ty, ///< The type to convert to
5090 const Twine &NameStr, ///< A name for the new instruction
5091 BasicBlock *InsertAtEnd ///< Where to insert the new instruction
5092 );
5093
5094 /// Methods for support type inquiry through isa, cast, and dyn_cast:
5095 static bool classof(const Instruction *I) {
5096 return I->getOpcode() == FPToUI;
5097 }
5098 static bool classof(const Value *V) {
5099 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5100 }
5101};
5102
5103//===----------------------------------------------------------------------===//
5104// FPToSIInst Class
5105//===----------------------------------------------------------------------===//
5106
5107/// This class represents a cast from floating point to signed integer.
5108class FPToSIInst : public CastInst {
5109protected:
5110 // Note: Instruction needs to be a friend here to call cloneImpl.
5111 friend class Instruction;
5112
5113 /// Clone an identical FPToSIInst
5114 FPToSIInst *cloneImpl() const;
5115
5116public:
5117 /// Constructor with insert-before-instruction semantics
5118 FPToSIInst(
5119 Value *S, ///< The value to be converted
5120 Type *Ty, ///< The type to convert to
5121 const Twine &NameStr = "", ///< A name for the new instruction
5122 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5123 );
5124
5125 /// Constructor with insert-at-end-of-block semantics
5126 FPToSIInst(
5127 Value *S, ///< The value to be converted
5128 Type *Ty, ///< The type to convert to
5129 const Twine &NameStr, ///< A name for the new instruction
5130 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5131 );
5132
5133 /// Methods for support type inquiry through isa, cast, and dyn_cast:
5134 static bool classof(const Instruction *I) {
5135 return I->getOpcode() == FPToSI;
5136 }
5137 static bool classof(const Value *V) {
5138 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5139 }
5140};
5141
5142//===----------------------------------------------------------------------===//
5143// IntToPtrInst Class
5144//===----------------------------------------------------------------------===//
5145
5146/// This class represents a cast from an integer to a pointer.
5147class IntToPtrInst : public CastInst {
5148public:
5149 // Note: Instruction needs to be a friend here to call cloneImpl.
5150 friend class Instruction;
5151
5152 /// Constructor with insert-before-instruction semantics
5153 IntToPtrInst(
5154 Value *S, ///< The value to be converted
5155 Type *Ty, ///< The type to convert to
5156 const Twine &NameStr = "", ///< A name for the new instruction
5157 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5158 );
5159
5160 /// Constructor with insert-at-end-of-block semantics
5161 IntToPtrInst(
5162 Value *S, ///< The value to be converted
5163 Type *Ty, ///< The type to convert to
5164 const Twine &NameStr, ///< A name for the new instruction
5165 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5166 );
5167
5168 /// Clone an identical IntToPtrInst.
5169 IntToPtrInst *cloneImpl() const;
5170
5171 /// Returns the address space of this instruction's pointer type.
5172 unsigned getAddressSpace() const {
5173 return getType()->getPointerAddressSpace();
5174 }
5175
5176 // Methods for support type inquiry through isa, cast, and dyn_cast:
5177 static bool classof(const Instruction *I) {
5178 return I->getOpcode() == IntToPtr;
5179 }
5180 static bool classof(const Value *V) {
5181 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5182 }
5183};
5184
5185//===----------------------------------------------------------------------===//
5186// PtrToIntInst Class
5187//===----------------------------------------------------------------------===//
5188
5189/// This class represents a cast from a pointer to an integer.
5190class PtrToIntInst : public CastInst {
5191protected:
5192 // Note: Instruction needs to be a friend here to call cloneImpl.
5193 friend class Instruction;
5194
5195 /// Clone an identical PtrToIntInst.
5196 PtrToIntInst *cloneImpl() const;
5197
5198public:
5199 /// Constructor with insert-before-instruction semantics
5200 PtrToIntInst(
5201 Value *S, ///< The value to be converted
5202 Type *Ty, ///< The type to convert to
5203 const Twine &NameStr = "", ///< A name for the new instruction
5204 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5205 );
5206
5207 /// Constructor with insert-at-end-of-block semantics
5208 PtrToIntInst(
5209 Value *S, ///< The value to be converted
5210 Type *Ty, ///< The type to convert to
5211 const Twine &NameStr, ///< A name for the new instruction
5212 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5213 );
5214
5215 /// Gets the pointer operand.
5216 Value *getPointerOperand() { return getOperand(0); }
5217 /// Gets the pointer operand.
5218 const Value *getPointerOperand() const { return getOperand(0); }
5219 /// Gets the operand index of the pointer operand.
5220 static unsigned getPointerOperandIndex() { return 0U; }
5221
5222 /// Returns the address space of the pointer operand.
5223 unsigned getPointerAddressSpace() const {
5224 return getPointerOperand()->getType()->getPointerAddressSpace();
5225 }
5226
5227 // Methods for support type inquiry through isa, cast, and dyn_cast:
5228 static bool classof(const Instruction *I) {
5229 return I->getOpcode() == PtrToInt;
5230 }
5231 static bool classof(const Value *V) {
5232 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5233 }
5234};
5235
5236//===----------------------------------------------------------------------===//
5237// BitCastInst Class
5238//===----------------------------------------------------------------------===//
5239
5240/// This class represents a no-op cast from one type to another.
5241class BitCastInst : public CastInst {
5242protected:
5243 // Note: Instruction needs to be a friend here to call cloneImpl.
5244 friend class Instruction;
5245
5246 /// Clone an identical BitCastInst.
5247 BitCastInst *cloneImpl() const;
5248
5249public:
5250 /// Constructor with insert-before-instruction semantics
5251 BitCastInst(
5252 Value *S, ///< The value to be casted
5253 Type *Ty, ///< The type to casted to
5254 const Twine &NameStr = "", ///< A name for the new instruction
5255 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5256 );
5257
5258 /// Constructor with insert-at-end-of-block semantics
5259 BitCastInst(
5260 Value *S, ///< The value to be casted
5261 Type *Ty, ///< The type to casted to
5262 const Twine &NameStr, ///< A name for the new instruction
5263 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5264 );
5265
5266 // Methods for support type inquiry through isa, cast, and dyn_cast:
5267 static bool classof(const Instruction *I) {
5268 return I->getOpcode() == BitCast;
5269 }
5270 static bool classof(const Value *V) {
5271 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5272 }
5273};
5274
5275//===----------------------------------------------------------------------===//
5276// AddrSpaceCastInst Class
5277//===----------------------------------------------------------------------===//
5278
5279/// This class represents a conversion between pointers from one address space
5280/// to another.
5281class AddrSpaceCastInst : public CastInst {
5282protected:
5283 // Note: Instruction needs to be a friend here to call cloneImpl.
5284 friend class Instruction;
5285
5286 /// Clone an identical AddrSpaceCastInst.
5287 AddrSpaceCastInst *cloneImpl() const;
5288
5289public:
5290 /// Constructor with insert-before-instruction semantics
5291 AddrSpaceCastInst(
5292 Value *S, ///< The value to be casted
5293 Type *Ty, ///< The type to casted to
5294 const Twine &NameStr = "", ///< A name for the new instruction
5295 Instruction *InsertBefore = nullptr ///< Where to insert the new instruction
5296 );
5297
5298 /// Constructor with insert-at-end-of-block semantics
5299 AddrSpaceCastInst(
5300 Value *S, ///< The value to be casted
5301 Type *Ty, ///< The type to casted to
5302 const Twine &NameStr, ///< A name for the new instruction
5303 BasicBlock *InsertAtEnd ///< The block to insert the instruction into
5304 );
5305
5306 // Methods for support type inquiry through isa, cast, and dyn_cast:
5307 static bool classof(const Instruction *I) {
5308 return I->getOpcode() == AddrSpaceCast;
5309 }
5310 static bool classof(const Value *V) {
5311 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5312 }
5313
5314 /// Gets the pointer operand.
5315 Value *getPointerOperand() {
5316 return getOperand(0);
5317 }
5318
5319 /// Gets the pointer operand.
5320 const Value *getPointerOperand() const {
5321 return getOperand(0);
5322 }
5323
5324 /// Gets the operand index of the pointer operand.
5325 static unsigned getPointerOperandIndex() {
5326 return 0U;
5327 }
5328
5329 /// Returns the address space of the pointer operand.
5330 unsigned getSrcAddressSpace() const {
5331 return getPointerOperand()->getType()->getPointerAddressSpace();
5332 }
5333
5334 /// Returns the address space of the result.
5335 unsigned getDestAddressSpace() const {
5336 return getType()->getPointerAddressSpace();
5337 }
5338};
5339
5340//===----------------------------------------------------------------------===//
5341// Helper functions
5342//===----------------------------------------------------------------------===//
5343
5344/// A helper function that returns the pointer operand of a load or store
5345/// instruction. Returns nullptr if not load or store.
5346inline const Value *getLoadStorePointerOperand(const Value *V) {
5347 if (auto *Load = dyn_cast<LoadInst>(V))
5348 return Load->getPointerOperand();
5349 if (auto *Store = dyn_cast<StoreInst>(V))
5350 return Store->getPointerOperand();
5351 return nullptr;
5352}
5353inline Value *getLoadStorePointerOperand(Value *V) {
5354 return const_cast<Value *>(
5355 getLoadStorePointerOperand(static_cast<const Value *>(V)));
5356}
5357
5358/// A helper function that returns the pointer operand of a load, store
5359/// or GEP instruction. Returns nullptr if not load, store, or GEP.
5360inline const Value *getPointerOperand(const Value *V) {
5361 if (auto *Ptr = getLoadStorePointerOperand(V))
5362 return Ptr;
5363 if (auto *Gep = dyn_cast<GetElementPtrInst>(V))
5364 return Gep->getPointerOperand();
5365 return nullptr;
5366}
5367inline Value *getPointerOperand(Value *V) {
5368 return const_cast<Value *>(getPointerOperand(static_cast<const Value *>(V)));
5369}
5370
5371/// A helper function that returns the alignment of load or store instruction.
5372inline Align getLoadStoreAlignment(Value *I) {
5373 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5374, __extension__ __PRETTY_FUNCTION__
))
5374 "Expected Load or Store instruction")(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5374, __extension__ __PRETTY_FUNCTION__
))
;
5375 if (auto *LI = dyn_cast<LoadInst>(I))
5376 return LI->getAlign();
5377 return cast<StoreInst>(I)->getAlign();
5378}
5379
5380/// A helper function that returns the address space of the pointer operand of
5381/// load or store instruction.
5382inline unsigned getLoadStoreAddressSpace(Value *I) {
5383 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5384, __extension__ __PRETTY_FUNCTION__
))
5384 "Expected Load or Store instruction")(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5384, __extension__ __PRETTY_FUNCTION__
))
;
5385 if (auto *LI = dyn_cast<LoadInst>(I))
5386 return LI->getPointerAddressSpace();
5387 return cast<StoreInst>(I)->getPointerAddressSpace();
5388}
5389
5390/// A helper function that returns the type of a load or store instruction.
5391inline Type *getLoadStoreType(Value *I) {
5392 assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5393, __extension__ __PRETTY_FUNCTION__
))
5393 "Expected Load or Store instruction")(static_cast <bool> ((isa<LoadInst>(I) || isa<
StoreInst>(I)) && "Expected Load or Store instruction"
) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected Load or Store instruction\""
, "llvm/include/llvm/IR/Instructions.h", 5393, __extension__ __PRETTY_FUNCTION__
))
;
5394 if (auto *LI = dyn_cast<LoadInst>(I))
5395 return LI->getType();
5396 return cast<StoreInst>(I)->getValueOperand()->getType();
5397}
5398
5399/// A helper function that returns an atomic operation's sync scope; returns
5400/// None if it is not an atomic operation.
5401inline Optional<SyncScope::ID> getAtomicSyncScopeID(const Instruction *I) {
5402 if (!I->isAtomic())
5403 return None;
5404 if (auto *AI = dyn_cast<LoadInst>(I))
5405 return AI->getSyncScopeID();
5406 if (auto *AI = dyn_cast<StoreInst>(I))
5407 return AI->getSyncScopeID();
5408 if (auto *AI = dyn_cast<FenceInst>(I))
5409 return AI->getSyncScopeID();
5410 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(I))
5411 return AI->getSyncScopeID();
5412 if (auto *AI = dyn_cast<AtomicRMWInst>(I))
5413 return AI->getSyncScopeID();
5414 llvm_unreachable("unhandled atomic operation")::llvm::llvm_unreachable_internal("unhandled atomic operation"
, "llvm/include/llvm/IR/Instructions.h", 5414)
;
5415}
5416
5417//===----------------------------------------------------------------------===//
5418// FreezeInst Class
5419//===----------------------------------------------------------------------===//
5420
5421/// This class represents a freeze function that returns random concrete
5422/// value if an operand is either a poison value or an undef value
5423class FreezeInst : public UnaryInstruction {
5424protected:
5425 // Note: Instruction needs to be a friend here to call cloneImpl.
5426 friend class Instruction;
5427
5428 /// Clone an identical FreezeInst
5429 FreezeInst *cloneImpl() const;
5430
5431public:
5432 explicit FreezeInst(Value *S,
5433 const Twine &NameStr = "",
5434 Instruction *InsertBefore = nullptr);
5435 FreezeInst(Value *S, const Twine &NameStr, BasicBlock *InsertAtEnd);
5436
5437 // Methods for support type inquiry through isa, cast, and dyn_cast:
5438 static inline bool classof(const Instruction *I) {
5439 return I->getOpcode() == Freeze;
5440 }
5441 static inline bool classof(const Value *V) {
5442 return isa<Instruction>(V) && classof(cast<Instruction>(V));
5443 }
5444};
5445
5446} // end namespace llvm
5447
5448#endif // LLVM_IR_INSTRUCTIONS_H