Bug Summary

File:llvm/lib/Transforms/Utils/SimplifyCFG.cpp
Warning:line 4144, column 24
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 -disable-llvm-verifier -discard-value-names -main-file-name SimplifyCFG.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -fhalf-no-semantic-interposition -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/lib/Transforms/Utils -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2021-01-24-223304-31662-1 -x c++ /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
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// Peephole optimize the CFG.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/ADT/APInt.h"
14#include "llvm/ADT/ArrayRef.h"
15#include "llvm/ADT/DenseMap.h"
16#include "llvm/ADT/MapVector.h"
17#include "llvm/ADT/Optional.h"
18#include "llvm/ADT/STLExtras.h"
19#include "llvm/ADT/ScopeExit.h"
20#include "llvm/ADT/Sequence.h"
21#include "llvm/ADT/SetOperations.h"
22#include "llvm/ADT/SetVector.h"
23#include "llvm/ADT/SmallPtrSet.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/Statistic.h"
26#include "llvm/ADT/StringRef.h"
27#include "llvm/Analysis/AssumptionCache.h"
28#include "llvm/Analysis/ConstantFolding.h"
29#include "llvm/Analysis/EHPersonalities.h"
30#include "llvm/Analysis/GuardUtils.h"
31#include "llvm/Analysis/InstructionSimplify.h"
32#include "llvm/Analysis/MemorySSA.h"
33#include "llvm/Analysis/MemorySSAUpdater.h"
34#include "llvm/Analysis/TargetTransformInfo.h"
35#include "llvm/Analysis/ValueTracking.h"
36#include "llvm/IR/Attributes.h"
37#include "llvm/IR/BasicBlock.h"
38#include "llvm/IR/CFG.h"
39#include "llvm/IR/Constant.h"
40#include "llvm/IR/ConstantRange.h"
41#include "llvm/IR/Constants.h"
42#include "llvm/IR/DataLayout.h"
43#include "llvm/IR/DerivedTypes.h"
44#include "llvm/IR/Function.h"
45#include "llvm/IR/GlobalValue.h"
46#include "llvm/IR/GlobalVariable.h"
47#include "llvm/IR/IRBuilder.h"
48#include "llvm/IR/InstrTypes.h"
49#include "llvm/IR/Instruction.h"
50#include "llvm/IR/Instructions.h"
51#include "llvm/IR/IntrinsicInst.h"
52#include "llvm/IR/Intrinsics.h"
53#include "llvm/IR/LLVMContext.h"
54#include "llvm/IR/MDBuilder.h"
55#include "llvm/IR/Metadata.h"
56#include "llvm/IR/Module.h"
57#include "llvm/IR/NoFolder.h"
58#include "llvm/IR/Operator.h"
59#include "llvm/IR/PatternMatch.h"
60#include "llvm/IR/Type.h"
61#include "llvm/IR/Use.h"
62#include "llvm/IR/User.h"
63#include "llvm/IR/Value.h"
64#include "llvm/IR/ValueHandle.h"
65#include "llvm/Support/Casting.h"
66#include "llvm/Support/CommandLine.h"
67#include "llvm/Support/Debug.h"
68#include "llvm/Support/ErrorHandling.h"
69#include "llvm/Support/KnownBits.h"
70#include "llvm/Support/MathExtras.h"
71#include "llvm/Support/raw_ostream.h"
72#include "llvm/Transforms/Utils/BasicBlockUtils.h"
73#include "llvm/Transforms/Utils/Local.h"
74#include "llvm/Transforms/Utils/SSAUpdater.h"
75#include "llvm/Transforms/Utils/ValueMapper.h"
76#include <algorithm>
77#include <cassert>
78#include <climits>
79#include <cstddef>
80#include <cstdint>
81#include <iterator>
82#include <map>
83#include <set>
84#include <tuple>
85#include <utility>
86#include <vector>
87
88using namespace llvm;
89using namespace PatternMatch;
90
91#define DEBUG_TYPE"simplifycfg" "simplifycfg"
92
93cl::opt<bool> llvm::RequireAndPreserveDomTree(
94 "simplifycfg-require-and-preserve-domtree", cl::Hidden, cl::ZeroOrMore,
95 cl::init(false),
96 cl::desc("Temorary development switch used to gradually uplift SimplifyCFG "
97 "into preserving DomTree,"));
98
99// Chosen as 2 so as to be cheap, but still to have enough power to fold
100// a select, so the "clamp" idiom (of a min followed by a max) will be caught.
101// To catch this, we need to fold a compare and a select, hence '2' being the
102// minimum reasonable default.
103static cl::opt<unsigned> PHINodeFoldingThreshold(
104 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
105 cl::desc(
106 "Control the amount of phi node folding to perform (default = 2)"));
107
108static cl::opt<unsigned> TwoEntryPHINodeFoldingThreshold(
109 "two-entry-phi-node-folding-threshold", cl::Hidden, cl::init(4),
110 cl::desc("Control the maximal total instruction cost that we are willing "
111 "to speculatively execute to fold a 2-entry PHI node into a "
112 "select (default = 4)"));
113
114static cl::opt<bool> DupRet(
115 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
116 cl::desc("Duplicate return instructions into unconditional branches"));
117
118static cl::opt<bool>
119 HoistCommon("simplifycfg-hoist-common", cl::Hidden, cl::init(true),
120 cl::desc("Hoist common instructions up to the parent block"));
121
122static cl::opt<bool>
123 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
124 cl::desc("Sink common instructions down to the end block"));
125
126static cl::opt<bool> HoistCondStores(
127 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
128 cl::desc("Hoist conditional stores if an unconditional store precedes"));
129
130static cl::opt<bool> MergeCondStores(
131 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
132 cl::desc("Hoist conditional stores even if an unconditional store does not "
133 "precede - hoist multiple conditional stores into a single "
134 "predicated store"));
135
136static cl::opt<bool> MergeCondStoresAggressively(
137 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
138 cl::desc("When merging conditional stores, do so even if the resultant "
139 "basic blocks are unlikely to be if-converted as a result"));
140
141static cl::opt<bool> SpeculateOneExpensiveInst(
142 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
143 cl::desc("Allow exactly one expensive instruction to be speculatively "
144 "executed"));
145
146static cl::opt<unsigned> MaxSpeculationDepth(
147 "max-speculation-depth", cl::Hidden, cl::init(10),
148 cl::desc("Limit maximum recursion depth when calculating costs of "
149 "speculatively executed instructions"));
150
151static cl::opt<int>
152MaxSmallBlockSize("simplifycfg-max-small-block-size", cl::Hidden, cl::init(10),
153 cl::desc("Max size of a block which is still considered "
154 "small enough to thread through"));
155
156// Two is chosen to allow one negation and a logical combine.
157static cl::opt<unsigned>
158 BranchFoldThreshold("simplifycfg-branch-fold-threshold", cl::Hidden,
159 cl::init(2),
160 cl::desc("Maximum cost of combining conditions when "
161 "folding branches"));
162
163STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps")static llvm::Statistic NumBitMaps = {"simplifycfg", "NumBitMaps"
, "Number of switch instructions turned into bitmaps"}
;
164STATISTIC(NumLinearMaps,static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping"}
165 "Number of switch instructions turned into linear mapping")static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping"}
;
166STATISTIC(NumLookupTables,static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables"}
167 "Number of switch instructions turned into lookup tables")static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables"}
;
168STATISTIC(static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
}
169 NumLookupTablesHoles,static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
}
170 "Number of switch instructions turned into lookup tables (holes checked)")static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
}
;
171STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares")static llvm::Statistic NumTableCmpReuses = {"simplifycfg", "NumTableCmpReuses"
, "Number of reused switch table lookup compares"}
;
172STATISTIC(NumFoldValueComparisonIntoPredecessors,static llvm::Statistic NumFoldValueComparisonIntoPredecessors
= {"simplifycfg", "NumFoldValueComparisonIntoPredecessors", "Number of value comparisons folded into predecessor basic blocks"
}
173 "Number of value comparisons folded into predecessor basic blocks")static llvm::Statistic NumFoldValueComparisonIntoPredecessors
= {"simplifycfg", "NumFoldValueComparisonIntoPredecessors", "Number of value comparisons folded into predecessor basic blocks"
}
;
174STATISTIC(NumFoldBranchToCommonDest,static llvm::Statistic NumFoldBranchToCommonDest = {"simplifycfg"
, "NumFoldBranchToCommonDest", "Number of branches folded into predecessor basic block"
}
175 "Number of branches folded into predecessor basic block")static llvm::Statistic NumFoldBranchToCommonDest = {"simplifycfg"
, "NumFoldBranchToCommonDest", "Number of branches folded into predecessor basic block"
}
;
176STATISTIC(static llvm::Statistic NumHoistCommonCode = {"simplifycfg", "NumHoistCommonCode"
, "Number of common instruction 'blocks' hoisted up to the begin block"
}
177 NumHoistCommonCode,static llvm::Statistic NumHoistCommonCode = {"simplifycfg", "NumHoistCommonCode"
, "Number of common instruction 'blocks' hoisted up to the begin block"
}
178 "Number of common instruction 'blocks' hoisted up to the begin block")static llvm::Statistic NumHoistCommonCode = {"simplifycfg", "NumHoistCommonCode"
, "Number of common instruction 'blocks' hoisted up to the begin block"
}
;
179STATISTIC(NumHoistCommonInstrs,static llvm::Statistic NumHoistCommonInstrs = {"simplifycfg",
"NumHoistCommonInstrs", "Number of common instructions hoisted up to the begin block"
}
180 "Number of common instructions hoisted up to the begin block")static llvm::Statistic NumHoistCommonInstrs = {"simplifycfg",
"NumHoistCommonInstrs", "Number of common instructions hoisted up to the begin block"
}
;
181STATISTIC(NumSinkCommonCode,static llvm::Statistic NumSinkCommonCode = {"simplifycfg", "NumSinkCommonCode"
, "Number of common instruction 'blocks' sunk down to the end block"
}
182 "Number of common instruction 'blocks' sunk down to the end block")static llvm::Statistic NumSinkCommonCode = {"simplifycfg", "NumSinkCommonCode"
, "Number of common instruction 'blocks' sunk down to the end block"
}
;
183STATISTIC(NumSinkCommonInstrs,static llvm::Statistic NumSinkCommonInstrs = {"simplifycfg", "NumSinkCommonInstrs"
, "Number of common instructions sunk down to the end block"}
184 "Number of common instructions sunk down to the end block")static llvm::Statistic NumSinkCommonInstrs = {"simplifycfg", "NumSinkCommonInstrs"
, "Number of common instructions sunk down to the end block"}
;
185STATISTIC(NumSpeculations, "Number of speculative executed instructions")static llvm::Statistic NumSpeculations = {"simplifycfg", "NumSpeculations"
, "Number of speculative executed instructions"}
;
186STATISTIC(NumInvokes,static llvm::Statistic NumInvokes = {"simplifycfg", "NumInvokes"
, "Number of invokes with empty resume blocks simplified into calls"
}
187 "Number of invokes with empty resume blocks simplified into calls")static llvm::Statistic NumInvokes = {"simplifycfg", "NumInvokes"
, "Number of invokes with empty resume blocks simplified into calls"
}
;
188
189namespace {
190
191// The first field contains the value that the switch produces when a certain
192// case group is selected, and the second field is a vector containing the
193// cases composing the case group.
194using SwitchCaseResultVectorTy =
195 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
196
197// The first field contains the phi node that generates a result of the switch
198// and the second field contains the value generated for a certain case in the
199// switch for that PHI.
200using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
201
202/// ValueEqualityComparisonCase - Represents a case of a switch.
203struct ValueEqualityComparisonCase {
204 ConstantInt *Value;
205 BasicBlock *Dest;
206
207 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
208 : Value(Value), Dest(Dest) {}
209
210 bool operator<(ValueEqualityComparisonCase RHS) const {
211 // Comparing pointers is ok as we only rely on the order for uniquing.
212 return Value < RHS.Value;
213 }
214
215 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
216};
217
218class SimplifyCFGOpt {
219 const TargetTransformInfo &TTI;
220 DomTreeUpdater *DTU;
221 const DataLayout &DL;
222 ArrayRef<WeakVH> LoopHeaders;
223 const SimplifyCFGOptions &Options;
224 bool Resimplify;
225
226 Value *isValueEqualityComparison(Instruction *TI);
227 BasicBlock *GetValueEqualityComparisonCases(
228 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
229 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
230 BasicBlock *Pred,
231 IRBuilder<> &Builder);
232 bool PerformValueComparisonIntoPredecessorFolding(Instruction *TI, Value *&CV,
233 Instruction *PTI,
234 IRBuilder<> &Builder);
235 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
236 IRBuilder<> &Builder);
237
238 bool simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
239 bool simplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
240 bool simplifySingleResume(ResumeInst *RI);
241 bool simplifyCommonResume(ResumeInst *RI);
242 bool simplifyCleanupReturn(CleanupReturnInst *RI);
243 bool simplifyUnreachable(UnreachableInst *UI);
244 bool simplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
245 bool simplifyIndirectBr(IndirectBrInst *IBI);
246 bool simplifyBranch(BranchInst *Branch, IRBuilder<> &Builder);
247 bool simplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
248 bool simplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
249 bool SimplifyCondBranchToTwoReturns(BranchInst *BI, IRBuilder<> &Builder);
250
251 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
252 IRBuilder<> &Builder);
253
254 bool HoistThenElseCodeToIf(BranchInst *BI, const TargetTransformInfo &TTI);
255 bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
256 const TargetTransformInfo &TTI);
257 bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
258 BasicBlock *TrueBB, BasicBlock *FalseBB,
259 uint32_t TrueWeight, uint32_t FalseWeight);
260 bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
261 const DataLayout &DL);
262 bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select);
263 bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI);
264 bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder);
265
266public:
267 SimplifyCFGOpt(const TargetTransformInfo &TTI, DomTreeUpdater *DTU,
268 const DataLayout &DL, ArrayRef<WeakVH> LoopHeaders,
269 const SimplifyCFGOptions &Opts)
270 : TTI(TTI), DTU(DTU), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {
271 assert((!DTU || !DTU->hasPostDomTree()) &&(((!DTU || !DTU->hasPostDomTree()) && "SimplifyCFG is not yet capable of maintaining validity of a "
"PostDomTree, so don't ask for it.") ? static_cast<void>
(0) : __assert_fail ("(!DTU || !DTU->hasPostDomTree()) && \"SimplifyCFG is not yet capable of maintaining validity of a \" \"PostDomTree, so don't ask for it.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 273, __PRETTY_FUNCTION__))
272 "SimplifyCFG is not yet capable of maintaining validity of a "(((!DTU || !DTU->hasPostDomTree()) && "SimplifyCFG is not yet capable of maintaining validity of a "
"PostDomTree, so don't ask for it.") ? static_cast<void>
(0) : __assert_fail ("(!DTU || !DTU->hasPostDomTree()) && \"SimplifyCFG is not yet capable of maintaining validity of a \" \"PostDomTree, so don't ask for it.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 273, __PRETTY_FUNCTION__))
273 "PostDomTree, so don't ask for it.")(((!DTU || !DTU->hasPostDomTree()) && "SimplifyCFG is not yet capable of maintaining validity of a "
"PostDomTree, so don't ask for it.") ? static_cast<void>
(0) : __assert_fail ("(!DTU || !DTU->hasPostDomTree()) && \"SimplifyCFG is not yet capable of maintaining validity of a \" \"PostDomTree, so don't ask for it.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 273, __PRETTY_FUNCTION__))
;
274 }
275
276 bool simplifyOnce(BasicBlock *BB);
277 bool simplifyOnceImpl(BasicBlock *BB);
278 bool run(BasicBlock *BB);
279
280 // Helper to set Resimplify and return change indication.
281 bool requestResimplify() {
282 Resimplify = true;
283 return true;
284 }
285};
286
287} // end anonymous namespace
288
289/// Return true if it is safe to merge these two
290/// terminator instructions together.
291static bool
292SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
293 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
294 if (SI1 == SI2)
295 return false; // Can't merge with self!
296
297 // It is not safe to merge these two switch instructions if they have a common
298 // successor, and if that successor has a PHI node, and if *that* PHI node has
299 // conflicting incoming values from the two switch blocks.
300 BasicBlock *SI1BB = SI1->getParent();
301 BasicBlock *SI2BB = SI2->getParent();
302
303 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
304 bool Fail = false;
305 for (BasicBlock *Succ : successors(SI2BB))
306 if (SI1Succs.count(Succ))
307 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
308 PHINode *PN = cast<PHINode>(BBI);
309 if (PN->getIncomingValueForBlock(SI1BB) !=
310 PN->getIncomingValueForBlock(SI2BB)) {
311 if (FailBlocks)
312 FailBlocks->insert(Succ);
313 Fail = true;
314 }
315 }
316
317 return !Fail;
318}
319
320/// Update PHI nodes in Succ to indicate that there will now be entries in it
321/// from the 'NewPred' block. The values that will be flowing into the PHI nodes
322/// will be the same as those coming in from ExistPred, an existing predecessor
323/// of Succ.
324static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
325 BasicBlock *ExistPred,
326 MemorySSAUpdater *MSSAU = nullptr) {
327 for (PHINode &PN : Succ->phis())
328 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
329 if (MSSAU)
330 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
331 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
332}
333
334/// Compute an abstract "cost" of speculating the given instruction,
335/// which is assumed to be safe to speculate. TCC_Free means cheap,
336/// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
337/// expensive.
338static unsigned ComputeSpeculationCost(const User *I,
339 const TargetTransformInfo &TTI) {
340 assert(isSafeToSpeculativelyExecute(I) &&((isSafeToSpeculativelyExecute(I) && "Instruction is not safe to speculatively execute!"
) ? static_cast<void> (0) : __assert_fail ("isSafeToSpeculativelyExecute(I) && \"Instruction is not safe to speculatively execute!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 341, __PRETTY_FUNCTION__))
341 "Instruction is not safe to speculatively execute!")((isSafeToSpeculativelyExecute(I) && "Instruction is not safe to speculatively execute!"
) ? static_cast<void> (0) : __assert_fail ("isSafeToSpeculativelyExecute(I) && \"Instruction is not safe to speculatively execute!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 341, __PRETTY_FUNCTION__))
;
342 return TTI.getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency);
343}
344
345/// If we have a merge point of an "if condition" as accepted above,
346/// return true if the specified value dominates the block. We
347/// don't handle the true generality of domination here, just a special case
348/// which works well enough for us.
349///
350/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
351/// see if V (which must be an instruction) and its recursive operands
352/// that do not dominate BB have a combined cost lower than CostRemaining and
353/// are non-trapping. If both are true, the instruction is inserted into the
354/// set and true is returned.
355///
356/// The cost for most non-trapping instructions is defined as 1 except for
357/// Select whose cost is 2.
358///
359/// After this function returns, CostRemaining is decreased by the cost of
360/// V plus its non-dominating operands. If that cost is greater than
361/// CostRemaining, false is returned and CostRemaining is undefined.
362static bool DominatesMergePoint(Value *V, BasicBlock *BB,
363 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
364 int &BudgetRemaining,
365 const TargetTransformInfo &TTI,
366 unsigned Depth = 0) {
367 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
368 // so limit the recursion depth.
369 // TODO: While this recursion limit does prevent pathological behavior, it
370 // would be better to track visited instructions to avoid cycles.
371 if (Depth == MaxSpeculationDepth)
372 return false;
373
374 Instruction *I = dyn_cast<Instruction>(V);
375 if (!I) {
376 // Non-instructions all dominate instructions, but not all constantexprs
377 // can be executed unconditionally.
378 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
379 if (C->canTrap())
380 return false;
381 return true;
382 }
383 BasicBlock *PBB = I->getParent();
384
385 // We don't want to allow weird loops that might have the "if condition" in
386 // the bottom of this block.
387 if (PBB == BB)
388 return false;
389
390 // If this instruction is defined in a block that contains an unconditional
391 // branch to BB, then it must be in the 'conditional' part of the "if
392 // statement". If not, it definitely dominates the region.
393 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
394 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
395 return true;
396
397 // If we have seen this instruction before, don't count it again.
398 if (AggressiveInsts.count(I))
399 return true;
400
401 // Okay, it looks like the instruction IS in the "condition". Check to
402 // see if it's a cheap instruction to unconditionally compute, and if it
403 // only uses stuff defined outside of the condition. If so, hoist it out.
404 if (!isSafeToSpeculativelyExecute(I))
405 return false;
406
407 BudgetRemaining -= ComputeSpeculationCost(I, TTI);
408
409 // Allow exactly one instruction to be speculated regardless of its cost
410 // (as long as it is safe to do so).
411 // This is intended to flatten the CFG even if the instruction is a division
412 // or other expensive operation. The speculation of an expensive instruction
413 // is expected to be undone in CodeGenPrepare if the speculation has not
414 // enabled further IR optimizations.
415 if (BudgetRemaining < 0 &&
416 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
417 return false;
418
419 // Okay, we can only really hoist these out if their operands do
420 // not take us over the cost threshold.
421 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
422 if (!DominatesMergePoint(*i, BB, AggressiveInsts, BudgetRemaining, TTI,
423 Depth + 1))
424 return false;
425 // Okay, it's safe to do this! Remember this instruction.
426 AggressiveInsts.insert(I);
427 return true;
428}
429
430/// Extract ConstantInt from value, looking through IntToPtr
431/// and PointerNullValue. Return NULL if value is not a constant int.
432static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
433 // Normal constant int.
434 ConstantInt *CI = dyn_cast<ConstantInt>(V);
435 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
436 return CI;
437
438 // This is some kind of pointer constant. Turn it into a pointer-sized
439 // ConstantInt if possible.
440 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
441
442 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
443 if (isa<ConstantPointerNull>(V))
444 return ConstantInt::get(PtrTy, 0);
445
446 // IntToPtr const int.
447 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
448 if (CE->getOpcode() == Instruction::IntToPtr)
449 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
450 // The constant is very likely to have the right type already.
451 if (CI->getType() == PtrTy)
452 return CI;
453 else
454 return cast<ConstantInt>(
455 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
456 }
457 return nullptr;
458}
459
460namespace {
461
462/// Given a chain of or (||) or and (&&) comparison of a value against a
463/// constant, this will try to recover the information required for a switch
464/// structure.
465/// It will depth-first traverse the chain of comparison, seeking for patterns
466/// like %a == 12 or %a < 4 and combine them to produce a set of integer
467/// representing the different cases for the switch.
468/// Note that if the chain is composed of '||' it will build the set of elements
469/// that matches the comparisons (i.e. any of this value validate the chain)
470/// while for a chain of '&&' it will build the set elements that make the test
471/// fail.
472struct ConstantComparesGatherer {
473 const DataLayout &DL;
474
475 /// Value found for the switch comparison
476 Value *CompValue = nullptr;
477
478 /// Extra clause to be checked before the switch
479 Value *Extra = nullptr;
480
481 /// Set of integers to match in switch
482 SmallVector<ConstantInt *, 8> Vals;
483
484 /// Number of comparisons matched in the and/or chain
485 unsigned UsedICmps = 0;
486
487 /// Construct and compute the result for the comparison instruction Cond
488 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
489 gather(Cond);
490 }
491
492 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
493 ConstantComparesGatherer &
494 operator=(const ConstantComparesGatherer &) = delete;
495
496private:
497 /// Try to set the current value used for the comparison, it succeeds only if
498 /// it wasn't set before or if the new value is the same as the old one
499 bool setValueOnce(Value *NewVal) {
500 if (CompValue && CompValue != NewVal)
501 return false;
502 CompValue = NewVal;
503 return (CompValue != nullptr);
504 }
505
506 /// Try to match Instruction "I" as a comparison against a constant and
507 /// populates the array Vals with the set of values that match (or do not
508 /// match depending on isEQ).
509 /// Return false on failure. On success, the Value the comparison matched
510 /// against is placed in CompValue.
511 /// If CompValue is already set, the function is expected to fail if a match
512 /// is found but the value compared to is different.
513 bool matchInstruction(Instruction *I, bool isEQ) {
514 // If this is an icmp against a constant, handle this as one of the cases.
515 ICmpInst *ICI;
516 ConstantInt *C;
517 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
518 (C = GetConstantInt(I->getOperand(1), DL)))) {
519 return false;
520 }
521
522 Value *RHSVal;
523 const APInt *RHSC;
524
525 // Pattern match a special case
526 // (x & ~2^z) == y --> x == y || x == y|2^z
527 // This undoes a transformation done by instcombine to fuse 2 compares.
528 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
529 // It's a little bit hard to see why the following transformations are
530 // correct. Here is a CVC3 program to verify them for 64-bit values:
531
532 /*
533 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
534 x : BITVECTOR(64);
535 y : BITVECTOR(64);
536 z : BITVECTOR(64);
537 mask : BITVECTOR(64) = BVSHL(ONE, z);
538 QUERY( (y & ~mask = y) =>
539 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
540 );
541 QUERY( (y | mask = y) =>
542 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
543 );
544 */
545
546 // Please note that each pattern must be a dual implication (<--> or
547 // iff). One directional implication can create spurious matches. If the
548 // implication is only one-way, an unsatisfiable condition on the left
549 // side can imply a satisfiable condition on the right side. Dual
550 // implication ensures that satisfiable conditions are transformed to
551 // other satisfiable conditions and unsatisfiable conditions are
552 // transformed to other unsatisfiable conditions.
553
554 // Here is a concrete example of a unsatisfiable condition on the left
555 // implying a satisfiable condition on the right:
556 //
557 // mask = (1 << z)
558 // (x & ~mask) == y --> (x == y || x == (y | mask))
559 //
560 // Substituting y = 3, z = 0 yields:
561 // (x & -2) == 3 --> (x == 3 || x == 2)
562
563 // Pattern match a special case:
564 /*
565 QUERY( (y & ~mask = y) =>
566 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
567 );
568 */
569 if (match(ICI->getOperand(0),
570 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
571 APInt Mask = ~*RHSC;
572 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
573 // If we already have a value for the switch, it has to match!
574 if (!setValueOnce(RHSVal))
575 return false;
576
577 Vals.push_back(C);
578 Vals.push_back(
579 ConstantInt::get(C->getContext(),
580 C->getValue() | Mask));
581 UsedICmps++;
582 return true;
583 }
584 }
585
586 // Pattern match a special case:
587 /*
588 QUERY( (y | mask = y) =>
589 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
590 );
591 */
592 if (match(ICI->getOperand(0),
593 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
594 APInt Mask = *RHSC;
595 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
596 // If we already have a value for the switch, it has to match!
597 if (!setValueOnce(RHSVal))
598 return false;
599
600 Vals.push_back(C);
601 Vals.push_back(ConstantInt::get(C->getContext(),
602 C->getValue() & ~Mask));
603 UsedICmps++;
604 return true;
605 }
606 }
607
608 // If we already have a value for the switch, it has to match!
609 if (!setValueOnce(ICI->getOperand(0)))
610 return false;
611
612 UsedICmps++;
613 Vals.push_back(C);
614 return ICI->getOperand(0);
615 }
616
617 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
618 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
619 ICI->getPredicate(), C->getValue());
620
621 // Shift the range if the compare is fed by an add. This is the range
622 // compare idiom as emitted by instcombine.
623 Value *CandidateVal = I->getOperand(0);
624 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
625 Span = Span.subtract(*RHSC);
626 CandidateVal = RHSVal;
627 }
628
629 // If this is an and/!= check, then we are looking to build the set of
630 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
631 // x != 0 && x != 1.
632 if (!isEQ)
633 Span = Span.inverse();
634
635 // If there are a ton of values, we don't want to make a ginormous switch.
636 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
637 return false;
638 }
639
640 // If we already have a value for the switch, it has to match!
641 if (!setValueOnce(CandidateVal))
642 return false;
643
644 // Add all values from the range to the set
645 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
646 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
647
648 UsedICmps++;
649 return true;
650 }
651
652 /// Given a potentially 'or'd or 'and'd together collection of icmp
653 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
654 /// the value being compared, and stick the list constants into the Vals
655 /// vector.
656 /// One "Extra" case is allowed to differ from the other.
657 void gather(Value *V) {
658 bool isEQ = match(V, m_LogicalOr(m_Value(), m_Value()));
659
660 // Keep a stack (SmallVector for efficiency) for depth-first traversal
661 SmallVector<Value *, 8> DFT;
662 SmallPtrSet<Value *, 8> Visited;
663
664 // Initialize
665 Visited.insert(V);
666 DFT.push_back(V);
667
668 while (!DFT.empty()) {
669 V = DFT.pop_back_val();
670
671 if (Instruction *I = dyn_cast<Instruction>(V)) {
672 // If it is a || (or && depending on isEQ), process the operands.
673 Value *Op0, *Op1;
674 if (isEQ ? match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1)))
675 : match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
676 if (Visited.insert(Op1).second)
677 DFT.push_back(Op1);
678 if (Visited.insert(Op0).second)
679 DFT.push_back(Op0);
680
681 continue;
682 }
683
684 // Try to match the current instruction
685 if (matchInstruction(I, isEQ))
686 // Match succeed, continue the loop
687 continue;
688 }
689
690 // One element of the sequence of || (or &&) could not be match as a
691 // comparison against the same value as the others.
692 // We allow only one "Extra" case to be checked before the switch
693 if (!Extra) {
694 Extra = V;
695 continue;
696 }
697 // Failed to parse a proper sequence, abort now
698 CompValue = nullptr;
699 break;
700 }
701 }
702};
703
704} // end anonymous namespace
705
706static void EraseTerminatorAndDCECond(Instruction *TI,
707 MemorySSAUpdater *MSSAU = nullptr) {
708 Instruction *Cond = nullptr;
709 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
710 Cond = dyn_cast<Instruction>(SI->getCondition());
711 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
712 if (BI->isConditional())
713 Cond = dyn_cast<Instruction>(BI->getCondition());
714 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
715 Cond = dyn_cast<Instruction>(IBI->getAddress());
716 }
717
718 TI->eraseFromParent();
719 if (Cond)
720 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
721}
722
723/// Return true if the specified terminator checks
724/// to see if a value is equal to constant integer value.
725Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
726 Value *CV = nullptr;
727 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
728 // Do not permit merging of large switch instructions into their
729 // predecessors unless there is only one predecessor.
730 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
731 CV = SI->getCondition();
732 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
733 if (BI->isConditional() && BI->getCondition()->hasOneUse())
734 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
735 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
736 CV = ICI->getOperand(0);
737 }
738
739 // Unwrap any lossless ptrtoint cast.
740 if (CV) {
741 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
742 Value *Ptr = PTII->getPointerOperand();
743 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
744 CV = Ptr;
745 }
746 }
747 return CV;
748}
749
750/// Given a value comparison instruction,
751/// decode all of the 'cases' that it represents and return the 'default' block.
752BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
753 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
754 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
755 Cases.reserve(SI->getNumCases());
756 for (auto Case : SI->cases())
757 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
758 Case.getCaseSuccessor()));
759 return SI->getDefaultDest();
760 }
761
762 BranchInst *BI = cast<BranchInst>(TI);
763 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
764 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
765 Cases.push_back(ValueEqualityComparisonCase(
766 GetConstantInt(ICI->getOperand(1), DL), Succ));
767 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
768}
769
770/// Given a vector of bb/value pairs, remove any entries
771/// in the list that match the specified block.
772static void
773EliminateBlockCases(BasicBlock *BB,
774 std::vector<ValueEqualityComparisonCase> &Cases) {
775 llvm::erase_value(Cases, BB);
776}
777
778/// Return true if there are any keys in C1 that exist in C2 as well.
779static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
780 std::vector<ValueEqualityComparisonCase> &C2) {
781 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
782
783 // Make V1 be smaller than V2.
784 if (V1->size() > V2->size())
785 std::swap(V1, V2);
786
787 if (V1->empty())
788 return false;
789 if (V1->size() == 1) {
790 // Just scan V2.
791 ConstantInt *TheVal = (*V1)[0].Value;
792 for (unsigned i = 0, e = V2->size(); i != e; ++i)
793 if (TheVal == (*V2)[i].Value)
794 return true;
795 }
796
797 // Otherwise, just sort both lists and compare element by element.
798 array_pod_sort(V1->begin(), V1->end());
799 array_pod_sort(V2->begin(), V2->end());
800 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
801 while (i1 != e1 && i2 != e2) {
802 if ((*V1)[i1].Value == (*V2)[i2].Value)
803 return true;
804 if ((*V1)[i1].Value < (*V2)[i2].Value)
805 ++i1;
806 else
807 ++i2;
808 }
809 return false;
810}
811
812// Set branch weights on SwitchInst. This sets the metadata if there is at
813// least one non-zero weight.
814static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
815 // Check that there is at least one non-zero weight. Otherwise, pass
816 // nullptr to setMetadata which will erase the existing metadata.
817 MDNode *N = nullptr;
818 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
819 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
820 SI->setMetadata(LLVMContext::MD_prof, N);
821}
822
823// Similar to the above, but for branch and select instructions that take
824// exactly 2 weights.
825static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
826 uint32_t FalseWeight) {
827 assert(isa<BranchInst>(I) || isa<SelectInst>(I))((isa<BranchInst>(I) || isa<SelectInst>(I)) ? static_cast
<void> (0) : __assert_fail ("isa<BranchInst>(I) || isa<SelectInst>(I)"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 827, __PRETTY_FUNCTION__))
;
828 // Check that there is at least one non-zero weight. Otherwise, pass
829 // nullptr to setMetadata which will erase the existing metadata.
830 MDNode *N = nullptr;
831 if (TrueWeight || FalseWeight)
832 N = MDBuilder(I->getParent()->getContext())
833 .createBranchWeights(TrueWeight, FalseWeight);
834 I->setMetadata(LLVMContext::MD_prof, N);
835}
836
837/// If TI is known to be a terminator instruction and its block is known to
838/// only have a single predecessor block, check to see if that predecessor is
839/// also a value comparison with the same value, and if that comparison
840/// determines the outcome of this comparison. If so, simplify TI. This does a
841/// very limited form of jump threading.
842bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
843 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
844 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
845 if (!PredVal)
846 return false; // Not a value comparison in predecessor.
847
848 Value *ThisVal = isValueEqualityComparison(TI);
849 assert(ThisVal && "This isn't a value comparison!!")((ThisVal && "This isn't a value comparison!!") ? static_cast
<void> (0) : __assert_fail ("ThisVal && \"This isn't a value comparison!!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 849, __PRETTY_FUNCTION__))
;
850 if (ThisVal != PredVal)
851 return false; // Different predicates.
852
853 // TODO: Preserve branch weight metadata, similarly to how
854 // FoldValueComparisonIntoPredecessors preserves it.
855
856 // Find out information about when control will move from Pred to TI's block.
857 std::vector<ValueEqualityComparisonCase> PredCases;
858 BasicBlock *PredDef =
859 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
860 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
861
862 // Find information about how control leaves this block.
863 std::vector<ValueEqualityComparisonCase> ThisCases;
864 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
865 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
866
867 // If TI's block is the default block from Pred's comparison, potentially
868 // simplify TI based on this knowledge.
869 if (PredDef == TI->getParent()) {
870 // If we are here, we know that the value is none of those cases listed in
871 // PredCases. If there are any cases in ThisCases that are in PredCases, we
872 // can simplify TI.
873 if (!ValuesOverlap(PredCases, ThisCases))
874 return false;
875
876 if (isa<BranchInst>(TI)) {
877 // Okay, one of the successors of this condbr is dead. Convert it to a
878 // uncond br.
879 assert(ThisCases.size() == 1 && "Branch can only have one case!")((ThisCases.size() == 1 && "Branch can only have one case!"
) ? static_cast<void> (0) : __assert_fail ("ThisCases.size() == 1 && \"Branch can only have one case!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 879, __PRETTY_FUNCTION__))
;
880 // Insert the new branch.
881 Instruction *NI = Builder.CreateBr(ThisDef);
882 (void)NI;
883
884 // Remove PHI node entries for the dead edge.
885 ThisCases[0].Dest->removePredecessor(PredDef);
886
887 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
888 << "Through successor TI: " << *TI << "Leaving: " << *NIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
889 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
890
891 EraseTerminatorAndDCECond(TI);
892
893 if (DTU)
894 DTU->applyUpdates(
895 {{DominatorTree::Delete, PredDef, ThisCases[0].Dest}});
896
897 return true;
898 }
899
900 SwitchInstProfUpdateWrapper SI = *cast<SwitchInst>(TI);
901 // Okay, TI has cases that are statically dead, prune them away.
902 SmallPtrSet<Constant *, 16> DeadCases;
903 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
904 DeadCases.insert(PredCases[i].Value);
905
906 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI; } } while (false)
907 << "Through successor TI: " << *TI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI; } } while (false)
;
908
909 SmallMapVector<BasicBlock *, int, 8> NumPerSuccessorCases;
910 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
911 --i;
912 auto *Successor = i->getCaseSuccessor();
913 ++NumPerSuccessorCases[Successor];
914 if (DeadCases.count(i->getCaseValue())) {
915 Successor->removePredecessor(PredDef);
916 SI.removeCase(i);
917 --NumPerSuccessorCases[Successor];
918 }
919 }
920
921 std::vector<DominatorTree::UpdateType> Updates;
922 for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
923 if (I.second == 0)
924 Updates.push_back({DominatorTree::Delete, PredDef, I.first});
925 if (DTU)
926 DTU->applyUpdates(Updates);
927
928 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Leaving: " << *TI <<
"\n"; } } while (false)
;
929 return true;
930 }
931
932 // Otherwise, TI's block must correspond to some matched value. Find out
933 // which value (or set of values) this is.
934 ConstantInt *TIV = nullptr;
935 BasicBlock *TIBB = TI->getParent();
936 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
937 if (PredCases[i].Dest == TIBB) {
938 if (TIV)
939 return false; // Cannot handle multiple values coming to this block.
940 TIV = PredCases[i].Value;
941 }
942 assert(TIV && "No edge from pred to succ?")((TIV && "No edge from pred to succ?") ? static_cast<
void> (0) : __assert_fail ("TIV && \"No edge from pred to succ?\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 942, __PRETTY_FUNCTION__))
;
943
944 // Okay, we found the one constant that our value can be if we get into TI's
945 // BB. Find out which successor will unconditionally be branched to.
946 BasicBlock *TheRealDest = nullptr;
947 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
948 if (ThisCases[i].Value == TIV) {
949 TheRealDest = ThisCases[i].Dest;
950 break;
951 }
952
953 // If not handled by any explicit cases, it is handled by the default case.
954 if (!TheRealDest)
955 TheRealDest = ThisDef;
956
957 SmallSetVector<BasicBlock *, 2> RemovedSuccs;
958
959 // Remove PHI node entries for dead edges.
960 BasicBlock *CheckEdge = TheRealDest;
961 for (BasicBlock *Succ : successors(TIBB))
962 if (Succ != CheckEdge) {
963 if (Succ != TheRealDest)
964 RemovedSuccs.insert(Succ);
965 Succ->removePredecessor(TIBB);
966 } else
967 CheckEdge = nullptr;
968
969 // Insert the new branch.
970 Instruction *NI = Builder.CreateBr(TheRealDest);
971 (void)NI;
972
973 LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
974 << "Through successor TI: " << *TI << "Leaving: " << *NIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
975 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
976
977 EraseTerminatorAndDCECond(TI);
978 if (DTU) {
979 SmallVector<DominatorTree::UpdateType, 2> Updates;
980 Updates.reserve(RemovedSuccs.size());
981 for (auto *RemovedSucc : RemovedSuccs)
982 Updates.push_back({DominatorTree::Delete, TIBB, RemovedSucc});
983 DTU->applyUpdates(Updates);
984 }
985 return true;
986}
987
988namespace {
989
990/// This class implements a stable ordering of constant
991/// integers that does not depend on their address. This is important for
992/// applications that sort ConstantInt's to ensure uniqueness.
993struct ConstantIntOrdering {
994 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
995 return LHS->getValue().ult(RHS->getValue());
996 }
997};
998
999} // end anonymous namespace
1000
1001static int ConstantIntSortPredicate(ConstantInt *const *P1,
1002 ConstantInt *const *P2) {
1003 const ConstantInt *LHS = *P1;
1004 const ConstantInt *RHS = *P2;
1005 if (LHS == RHS)
1006 return 0;
1007 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
1008}
1009
1010static inline bool HasBranchWeights(const Instruction *I) {
1011 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
1012 if (ProfMD && ProfMD->getOperand(0))
1013 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
1014 return MDS->getString().equals("branch_weights");
1015
1016 return false;
1017}
1018
1019/// Get Weights of a given terminator, the default weight is at the front
1020/// of the vector. If TI is a conditional eq, we need to swap the branch-weight
1021/// metadata.
1022static void GetBranchWeights(Instruction *TI,
1023 SmallVectorImpl<uint64_t> &Weights) {
1024 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
1025 assert(MD)((MD) ? static_cast<void> (0) : __assert_fail ("MD", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1025, __PRETTY_FUNCTION__))
;
1026 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
1027 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
1028 Weights.push_back(CI->getValue().getZExtValue());
1029 }
1030
1031 // If TI is a conditional eq, the default case is the false case,
1032 // and the corresponding branch-weight data is at index 2. We swap the
1033 // default weight to be the first entry.
1034 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1035 assert(Weights.size() == 2)((Weights.size() == 2) ? static_cast<void> (0) : __assert_fail
("Weights.size() == 2", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1035, __PRETTY_FUNCTION__))
;
1036 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
1037 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
1038 std::swap(Weights.front(), Weights.back());
1039 }
1040}
1041
1042/// Keep halving the weights until all can fit in uint32_t.
1043static void FitWeights(MutableArrayRef<uint64_t> Weights) {
1044 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1045 if (Max > UINT_MAX(2147483647 *2U +1U)) {
1046 unsigned Offset = 32 - countLeadingZeros(Max);
1047 for (uint64_t &I : Weights)
1048 I >>= Offset;
1049 }
1050}
1051
1052static void CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(
1053 BasicBlock *BB, BasicBlock *PredBlock, ValueToValueMapTy &VMap) {
1054 Instruction *PTI = PredBlock->getTerminator();
1055
1056 // If we have bonus instructions, clone them into the predecessor block.
1057 // Note that there may be multiple predecessor blocks, so we cannot move
1058 // bonus instructions to a predecessor block.
1059 for (Instruction &BonusInst : *BB) {
1060 if (isa<DbgInfoIntrinsic>(BonusInst) || BonusInst.isTerminator())
1061 continue;
1062
1063 Instruction *NewBonusInst = BonusInst.clone();
1064
1065 if (PTI->getDebugLoc() != NewBonusInst->getDebugLoc()) {
1066 // Unless the instruction has the same !dbg location as the original
1067 // branch, drop it. When we fold the bonus instructions we want to make
1068 // sure we reset their debug locations in order to avoid stepping on
1069 // dead code caused by folding dead branches.
1070 NewBonusInst->setDebugLoc(DebugLoc());
1071 }
1072
1073 RemapInstruction(NewBonusInst, VMap,
1074 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1075 VMap[&BonusInst] = NewBonusInst;
1076
1077 // If we moved a load, we cannot any longer claim any knowledge about
1078 // its potential value. The previous information might have been valid
1079 // only given the branch precondition.
1080 // For an analogous reason, we must also drop all the metadata whose
1081 // semantics we don't understand. We *can* preserve !annotation, because
1082 // it is tied to the instruction itself, not the value or position.
1083 NewBonusInst->dropUnknownNonDebugMetadata(LLVMContext::MD_annotation);
1084
1085 PredBlock->getInstList().insert(PTI->getIterator(), NewBonusInst);
1086 NewBonusInst->takeName(&BonusInst);
1087 BonusInst.setName(NewBonusInst->getName() + ".old");
1088
1089 // Update (liveout) uses of bonus instructions,
1090 // now that the bonus instruction has been cloned into predecessor.
1091 SSAUpdater SSAUpdate;
1092 SSAUpdate.Initialize(BonusInst.getType(),
1093 (NewBonusInst->getName() + ".merge").str());
1094 SSAUpdate.AddAvailableValue(BB, &BonusInst);
1095 SSAUpdate.AddAvailableValue(PredBlock, NewBonusInst);
1096 for (Use &U : make_early_inc_range(BonusInst.uses()))
1097 SSAUpdate.RewriteUseAfterInsertions(U);
1098 }
1099}
1100
1101bool SimplifyCFGOpt::PerformValueComparisonIntoPredecessorFolding(
1102 Instruction *TI, Value *&CV, Instruction *PTI, IRBuilder<> &Builder) {
1103 BasicBlock *BB = TI->getParent();
1104 BasicBlock *Pred = PTI->getParent();
1105
1106 std::vector<DominatorTree::UpdateType> Updates;
1107
1108 // Figure out which 'cases' to copy from SI to PSI.
1109 std::vector<ValueEqualityComparisonCase> BBCases;
1110 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1111
1112 std::vector<ValueEqualityComparisonCase> PredCases;
1113 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1114
1115 // Based on whether the default edge from PTI goes to BB or not, fill in
1116 // PredCases and PredDefault with the new switch cases we would like to
1117 // build.
1118 SmallMapVector<BasicBlock *, int, 8> NewSuccessors;
1119
1120 // Update the branch weight metadata along the way
1121 SmallVector<uint64_t, 8> Weights;
1122 bool PredHasWeights = HasBranchWeights(PTI);
1123 bool SuccHasWeights = HasBranchWeights(TI);
1124
1125 if (PredHasWeights) {
1126 GetBranchWeights(PTI, Weights);
1127 // branch-weight metadata is inconsistent here.
1128 if (Weights.size() != 1 + PredCases.size())
1129 PredHasWeights = SuccHasWeights = false;
1130 } else if (SuccHasWeights)
1131 // If there are no predecessor weights but there are successor weights,
1132 // populate Weights with 1, which will later be scaled to the sum of
1133 // successor's weights
1134 Weights.assign(1 + PredCases.size(), 1);
1135
1136 SmallVector<uint64_t, 8> SuccWeights;
1137 if (SuccHasWeights) {
1138 GetBranchWeights(TI, SuccWeights);
1139 // branch-weight metadata is inconsistent here.
1140 if (SuccWeights.size() != 1 + BBCases.size())
1141 PredHasWeights = SuccHasWeights = false;
1142 } else if (PredHasWeights)
1143 SuccWeights.assign(1 + BBCases.size(), 1);
1144
1145 if (PredDefault == BB) {
1146 // If this is the default destination from PTI, only the edges in TI
1147 // that don't occur in PTI, or that branch to BB will be activated.
1148 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1149 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1150 if (PredCases[i].Dest != BB)
1151 PTIHandled.insert(PredCases[i].Value);
1152 else {
1153 // The default destination is BB, we don't need explicit targets.
1154 std::swap(PredCases[i], PredCases.back());
1155
1156 if (PredHasWeights || SuccHasWeights) {
1157 // Increase weight for the default case.
1158 Weights[0] += Weights[i + 1];
1159 std::swap(Weights[i + 1], Weights.back());
1160 Weights.pop_back();
1161 }
1162
1163 PredCases.pop_back();
1164 --i;
1165 --e;
1166 }
1167
1168 // Reconstruct the new switch statement we will be building.
1169 if (PredDefault != BBDefault) {
1170 PredDefault->removePredecessor(Pred);
1171 if (PredDefault != BB)
1172 Updates.push_back({DominatorTree::Delete, Pred, PredDefault});
1173 PredDefault = BBDefault;
1174 ++NewSuccessors[BBDefault];
1175 }
1176
1177 unsigned CasesFromPred = Weights.size();
1178 uint64_t ValidTotalSuccWeight = 0;
1179 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1180 if (!PTIHandled.count(BBCases[i].Value) && BBCases[i].Dest != BBDefault) {
1181 PredCases.push_back(BBCases[i]);
1182 ++NewSuccessors[BBCases[i].Dest];
1183 if (SuccHasWeights || PredHasWeights) {
1184 // The default weight is at index 0, so weight for the ith case
1185 // should be at index i+1. Scale the cases from successor by
1186 // PredDefaultWeight (Weights[0]).
1187 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1188 ValidTotalSuccWeight += SuccWeights[i + 1];
1189 }
1190 }
1191
1192 if (SuccHasWeights || PredHasWeights) {
1193 ValidTotalSuccWeight += SuccWeights[0];
1194 // Scale the cases from predecessor by ValidTotalSuccWeight.
1195 for (unsigned i = 1; i < CasesFromPred; ++i)
1196 Weights[i] *= ValidTotalSuccWeight;
1197 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1198 Weights[0] *= SuccWeights[0];
1199 }
1200 } else {
1201 // If this is not the default destination from PSI, only the edges
1202 // in SI that occur in PSI with a destination of BB will be
1203 // activated.
1204 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1205 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1206 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1207 if (PredCases[i].Dest == BB) {
1208 PTIHandled.insert(PredCases[i].Value);
1209
1210 if (PredHasWeights || SuccHasWeights) {
1211 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1212 std::swap(Weights[i + 1], Weights.back());
1213 Weights.pop_back();
1214 }
1215
1216 std::swap(PredCases[i], PredCases.back());
1217 PredCases.pop_back();
1218 --i;
1219 --e;
1220 }
1221
1222 // Okay, now we know which constants were sent to BB from the
1223 // predecessor. Figure out where they will all go now.
1224 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1225 if (PTIHandled.count(BBCases[i].Value)) {
1226 // If this is one we are capable of getting...
1227 if (PredHasWeights || SuccHasWeights)
1228 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1229 PredCases.push_back(BBCases[i]);
1230 ++NewSuccessors[BBCases[i].Dest];
1231 PTIHandled.erase(BBCases[i].Value); // This constant is taken care of
1232 }
1233
1234 // If there are any constants vectored to BB that TI doesn't handle,
1235 // they must go to the default destination of TI.
1236 for (ConstantInt *I : PTIHandled) {
1237 if (PredHasWeights || SuccHasWeights)
1238 Weights.push_back(WeightsForHandled[I]);
1239 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1240 ++NewSuccessors[BBDefault];
1241 }
1242 }
1243
1244 // Okay, at this point, we know which new successor Pred will get. Make
1245 // sure we update the number of entries in the PHI nodes for these
1246 // successors.
1247 for (const std::pair<BasicBlock *, int /*Num*/> &NewSuccessor :
1248 NewSuccessors) {
1249 for (auto I : seq(0, NewSuccessor.second)) {
1250 (void)I;
1251 AddPredecessorToBlock(NewSuccessor.first, Pred, BB);
1252 }
1253 if (!is_contained(successors(Pred), NewSuccessor.first))
1254 Updates.push_back({DominatorTree::Insert, Pred, NewSuccessor.first});
1255 }
1256
1257 Builder.SetInsertPoint(PTI);
1258 // Convert pointer to int before we switch.
1259 if (CV->getType()->isPointerTy()) {
1260 CV =
1261 Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()), "magicptr");
1262 }
1263
1264 // Now that the successors are updated, create the new Switch instruction.
1265 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1266 NewSI->setDebugLoc(PTI->getDebugLoc());
1267 for (ValueEqualityComparisonCase &V : PredCases)
1268 NewSI->addCase(V.Value, V.Dest);
1269
1270 if (PredHasWeights || SuccHasWeights) {
1271 // Halve the weights if any of them cannot fit in an uint32_t
1272 FitWeights(Weights);
1273
1274 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1275
1276 setBranchWeights(NewSI, MDWeights);
1277 }
1278
1279 EraseTerminatorAndDCECond(PTI);
1280
1281 // Okay, last check. If BB is still a successor of PSI, then we must
1282 // have an infinite loop case. If so, add an infinitely looping block
1283 // to handle the case to preserve the behavior of the code.
1284 BasicBlock *InfLoopBlock = nullptr;
1285 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1286 if (NewSI->getSuccessor(i) == BB) {
1287 if (!InfLoopBlock) {
1288 // Insert it at the end of the function, because it's either code,
1289 // or it won't matter if it's hot. :)
1290 InfLoopBlock =
1291 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
1292 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1293 Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
1294 }
1295 NewSI->setSuccessor(i, InfLoopBlock);
1296 }
1297
1298 if (InfLoopBlock)
1299 Updates.push_back({DominatorTree::Insert, Pred, InfLoopBlock});
1300
1301 Updates.push_back({DominatorTree::Delete, Pred, BB});
1302
1303 if (DTU)
1304 DTU->applyUpdates(Updates);
1305
1306 ++NumFoldValueComparisonIntoPredecessors;
1307 return true;
1308}
1309
1310/// The specified terminator is a value equality comparison instruction
1311/// (either a switch or a branch on "X == c").
1312/// See if any of the predecessors of the terminator block are value comparisons
1313/// on the same value. If so, and if safe to do so, fold them together.
1314bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1315 IRBuilder<> &Builder) {
1316 BasicBlock *BB = TI->getParent();
1317 Value *CV = isValueEqualityComparison(TI); // CondVal
1318 assert(CV && "Not a comparison?")((CV && "Not a comparison?") ? static_cast<void>
(0) : __assert_fail ("CV && \"Not a comparison?\"", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1318, __PRETTY_FUNCTION__))
;
1319
1320 bool Changed = false;
1321
1322 SmallSetVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1323 while (!Preds.empty()) {
1324 BasicBlock *Pred = Preds.pop_back_val();
1325 Instruction *PTI = Pred->getTerminator();
1326
1327 // Don't try to fold into itself.
1328 if (Pred == BB)
1329 continue;
1330
1331 // See if the predecessor is a comparison with the same value.
1332 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1333 if (PCV != CV)
1334 continue;
1335
1336 SmallSetVector<BasicBlock *, 4> FailBlocks;
1337 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1338 for (auto *Succ : FailBlocks) {
1339 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split", DTU))
1340 return false;
1341 }
1342 }
1343
1344 PerformValueComparisonIntoPredecessorFolding(TI, CV, PTI, Builder);
1345 Changed = true;
1346 }
1347 return Changed;
1348}
1349
1350// If we would need to insert a select that uses the value of this invoke
1351// (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1352// can't hoist the invoke, as there is nowhere to put the select in this case.
1353static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1354 Instruction *I1, Instruction *I2) {
1355 for (BasicBlock *Succ : successors(BB1)) {
1356 for (const PHINode &PN : Succ->phis()) {
1357 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1358 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1359 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1360 return false;
1361 }
1362 }
1363 }
1364 return true;
1365}
1366
1367static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I, bool PtrValueMayBeModified = false);
1368
1369/// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1370/// in the two blocks up into the branch block. The caller of this function
1371/// guarantees that BI's block dominates BB1 and BB2.
1372bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
1373 const TargetTransformInfo &TTI) {
1374 // This does very trivial matching, with limited scanning, to find identical
1375 // instructions in the two blocks. In particular, we don't want to get into
1376 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1377 // such, we currently just scan for obviously identical instructions in an
1378 // identical order.
1379 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1380 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1381
1382 BasicBlock::iterator BB1_Itr = BB1->begin();
1383 BasicBlock::iterator BB2_Itr = BB2->begin();
1384
1385 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1386 // Skip debug info if it is not identical.
1387 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1388 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1389 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1390 while (isa<DbgInfoIntrinsic>(I1))
1391 I1 = &*BB1_Itr++;
1392 while (isa<DbgInfoIntrinsic>(I2))
1393 I2 = &*BB2_Itr++;
1394 }
1395 // FIXME: Can we define a safety predicate for CallBr?
1396 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1397 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1398 isa<CallBrInst>(I1))
1399 return false;
1400
1401 BasicBlock *BIParent = BI->getParent();
1402
1403 bool Changed = false;
1404
1405 auto _ = make_scope_exit([&]() {
1406 if (Changed)
1407 ++NumHoistCommonCode;
1408 });
1409
1410 do {
1411 // If we are hoisting the terminator instruction, don't move one (making a
1412 // broken BB), instead clone it, and remove BI.
1413 if (I1->isTerminator())
1414 goto HoistTerminator;
1415
1416 // If we're going to hoist a call, make sure that the two instructions we're
1417 // commoning/hoisting are both marked with musttail, or neither of them is
1418 // marked as such. Otherwise, we might end up in a situation where we hoist
1419 // from a block where the terminator is a `ret` to a block where the terminator
1420 // is a `br`, and `musttail` calls expect to be followed by a return.
1421 auto *C1 = dyn_cast<CallInst>(I1);
1422 auto *C2 = dyn_cast<CallInst>(I2);
1423 if (C1 && C2)
1424 if (C1->isMustTailCall() != C2->isMustTailCall())
1425 return Changed;
1426
1427 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1428 return Changed;
1429
1430 // If any of the two call sites has nomerge attribute, stop hoisting.
1431 if (const auto *CB1 = dyn_cast<CallBase>(I1))
1432 if (CB1->cannotMerge())
1433 return Changed;
1434 if (const auto *CB2 = dyn_cast<CallBase>(I2))
1435 if (CB2->cannotMerge())
1436 return Changed;
1437
1438 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1439 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2))((isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic
>(I2)) ? static_cast<void> (0) : __assert_fail ("isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2)"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1439, __PRETTY_FUNCTION__))
;
1440 // The debug location is an integral part of a debug info intrinsic
1441 // and can't be separated from it or replaced. Instead of attempting
1442 // to merge locations, simply hoist both copies of the intrinsic.
1443 BIParent->getInstList().splice(BI->getIterator(),
1444 BB1->getInstList(), I1);
1445 BIParent->getInstList().splice(BI->getIterator(),
1446 BB2->getInstList(), I2);
1447 Changed = true;
1448 } else {
1449 // For a normal instruction, we just move one to right before the branch,
1450 // then replace all uses of the other with the first. Finally, we remove
1451 // the now redundant second instruction.
1452 BIParent->getInstList().splice(BI->getIterator(),
1453 BB1->getInstList(), I1);
1454 if (!I2->use_empty())
1455 I2->replaceAllUsesWith(I1);
1456 I1->andIRFlags(I2);
1457 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1458 LLVMContext::MD_range,
1459 LLVMContext::MD_fpmath,
1460 LLVMContext::MD_invariant_load,
1461 LLVMContext::MD_nonnull,
1462 LLVMContext::MD_invariant_group,
1463 LLVMContext::MD_align,
1464 LLVMContext::MD_dereferenceable,
1465 LLVMContext::MD_dereferenceable_or_null,
1466 LLVMContext::MD_mem_parallel_loop_access,
1467 LLVMContext::MD_access_group,
1468 LLVMContext::MD_preserve_access_index};
1469 combineMetadata(I1, I2, KnownIDs, true);
1470
1471 // I1 and I2 are being combined into a single instruction. Its debug
1472 // location is the merged locations of the original instructions.
1473 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1474
1475 I2->eraseFromParent();
1476 Changed = true;
1477 }
1478 ++NumHoistCommonInstrs;
1479
1480 I1 = &*BB1_Itr++;
1481 I2 = &*BB2_Itr++;
1482 // Skip debug info if it is not identical.
1483 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1484 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1485 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1486 while (isa<DbgInfoIntrinsic>(I1))
1487 I1 = &*BB1_Itr++;
1488 while (isa<DbgInfoIntrinsic>(I2))
1489 I2 = &*BB2_Itr++;
1490 }
1491 } while (I1->isIdenticalToWhenDefined(I2));
1492
1493 return true;
1494
1495HoistTerminator:
1496 // It may not be possible to hoist an invoke.
1497 // FIXME: Can we define a safety predicate for CallBr?
1498 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1499 return Changed;
1500
1501 // TODO: callbr hoisting currently disabled pending further study.
1502 if (isa<CallBrInst>(I1))
1503 return Changed;
1504
1505 for (BasicBlock *Succ : successors(BB1)) {
1506 for (PHINode &PN : Succ->phis()) {
1507 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1508 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1509 if (BB1V == BB2V)
1510 continue;
1511
1512 // Check for passingValueIsAlwaysUndefined here because we would rather
1513 // eliminate undefined control flow then converting it to a select.
1514 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1515 passingValueIsAlwaysUndefined(BB2V, &PN))
1516 return Changed;
1517
1518 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1519 return Changed;
1520 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1521 return Changed;
1522 }
1523 }
1524
1525 // Okay, it is safe to hoist the terminator.
1526 Instruction *NT = I1->clone();
1527 BIParent->getInstList().insert(BI->getIterator(), NT);
1528 if (!NT->getType()->isVoidTy()) {
1529 I1->replaceAllUsesWith(NT);
1530 I2->replaceAllUsesWith(NT);
1531 NT->takeName(I1);
1532 }
1533 Changed = true;
1534 ++NumHoistCommonInstrs;
1535
1536 // Ensure terminator gets a debug location, even an unknown one, in case
1537 // it involves inlinable calls.
1538 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1539
1540 // PHIs created below will adopt NT's merged DebugLoc.
1541 IRBuilder<NoFolder> Builder(NT);
1542
1543 // Hoisting one of the terminators from our successor is a great thing.
1544 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1545 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1546 // nodes, so we insert select instruction to compute the final result.
1547 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1548 for (BasicBlock *Succ : successors(BB1)) {
1549 for (PHINode &PN : Succ->phis()) {
1550 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1551 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1552 if (BB1V == BB2V)
1553 continue;
1554
1555 // These values do not agree. Insert a select instruction before NT
1556 // that determines the right value.
1557 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1558 if (!SI) {
1559 // Propagate fast-math-flags from phi node to its replacement select.
1560 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
1561 if (isa<FPMathOperator>(PN))
1562 Builder.setFastMathFlags(PN.getFastMathFlags());
1563
1564 SI = cast<SelectInst>(
1565 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1566 BB1V->getName() + "." + BB2V->getName(), BI));
1567 }
1568
1569 // Make the PHI node use the select for all incoming values for BB1/BB2
1570 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1571 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1572 PN.setIncomingValue(i, SI);
1573 }
1574 }
1575
1576 SmallVector<DominatorTree::UpdateType, 4> Updates;
1577
1578 // Update any PHI nodes in our new successors.
1579 for (BasicBlock *Succ : successors(BB1)) {
1580 AddPredecessorToBlock(Succ, BIParent, BB1);
1581 Updates.push_back({DominatorTree::Insert, BIParent, Succ});
1582 }
1583 for (BasicBlock *Succ : successors(BI))
1584 Updates.push_back({DominatorTree::Delete, BIParent, Succ});
1585
1586 EraseTerminatorAndDCECond(BI);
1587 if (DTU)
1588 DTU->applyUpdates(Updates);
1589 return Changed;
1590}
1591
1592// Check lifetime markers.
1593static bool isLifeTimeMarker(const Instruction *I) {
1594 if (auto II = dyn_cast<IntrinsicInst>(I)) {
1595 switch (II->getIntrinsicID()) {
1596 default:
1597 break;
1598 case Intrinsic::lifetime_start:
1599 case Intrinsic::lifetime_end:
1600 return true;
1601 }
1602 }
1603 return false;
1604}
1605
1606// TODO: Refine this. This should avoid cases like turning constant memcpy sizes
1607// into variables.
1608static bool replacingOperandWithVariableIsCheap(const Instruction *I,
1609 int OpIdx) {
1610 return !isa<IntrinsicInst>(I);
1611}
1612
1613// All instructions in Insts belong to different blocks that all unconditionally
1614// branch to a common successor. Analyze each instruction and return true if it
1615// would be possible to sink them into their successor, creating one common
1616// instruction instead. For every value that would be required to be provided by
1617// PHI node (because an operand varies in each input block), add to PHIOperands.
1618static bool canSinkInstructions(
1619 ArrayRef<Instruction *> Insts,
1620 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1621 // Prune out obviously bad instructions to move. Each instruction must have
1622 // exactly zero or one use, and we check later that use is by a single, common
1623 // PHI instruction in the successor.
1624 bool HasUse = !Insts.front()->user_empty();
1625 for (auto *I : Insts) {
1626 // These instructions may change or break semantics if moved.
1627 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1628 I->getType()->isTokenTy())
1629 return false;
1630
1631 // Conservatively return false if I is an inline-asm instruction. Sinking
1632 // and merging inline-asm instructions can potentially create arguments
1633 // that cannot satisfy the inline-asm constraints.
1634 // If the instruction has nomerge attribute, return false.
1635 if (const auto *C = dyn_cast<CallBase>(I))
1636 if (C->isInlineAsm() || C->cannotMerge())
1637 return false;
1638
1639 // Each instruction must have zero or one use.
1640 if (HasUse && !I->hasOneUse())
1641 return false;
1642 if (!HasUse && !I->user_empty())
1643 return false;
1644 }
1645
1646 const Instruction *I0 = Insts.front();
1647 for (auto *I : Insts)
1648 if (!I->isSameOperationAs(I0))
1649 return false;
1650
1651 // All instructions in Insts are known to be the same opcode. If they have a
1652 // use, check that the only user is a PHI or in the same block as the
1653 // instruction, because if a user is in the same block as an instruction we're
1654 // contemplating sinking, it must already be determined to be sinkable.
1655 if (HasUse) {
1656 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1657 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1658 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1659 auto *U = cast<Instruction>(*I->user_begin());
1660 return (PNUse &&
1661 PNUse->getParent() == Succ &&
1662 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1663 U->getParent() == I->getParent();
1664 }))
1665 return false;
1666 }
1667
1668 // Because SROA can't handle speculating stores of selects, try not to sink
1669 // loads, stores or lifetime markers of allocas when we'd have to create a
1670 // PHI for the address operand. Also, because it is likely that loads or
1671 // stores of allocas will disappear when Mem2Reg/SROA is run, don't sink
1672 // them.
1673 // This can cause code churn which can have unintended consequences down
1674 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1675 // FIXME: This is a workaround for a deficiency in SROA - see
1676 // https://llvm.org/bugs/show_bug.cgi?id=30188
1677 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1678 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1679 }))
1680 return false;
1681 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1682 return isa<AllocaInst>(I->getOperand(0)->stripPointerCasts());
1683 }))
1684 return false;
1685 if (isLifeTimeMarker(I0) && any_of(Insts, [](const Instruction *I) {
1686 return isa<AllocaInst>(I->getOperand(1)->stripPointerCasts());
1687 }))
1688 return false;
1689
1690 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1691 Value *Op = I0->getOperand(OI);
1692 if (Op->getType()->isTokenTy())
1693 // Don't touch any operand of token type.
1694 return false;
1695
1696 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1697 assert(I->getNumOperands() == I0->getNumOperands())((I->getNumOperands() == I0->getNumOperands()) ? static_cast
<void> (0) : __assert_fail ("I->getNumOperands() == I0->getNumOperands()"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1697, __PRETTY_FUNCTION__))
;
1698 return I->getOperand(OI) == I0->getOperand(OI);
1699 };
1700 if (!all_of(Insts, SameAsI0)) {
1701 if ((isa<Constant>(Op) && !replacingOperandWithVariableIsCheap(I0, OI)) ||
1702 !canReplaceOperandWithVariable(I0, OI))
1703 // We can't create a PHI from this GEP.
1704 return false;
1705 // Don't create indirect calls! The called value is the final operand.
1706 if (isa<CallBase>(I0) && OI == OE - 1) {
1707 // FIXME: if the call was *already* indirect, we should do this.
1708 return false;
1709 }
1710 for (auto *I : Insts)
1711 PHIOperands[I].push_back(I->getOperand(OI));
1712 }
1713 }
1714 return true;
1715}
1716
1717// Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1718// instruction of every block in Blocks to their common successor, commoning
1719// into one instruction.
1720static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1721 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1722
1723 // canSinkLastInstruction returning true guarantees that every block has at
1724 // least one non-terminator instruction.
1725 SmallVector<Instruction*,4> Insts;
1726 for (auto *BB : Blocks) {
1727 Instruction *I = BB->getTerminator();
1728 do {
1729 I = I->getPrevNode();
1730 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1731 if (!isa<DbgInfoIntrinsic>(I))
1732 Insts.push_back(I);
1733 }
1734
1735 // The only checking we need to do now is that all users of all instructions
1736 // are the same PHI node. canSinkLastInstruction should have checked this but
1737 // it is slightly over-aggressive - it gets confused by commutative instructions
1738 // so double-check it here.
1739 Instruction *I0 = Insts.front();
1740 if (!I0->user_empty()) {
1741 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1742 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1743 auto *U = cast<Instruction>(*I->user_begin());
1744 return U == PNUse;
1745 }))
1746 return false;
1747 }
1748
1749 // We don't need to do any more checking here; canSinkLastInstruction should
1750 // have done it all for us.
1751 SmallVector<Value*, 4> NewOperands;
1752 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1753 // This check is different to that in canSinkLastInstruction. There, we
1754 // cared about the global view once simplifycfg (and instcombine) have
1755 // completed - it takes into account PHIs that become trivially
1756 // simplifiable. However here we need a more local view; if an operand
1757 // differs we create a PHI and rely on instcombine to clean up the very
1758 // small mess we may make.
1759 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1760 return I->getOperand(O) != I0->getOperand(O);
1761 });
1762 if (!NeedPHI) {
1763 NewOperands.push_back(I0->getOperand(O));
1764 continue;
1765 }
1766
1767 // Create a new PHI in the successor block and populate it.
1768 auto *Op = I0->getOperand(O);
1769 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!")((!Op->getType()->isTokenTy() && "Can't PHI tokens!"
) ? static_cast<void> (0) : __assert_fail ("!Op->getType()->isTokenTy() && \"Can't PHI tokens!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1769, __PRETTY_FUNCTION__))
;
1770 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1771 Op->getName() + ".sink", &BBEnd->front());
1772 for (auto *I : Insts)
1773 PN->addIncoming(I->getOperand(O), I->getParent());
1774 NewOperands.push_back(PN);
1775 }
1776
1777 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1778 // and move it to the start of the successor block.
1779 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1780 I0->getOperandUse(O).set(NewOperands[O]);
1781 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1782
1783 // Update metadata and IR flags, and merge debug locations.
1784 for (auto *I : Insts)
1785 if (I != I0) {
1786 // The debug location for the "common" instruction is the merged locations
1787 // of all the commoned instructions. We start with the original location
1788 // of the "common" instruction and iteratively merge each location in the
1789 // loop below.
1790 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1791 // However, as N-way merge for CallInst is rare, so we use simplified API
1792 // instead of using complex API for N-way merge.
1793 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1794 combineMetadataForCSE(I0, I, true);
1795 I0->andIRFlags(I);
1796 }
1797
1798 if (!I0->user_empty()) {
1799 // canSinkLastInstruction checked that all instructions were used by
1800 // one and only one PHI node. Find that now, RAUW it to our common
1801 // instruction and nuke it.
1802 auto *PN = cast<PHINode>(*I0->user_begin());
1803 PN->replaceAllUsesWith(I0);
1804 PN->eraseFromParent();
1805 }
1806
1807 // Finally nuke all instructions apart from the common instruction.
1808 for (auto *I : Insts)
1809 if (I != I0)
1810 I->eraseFromParent();
1811
1812 return true;
1813}
1814
1815namespace {
1816
1817 // LockstepReverseIterator - Iterates through instructions
1818 // in a set of blocks in reverse order from the first non-terminator.
1819 // For example (assume all blocks have size n):
1820 // LockstepReverseIterator I([B1, B2, B3]);
1821 // *I-- = [B1[n], B2[n], B3[n]];
1822 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1823 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1824 // ...
1825 class LockstepReverseIterator {
1826 ArrayRef<BasicBlock*> Blocks;
1827 SmallVector<Instruction*,4> Insts;
1828 bool Fail;
1829
1830 public:
1831 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1832 reset();
1833 }
1834
1835 void reset() {
1836 Fail = false;
1837 Insts.clear();
1838 for (auto *BB : Blocks) {
1839 Instruction *Inst = BB->getTerminator();
1840 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1841 Inst = Inst->getPrevNode();
1842 if (!Inst) {
1843 // Block wasn't big enough.
1844 Fail = true;
1845 return;
1846 }
1847 Insts.push_back(Inst);
1848 }
1849 }
1850
1851 bool isValid() const {
1852 return !Fail;
1853 }
1854
1855 void operator--() {
1856 if (Fail)
1857 return;
1858 for (auto *&Inst : Insts) {
1859 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1860 Inst = Inst->getPrevNode();
1861 // Already at beginning of block.
1862 if (!Inst) {
1863 Fail = true;
1864 return;
1865 }
1866 }
1867 }
1868
1869 ArrayRef<Instruction*> operator * () const {
1870 return Insts;
1871 }
1872 };
1873
1874} // end anonymous namespace
1875
1876/// Check whether BB's predecessors end with unconditional branches. If it is
1877/// true, sink any common code from the predecessors to BB.
1878/// We also allow one predecessor to end with conditional branch (but no more
1879/// than one).
1880static bool SinkCommonCodeFromPredecessors(BasicBlock *BB,
1881 DomTreeUpdater *DTU) {
1882 // We support two situations:
1883 // (1) all incoming arcs are unconditional
1884 // (2) one incoming arc is conditional
1885 //
1886 // (2) is very common in switch defaults and
1887 // else-if patterns;
1888 //
1889 // if (a) f(1);
1890 // else if (b) f(2);
1891 //
1892 // produces:
1893 //
1894 // [if]
1895 // / \
1896 // [f(1)] [if]
1897 // | | \
1898 // | | |
1899 // | [f(2)]|
1900 // \ | /
1901 // [ end ]
1902 //
1903 // [end] has two unconditional predecessor arcs and one conditional. The
1904 // conditional refers to the implicit empty 'else' arc. This conditional
1905 // arc can also be caused by an empty default block in a switch.
1906 //
1907 // In this case, we attempt to sink code from all *unconditional* arcs.
1908 // If we can sink instructions from these arcs (determined during the scan
1909 // phase below) we insert a common successor for all unconditional arcs and
1910 // connect that to [end], to enable sinking:
1911 //
1912 // [if]
1913 // / \
1914 // [x(1)] [if]
1915 // | | \
1916 // | | \
1917 // | [x(2)] |
1918 // \ / |
1919 // [sink.split] |
1920 // \ /
1921 // [ end ]
1922 //
1923 SmallVector<BasicBlock*,4> UnconditionalPreds;
1924 Instruction *Cond = nullptr;
1925 for (auto *B : predecessors(BB)) {
1926 auto *T = B->getTerminator();
1927 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1928 UnconditionalPreds.push_back(B);
1929 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1930 Cond = T;
1931 else
1932 return false;
1933 }
1934 if (UnconditionalPreds.size() < 2)
1935 return false;
1936
1937 // We take a two-step approach to tail sinking. First we scan from the end of
1938 // each block upwards in lockstep. If the n'th instruction from the end of each
1939 // block can be sunk, those instructions are added to ValuesToSink and we
1940 // carry on. If we can sink an instruction but need to PHI-merge some operands
1941 // (because they're not identical in each instruction) we add these to
1942 // PHIOperands.
1943 unsigned ScanIdx = 0;
1944 SmallPtrSet<Value*,4> InstructionsToSink;
1945 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1946 LockstepReverseIterator LRI(UnconditionalPreds);
1947 while (LRI.isValid() &&
1948 canSinkInstructions(*LRI, PHIOperands)) {
1949 LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: instruction can be sunk: "
<< *(*LRI)[0] << "\n"; } } while (false)
1950 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: instruction can be sunk: "
<< *(*LRI)[0] << "\n"; } } while (false)
;
1951 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1952 ++ScanIdx;
1953 --LRI;
1954 }
1955
1956 // If no instructions can be sunk, early-return.
1957 if (ScanIdx == 0)
1958 return false;
1959
1960 bool Changed = false;
1961
1962 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1963 unsigned NumPHIdValues = 0;
1964 for (auto *I : *LRI)
1965 for (auto *V : PHIOperands[I])
1966 if (InstructionsToSink.count(V) == 0)
1967 ++NumPHIdValues;
1968 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: #phid values: " <<
NumPHIdValues << "\n"; } } while (false)
;
1969 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1970 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1971 NumPHIInsts++;
1972
1973 return NumPHIInsts <= 1;
1974 };
1975
1976 if (Cond) {
1977 // Check if we would actually sink anything first! This mutates the CFG and
1978 // adds an extra block. The goal in doing this is to allow instructions that
1979 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1980 // (such as trunc, add) can be sunk and predicated already. So we check that
1981 // we're going to sink at least one non-speculatable instruction.
1982 LRI.reset();
1983 unsigned Idx = 0;
1984 bool Profitable = false;
1985 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1986 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1987 Profitable = true;
1988 break;
1989 }
1990 --LRI;
1991 ++Idx;
1992 }
1993 if (!Profitable)
1994 return false;
1995
1996 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Splitting edge\n"; }
} while (false)
;
1997 // We have a conditional edge and we're going to sink some instructions.
1998 // Insert a new block postdominating all blocks we're going to sink from.
1999 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split", DTU))
2000 // Edges couldn't be split.
2001 return false;
2002 Changed = true;
2003 }
2004
2005 // Now that we've analyzed all potential sinking candidates, perform the
2006 // actual sink. We iteratively sink the last non-terminator of the source
2007 // blocks into their common successor unless doing so would require too
2008 // many PHI instructions to be generated (currently only one PHI is allowed
2009 // per sunk instruction).
2010 //
2011 // We can use InstructionsToSink to discount values needing PHI-merging that will
2012 // actually be sunk in a later iteration. This allows us to be more
2013 // aggressive in what we sink. This does allow a false positive where we
2014 // sink presuming a later value will also be sunk, but stop half way through
2015 // and never actually sink it which means we produce more PHIs than intended.
2016 // This is unlikely in practice though.
2017 unsigned SinkIdx = 0;
2018 for (; SinkIdx != ScanIdx; ++SinkIdx) {
2019 LLVM_DEBUG(dbgs() << "SINK: Sink: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
2020 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
2021 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
;
2022
2023 // Because we've sunk every instruction in turn, the current instruction to
2024 // sink is always at index 0.
2025 LRI.reset();
2026 if (!ProfitableToSinkInstruction(LRI)) {
2027 // Too many PHIs would be created.
2028 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, too many PHIs would be created!\n"
; } } while (false)
2029 dbgs() << "SINK: stopping here, too many PHIs would be created!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, too many PHIs would be created!\n"
; } } while (false)
;
2030 break;
2031 }
2032
2033 if (!sinkLastInstruction(UnconditionalPreds)) {
2034 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, failed to actually sink instruction!\n"
; } } while (false)
2035 dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, failed to actually sink instruction!\n"
; } } while (false)
2036 << "SINK: stopping here, failed to actually sink instruction!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, failed to actually sink instruction!\n"
; } } while (false)
;
2037 break;
2038 }
2039
2040 NumSinkCommonInstrs++;
2041 Changed = true;
2042 }
2043 if (SinkIdx != 0)
2044 ++NumSinkCommonCode;
2045 return Changed;
2046}
2047
2048/// Determine if we can hoist sink a sole store instruction out of a
2049/// conditional block.
2050///
2051/// We are looking for code like the following:
2052/// BrBB:
2053/// store i32 %add, i32* %arrayidx2
2054/// ... // No other stores or function calls (we could be calling a memory
2055/// ... // function).
2056/// %cmp = icmp ult %x, %y
2057/// br i1 %cmp, label %EndBB, label %ThenBB
2058/// ThenBB:
2059/// store i32 %add5, i32* %arrayidx2
2060/// br label EndBB
2061/// EndBB:
2062/// ...
2063/// We are going to transform this into:
2064/// BrBB:
2065/// store i32 %add, i32* %arrayidx2
2066/// ... //
2067/// %cmp = icmp ult %x, %y
2068/// %add.add5 = select i1 %cmp, i32 %add, %add5
2069/// store i32 %add.add5, i32* %arrayidx2
2070/// ...
2071///
2072/// \return The pointer to the value of the previous store if the store can be
2073/// hoisted into the predecessor block. 0 otherwise.
2074static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
2075 BasicBlock *StoreBB, BasicBlock *EndBB) {
2076 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
2077 if (!StoreToHoist)
2078 return nullptr;
2079
2080 // Volatile or atomic.
2081 if (!StoreToHoist->isSimple())
2082 return nullptr;
2083
2084 Value *StorePtr = StoreToHoist->getPointerOperand();
2085
2086 // Look for a store to the same pointer in BrBB.
2087 unsigned MaxNumInstToLookAt = 9;
2088 // Skip pseudo probe intrinsic calls which are not really killing any memory
2089 // accesses.
2090 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug(true))) {
2091 if (!MaxNumInstToLookAt)
2092 break;
2093 --MaxNumInstToLookAt;
2094
2095 // Could be calling an instruction that affects memory like free().
2096 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
2097 return nullptr;
2098
2099 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
2100 // Found the previous store make sure it stores to the same location.
2101 if (SI->getPointerOperand() == StorePtr)
2102 // Found the previous store, return its value operand.
2103 return SI->getValueOperand();
2104 return nullptr; // Unknown store.
2105 }
2106 }
2107
2108 return nullptr;
2109}
2110
2111/// Estimate the cost of the insertion(s) and check that the PHI nodes can be
2112/// converted to selects.
2113static bool validateAndCostRequiredSelects(BasicBlock *BB, BasicBlock *ThenBB,
2114 BasicBlock *EndBB,
2115 unsigned &SpeculatedInstructions,
2116 int &BudgetRemaining,
2117 const TargetTransformInfo &TTI) {
2118 TargetTransformInfo::TargetCostKind CostKind =
2119 BB->getParent()->hasMinSize()
2120 ? TargetTransformInfo::TCK_CodeSize
2121 : TargetTransformInfo::TCK_SizeAndLatency;
2122
2123 bool HaveRewritablePHIs = false;
2124 for (PHINode &PN : EndBB->phis()) {
2125 Value *OrigV = PN.getIncomingValueForBlock(BB);
2126 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2127
2128 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2129 // Skip PHIs which are trivial.
2130 if (ThenV == OrigV)
2131 continue;
2132
2133 BudgetRemaining -=
2134 TTI.getCmpSelInstrCost(Instruction::Select, PN.getType(), nullptr,
2135 CmpInst::BAD_ICMP_PREDICATE, CostKind);
2136
2137 // Don't convert to selects if we could remove undefined behavior instead.
2138 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2139 passingValueIsAlwaysUndefined(ThenV, &PN))
2140 return false;
2141
2142 HaveRewritablePHIs = true;
2143 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2144 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2145 if (!OrigCE && !ThenCE)
2146 continue; // Known safe and cheap.
2147
2148 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2149 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2150 return false;
2151 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2152 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2153 unsigned MaxCost =
2154 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2155 if (OrigCost + ThenCost > MaxCost)
2156 return false;
2157
2158 // Account for the cost of an unfolded ConstantExpr which could end up
2159 // getting expanded into Instructions.
2160 // FIXME: This doesn't account for how many operations are combined in the
2161 // constant expression.
2162 ++SpeculatedInstructions;
2163 if (SpeculatedInstructions > 1)
2164 return false;
2165 }
2166
2167 return HaveRewritablePHIs;
2168}
2169
2170/// Speculate a conditional basic block flattening the CFG.
2171///
2172/// Note that this is a very risky transform currently. Speculating
2173/// instructions like this is most often not desirable. Instead, there is an MI
2174/// pass which can do it with full awareness of the resource constraints.
2175/// However, some cases are "obvious" and we should do directly. An example of
2176/// this is speculating a single, reasonably cheap instruction.
2177///
2178/// There is only one distinct advantage to flattening the CFG at the IR level:
2179/// it makes very common but simplistic optimizations such as are common in
2180/// instcombine and the DAG combiner more powerful by removing CFG edges and
2181/// modeling their effects with easier to reason about SSA value graphs.
2182///
2183///
2184/// An illustration of this transform is turning this IR:
2185/// \code
2186/// BB:
2187/// %cmp = icmp ult %x, %y
2188/// br i1 %cmp, label %EndBB, label %ThenBB
2189/// ThenBB:
2190/// %sub = sub %x, %y
2191/// br label BB2
2192/// EndBB:
2193/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
2194/// ...
2195/// \endcode
2196///
2197/// Into this IR:
2198/// \code
2199/// BB:
2200/// %cmp = icmp ult %x, %y
2201/// %sub = sub %x, %y
2202/// %cond = select i1 %cmp, 0, %sub
2203/// ...
2204/// \endcode
2205///
2206/// \returns true if the conditional block is removed.
2207bool SimplifyCFGOpt::SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
2208 const TargetTransformInfo &TTI) {
2209 // Be conservative for now. FP select instruction can often be expensive.
2210 Value *BrCond = BI->getCondition();
2211 if (isa<FCmpInst>(BrCond))
2212 return false;
2213
2214 BasicBlock *BB = BI->getParent();
2215 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
2216 int BudgetRemaining =
2217 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2218
2219 // If ThenBB is actually on the false edge of the conditional branch, remember
2220 // to swap the select operands later.
2221 bool Invert = false;
2222 if (ThenBB != BI->getSuccessor(0)) {
2223 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?")((ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?"
) ? static_cast<void> (0) : __assert_fail ("ThenBB == BI->getSuccessor(1) && \"No edge from 'if' block?\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2223, __PRETTY_FUNCTION__))
;
2224 Invert = true;
2225 }
2226 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block")((EndBB == BI->getSuccessor(!Invert) && "No edge from to end block"
) ? static_cast<void> (0) : __assert_fail ("EndBB == BI->getSuccessor(!Invert) && \"No edge from to end block\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2226, __PRETTY_FUNCTION__))
;
2227
2228 // Keep a count of how many times instructions are used within ThenBB when
2229 // they are candidates for sinking into ThenBB. Specifically:
2230 // - They are defined in BB, and
2231 // - They have no side effects, and
2232 // - All of their uses are in ThenBB.
2233 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
2234
2235 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
2236
2237 unsigned SpeculatedInstructions = 0;
2238 Value *SpeculatedStoreValue = nullptr;
2239 StoreInst *SpeculatedStore = nullptr;
2240 for (BasicBlock::iterator BBI = ThenBB->begin(),
2241 BBE = std::prev(ThenBB->end());
2242 BBI != BBE; ++BBI) {
2243 Instruction *I = &*BBI;
2244 // Skip debug info.
2245 if (isa<DbgInfoIntrinsic>(I)) {
2246 SpeculatedDbgIntrinsics.push_back(I);
2247 continue;
2248 }
2249
2250 // Skip pseudo probes. The consequence is we lose track of the branch
2251 // probability for ThenBB, which is fine since the optimization here takes
2252 // place regardless of the branch probability.
2253 if (isa<PseudoProbeInst>(I)) {
2254 SpeculatedDbgIntrinsics.push_back(I);
2255 continue;
2256 }
2257
2258 // Only speculatively execute a single instruction (not counting the
2259 // terminator) for now.
2260 ++SpeculatedInstructions;
2261 if (SpeculatedInstructions > 1)
2262 return false;
2263
2264 // Don't hoist the instruction if it's unsafe or expensive.
2265 if (!isSafeToSpeculativelyExecute(I) &&
2266 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2267 I, BB, ThenBB, EndBB))))
2268 return false;
2269 if (!SpeculatedStoreValue &&
2270 ComputeSpeculationCost(I, TTI) >
2271 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2272 return false;
2273
2274 // Store the store speculation candidate.
2275 if (SpeculatedStoreValue)
2276 SpeculatedStore = cast<StoreInst>(I);
2277
2278 // Do not hoist the instruction if any of its operands are defined but not
2279 // used in BB. The transformation will prevent the operand from
2280 // being sunk into the use block.
2281 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2282 Instruction *OpI = dyn_cast<Instruction>(*i);
2283 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2284 continue; // Not a candidate for sinking.
2285
2286 ++SinkCandidateUseCounts[OpI];
2287 }
2288 }
2289
2290 // Consider any sink candidates which are only used in ThenBB as costs for
2291 // speculation. Note, while we iterate over a DenseMap here, we are summing
2292 // and so iteration order isn't significant.
2293 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2294 I = SinkCandidateUseCounts.begin(),
2295 E = SinkCandidateUseCounts.end();
2296 I != E; ++I)
2297 if (I->first->hasNUses(I->second)) {
2298 ++SpeculatedInstructions;
2299 if (SpeculatedInstructions > 1)
2300 return false;
2301 }
2302
2303 // Check that we can insert the selects and that it's not too expensive to do
2304 // so.
2305 bool Convert = SpeculatedStore != nullptr;
2306 Convert |= validateAndCostRequiredSelects(BB, ThenBB, EndBB,
2307 SpeculatedInstructions,
2308 BudgetRemaining, TTI);
2309 if (!Convert || BudgetRemaining < 0)
2310 return false;
2311
2312 // If we get here, we can hoist the instruction and if-convert.
2313 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SPECULATIVELY EXECUTING BB"
<< *ThenBB << "\n";; } } while (false)
;
2314
2315 // Insert a select of the value of the speculated store.
2316 if (SpeculatedStoreValue) {
2317 IRBuilder<NoFolder> Builder(BI);
2318 Value *TrueV = SpeculatedStore->getValueOperand();
2319 Value *FalseV = SpeculatedStoreValue;
2320 if (Invert)
2321 std::swap(TrueV, FalseV);
2322 Value *S = Builder.CreateSelect(
2323 BrCond, TrueV, FalseV, "spec.store.select", BI);
2324 SpeculatedStore->setOperand(0, S);
2325 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2326 SpeculatedStore->getDebugLoc());
2327 }
2328
2329 // Metadata can be dependent on the condition we are hoisting above.
2330 // Conservatively strip all metadata on the instruction. Drop the debug loc
2331 // to avoid making it appear as if the condition is a constant, which would
2332 // be misleading while debugging.
2333 for (auto &I : *ThenBB) {
2334 if (!SpeculatedStoreValue || &I != SpeculatedStore)
2335 I.setDebugLoc(DebugLoc());
2336 I.dropUnknownNonDebugMetadata();
2337 }
2338
2339 // Hoist the instructions.
2340 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2341 ThenBB->begin(), std::prev(ThenBB->end()));
2342
2343 // Insert selects and rewrite the PHI operands.
2344 IRBuilder<NoFolder> Builder(BI);
2345 for (PHINode &PN : EndBB->phis()) {
2346 unsigned OrigI = PN.getBasicBlockIndex(BB);
2347 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2348 Value *OrigV = PN.getIncomingValue(OrigI);
2349 Value *ThenV = PN.getIncomingValue(ThenI);
2350
2351 // Skip PHIs which are trivial.
2352 if (OrigV == ThenV)
2353 continue;
2354
2355 // Create a select whose true value is the speculatively executed value and
2356 // false value is the pre-existing value. Swap them if the branch
2357 // destinations were inverted.
2358 Value *TrueV = ThenV, *FalseV = OrigV;
2359 if (Invert)
2360 std::swap(TrueV, FalseV);
2361 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV, "spec.select", BI);
2362 PN.setIncomingValue(OrigI, V);
2363 PN.setIncomingValue(ThenI, V);
2364 }
2365
2366 // Remove speculated dbg intrinsics.
2367 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2368 // dbg value for the different flows and inserting it after the select.
2369 for (Instruction *I : SpeculatedDbgIntrinsics)
2370 I->eraseFromParent();
2371
2372 ++NumSpeculations;
2373 return true;
2374}
2375
2376/// Return true if we can thread a branch across this block.
2377static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2378 int Size = 0;
2379
2380 for (Instruction &I : BB->instructionsWithoutDebug()) {
2381 if (Size > MaxSmallBlockSize)
2382 return false; // Don't clone large BB's.
2383
2384 // Can't fold blocks that contain noduplicate or convergent calls.
2385 if (CallInst *CI = dyn_cast<CallInst>(&I))
2386 if (CI->cannotDuplicate() || CI->isConvergent())
2387 return false;
2388
2389 // We will delete Phis while threading, so Phis should not be accounted in
2390 // block's size
2391 if (!isa<PHINode>(I))
2392 ++Size;
2393
2394 // We can only support instructions that do not define values that are
2395 // live outside of the current basic block.
2396 for (User *U : I.users()) {
2397 Instruction *UI = cast<Instruction>(U);
2398 if (UI->getParent() != BB || isa<PHINode>(UI))
2399 return false;
2400 }
2401
2402 // Looks ok, continue checking.
2403 }
2404
2405 return true;
2406}
2407
2408/// If we have a conditional branch on a PHI node value that is defined in the
2409/// same block as the branch and if any PHI entries are constants, thread edges
2410/// corresponding to that entry to be branches to their ultimate destination.
2411static bool FoldCondBranchOnPHI(BranchInst *BI, DomTreeUpdater *DTU,
2412 const DataLayout &DL, AssumptionCache *AC) {
2413 BasicBlock *BB = BI->getParent();
2414 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2415 // NOTE: we currently cannot transform this case if the PHI node is used
2416 // outside of the block.
2417 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2418 return false;
2419
2420 // Degenerate case of a single entry PHI.
2421 if (PN->getNumIncomingValues() == 1) {
2422 FoldSingleEntryPHINodes(PN->getParent());
2423 return true;
2424 }
2425
2426 // Now we know that this block has multiple preds and two succs.
2427 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2428 return false;
2429
2430 // Okay, this is a simple enough basic block. See if any phi values are
2431 // constants.
2432 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2433 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2434 if (!CB || !CB->getType()->isIntegerTy(1))
2435 continue;
2436
2437 // Okay, we now know that all edges from PredBB should be revectored to
2438 // branch to RealDest.
2439 BasicBlock *PredBB = PN->getIncomingBlock(i);
2440 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2441
2442 if (RealDest == BB)
2443 continue; // Skip self loops.
2444 // Skip if the predecessor's terminator is an indirect branch.
2445 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2446 continue;
2447
2448 SmallVector<DominatorTree::UpdateType, 3> Updates;
2449
2450 // The dest block might have PHI nodes, other predecessors and other
2451 // difficult cases. Instead of being smart about this, just insert a new
2452 // block that jumps to the destination block, effectively splitting
2453 // the edge we are about to create.
2454 BasicBlock *EdgeBB =
2455 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2456 RealDest->getParent(), RealDest);
2457 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2458 Updates.push_back({DominatorTree::Insert, EdgeBB, RealDest});
2459 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2460
2461 // Update PHI nodes.
2462 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2463
2464 // BB may have instructions that are being threaded over. Clone these
2465 // instructions into EdgeBB. We know that there will be no uses of the
2466 // cloned instructions outside of EdgeBB.
2467 BasicBlock::iterator InsertPt = EdgeBB->begin();
2468 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2469 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2470 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2471 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2472 continue;
2473 }
2474 // Clone the instruction.
2475 Instruction *N = BBI->clone();
2476 if (BBI->hasName())
2477 N->setName(BBI->getName() + ".c");
2478
2479 // Update operands due to translation.
2480 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2481 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2482 if (PI != TranslateMap.end())
2483 *i = PI->second;
2484 }
2485
2486 // Check for trivial simplification.
2487 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2488 if (!BBI->use_empty())
2489 TranslateMap[&*BBI] = V;
2490 if (!N->mayHaveSideEffects()) {
2491 N->deleteValue(); // Instruction folded away, don't need actual inst
2492 N = nullptr;
2493 }
2494 } else {
2495 if (!BBI->use_empty())
2496 TranslateMap[&*BBI] = N;
2497 }
2498 if (N) {
2499 // Insert the new instruction into its new home.
2500 EdgeBB->getInstList().insert(InsertPt, N);
2501
2502 // Register the new instruction with the assumption cache if necessary.
2503 if (AC && match(N, m_Intrinsic<Intrinsic::assume>()))
2504 AC->registerAssumption(cast<IntrinsicInst>(N));
2505 }
2506 }
2507
2508 // Loop over all of the edges from PredBB to BB, changing them to branch
2509 // to EdgeBB instead.
2510 Instruction *PredBBTI = PredBB->getTerminator();
2511 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2512 if (PredBBTI->getSuccessor(i) == BB) {
2513 BB->removePredecessor(PredBB);
2514 PredBBTI->setSuccessor(i, EdgeBB);
2515 }
2516
2517 Updates.push_back({DominatorTree::Insert, PredBB, EdgeBB});
2518 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2519
2520 if (DTU)
2521 DTU->applyUpdates(Updates);
2522
2523 // Recurse, simplifying any other constants.
2524 return FoldCondBranchOnPHI(BI, DTU, DL, AC) || true;
2525 }
2526
2527 return false;
2528}
2529
2530/// Given a BB that starts with the specified two-entry PHI node,
2531/// see if we can eliminate it.
2532static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2533 DomTreeUpdater *DTU, const DataLayout &DL) {
2534 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2535 // statement", which has a very simple dominance structure. Basically, we
2536 // are trying to find the condition that is being branched on, which
2537 // subsequently causes this merge to happen. We really want control
2538 // dependence information for this check, but simplifycfg can't keep it up
2539 // to date, and this catches most of the cases we care about anyway.
2540 BasicBlock *BB = PN->getParent();
2541
2542 BasicBlock *IfTrue, *IfFalse;
2543 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2544 if (!IfCond ||
2545 // Don't bother if the branch will be constant folded trivially.
2546 isa<ConstantInt>(IfCond))
2547 return false;
2548
2549 // Okay, we found that we can merge this two-entry phi node into a select.
2550 // Doing so would require us to fold *all* two entry phi nodes in this block.
2551 // At some point this becomes non-profitable (particularly if the target
2552 // doesn't support cmov's). Only do this transformation if there are two or
2553 // fewer PHI nodes in this block.
2554 unsigned NumPhis = 0;
2555 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2556 if (NumPhis > 2)
2557 return false;
2558
2559 // Loop over the PHI's seeing if we can promote them all to select
2560 // instructions. While we are at it, keep track of the instructions
2561 // that need to be moved to the dominating block.
2562 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2563 int BudgetRemaining =
2564 TwoEntryPHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2565
2566 bool Changed = false;
2567 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2568 PHINode *PN = cast<PHINode>(II++);
2569 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2570 PN->replaceAllUsesWith(V);
2571 PN->eraseFromParent();
2572 Changed = true;
2573 continue;
2574 }
2575
2576 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2577 BudgetRemaining, TTI) ||
2578 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2579 BudgetRemaining, TTI))
2580 return Changed;
2581 }
2582
2583 // If we folded the first phi, PN dangles at this point. Refresh it. If
2584 // we ran out of PHIs then we simplified them all.
2585 PN = dyn_cast<PHINode>(BB->begin());
2586 if (!PN)
2587 return true;
2588
2589 // Return true if at least one of these is a 'not', and another is either
2590 // a 'not' too, or a constant.
2591 auto CanHoistNotFromBothValues = [](Value *V0, Value *V1) {
2592 if (!match(V0, m_Not(m_Value())))
2593 std::swap(V0, V1);
2594 auto Invertible = m_CombineOr(m_Not(m_Value()), m_AnyIntegralConstant());
2595 return match(V0, m_Not(m_Value())) && match(V1, Invertible);
2596 };
2597
2598 // Don't fold i1 branches on PHIs which contain binary operators, unless one
2599 // of the incoming values is an 'not' and another one is freely invertible.
2600 // These can often be turned into switches and other things.
2601 if (PN->getType()->isIntegerTy(1) &&
2602 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2603 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2604 isa<BinaryOperator>(IfCond)) &&
2605 !CanHoistNotFromBothValues(PN->getIncomingValue(0),
2606 PN->getIncomingValue(1)))
2607 return Changed;
2608
2609 // If all PHI nodes are promotable, check to make sure that all instructions
2610 // in the predecessor blocks can be promoted as well. If not, we won't be able
2611 // to get rid of the control flow, so it's not worth promoting to select
2612 // instructions.
2613 BasicBlock *DomBlock = nullptr;
2614 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2615 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2616 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2617 IfBlock1 = nullptr;
2618 } else {
2619 DomBlock = *pred_begin(IfBlock1);
2620 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2621 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I) &&
2622 !isa<PseudoProbeInst>(I)) {
2623 // This is not an aggressive instruction that we can promote.
2624 // Because of this, we won't be able to get rid of the control flow, so
2625 // the xform is not worth it.
2626 return Changed;
2627 }
2628 }
2629
2630 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2631 IfBlock2 = nullptr;
2632 } else {
2633 DomBlock = *pred_begin(IfBlock2);
2634 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2635 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I) &&
2636 !isa<PseudoProbeInst>(I)) {
2637 // This is not an aggressive instruction that we can promote.
2638 // Because of this, we won't be able to get rid of the control flow, so
2639 // the xform is not worth it.
2640 return Changed;
2641 }
2642 }
2643 assert(DomBlock && "Failed to find root DomBlock")((DomBlock && "Failed to find root DomBlock") ? static_cast
<void> (0) : __assert_fail ("DomBlock && \"Failed to find root DomBlock\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2643, __PRETTY_FUNCTION__))
;
2644
2645 LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfConddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
2646 << " T: " << IfTrue->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
2647 << " F: " << IfFalse->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
;
2648
2649 // If we can still promote the PHI nodes after this gauntlet of tests,
2650 // do all of the PHI's now.
2651 Instruction *InsertPt = DomBlock->getTerminator();
2652 IRBuilder<NoFolder> Builder(InsertPt);
2653
2654 // Move all 'aggressive' instructions, which are defined in the
2655 // conditional parts of the if's up to the dominating block.
2656 if (IfBlock1)
2657 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2658 if (IfBlock2)
2659 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2660
2661 // Propagate fast-math-flags from phi nodes to replacement selects.
2662 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
2663 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2664 if (isa<FPMathOperator>(PN))
2665 Builder.setFastMathFlags(PN->getFastMathFlags());
2666
2667 // Change the PHI node into a select instruction.
2668 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2669 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2670
2671 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2672 PN->replaceAllUsesWith(Sel);
2673 Sel->takeName(PN);
2674 PN->eraseFromParent();
2675 }
2676
2677 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2678 // has been flattened. Change DomBlock to jump directly to our new block to
2679 // avoid other simplifycfg's kicking in on the diamond.
2680 Instruction *OldTI = DomBlock->getTerminator();
2681 Builder.SetInsertPoint(OldTI);
2682 Builder.CreateBr(BB);
2683
2684 SmallVector<DominatorTree::UpdateType, 3> Updates;
2685 if (DTU) {
2686 Updates.push_back({DominatorTree::Insert, DomBlock, BB});
2687 for (auto *Successor : successors(DomBlock))
2688 Updates.push_back({DominatorTree::Delete, DomBlock, Successor});
2689 }
2690
2691 OldTI->eraseFromParent();
2692 if (DTU)
2693 DTU->applyUpdates(Updates);
2694
2695 return true;
2696}
2697
2698/// If we found a conditional branch that goes to two returning blocks,
2699/// try to merge them together into one return,
2700/// introducing a select if the return values disagree.
2701bool SimplifyCFGOpt::SimplifyCondBranchToTwoReturns(BranchInst *BI,
2702 IRBuilder<> &Builder) {
2703 auto *BB = BI->getParent();
2704 assert(BI->isConditional() && "Must be a conditional branch")((BI->isConditional() && "Must be a conditional branch"
) ? static_cast<void> (0) : __assert_fail ("BI->isConditional() && \"Must be a conditional branch\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2704, __PRETTY_FUNCTION__))
;
2705 BasicBlock *TrueSucc = BI->getSuccessor(0);
2706 BasicBlock *FalseSucc = BI->getSuccessor(1);
2707 // NOTE: destinations may match, this could be degenerate uncond branch.
2708 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2709 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2710
2711 // Check to ensure both blocks are empty (just a return) or optionally empty
2712 // with PHI nodes. If there are other instructions, merging would cause extra
2713 // computation on one path or the other.
2714 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2715 return false;
2716 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2717 return false;
2718
2719 Builder.SetInsertPoint(BI);
2720 // Okay, we found a branch that is going to two return nodes. If
2721 // there is no return value for this function, just change the
2722 // branch into a return.
2723 if (FalseRet->getNumOperands() == 0) {
2724 TrueSucc->removePredecessor(BB);
2725 FalseSucc->removePredecessor(BB);
2726 Builder.CreateRetVoid();
2727 EraseTerminatorAndDCECond(BI);
2728 if (DTU) {
2729 SmallVector<DominatorTree::UpdateType, 2> Updates;
2730 Updates.push_back({DominatorTree::Delete, BB, TrueSucc});
2731 if (TrueSucc != FalseSucc)
2732 Updates.push_back({DominatorTree::Delete, BB, FalseSucc});
2733 DTU->applyUpdates(Updates);
2734 }
2735 return true;
2736 }
2737
2738 // Otherwise, figure out what the true and false return values are
2739 // so we can insert a new select instruction.
2740 Value *TrueValue = TrueRet->getReturnValue();
2741 Value *FalseValue = FalseRet->getReturnValue();
2742
2743 // Unwrap any PHI nodes in the return blocks.
2744 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2745 if (TVPN->getParent() == TrueSucc)
2746 TrueValue = TVPN->getIncomingValueForBlock(BB);
2747 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2748 if (FVPN->getParent() == FalseSucc)
2749 FalseValue = FVPN->getIncomingValueForBlock(BB);
2750
2751 // In order for this transformation to be safe, we must be able to
2752 // unconditionally execute both operands to the return. This is
2753 // normally the case, but we could have a potentially-trapping
2754 // constant expression that prevents this transformation from being
2755 // safe.
2756 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2757 if (TCV->canTrap())
2758 return false;
2759 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2760 if (FCV->canTrap())
2761 return false;
2762
2763 // Okay, we collected all the mapped values and checked them for sanity, and
2764 // defined to really do this transformation. First, update the CFG.
2765 TrueSucc->removePredecessor(BB);
2766 FalseSucc->removePredecessor(BB);
2767
2768 // Insert select instructions where needed.
2769 Value *BrCond = BI->getCondition();
2770 if (TrueValue) {
2771 // Insert a select if the results differ.
2772 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2773 } else if (isa<UndefValue>(TrueValue)) {
2774 TrueValue = FalseValue;
2775 } else {
2776 TrueValue =
2777 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2778 }
2779 }
2780
2781 Value *RI =
2782 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2783
2784 (void)RI;
2785
2786 LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "\nNewRet = " <<
*RI << "\nTRUEBLOCK: " << *TrueSucc << "\nFALSEBLOCK: "
<< *FalseSucc; } } while (false)
2787 << "\n " << *BI << "\nNewRet = " << *RI << "\nTRUEBLOCK: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "\nNewRet = " <<
*RI << "\nTRUEBLOCK: " << *TrueSucc << "\nFALSEBLOCK: "
<< *FalseSucc; } } while (false)
2788 << *TrueSucc << "\nFALSEBLOCK: " << *FalseSucc)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "\nNewRet = " <<
*RI << "\nTRUEBLOCK: " << *TrueSucc << "\nFALSEBLOCK: "
<< *FalseSucc; } } while (false)
;
2789
2790 EraseTerminatorAndDCECond(BI);
2791 if (DTU) {
2792 SmallVector<DominatorTree::UpdateType, 2> Updates;
2793 Updates.push_back({DominatorTree::Delete, BB, TrueSucc});
2794 if (TrueSucc != FalseSucc)
2795 Updates.push_back({DominatorTree::Delete, BB, FalseSucc});
2796 DTU->applyUpdates(Updates);
2797 }
2798
2799 return true;
2800}
2801
2802/// Return true if either PBI or BI has branch weight available, and store
2803/// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2804/// not have branch weight, use 1:1 as its weight.
2805static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2806 uint64_t &PredTrueWeight,
2807 uint64_t &PredFalseWeight,
2808 uint64_t &SuccTrueWeight,
2809 uint64_t &SuccFalseWeight) {
2810 bool PredHasWeights =
2811 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2812 bool SuccHasWeights =
2813 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2814 if (PredHasWeights || SuccHasWeights) {
2815 if (!PredHasWeights)
2816 PredTrueWeight = PredFalseWeight = 1;
2817 if (!SuccHasWeights)
2818 SuccTrueWeight = SuccFalseWeight = 1;
2819 return true;
2820 } else {
2821 return false;
2822 }
2823}
2824
2825// Determine if the two branches share a common destination,
2826// and deduce a glue that we need to use to join branch's conditions
2827// to arrive at the common destination.
2828static Optional<std::pair<Instruction::BinaryOps, bool>>
2829CheckIfCondBranchesShareCommonDestination(BranchInst *BI, BranchInst *PBI) {
2830 assert(BI && PBI && BI->isConditional() && PBI->isConditional() &&((BI && PBI && BI->isConditional() &&
PBI->isConditional() && "Both blocks must end with a conditional branches."
) ? static_cast<void> (0) : __assert_fail ("BI && PBI && BI->isConditional() && PBI->isConditional() && \"Both blocks must end with a conditional branches.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2831, __PRETTY_FUNCTION__))
2831 "Both blocks must end with a conditional branches.")((BI && PBI && BI->isConditional() &&
PBI->isConditional() && "Both blocks must end with a conditional branches."
) ? static_cast<void> (0) : __assert_fail ("BI && PBI && BI->isConditional() && PBI->isConditional() && \"Both blocks must end with a conditional branches.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2831, __PRETTY_FUNCTION__))
;
2832 assert(is_contained(predecessors(BI->getParent()), PBI->getParent()) &&((is_contained(predecessors(BI->getParent()), PBI->getParent
()) && "PredBB must be a predecessor of BB.") ? static_cast
<void> (0) : __assert_fail ("is_contained(predecessors(BI->getParent()), PBI->getParent()) && \"PredBB must be a predecessor of BB.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2833, __PRETTY_FUNCTION__))
2833 "PredBB must be a predecessor of BB.")((is_contained(predecessors(BI->getParent()), PBI->getParent
()) && "PredBB must be a predecessor of BB.") ? static_cast
<void> (0) : __assert_fail ("is_contained(predecessors(BI->getParent()), PBI->getParent()) && \"PredBB must be a predecessor of BB.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2833, __PRETTY_FUNCTION__))
;
2834
2835 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2836 return {{Instruction::Or, false}};
2837 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2838 return {{Instruction::And, false}};
2839 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2840 return {{Instruction::And, true}};
2841 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2842 return {{Instruction::Or, true}};
2843 return None;
2844}
2845
2846static bool PerformBranchToCommonDestFolding(BranchInst *BI, BranchInst *PBI,
2847 DomTreeUpdater *DTU,
2848 MemorySSAUpdater *MSSAU) {
2849 BasicBlock *BB = BI->getParent();
2850 BasicBlock *PredBlock = PBI->getParent();
2851
2852 // Determine if the two branches share a common destination.
2853 Instruction::BinaryOps Opc;
2854 bool InvertPredCond;
2855 std::tie(Opc, InvertPredCond) =
2856 *CheckIfCondBranchesShareCommonDestination(BI, PBI);
2857
2858 LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRANCH TO COMMON DEST:\n"
<< *PBI << *BB; } } while (false)
;
2859
2860 IRBuilder<> Builder(PBI);
2861 // The builder is used to create instructions to eliminate the branch in BB.
2862 // If BB's terminator has !annotation metadata, add it to the new
2863 // instructions.
2864 Builder.CollectMetadataToCopy(BB->getTerminator(),
2865 {LLVMContext::MD_annotation});
2866
2867 // If we need to invert the condition in the pred block to match, do so now.
2868 if (InvertPredCond) {
2869 Value *NewCond = PBI->getCondition();
2870 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2871 CmpInst *CI = cast<CmpInst>(NewCond);
2872 CI->setPredicate(CI->getInversePredicate());
2873 } else {
2874 NewCond =
2875 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2876 }
2877
2878 PBI->setCondition(NewCond);
2879 PBI->swapSuccessors();
2880 }
2881
2882 BasicBlock *UniqueSucc =
2883 PBI->getSuccessor(0) == BB ? BI->getSuccessor(0) : BI->getSuccessor(1);
2884
2885 // Before cloning instructions, notify the successor basic block that it
2886 // is about to have a new predecessor. This will update PHI nodes,
2887 // which will allow us to update live-out uses of bonus instructions.
2888 AddPredecessorToBlock(UniqueSucc, PredBlock, BB, MSSAU);
2889
2890 // Try to update branch weights.
2891 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2892 if (extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2893 SuccTrueWeight, SuccFalseWeight)) {
2894 SmallVector<uint64_t, 8> NewWeights;
2895
2896 if (PBI->getSuccessor(0) == BB) {
2897 // PBI: br i1 %x, BB, FalseDest
2898 // BI: br i1 %y, UniqueSucc, FalseDest
2899 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2900 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2901 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2902 // TrueWeight for PBI * FalseWeight for BI.
2903 // We assume that total weights of a BranchInst can fit into 32 bits.
2904 // Therefore, we will not have overflow using 64-bit arithmetic.
2905 NewWeights.push_back(PredFalseWeight *
2906 (SuccFalseWeight + SuccTrueWeight) +
2907 PredTrueWeight * SuccFalseWeight);
2908 } else {
2909 // PBI: br i1 %x, TrueDest, BB
2910 // BI: br i1 %y, TrueDest, UniqueSucc
2911 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2912 // FalseWeight for PBI * TrueWeight for BI.
2913 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight + SuccTrueWeight) +
2914 PredFalseWeight * SuccTrueWeight);
2915 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2916 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2917 }
2918
2919 // Halve the weights if any of them cannot fit in an uint32_t
2920 FitWeights(NewWeights);
2921
2922 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(), NewWeights.end());
2923 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2924
2925 // TODO: If BB is reachable from all paths through PredBlock, then we
2926 // could replace PBI's branch probabilities with BI's.
2927 } else
2928 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2929
2930 // Now, update the CFG.
2931 PBI->setSuccessor(PBI->getSuccessor(0) != BB, UniqueSucc);
2932
2933 if (DTU)
2934 DTU->applyUpdates({{DominatorTree::Insert, PredBlock, UniqueSucc},
2935 {DominatorTree::Delete, PredBlock, BB}});
2936
2937 // If BI was a loop latch, it may have had associated loop metadata.
2938 // We need to copy it to the new latch, that is, PBI.
2939 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2940 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2941
2942 ValueToValueMapTy VMap; // maps original values to cloned values
2943 CloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(BB, PredBlock, VMap);
2944
2945 // Now that the Cond was cloned into the predecessor basic block,
2946 // or/and the two conditions together.
2947 Instruction *NewCond = cast<Instruction>(Builder.CreateBinOp(
2948 Opc, PBI->getCondition(), VMap[BI->getCondition()], "or.cond"));
2949 PBI->setCondition(NewCond);
2950
2951 // Copy any debug value intrinsics into the end of PredBlock.
2952 for (Instruction &I : *BB) {
2953 if (isa<DbgInfoIntrinsic>(I)) {
2954 Instruction *NewI = I.clone();
2955 RemapInstruction(NewI, VMap,
2956 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2957 NewI->insertBefore(PBI);
2958 }
2959 }
2960
2961 ++NumFoldBranchToCommonDest;
2962 return true;
2963}
2964
2965/// If this basic block is simple enough, and if a predecessor branches to us
2966/// and one of our successors, fold the block into the predecessor and use
2967/// logical operations to pick the right destination.
2968bool llvm::FoldBranchToCommonDest(BranchInst *BI, DomTreeUpdater *DTU,
2969 MemorySSAUpdater *MSSAU,
2970 const TargetTransformInfo *TTI,
2971 unsigned BonusInstThreshold) {
2972 // If this block ends with an unconditional branch,
2973 // let SpeculativelyExecuteBB() deal with it.
2974 if (!BI->isConditional())
2975 return false;
2976
2977 BasicBlock *BB = BI->getParent();
2978
2979 const unsigned PredCount = pred_size(BB);
2980
2981 bool Changed = false;
2982
2983 TargetTransformInfo::TargetCostKind CostKind =
2984 BB->getParent()->hasMinSize() ? TargetTransformInfo::TCK_CodeSize
2985 : TargetTransformInfo::TCK_SizeAndLatency;
2986
2987 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2988
2989 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2990 Cond->getParent() != BB || !Cond->hasOneUse())
2991 return Changed;
2992
2993 // Only allow this transformation if computing the condition doesn't involve
2994 // too many instructions and these involved instructions can be executed
2995 // unconditionally. We denote all involved instructions except the condition
2996 // as "bonus instructions", and only allow this transformation when the
2997 // number of the bonus instructions we'll need to create when cloning into
2998 // each predecessor does not exceed a certain threshold.
2999 unsigned NumBonusInsts = 0;
3000 for (Instruction &I : *BB) {
3001 // Don't check the branch condition comparison itself.
3002 if (&I == Cond)
3003 continue;
3004 // Ignore dbg intrinsics, and the terminator.
3005 if (isa<DbgInfoIntrinsic>(I) || isa<BranchInst>(I))
3006 continue;
3007 // I must be safe to execute unconditionally.
3008 if (!isSafeToSpeculativelyExecute(&I))
3009 return Changed;
3010
3011 // Account for the cost of duplicating this instruction into each
3012 // predecessor.
3013 NumBonusInsts += PredCount;
3014 // Early exits once we reach the limit.
3015 if (NumBonusInsts > BonusInstThreshold)
3016 return Changed;
3017 }
3018
3019 // Cond is known to be a compare or binary operator. Check to make sure that
3020 // neither operand is a potentially-trapping constant expression.
3021 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
3022 if (CE->canTrap())
3023 return Changed;
3024 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
3025 if (CE->canTrap())
3026 return Changed;
3027
3028 // Finally, don't infinitely unroll conditional loops.
3029 if (is_contained(successors(BB), BB))
3030 return Changed;
3031
3032 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3033 BasicBlock *PredBlock = *PI;
3034 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
3035
3036 // Check that we have two conditional branches. If there is a PHI node in
3037 // the common successor, verify that the same value flows in from both
3038 // blocks.
3039 if (!PBI || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
3040 continue;
3041
3042 // Determine if the two branches share a common destination.
3043 Instruction::BinaryOps Opc;
3044 bool InvertPredCond;
3045 if (auto Recepie = CheckIfCondBranchesShareCommonDestination(BI, PBI))
3046 std::tie(Opc, InvertPredCond) = *Recepie;
3047 else
3048 continue;
3049
3050 // Check the cost of inserting the necessary logic before performing the
3051 // transformation.
3052 if (TTI) {
3053 Type *Ty = BI->getCondition()->getType();
3054 unsigned Cost = TTI->getArithmeticInstrCost(Opc, Ty, CostKind);
3055 if (InvertPredCond && (!PBI->getCondition()->hasOneUse() ||
3056 !isa<CmpInst>(PBI->getCondition())))
3057 Cost += TTI->getArithmeticInstrCost(Instruction::Xor, Ty, CostKind);
3058
3059 if (Cost > BranchFoldThreshold)
3060 continue;
3061 }
3062
3063 return PerformBranchToCommonDestFolding(BI, PBI, DTU, MSSAU);
3064 }
3065 return Changed;
3066}
3067
3068// If there is only one store in BB1 and BB2, return it, otherwise return
3069// nullptr.
3070static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
3071 StoreInst *S = nullptr;
3072 for (auto *BB : {BB1, BB2}) {
3073 if (!BB)
3074 continue;
3075 for (auto &I : *BB)
3076 if (auto *SI = dyn_cast<StoreInst>(&I)) {
3077 if (S)
3078 // Multiple stores seen.
3079 return nullptr;
3080 else
3081 S = SI;
3082 }
3083 }
3084 return S;
3085}
3086
3087static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
3088 Value *AlternativeV = nullptr) {
3089 // PHI is going to be a PHI node that allows the value V that is defined in
3090 // BB to be referenced in BB's only successor.
3091 //
3092 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
3093 // doesn't matter to us what the other operand is (it'll never get used). We
3094 // could just create a new PHI with an undef incoming value, but that could
3095 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
3096 // other PHI. So here we directly look for some PHI in BB's successor with V
3097 // as an incoming operand. If we find one, we use it, else we create a new
3098 // one.
3099 //
3100 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
3101 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
3102 // where OtherBB is the single other predecessor of BB's only successor.
3103 PHINode *PHI = nullptr;
3104 BasicBlock *Succ = BB->getSingleSuccessor();
3105
3106 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
3107 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
3108 PHI = cast<PHINode>(I);
3109 if (!AlternativeV)
3110 break;
3111
3112 assert(Succ->hasNPredecessors(2))((Succ->hasNPredecessors(2)) ? static_cast<void> (0)
: __assert_fail ("Succ->hasNPredecessors(2)", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3112, __PRETTY_FUNCTION__))
;
3113 auto PredI = pred_begin(Succ);
3114 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
3115 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
3116 break;
3117 PHI = nullptr;
3118 }
3119 if (PHI)
3120 return PHI;
3121
3122 // If V is not an instruction defined in BB, just return it.
3123 if (!AlternativeV &&
3124 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
3125 return V;
3126
3127 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
3128 PHI->addIncoming(V, BB);
3129 for (BasicBlock *PredBB : predecessors(Succ))
3130 if (PredBB != BB)
3131 PHI->addIncoming(
3132 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
3133 return PHI;
3134}
3135
3136static bool mergeConditionalStoreToAddress(
3137 BasicBlock *PTB, BasicBlock *PFB, BasicBlock *QTB, BasicBlock *QFB,
3138 BasicBlock *PostBB, Value *Address, bool InvertPCond, bool InvertQCond,
3139 DomTreeUpdater *DTU, const DataLayout &DL, const TargetTransformInfo &TTI) {
3140 // For every pointer, there must be exactly two stores, one coming from
3141 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
3142 // store (to any address) in PTB,PFB or QTB,QFB.
3143 // FIXME: We could relax this restriction with a bit more work and performance
3144 // testing.
3145 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
3146 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
3147 if (!PStore || !QStore)
3148 return false;
3149
3150 // Now check the stores are compatible.
3151 if (!QStore->isUnordered() || !PStore->isUnordered())
3152 return false;
3153
3154 // Check that sinking the store won't cause program behavior changes. Sinking
3155 // the store out of the Q blocks won't change any behavior as we're sinking
3156 // from a block to its unconditional successor. But we're moving a store from
3157 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
3158 // So we need to check that there are no aliasing loads or stores in
3159 // QBI, QTB and QFB. We also need to check there are no conflicting memory
3160 // operations between PStore and the end of its parent block.
3161 //
3162 // The ideal way to do this is to query AliasAnalysis, but we don't
3163 // preserve AA currently so that is dangerous. Be super safe and just
3164 // check there are no other memory operations at all.
3165 for (auto &I : *QFB->getSinglePredecessor())
3166 if (I.mayReadOrWriteMemory())
3167 return false;
3168 for (auto &I : *QFB)
3169 if (&I != QStore && I.mayReadOrWriteMemory())
3170 return false;
3171 if (QTB)
3172 for (auto &I : *QTB)
3173 if (&I != QStore && I.mayReadOrWriteMemory())
3174 return false;
3175 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3176 I != E; ++I)
3177 if (&*I != PStore && I->mayReadOrWriteMemory())
3178 return false;
3179
3180 // If we're not in aggressive mode, we only optimize if we have some
3181 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
3182 auto IsWorthwhile = [&](BasicBlock *BB, ArrayRef<StoreInst *> FreeStores) {
3183 if (!BB)
3184 return true;
3185 // Heuristic: if the block can be if-converted/phi-folded and the
3186 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
3187 // thread this store.
3188 int BudgetRemaining =
3189 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
3190 for (auto &I : BB->instructionsWithoutDebug()) {
3191 // Consider terminator instruction to be free.
3192 if (I.isTerminator())
3193 continue;
3194 // If this is one the stores that we want to speculate out of this BB,
3195 // then don't count it's cost, consider it to be free.
3196 if (auto *S = dyn_cast<StoreInst>(&I))
3197 if (llvm::find(FreeStores, S))
3198 continue;
3199 // Else, we have a white-list of instructions that we are ak speculating.
3200 if (!isa<BinaryOperator>(I) && !isa<GetElementPtrInst>(I))
3201 return false; // Not in white-list - not worthwhile folding.
3202 // And finally, if this is a non-free instruction that we are okay
3203 // speculating, ensure that we consider the speculation budget.
3204 BudgetRemaining -= TTI.getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency);
3205 if (BudgetRemaining < 0)
3206 return false; // Eagerly refuse to fold as soon as we're out of budget.
3207 }
3208 assert(BudgetRemaining >= 0 &&((BudgetRemaining >= 0 && "When we run out of budget we will eagerly return from within the "
"per-instruction loop.") ? static_cast<void> (0) : __assert_fail
("BudgetRemaining >= 0 && \"When we run out of budget we will eagerly return from within the \" \"per-instruction loop.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3210, __PRETTY_FUNCTION__))
3209 "When we run out of budget we will eagerly return from within the "((BudgetRemaining >= 0 && "When we run out of budget we will eagerly return from within the "
"per-instruction loop.") ? static_cast<void> (0) : __assert_fail
("BudgetRemaining >= 0 && \"When we run out of budget we will eagerly return from within the \" \"per-instruction loop.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3210, __PRETTY_FUNCTION__))
3210 "per-instruction loop.")((BudgetRemaining >= 0 && "When we run out of budget we will eagerly return from within the "
"per-instruction loop.") ? static_cast<void> (0) : __assert_fail
("BudgetRemaining >= 0 && \"When we run out of budget we will eagerly return from within the \" \"per-instruction loop.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3210, __PRETTY_FUNCTION__))
;
3211 return true;
3212 };
3213
3214 const std::array<StoreInst *, 2> FreeStores = {PStore, QStore};
3215 if (!MergeCondStoresAggressively &&
3216 (!IsWorthwhile(PTB, FreeStores) || !IsWorthwhile(PFB, FreeStores) ||
3217 !IsWorthwhile(QTB, FreeStores) || !IsWorthwhile(QFB, FreeStores)))
3218 return false;
3219
3220 // If PostBB has more than two predecessors, we need to split it so we can
3221 // sink the store.
3222 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3223 // We know that QFB's only successor is PostBB. And QFB has a single
3224 // predecessor. If QTB exists, then its only successor is also PostBB.
3225 // If QTB does not exist, then QFB's only predecessor has a conditional
3226 // branch to QFB and PostBB.
3227 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3228 BasicBlock *NewBB =
3229 SplitBlockPredecessors(PostBB, {QFB, TruePred}, "condstore.split", DTU);
3230 if (!NewBB)
3231 return false;
3232 PostBB = NewBB;
3233 }
3234
3235 // OK, we're going to sink the stores to PostBB. The store has to be
3236 // conditional though, so first create the predicate.
3237 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3238 ->getCondition();
3239 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3240 ->getCondition();
3241
3242 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3243 PStore->getParent());
3244 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3245 QStore->getParent(), PPHI);
3246
3247 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3248
3249 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3250 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3251
3252 if (InvertPCond)
3253 PPred = QB.CreateNot(PPred);
3254 if (InvertQCond)
3255 QPred = QB.CreateNot(QPred);
3256 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3257
3258 auto *T = SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(),
3259 /*Unreachable=*/false,
3260 /*BranchWeights=*/nullptr, DTU);
3261 QB.SetInsertPoint(T);
3262 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3263 AAMDNodes AAMD;
3264 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3265 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3266 SI->setAAMetadata(AAMD);
3267 // Choose the minimum alignment. If we could prove both stores execute, we
3268 // could use biggest one. In this case, though, we only know that one of the
3269 // stores executes. And we don't know it's safe to take the alignment from a
3270 // store that doesn't execute.
3271 SI->setAlignment(std::min(PStore->getAlign(), QStore->getAlign()));
3272
3273 QStore->eraseFromParent();
3274 PStore->eraseFromParent();
3275
3276 return true;
3277}
3278
3279static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3280 DomTreeUpdater *DTU, const DataLayout &DL,
3281 const TargetTransformInfo &TTI) {
3282 // The intention here is to find diamonds or triangles (see below) where each
3283 // conditional block contains a store to the same address. Both of these
3284 // stores are conditional, so they can't be unconditionally sunk. But it may
3285 // be profitable to speculatively sink the stores into one merged store at the
3286 // end, and predicate the merged store on the union of the two conditions of
3287 // PBI and QBI.
3288 //
3289 // This can reduce the number of stores executed if both of the conditions are
3290 // true, and can allow the blocks to become small enough to be if-converted.
3291 // This optimization will also chain, so that ladders of test-and-set
3292 // sequences can be if-converted away.
3293 //
3294 // We only deal with simple diamonds or triangles:
3295 //
3296 // PBI or PBI or a combination of the two
3297 // / \ | \
3298 // PTB PFB | PFB
3299 // \ / | /
3300 // QBI QBI
3301 // / \ | \
3302 // QTB QFB | QFB
3303 // \ / | /
3304 // PostBB PostBB
3305 //
3306 // We model triangles as a type of diamond with a nullptr "true" block.
3307 // Triangles are canonicalized so that the fallthrough edge is represented by
3308 // a true condition, as in the diagram above.
3309 BasicBlock *PTB = PBI->getSuccessor(0);
3310 BasicBlock *PFB = PBI->getSuccessor(1);
3311 BasicBlock *QTB = QBI->getSuccessor(0);
3312 BasicBlock *QFB = QBI->getSuccessor(1);
3313 BasicBlock *PostBB = QFB->getSingleSuccessor();
3314
3315 // Make sure we have a good guess for PostBB. If QTB's only successor is
3316 // QFB, then QFB is a better PostBB.
3317 if (QTB->getSingleSuccessor() == QFB)
3318 PostBB = QFB;
3319
3320 // If we couldn't find a good PostBB, stop.
3321 if (!PostBB)
3322 return false;
3323
3324 bool InvertPCond = false, InvertQCond = false;
3325 // Canonicalize fallthroughs to the true branches.
3326 if (PFB == QBI->getParent()) {
3327 std::swap(PFB, PTB);
3328 InvertPCond = true;
3329 }
3330 if (QFB == PostBB) {
3331 std::swap(QFB, QTB);
3332 InvertQCond = true;
3333 }
3334
3335 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3336 // and QFB may not. Model fallthroughs as a nullptr block.
3337 if (PTB == QBI->getParent())
3338 PTB = nullptr;
3339 if (QTB == PostBB)
3340 QTB = nullptr;
3341
3342 // Legality bailouts. We must have at least the non-fallthrough blocks and
3343 // the post-dominating block, and the non-fallthroughs must only have one
3344 // predecessor.
3345 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3346 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3347 };
3348 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3349 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3350 return false;
3351 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3352 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3353 return false;
3354 if (!QBI->getParent()->hasNUses(2))
3355 return false;
3356
3357 // OK, this is a sequence of two diamonds or triangles.
3358 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3359 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3360 for (auto *BB : {PTB, PFB}) {
3361 if (!BB)
3362 continue;
3363 for (auto &I : *BB)
3364 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3365 PStoreAddresses.insert(SI->getPointerOperand());
3366 }
3367 for (auto *BB : {QTB, QFB}) {
3368 if (!BB)
3369 continue;
3370 for (auto &I : *BB)
3371 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3372 QStoreAddresses.insert(SI->getPointerOperand());
3373 }
3374
3375 set_intersect(PStoreAddresses, QStoreAddresses);
3376 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3377 // clear what it contains.
3378 auto &CommonAddresses = PStoreAddresses;
3379
3380 bool Changed = false;
3381 for (auto *Address : CommonAddresses)
3382 Changed |=
3383 mergeConditionalStoreToAddress(PTB, PFB, QTB, QFB, PostBB, Address,
3384 InvertPCond, InvertQCond, DTU, DL, TTI);
3385 return Changed;
3386}
3387
3388/// If the previous block ended with a widenable branch, determine if reusing
3389/// the target block is profitable and legal. This will have the effect of
3390/// "widening" PBI, but doesn't require us to reason about hosting safety.
3391static bool tryWidenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3392 DomTreeUpdater *DTU) {
3393 // TODO: This can be generalized in two important ways:
3394 // 1) We can allow phi nodes in IfFalseBB and simply reuse all the input
3395 // values from the PBI edge.
3396 // 2) We can sink side effecting instructions into BI's fallthrough
3397 // successor provided they doesn't contribute to computation of
3398 // BI's condition.
3399 Value *CondWB, *WC;
3400 BasicBlock *IfTrueBB, *IfFalseBB;
3401 if (!parseWidenableBranch(PBI, CondWB, WC, IfTrueBB, IfFalseBB) ||
3402 IfTrueBB != BI->getParent() || !BI->getParent()->getSinglePredecessor())
3403 return false;
3404 if (!IfFalseBB->phis().empty())
3405 return false; // TODO
3406 // Use lambda to lazily compute expensive condition after cheap ones.
3407 auto NoSideEffects = [](BasicBlock &BB) {
3408 return !llvm::any_of(BB, [](const Instruction &I) {
3409 return I.mayWriteToMemory() || I.mayHaveSideEffects();
3410 });
3411 };
3412 if (BI->getSuccessor(1) != IfFalseBB && // no inf looping
3413 BI->getSuccessor(1)->getTerminatingDeoptimizeCall() && // profitability
3414 NoSideEffects(*BI->getParent())) {
3415 auto *OldSuccessor = BI->getSuccessor(1);
3416 OldSuccessor->removePredecessor(BI->getParent());
3417 BI->setSuccessor(1, IfFalseBB);
3418 if (DTU)
3419 DTU->applyUpdates(
3420 {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3421 {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3422 return true;
3423 }
3424 if (BI->getSuccessor(0) != IfFalseBB && // no inf looping
3425 BI->getSuccessor(0)->getTerminatingDeoptimizeCall() && // profitability
3426 NoSideEffects(*BI->getParent())) {
3427 auto *OldSuccessor = BI->getSuccessor(0);
3428 OldSuccessor->removePredecessor(BI->getParent());
3429 BI->setSuccessor(0, IfFalseBB);
3430 if (DTU)
3431 DTU->applyUpdates(
3432 {{DominatorTree::Insert, BI->getParent(), IfFalseBB},
3433 {DominatorTree::Delete, BI->getParent(), OldSuccessor}});
3434 return true;
3435 }
3436 return false;
3437}
3438
3439/// If we have a conditional branch as a predecessor of another block,
3440/// this function tries to simplify it. We know
3441/// that PBI and BI are both conditional branches, and BI is in one of the
3442/// successor blocks of PBI - PBI branches to BI.
3443static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3444 DomTreeUpdater *DTU,
3445 const DataLayout &DL,
3446 const TargetTransformInfo &TTI) {
3447 assert(PBI->isConditional() && BI->isConditional())((PBI->isConditional() && BI->isConditional()) ?
static_cast<void> (0) : __assert_fail ("PBI->isConditional() && BI->isConditional()"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3447, __PRETTY_FUNCTION__))
;
3448 BasicBlock *BB = BI->getParent();
3449
3450 // If this block ends with a branch instruction, and if there is a
3451 // predecessor that ends on a branch of the same condition, make
3452 // this conditional branch redundant.
3453 if (PBI->getCondition() == BI->getCondition() &&
3454 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3455 // Okay, the outcome of this conditional branch is statically
3456 // knowable. If this block had a single pred, handle specially.
3457 if (BB->getSinglePredecessor()) {
3458 // Turn this into a branch on constant.
3459 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3460 BI->setCondition(
3461 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3462 return true; // Nuke the branch on constant.
3463 }
3464
3465 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3466 // in the constant and simplify the block result. Subsequent passes of
3467 // simplifycfg will thread the block.
3468 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3469 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3470 PHINode *NewPN = PHINode::Create(
3471 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3472 BI->getCondition()->getName() + ".pr", &BB->front());
3473 // Okay, we're going to insert the PHI node. Since PBI is not the only
3474 // predecessor, compute the PHI'd conditional value for all of the preds.
3475 // Any predecessor where the condition is not computable we keep symbolic.
3476 for (pred_iterator PI = PB; PI != PE; ++PI) {
3477 BasicBlock *P = *PI;
3478 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3479 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3480 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3481 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3482 NewPN->addIncoming(
3483 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3484 P);
3485 } else {
3486 NewPN->addIncoming(BI->getCondition(), P);
3487 }
3488 }
3489
3490 BI->setCondition(NewPN);
3491 return true;
3492 }
3493 }
3494
3495 // If the previous block ended with a widenable branch, determine if reusing
3496 // the target block is profitable and legal. This will have the effect of
3497 // "widening" PBI, but doesn't require us to reason about hosting safety.
3498 if (tryWidenCondBranchToCondBranch(PBI, BI, DTU))
3499 return true;
3500
3501 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3502 if (CE->canTrap())
3503 return false;
3504
3505 // If both branches are conditional and both contain stores to the same
3506 // address, remove the stores from the conditionals and create a conditional
3507 // merged store at the end.
3508 if (MergeCondStores && mergeConditionalStores(PBI, BI, DTU, DL, TTI))
3509 return true;
3510
3511 // If this is a conditional branch in an empty block, and if any
3512 // predecessors are a conditional branch to one of our destinations,
3513 // fold the conditions into logical ops and one cond br.
3514
3515 // Ignore dbg intrinsics.
3516 if (&*BB->instructionsWithoutDebug().begin() != BI)
3517 return false;
3518
3519 int PBIOp, BIOp;
3520 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3521 PBIOp = 0;
3522 BIOp = 0;
3523 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3524 PBIOp = 0;
3525 BIOp = 1;
3526 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3527 PBIOp = 1;
3528 BIOp = 0;
3529 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3530 PBIOp = 1;
3531 BIOp = 1;
3532 } else {
3533 return false;
3534 }
3535
3536 // Check to make sure that the other destination of this branch
3537 // isn't BB itself. If so, this is an infinite loop that will
3538 // keep getting unwound.
3539 if (PBI->getSuccessor(PBIOp) == BB)
3540 return false;
3541
3542 // Do not perform this transformation if it would require
3543 // insertion of a large number of select instructions. For targets
3544 // without predication/cmovs, this is a big pessimization.
3545
3546 // Also do not perform this transformation if any phi node in the common
3547 // destination block can trap when reached by BB or PBB (PR17073). In that
3548 // case, it would be unsafe to hoist the operation into a select instruction.
3549
3550 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3551 BasicBlock *RemovedDest = PBI->getSuccessor(PBIOp ^ 1);
3552 unsigned NumPhis = 0;
3553 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3554 ++II, ++NumPhis) {
3555 if (NumPhis > 2) // Disable this xform.
3556 return false;
3557
3558 PHINode *PN = cast<PHINode>(II);
3559 Value *BIV = PN->getIncomingValueForBlock(BB);
3560 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3561 if (CE->canTrap())
3562 return false;
3563
3564 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3565 Value *PBIV = PN->getIncomingValue(PBBIdx);
3566 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3567 if (CE->canTrap())
3568 return false;
3569 }
3570
3571 // Finally, if everything is ok, fold the branches to logical ops.
3572 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3573
3574 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
3575 << "AND: " << *BI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
;
3576
3577 SmallVector<DominatorTree::UpdateType, 5> Updates;
3578
3579 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3580 // branch in it, where one edge (OtherDest) goes back to itself but the other
3581 // exits. We don't *know* that the program avoids the infinite loop
3582 // (even though that seems likely). If we do this xform naively, we'll end up
3583 // recursively unpeeling the loop. Since we know that (after the xform is
3584 // done) that the block *is* infinite if reached, we just make it an obviously
3585 // infinite loop with no cond branch.
3586 if (OtherDest == BB) {
3587 // Insert it at the end of the function, because it's either code,
3588 // or it won't matter if it's hot. :)
3589 BasicBlock *InfLoopBlock =
3590 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3591 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3592 Updates.push_back({DominatorTree::Insert, InfLoopBlock, InfLoopBlock});
3593 OtherDest = InfLoopBlock;
3594 }
3595
3596 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3597
3598 // BI may have other predecessors. Because of this, we leave
3599 // it alone, but modify PBI.
3600
3601 // Make sure we get to CommonDest on True&True directions.
3602 Value *PBICond = PBI->getCondition();
3603 IRBuilder<NoFolder> Builder(PBI);
3604 if (PBIOp)
3605 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3606
3607 Value *BICond = BI->getCondition();
3608 if (BIOp)
3609 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3610
3611 // Merge the conditions.
3612 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3613
3614 // Modify PBI to branch on the new condition to the new dests.
3615 PBI->setCondition(Cond);
3616 PBI->setSuccessor(0, CommonDest);
3617 PBI->setSuccessor(1, OtherDest);
3618
3619 Updates.push_back({DominatorTree::Insert, PBI->getParent(), OtherDest});
3620 Updates.push_back({DominatorTree::Delete, PBI->getParent(), RemovedDest});
3621
3622 if (DTU)
3623 DTU->applyUpdates(Updates);
3624
3625 // Update branch weight for PBI.
3626 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3627 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3628 bool HasWeights =
3629 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3630 SuccTrueWeight, SuccFalseWeight);
3631 if (HasWeights) {
3632 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3633 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3634 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3635 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3636 // The weight to CommonDest should be PredCommon * SuccTotal +
3637 // PredOther * SuccCommon.
3638 // The weight to OtherDest should be PredOther * SuccOther.
3639 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3640 PredOther * SuccCommon,
3641 PredOther * SuccOther};
3642 // Halve the weights if any of them cannot fit in an uint32_t
3643 FitWeights(NewWeights);
3644
3645 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3646 }
3647
3648 // OtherDest may have phi nodes. If so, add an entry from PBI's
3649 // block that are identical to the entries for BI's block.
3650 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3651
3652 // We know that the CommonDest already had an edge from PBI to
3653 // it. If it has PHIs though, the PHIs may have different
3654 // entries for BB and PBI's BB. If so, insert a select to make
3655 // them agree.
3656 for (PHINode &PN : CommonDest->phis()) {
3657 Value *BIV = PN.getIncomingValueForBlock(BB);
3658 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3659 Value *PBIV = PN.getIncomingValue(PBBIdx);
3660 if (BIV != PBIV) {
3661 // Insert a select in PBI to pick the right value.
3662 SelectInst *NV = cast<SelectInst>(
3663 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3664 PN.setIncomingValue(PBBIdx, NV);
3665 // Although the select has the same condition as PBI, the original branch
3666 // weights for PBI do not apply to the new select because the select's
3667 // 'logical' edges are incoming edges of the phi that is eliminated, not
3668 // the outgoing edges of PBI.
3669 if (HasWeights) {
3670 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3671 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3672 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3673 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3674 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3675 // The weight to PredOtherDest should be PredOther * SuccCommon.
3676 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3677 PredOther * SuccCommon};
3678
3679 FitWeights(NewWeights);
3680
3681 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3682 }
3683 }
3684 }
3685
3686 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "INTO: " << *PBI->
getParent(); } } while (false)
;
3687 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3688
3689 // This basic block is probably dead. We know it has at least
3690 // one fewer predecessor.
3691 return true;
3692}
3693
3694// Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3695// true or to FalseBB if Cond is false.
3696// Takes care of updating the successors and removing the old terminator.
3697// Also makes sure not to introduce new successors by assuming that edges to
3698// non-successor TrueBBs and FalseBBs aren't reachable.
3699bool SimplifyCFGOpt::SimplifyTerminatorOnSelect(Instruction *OldTerm,
3700 Value *Cond, BasicBlock *TrueBB,
3701 BasicBlock *FalseBB,
3702 uint32_t TrueWeight,
3703 uint32_t FalseWeight) {
3704 auto *BB = OldTerm->getParent();
3705 // Remove any superfluous successor edges from the CFG.
3706 // First, figure out which successors to preserve.
3707 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3708 // successor.
3709 BasicBlock *KeepEdge1 = TrueBB;
3710 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3711
3712 SmallSetVector<BasicBlock *, 2> RemovedSuccessors;
3713
3714 // Then remove the rest.
3715 for (BasicBlock *Succ : successors(OldTerm)) {
3716 // Make sure only to keep exactly one copy of each edge.
3717 if (Succ == KeepEdge1)
3718 KeepEdge1 = nullptr;
3719 else if (Succ == KeepEdge2)
3720 KeepEdge2 = nullptr;
3721 else {
3722 Succ->removePredecessor(BB,
3723 /*KeepOneInputPHIs=*/true);
3724
3725 if (Succ != TrueBB && Succ != FalseBB)
3726 RemovedSuccessors.insert(Succ);
3727 }
3728 }
3729
3730 IRBuilder<> Builder(OldTerm);
3731 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3732
3733 // Insert an appropriate new terminator.
3734 if (!KeepEdge1 && !KeepEdge2) {
3735 if (TrueBB == FalseBB) {
3736 // We were only looking for one successor, and it was present.
3737 // Create an unconditional branch to it.
3738 Builder.CreateBr(TrueBB);
3739 } else {
3740 // We found both of the successors we were looking for.
3741 // Create a conditional branch sharing the condition of the select.
3742 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3743 if (TrueWeight != FalseWeight)
3744 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3745 }
3746 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3747 // Neither of the selected blocks were successors, so this
3748 // terminator must be unreachable.
3749 new UnreachableInst(OldTerm->getContext(), OldTerm);
3750 } else {
3751 // One of the selected values was a successor, but the other wasn't.
3752 // Insert an unconditional branch to the one that was found;
3753 // the edge to the one that wasn't must be unreachable.
3754 if (!KeepEdge1) {
3755 // Only TrueBB was found.
3756 Builder.CreateBr(TrueBB);
3757 } else {
3758 // Only FalseBB was found.
3759 Builder.CreateBr(FalseBB);
3760 }
3761 }
3762
3763 EraseTerminatorAndDCECond(OldTerm);
3764
3765 if (DTU) {
3766 SmallVector<DominatorTree::UpdateType, 2> Updates;
3767 Updates.reserve(RemovedSuccessors.size());
3768 for (auto *RemovedSuccessor : RemovedSuccessors)
3769 Updates.push_back({DominatorTree::Delete, BB, RemovedSuccessor});
3770 DTU->applyUpdates(Updates);
3771 }
3772
3773 return true;
3774}
3775
3776// Replaces
3777// (switch (select cond, X, Y)) on constant X, Y
3778// with a branch - conditional if X and Y lead to distinct BBs,
3779// unconditional otherwise.
3780bool SimplifyCFGOpt::SimplifySwitchOnSelect(SwitchInst *SI,
3781 SelectInst *Select) {
3782 // Check for constant integer values in the select.
3783 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3784 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3785 if (!TrueVal || !FalseVal)
3786 return false;
3787
3788 // Find the relevant condition and destinations.
3789 Value *Condition = Select->getCondition();
3790 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3791 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3792
3793 // Get weight for TrueBB and FalseBB.
3794 uint32_t TrueWeight = 0, FalseWeight = 0;
3795 SmallVector<uint64_t, 8> Weights;
3796 bool HasWeights = HasBranchWeights(SI);
3797 if (HasWeights) {
3798 GetBranchWeights(SI, Weights);
3799 if (Weights.size() == 1 + SI->getNumCases()) {
3800 TrueWeight =
3801 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3802 FalseWeight =
3803 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3804 }
3805 }
3806
3807 // Perform the actual simplification.
3808 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3809 FalseWeight);
3810}
3811
3812// Replaces
3813// (indirectbr (select cond, blockaddress(@fn, BlockA),
3814// blockaddress(@fn, BlockB)))
3815// with
3816// (br cond, BlockA, BlockB).
3817bool SimplifyCFGOpt::SimplifyIndirectBrOnSelect(IndirectBrInst *IBI,
3818 SelectInst *SI) {
3819 // Check that both operands of the select are block addresses.
3820 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3821 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3822 if (!TBA || !FBA)
3823 return false;
3824
3825 // Extract the actual blocks.
3826 BasicBlock *TrueBB = TBA->getBasicBlock();
3827 BasicBlock *FalseBB = FBA->getBasicBlock();
3828
3829 // Perform the actual simplification.
3830 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3831 0);
3832}
3833
3834/// This is called when we find an icmp instruction
3835/// (a seteq/setne with a constant) as the only instruction in a
3836/// block that ends with an uncond branch. We are looking for a very specific
3837/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3838/// this case, we merge the first two "or's of icmp" into a switch, but then the
3839/// default value goes to an uncond block with a seteq in it, we get something
3840/// like:
3841///
3842/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3843/// DEFAULT:
3844/// %tmp = icmp eq i8 %A, 92
3845/// br label %end
3846/// end:
3847/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3848///
3849/// We prefer to split the edge to 'end' so that there is a true/false entry to
3850/// the PHI, merging the third icmp into the switch.
3851bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3852 ICmpInst *ICI, IRBuilder<> &Builder) {
3853 BasicBlock *BB = ICI->getParent();
3854
3855 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3856 // complex.
3857 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3858 return false;
3859
3860 Value *V = ICI->getOperand(0);
3861 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3862
3863 // The pattern we're looking for is where our only predecessor is a switch on
3864 // 'V' and this block is the default case for the switch. In this case we can
3865 // fold the compared value into the switch to simplify things.
3866 BasicBlock *Pred = BB->getSinglePredecessor();
3867 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3868 return false;
3869
3870 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3871 if (SI->getCondition() != V)
3872 return false;
3873
3874 // If BB is reachable on a non-default case, then we simply know the value of
3875 // V in this block. Substitute it and constant fold the icmp instruction
3876 // away.
3877 if (SI->getDefaultDest() != BB) {
3878 ConstantInt *VVal = SI->findCaseDest(BB);
3879 assert(VVal && "Should have a unique destination value")((VVal && "Should have a unique destination value") ?
static_cast<void> (0) : __assert_fail ("VVal && \"Should have a unique destination value\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3879, __PRETTY_FUNCTION__))
;
3880 ICI->setOperand(0, VVal);
3881
3882 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3883 ICI->replaceAllUsesWith(V);
3884 ICI->eraseFromParent();
3885 }
3886 // BB is now empty, so it is likely to simplify away.
3887 return requestResimplify();
3888 }
3889
3890 // Ok, the block is reachable from the default dest. If the constant we're
3891 // comparing exists in one of the other edges, then we can constant fold ICI
3892 // and zap it.
3893 if (SI->findCaseValue(Cst) != SI->case_default()) {
3894 Value *V;
3895 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3896 V = ConstantInt::getFalse(BB->getContext());
3897 else
3898 V = ConstantInt::getTrue(BB->getContext());
3899
3900 ICI->replaceAllUsesWith(V);
3901 ICI->eraseFromParent();
3902 // BB is now empty, so it is likely to simplify away.
3903 return requestResimplify();
3904 }
3905
3906 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3907 // the block.
3908 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3909 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3910 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3911 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3912 return false;
3913
3914 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3915 // true in the PHI.
3916 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3917 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3918
3919 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3920 std::swap(DefaultCst, NewCst);
3921
3922 // Replace ICI (which is used by the PHI for the default value) with true or
3923 // false depending on if it is EQ or NE.
3924 ICI->replaceAllUsesWith(DefaultCst);
3925 ICI->eraseFromParent();
3926
3927 SmallVector<DominatorTree::UpdateType, 2> Updates;
3928
3929 // Okay, the switch goes to this block on a default value. Add an edge from
3930 // the switch to the merge point on the compared value.
3931 BasicBlock *NewBB =
3932 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3933 {
3934 SwitchInstProfUpdateWrapper SIW(*SI);
3935 auto W0 = SIW.getSuccessorWeight(0);
3936 SwitchInstProfUpdateWrapper::CaseWeightOpt NewW;
3937 if (W0) {
3938 NewW = ((uint64_t(*W0) + 1) >> 1);
3939 SIW.setSuccessorWeight(0, *NewW);
3940 }
3941 SIW.addCase(Cst, NewBB, NewW);
3942 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
3943 }
3944
3945 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3946 Builder.SetInsertPoint(NewBB);
3947 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3948 Builder.CreateBr(SuccBlock);
3949 Updates.push_back({DominatorTree::Insert, NewBB, SuccBlock});
3950 PHIUse->addIncoming(NewCst, NewBB);
3951 if (DTU)
3952 DTU->applyUpdates(Updates);
3953 return true;
3954}
3955
3956/// The specified branch is a conditional branch.
3957/// Check to see if it is branching on an or/and chain of icmp instructions, and
3958/// fold it into a switch instruction if so.
3959bool SimplifyCFGOpt::SimplifyBranchOnICmpChain(BranchInst *BI,
3960 IRBuilder<> &Builder,
3961 const DataLayout &DL) {
3962 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3963 if (!Cond)
3964 return false;
3965
3966 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3967 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3968 // 'setne's and'ed together, collect them.
3969
3970 // Try to gather values from a chain of and/or to be turned into a switch
3971 ConstantComparesGatherer ConstantCompare(Cond, DL);
3972 // Unpack the result
3973 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3974 Value *CompVal = ConstantCompare.CompValue;
3975 unsigned UsedICmps = ConstantCompare.UsedICmps;
3976 Value *ExtraCase = ConstantCompare.Extra;
3977
3978 // If we didn't have a multiply compared value, fail.
3979 if (!CompVal)
3980 return false;
3981
3982 // Avoid turning single icmps into a switch.
3983 if (UsedICmps <= 1)
3984 return false;
3985
3986 bool TrueWhenEqual = match(Cond, m_LogicalOr(m_Value(), m_Value()));
3987
3988 // There might be duplicate constants in the list, which the switch
3989 // instruction can't handle, remove them now.
3990 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3991 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3992
3993 // If Extra was used, we require at least two switch values to do the
3994 // transformation. A switch with one value is just a conditional branch.
3995 if (ExtraCase && Values.size() < 2)
3996 return false;
3997
3998 // TODO: Preserve branch weight metadata, similarly to how
3999 // FoldValueComparisonIntoPredecessors preserves it.
4000
4001 // Figure out which block is which destination.
4002 BasicBlock *DefaultBB = BI->getSuccessor(1);
4003 BasicBlock *EdgeBB = BI->getSuccessor(0);
4004 if (!TrueWhenEqual)
4005 std::swap(DefaultBB, EdgeBB);
4006
4007 BasicBlock *BB = BI->getParent();
4008
4009 // MSAN does not like undefs as branch condition which can be introduced
4010 // with "explicit branch".
4011 if (ExtraCase && BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
4012 return false;
4013
4014 LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
4015 << " cases into SWITCH. BB is:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
4016 << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
;
4017
4018 SmallVector<DominatorTree::UpdateType, 2> Updates;
4019
4020 // If there are any extra values that couldn't be folded into the switch
4021 // then we evaluate them with an explicit branch first. Split the block
4022 // right before the condbr to handle it.
4023 if (ExtraCase) {
4024 BasicBlock *NewBB = SplitBlock(BB, BI, DTU, /*LI=*/nullptr,
4025 /*MSSAU=*/nullptr, "switch.early.test");
4026
4027 // Remove the uncond branch added to the old block.
4028 Instruction *OldTI = BB->getTerminator();
4029 Builder.SetInsertPoint(OldTI);
4030
4031 if (TrueWhenEqual)
4032 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
4033 else
4034 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
4035
4036 OldTI->eraseFromParent();
4037
4038 Updates.push_back({DominatorTree::Insert, BB, EdgeBB});
4039
4040 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
4041 // for the edge we just added.
4042 AddPredecessorToBlock(EdgeBB, BB, NewBB);
4043
4044 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCasedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
4045 << "\nEXTRABB = " << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
;
4046 BB = NewBB;
4047 }
4048
4049 Builder.SetInsertPoint(BI);
4050 // Convert pointer to int before we switch.
4051 if (CompVal->getType()->isPointerTy()) {
4052 CompVal = Builder.CreatePtrToInt(
4053 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
4054 }
4055
4056 // Create the new switch instruction now.
4057 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
4058
4059 // Add all of the 'cases' to the switch instruction.
4060 for (unsigned i = 0, e = Values.size(); i != e; ++i)
4061 New->addCase(Values[i], EdgeBB);
4062
4063 // We added edges from PI to the EdgeBB. As such, if there were any
4064 // PHI nodes in EdgeBB, they need entries to be added corresponding to
4065 // the number of edges added.
4066 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
4067 PHINode *PN = cast<PHINode>(BBI);
4068 Value *InVal = PN->getIncomingValueForBlock(BB);
4069 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
4070 PN->addIncoming(InVal, BB);
4071 }
4072
4073 // Erase the old branch instruction.
4074 EraseTerminatorAndDCECond(BI);
4075 if (DTU)
4076 DTU->applyUpdates(Updates);
4077
4078 LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain result is:\n"
<< *BB << '\n'; } } while (false)
;
4079 return true;
4080}
4081
4082bool SimplifyCFGOpt::simplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
4083 if (isa<PHINode>(RI->getValue()))
4084 return simplifyCommonResume(RI);
4085 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
4086 RI->getValue() == RI->getParent()->getFirstNonPHI())
4087 // The resume must unwind the exception that caused control to branch here.
4088 return simplifySingleResume(RI);
4089
4090 return false;
4091}
4092
4093// Check if cleanup block is empty
4094static bool isCleanupBlockEmpty(iterator_range<BasicBlock::iterator> R) {
4095 for (Instruction &I : R) {
4096 auto *II = dyn_cast<IntrinsicInst>(&I);
4097 if (!II)
4098 return false;
4099
4100 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
4101 switch (IntrinsicID) {
4102 case Intrinsic::dbg_declare:
4103 case Intrinsic::dbg_value:
4104 case Intrinsic::dbg_label:
4105 case Intrinsic::lifetime_end:
4106 break;
4107 default:
4108 return false;
4109 }
4110 }
4111 return true;
4112}
4113
4114// Simplify resume that is shared by several landing pads (phi of landing pad).
4115bool SimplifyCFGOpt::simplifyCommonResume(ResumeInst *RI) {
4116 BasicBlock *BB = RI->getParent();
4117
4118 // Check that there are no other instructions except for debug and lifetime
4119 // intrinsics between the phi's and resume instruction.
4120 if (!isCleanupBlockEmpty(
1
Assuming the condition is false
2
Taking false branch
4121 make_range(RI->getParent()->getFirstNonPHI(), BB->getTerminator())))
4122 return false;
4123
4124 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
4125 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3
The object is a 'PHINode'
4126
4127 // Check incoming blocks to see if any of them are trivial.
4128 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
4
Assuming 'Idx' is not equal to 'End'
5
Loop condition is true. Entering loop body
4129 Idx++) {
4130 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
4131 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
4132
4133 // If the block has other successors, we can not delete it because
4134 // it has other dependents.
4135 if (IncomingBB->getUniqueSuccessor() != BB)
6
Assuming the condition is false
7
Taking false branch
4136 continue;
4137
4138 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
8
Assuming the object is not a 'LandingPadInst'
9
'LandingPad' initialized to a null pointer value
4139 // Not the landing pad that caused the control to branch here.
4140 if (IncomingValue
9.1
'IncomingValue' is equal to 'LandingPad'
!= LandingPad)
10
Taking false branch
4141 continue;
4142
4143 if (isCleanupBlockEmpty(
4144 make_range(LandingPad->getNextNode(), IncomingBB->getTerminator())))
11
Called C++ object pointer is null
4145 TrivialUnwindBlocks.insert(IncomingBB);
4146 }
4147
4148 // If no trivial unwind blocks, don't do any simplifications.
4149 if (TrivialUnwindBlocks.empty())
4150 return false;
4151
4152 // Turn all invokes that unwind here into calls.
4153 for (auto *TrivialBB : TrivialUnwindBlocks) {
4154 // Blocks that will be simplified should be removed from the phi node.
4155 // Note there could be multiple edges to the resume block, and we need
4156 // to remove them all.
4157 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
4158 BB->removePredecessor(TrivialBB, true);
4159
4160 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
4161 PI != PE;) {
4162 BasicBlock *Pred = *PI++;
4163 removeUnwindEdge(Pred, DTU);
4164 ++NumInvokes;
4165 }
4166
4167 // In each SimplifyCFG run, only the current processed block can be erased.
4168 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
4169 // of erasing TrivialBB, we only remove the branch to the common resume
4170 // block so that we can later erase the resume block since it has no
4171 // predecessors.
4172 TrivialBB->getTerminator()->eraseFromParent();
4173 new UnreachableInst(RI->getContext(), TrivialBB);
4174 if (DTU)
4175 DTU->applyUpdates({{DominatorTree::Delete, TrivialBB, BB}});
4176 }
4177
4178 // Delete the resume block if all its predecessors have been removed.
4179 if (pred_empty(BB)) {
4180 if (DTU)
4181 DTU->deleteBB(BB);
4182 else
4183 BB->eraseFromParent();
4184 }
4185
4186 return !TrivialUnwindBlocks.empty();
4187}
4188
4189// Simplify resume that is only used by a single (non-phi) landing pad.
4190bool SimplifyCFGOpt::simplifySingleResume(ResumeInst *RI) {
4191 BasicBlock *BB = RI->getParent();
4192 auto *LPInst = cast<LandingPadInst>(BB->getFirstNonPHI());
4193 assert(RI->getValue() == LPInst &&((RI->getValue() == LPInst && "Resume must unwind the exception that caused control to here"
) ? static_cast<void> (0) : __assert_fail ("RI->getValue() == LPInst && \"Resume must unwind the exception that caused control to here\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4194, __PRETTY_FUNCTION__))
4194 "Resume must unwind the exception that caused control to here")((RI->getValue() == LPInst && "Resume must unwind the exception that caused control to here"
) ? static_cast<void> (0) : __assert_fail ("RI->getValue() == LPInst && \"Resume must unwind the exception that caused control to here\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4194, __PRETTY_FUNCTION__))
;
4195
4196 // Check that there are no other instructions except for debug intrinsics.
4197 if (!isCleanupBlockEmpty(
4198 make_range<Instruction *>(LPInst->getNextNode(), RI)))
4199 return false;
4200
4201 // Turn all invokes that unwind here into calls and delete the basic block.
4202 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4203 BasicBlock *Pred = *PI++;
4204 removeUnwindEdge(Pred, DTU);
4205 ++NumInvokes;
4206 }
4207
4208 // The landingpad is now unreachable. Zap it.
4209 if (DTU)
4210 DTU->deleteBB(BB);
4211 else
4212 BB->eraseFromParent();
4213 return true;
4214}
4215
4216static bool removeEmptyCleanup(CleanupReturnInst *RI, DomTreeUpdater *DTU) {
4217 // If this is a trivial cleanup pad that executes no instructions, it can be
4218 // eliminated. If the cleanup pad continues to the caller, any predecessor
4219 // that is an EH pad will be updated to continue to the caller and any
4220 // predecessor that terminates with an invoke instruction will have its invoke
4221 // instruction converted to a call instruction. If the cleanup pad being
4222 // simplified does not continue to the caller, each predecessor will be
4223 // updated to continue to the unwind destination of the cleanup pad being
4224 // simplified.
4225 BasicBlock *BB = RI->getParent();
4226 CleanupPadInst *CPInst = RI->getCleanupPad();
4227 if (CPInst->getParent() != BB)
4228 // This isn't an empty cleanup.
4229 return false;
4230
4231 // We cannot kill the pad if it has multiple uses. This typically arises
4232 // from unreachable basic blocks.
4233 if (!CPInst->hasOneUse())
4234 return false;
4235
4236 // Check that there are no other instructions except for benign intrinsics.
4237 if (!isCleanupBlockEmpty(
4238 make_range<Instruction *>(CPInst->getNextNode(), RI)))
4239 return false;
4240
4241 // If the cleanup return we are simplifying unwinds to the caller, this will
4242 // set UnwindDest to nullptr.
4243 BasicBlock *UnwindDest = RI->getUnwindDest();
4244 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
4245
4246 // We're about to remove BB from the control flow. Before we do, sink any
4247 // PHINodes into the unwind destination. Doing this before changing the
4248 // control flow avoids some potentially slow checks, since we can currently
4249 // be certain that UnwindDest and BB have no common predecessors (since they
4250 // are both EH pads).
4251 if (UnwindDest) {
4252 // First, go through the PHI nodes in UnwindDest and update any nodes that
4253 // reference the block we are removing
4254 for (BasicBlock::iterator I = UnwindDest->begin(),
4255 IE = DestEHPad->getIterator();
4256 I != IE; ++I) {
4257 PHINode *DestPN = cast<PHINode>(I);
4258
4259 int Idx = DestPN->getBasicBlockIndex(BB);
4260 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
4261 assert(Idx != -1)((Idx != -1) ? static_cast<void> (0) : __assert_fail ("Idx != -1"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4261, __PRETTY_FUNCTION__))
;
4262 // This PHI node has an incoming value that corresponds to a control
4263 // path through the cleanup pad we are removing. If the incoming
4264 // value is in the cleanup pad, it must be a PHINode (because we
4265 // verified above that the block is otherwise empty). Otherwise, the
4266 // value is either a constant or a value that dominates the cleanup
4267 // pad being removed.
4268 //
4269 // Because BB and UnwindDest are both EH pads, all of their
4270 // predecessors must unwind to these blocks, and since no instruction
4271 // can have multiple unwind destinations, there will be no overlap in
4272 // incoming blocks between SrcPN and DestPN.
4273 Value *SrcVal = DestPN->getIncomingValue(Idx);
4274 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
4275
4276 // Remove the entry for the block we are deleting.
4277 DestPN->removeIncomingValue(Idx, false);
4278
4279 if (SrcPN && SrcPN->getParent() == BB) {
4280 // If the incoming value was a PHI node in the cleanup pad we are
4281 // removing, we need to merge that PHI node's incoming values into
4282 // DestPN.
4283 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
4284 SrcIdx != SrcE; ++SrcIdx) {
4285 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
4286 SrcPN->getIncomingBlock(SrcIdx));
4287 }
4288 } else {
4289 // Otherwise, the incoming value came from above BB and
4290 // so we can just reuse it. We must associate all of BB's
4291 // predecessors with this value.
4292 for (auto *pred : predecessors(BB)) {
4293 DestPN->addIncoming(SrcVal, pred);
4294 }
4295 }
4296 }
4297
4298 // Sink any remaining PHI nodes directly into UnwindDest.
4299 Instruction *InsertPt = DestEHPad;
4300 for (BasicBlock::iterator I = BB->begin(),
4301 IE = BB->getFirstNonPHI()->getIterator();
4302 I != IE;) {
4303 // The iterator must be incremented here because the instructions are
4304 // being moved to another block.
4305 PHINode *PN = cast<PHINode>(I++);
4306 if (PN->use_empty() || !PN->isUsedOutsideOfBlock(BB))
4307 // If the PHI node has no uses or all of its uses are in this basic
4308 // block (meaning they are debug or lifetime intrinsics), just leave
4309 // it. It will be erased when we erase BB below.
4310 continue;
4311
4312 // Otherwise, sink this PHI node into UnwindDest.
4313 // Any predecessors to UnwindDest which are not already represented
4314 // must be back edges which inherit the value from the path through
4315 // BB. In this case, the PHI value must reference itself.
4316 for (auto *pred : predecessors(UnwindDest))
4317 if (pred != BB)
4318 PN->addIncoming(PN, pred);
4319 PN->moveBefore(InsertPt);
4320 }
4321 }
4322
4323 std::vector<DominatorTree::UpdateType> Updates;
4324
4325 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4326 // The iterator must be updated here because we are removing this pred.
4327 BasicBlock *PredBB = *PI++;
4328 if (UnwindDest == nullptr) {
4329 if (DTU)
4330 DTU->applyUpdates(Updates);
4331 Updates.clear();
4332 removeUnwindEdge(PredBB, DTU);
4333 ++NumInvokes;
4334 } else {
4335 Instruction *TI = PredBB->getTerminator();
4336 TI->replaceUsesOfWith(BB, UnwindDest);
4337 Updates.push_back({DominatorTree::Insert, PredBB, UnwindDest});
4338 Updates.push_back({DominatorTree::Delete, PredBB, BB});
4339 }
4340 }
4341
4342 if (DTU) {
4343 DTU->applyUpdates(Updates);
4344 DTU->deleteBB(BB);
4345 } else
4346 // The cleanup pad is now unreachable. Zap it.
4347 BB->eraseFromParent();
4348
4349 return true;
4350}
4351
4352// Try to merge two cleanuppads together.
4353static bool mergeCleanupPad(CleanupReturnInst *RI) {
4354 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4355 // with.
4356 BasicBlock *UnwindDest = RI->getUnwindDest();
4357 if (!UnwindDest)
4358 return false;
4359
4360 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4361 // be safe to merge without code duplication.
4362 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4363 return false;
4364
4365 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4366 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4367 if (!SuccessorCleanupPad)
4368 return false;
4369
4370 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4371 // Replace any uses of the successor cleanupad with the predecessor pad
4372 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4373 // funclet bundle operands.
4374 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4375 // Remove the old cleanuppad.
4376 SuccessorCleanupPad->eraseFromParent();
4377 // Now, we simply replace the cleanupret with a branch to the unwind
4378 // destination.
4379 BranchInst::Create(UnwindDest, RI->getParent());
4380 RI->eraseFromParent();
4381
4382 return true;
4383}
4384
4385bool SimplifyCFGOpt::simplifyCleanupReturn(CleanupReturnInst *RI) {
4386 // It is possible to transiantly have an undef cleanuppad operand because we
4387 // have deleted some, but not all, dead blocks.
4388 // Eventually, this block will be deleted.
4389 if (isa<UndefValue>(RI->getOperand(0)))
4390 return false;
4391
4392 if (mergeCleanupPad(RI))
4393 return true;
4394
4395 if (removeEmptyCleanup(RI, DTU))
4396 return true;
4397
4398 return false;
4399}
4400
4401bool SimplifyCFGOpt::simplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4402 BasicBlock *BB = RI->getParent();
4403 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4404 return false;
4405
4406 // Find predecessors that end with branches.
4407 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4408 SmallVector<BranchInst *, 8> CondBranchPreds;
4409 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4410 BasicBlock *P = *PI;
4411 Instruction *PTI = P->getTerminator();
4412 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4413 if (BI->isUnconditional())
4414 UncondBranchPreds.push_back(P);
4415 else
4416 CondBranchPreds.push_back(BI);
4417 }
4418 }
4419
4420 // If we found some, do the transformation!
4421 if (!UncondBranchPreds.empty() && DupRet) {
4422 while (!UncondBranchPreds.empty()) {
4423 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4424 LLVM_DEBUG(dbgs() << "FOLDING: " << *BBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
4425 << "INTO UNCOND BRANCH PRED: " << *Pred)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
;
4426 (void)FoldReturnIntoUncondBranch(RI, BB, Pred, DTU);
4427 }
4428
4429 // If we eliminated all predecessors of the block, delete the block now.
4430 if (pred_empty(BB)) {
4431 // We know there are no successors, so just nuke the block.
4432 if (DTU)
4433 DTU->deleteBB(BB);
4434 else
4435 BB->eraseFromParent();
4436 }
4437
4438 return true;
4439 }
4440
4441 // Check out all of the conditional branches going to this return
4442 // instruction. If any of them just select between returns, change the
4443 // branch itself into a select/return pair.
4444 while (!CondBranchPreds.empty()) {
4445 BranchInst *BI = CondBranchPreds.pop_back_val();
4446
4447 // Check to see if the non-BB successor is also a return block.
4448 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4449 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4450 SimplifyCondBranchToTwoReturns(BI, Builder))
4451 return true;
4452 }
4453 return false;
4454}
4455
4456bool SimplifyCFGOpt::simplifyUnreachable(UnreachableInst *UI) {
4457 BasicBlock *BB = UI->getParent();
4458
4459 bool Changed = false;
4460
4461 // If there are any instructions immediately before the unreachable that can
4462 // be removed, do so.
4463 while (UI->getIterator() != BB->begin()) {
4464 BasicBlock::iterator BBI = UI->getIterator();
4465 --BBI;
4466 // Do not delete instructions that can have side effects which might cause
4467 // the unreachable to not be reachable; specifically, calls and volatile
4468 // operations may have this effect.
4469 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4470 break;
4471
4472 if (BBI->mayHaveSideEffects()) {
4473 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4474 if (SI->isVolatile())
4475 break;
4476 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4477 if (LI->isVolatile())
4478 break;
4479 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4480 if (RMWI->isVolatile())
4481 break;
4482 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4483 if (CXI->isVolatile())
4484 break;
4485 } else if (isa<CatchPadInst>(BBI)) {
4486 // A catchpad may invoke exception object constructors and such, which
4487 // in some languages can be arbitrary code, so be conservative by
4488 // default.
4489 // For CoreCLR, it just involves a type test, so can be removed.
4490 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4491 EHPersonality::CoreCLR)
4492 break;
4493 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4494 !isa<LandingPadInst>(BBI)) {
4495 break;
4496 }
4497 // Note that deleting LandingPad's here is in fact okay, although it
4498 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4499 // all the predecessors of this block will be the unwind edges of Invokes,
4500 // and we can therefore guarantee this block will be erased.
4501 }
4502
4503 // Delete this instruction (any uses are guaranteed to be dead)
4504 if (!BBI->use_empty())
4505 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4506 BBI->eraseFromParent();
4507 Changed = true;
4508 }
4509
4510 // If the unreachable instruction is the first in the block, take a gander
4511 // at all of the predecessors of this instruction, and simplify them.
4512 if (&BB->front() != UI)
4513 return Changed;
4514
4515 std::vector<DominatorTree::UpdateType> Updates;
4516
4517 SmallSetVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4518 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4519 auto *Predecessor = Preds[i];
4520 Instruction *TI = Predecessor->getTerminator();
4521 IRBuilder<> Builder(TI);
4522 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4523 // We could either have a proper unconditional branch,
4524 // or a degenerate conditional branch with matching destinations.
4525 if (all_of(BI->successors(),
4526 [BB](auto *Successor) { return Successor == BB; })) {
4527 new UnreachableInst(TI->getContext(), TI);
4528 TI->eraseFromParent();
4529 Changed = true;
4530 } else {
4531 assert(BI->isConditional() && "Can't get here with an uncond branch.")((BI->isConditional() && "Can't get here with an uncond branch."
) ? static_cast<void> (0) : __assert_fail ("BI->isConditional() && \"Can't get here with an uncond branch.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4531, __PRETTY_FUNCTION__))
;
4532 Value* Cond = BI->getCondition();
4533 assert(BI->getSuccessor(0) != BI->getSuccessor(1) &&((BI->getSuccessor(0) != BI->getSuccessor(1) &&
"The destinations are guaranteed to be different here.") ? static_cast
<void> (0) : __assert_fail ("BI->getSuccessor(0) != BI->getSuccessor(1) && \"The destinations are guaranteed to be different here.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4534, __PRETTY_FUNCTION__))
4534 "The destinations are guaranteed to be different here.")((BI->getSuccessor(0) != BI->getSuccessor(1) &&
"The destinations are guaranteed to be different here.") ? static_cast
<void> (0) : __assert_fail ("BI->getSuccessor(0) != BI->getSuccessor(1) && \"The destinations are guaranteed to be different here.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4534, __PRETTY_FUNCTION__))
;
4535 if (BI->getSuccessor(0) == BB) {
4536 Builder.CreateAssumption(Builder.CreateNot(Cond));
4537 Builder.CreateBr(BI->getSuccessor(1));
4538 } else {
4539 assert(BI->getSuccessor(1) == BB && "Incorrect CFG")((BI->getSuccessor(1) == BB && "Incorrect CFG") ? static_cast
<void> (0) : __assert_fail ("BI->getSuccessor(1) == BB && \"Incorrect CFG\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4539, __PRETTY_FUNCTION__))
;
4540 Builder.CreateAssumption(Cond);
4541 Builder.CreateBr(BI->getSuccessor(0));
4542 }
4543 EraseTerminatorAndDCECond(BI);
4544 Changed = true;
4545 }
4546 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4547 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4548 SwitchInstProfUpdateWrapper SU(*SI);
4549 for (auto i = SU->case_begin(), e = SU->case_end(); i != e;) {
4550 if (i->getCaseSuccessor() != BB) {
4551 ++i;
4552 continue;
4553 }
4554 BB->removePredecessor(SU->getParent());
4555 i = SU.removeCase(i);
4556 e = SU->case_end();
4557 Changed = true;
4558 }
4559 // Note that the default destination can't be removed!
4560 if (SI->getDefaultDest() != BB)
4561 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4562 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4563 if (II->getUnwindDest() == BB) {
4564 if (DTU)
4565 DTU->applyUpdates(Updates);
4566 Updates.clear();
4567 removeUnwindEdge(TI->getParent(), DTU);
4568 Changed = true;
4569 }
4570 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4571 if (CSI->getUnwindDest() == BB) {
4572 if (DTU)
4573 DTU->applyUpdates(Updates);
4574 Updates.clear();
4575 removeUnwindEdge(TI->getParent(), DTU);
4576 Changed = true;
4577 continue;
4578 }
4579
4580 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4581 E = CSI->handler_end();
4582 I != E; ++I) {
4583 if (*I == BB) {
4584 CSI->removeHandler(I);
4585 --I;
4586 --E;
4587 Changed = true;
4588 }
4589 }
4590 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4591 if (CSI->getNumHandlers() == 0) {
4592 if (CSI->hasUnwindDest()) {
4593 // Redirect all predecessors of the block containing CatchSwitchInst
4594 // to instead branch to the CatchSwitchInst's unwind destination.
4595 for (auto *PredecessorOfPredecessor : predecessors(Predecessor)) {
4596 Updates.push_back({DominatorTree::Insert, PredecessorOfPredecessor,
4597 CSI->getUnwindDest()});
4598 Updates.push_back(
4599 {DominatorTree::Delete, PredecessorOfPredecessor, Predecessor});
4600 }
4601 Predecessor->replaceAllUsesWith(CSI->getUnwindDest());
4602 } else {
4603 // Rewrite all preds to unwind to caller (or from invoke to call).
4604 if (DTU)
4605 DTU->applyUpdates(Updates);
4606 Updates.clear();
4607 SmallVector<BasicBlock *, 8> EHPreds(predecessors(Predecessor));
4608 for (BasicBlock *EHPred : EHPreds)
4609 removeUnwindEdge(EHPred, DTU);
4610 }
4611 // The catchswitch is no longer reachable.
4612 new UnreachableInst(CSI->getContext(), CSI);
4613 CSI->eraseFromParent();
4614 Changed = true;
4615 }
4616 } else if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
4617 (void)CRI;
4618 assert(CRI->hasUnwindDest() && CRI->getUnwindDest() == BB &&((CRI->hasUnwindDest() && CRI->getUnwindDest() ==
BB && "Expected to always have an unwind to BB.") ? static_cast
<void> (0) : __assert_fail ("CRI->hasUnwindDest() && CRI->getUnwindDest() == BB && \"Expected to always have an unwind to BB.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4619, __PRETTY_FUNCTION__))
4619 "Expected to always have an unwind to BB.")((CRI->hasUnwindDest() && CRI->getUnwindDest() ==
BB && "Expected to always have an unwind to BB.") ? static_cast
<void> (0) : __assert_fail ("CRI->hasUnwindDest() && CRI->getUnwindDest() == BB && \"Expected to always have an unwind to BB.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4619, __PRETTY_FUNCTION__))
;
4620 Updates.push_back({DominatorTree::Delete, Predecessor, BB});
4621 new UnreachableInst(TI->getContext(), TI);
4622 TI->eraseFromParent();
4623 Changed = true;
4624 }
4625 }
4626
4627 if (DTU)
4628 DTU->applyUpdates(Updates);
4629
4630 // If this block is now dead, remove it.
4631 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4632 // We know there are no successors, so just nuke the block.
4633 if (DTU)
4634 DTU->deleteBB(BB);
4635 else
4636 BB->eraseFromParent();
4637 return true;
4638 }
4639
4640 return Changed;
4641}
4642
4643static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4644 assert(Cases.size() >= 1)((Cases.size() >= 1) ? static_cast<void> (0) : __assert_fail
("Cases.size() >= 1", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4644, __PRETTY_FUNCTION__))
;
4645
4646 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4647 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4648 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4649 return false;
4650 }
4651 return true;
4652}
4653
4654static void createUnreachableSwitchDefault(SwitchInst *Switch,
4655 DomTreeUpdater *DTU) {
4656 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch default is dead.\n"
; } } while (false)
;
4657 auto *BB = Switch->getParent();
4658 BasicBlock *NewDefaultBlock = SplitBlockPredecessors(
4659 Switch->getDefaultDest(), Switch->getParent(), "", DTU);
4660 auto *OrigDefaultBlock = Switch->getDefaultDest();
4661 Switch->setDefaultDest(&*NewDefaultBlock);
4662 if (DTU)
4663 DTU->applyUpdates({{DominatorTree::Insert, BB, &*NewDefaultBlock},
4664 {DominatorTree::Delete, BB, OrigDefaultBlock}});
4665 SplitBlock(&*NewDefaultBlock, &NewDefaultBlock->front(), DTU);
4666 SmallVector<DominatorTree::UpdateType, 2> Updates;
4667 for (auto *Successor : successors(NewDefaultBlock))
4668 Updates.push_back({DominatorTree::Delete, NewDefaultBlock, Successor});
4669 auto *NewTerminator = NewDefaultBlock->getTerminator();
4670 new UnreachableInst(Switch->getContext(), NewTerminator);
4671 EraseTerminatorAndDCECond(NewTerminator);
4672 if (DTU)
4673 DTU->applyUpdates(Updates);
4674}
4675
4676/// Turn a switch with two reachable destinations into an integer range
4677/// comparison and branch.
4678bool SimplifyCFGOpt::TurnSwitchRangeIntoICmp(SwitchInst *SI,
4679 IRBuilder<> &Builder) {
4680 assert(SI->getNumCases() > 1 && "Degenerate switch?")((SI->getNumCases() > 1 && "Degenerate switch?"
) ? static_cast<void> (0) : __assert_fail ("SI->getNumCases() > 1 && \"Degenerate switch?\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4680, __PRETTY_FUNCTION__))
;
4681
4682 bool HasDefault =
4683 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4684
4685 auto *BB = SI->getParent();
4686
4687 // Partition the cases into two sets with different destinations.
4688 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4689 BasicBlock *DestB = nullptr;
4690 SmallVector<ConstantInt *, 16> CasesA;
4691 SmallVector<ConstantInt *, 16> CasesB;
4692
4693 for (auto Case : SI->cases()) {
4694 BasicBlock *Dest = Case.getCaseSuccessor();
4695 if (!DestA)
4696 DestA = Dest;
4697 if (Dest == DestA) {
4698 CasesA.push_back(Case.getCaseValue());
4699 continue;
4700 }
4701 if (!DestB)
4702 DestB = Dest;
4703 if (Dest == DestB) {
4704 CasesB.push_back(Case.getCaseValue());
4705 continue;
4706 }
4707 return false; // More than two destinations.
4708 }
4709
4710 assert(DestA && DestB &&((DestA && DestB && "Single-destination switch should have been folded."
) ? static_cast<void> (0) : __assert_fail ("DestA && DestB && \"Single-destination switch should have been folded.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4711, __PRETTY_FUNCTION__))
4711 "Single-destination switch should have been folded.")((DestA && DestB && "Single-destination switch should have been folded."
) ? static_cast<void> (0) : __assert_fail ("DestA && DestB && \"Single-destination switch should have been folded.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4711, __PRETTY_FUNCTION__))
;
4712 assert(DestA != DestB)((DestA != DestB) ? static_cast<void> (0) : __assert_fail
("DestA != DestB", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4712, __PRETTY_FUNCTION__))
;
4713 assert(DestB != SI->getDefaultDest())((DestB != SI->getDefaultDest()) ? static_cast<void>
(0) : __assert_fail ("DestB != SI->getDefaultDest()", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4713, __PRETTY_FUNCTION__))
;
4714 assert(!CasesB.empty() && "There must be non-default cases.")((!CasesB.empty() && "There must be non-default cases."
) ? static_cast<void> (0) : __assert_fail ("!CasesB.empty() && \"There must be non-default cases.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4714, __PRETTY_FUNCTION__))
;
4715 assert(!CasesA.empty() || HasDefault)((!CasesA.empty() || HasDefault) ? static_cast<void> (0
) : __assert_fail ("!CasesA.empty() || HasDefault", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4715, __PRETTY_FUNCTION__))
;
4716
4717 // Figure out if one of the sets of cases form a contiguous range.
4718 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4719 BasicBlock *ContiguousDest = nullptr;
4720 BasicBlock *OtherDest = nullptr;
4721 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4722 ContiguousCases = &CasesA;
4723 ContiguousDest = DestA;
4724 OtherDest = DestB;
4725 } else if (CasesAreContiguous(CasesB)) {
4726 ContiguousCases = &CasesB;
4727 ContiguousDest = DestB;
4728 OtherDest = DestA;
4729 } else
4730 return false;
4731
4732 // Start building the compare and branch.
4733
4734 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4735 Constant *NumCases =
4736 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4737
4738 Value *Sub = SI->getCondition();
4739 if (!Offset->isNullValue())
4740 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4741
4742 Value *Cmp;
4743 // If NumCases overflowed, then all possible values jump to the successor.
4744 if (NumCases->isNullValue() && !ContiguousCases->empty())
4745 Cmp = ConstantInt::getTrue(SI->getContext());
4746 else
4747 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4748 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4749
4750 // Update weight for the newly-created conditional branch.
4751 if (HasBranchWeights(SI)) {
4752 SmallVector<uint64_t, 8> Weights;
4753 GetBranchWeights(SI, Weights);
4754 if (Weights.size() == 1 + SI->getNumCases()) {
4755 uint64_t TrueWeight = 0;
4756 uint64_t FalseWeight = 0;
4757 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4758 if (SI->getSuccessor(I) == ContiguousDest)
4759 TrueWeight += Weights[I];
4760 else
4761 FalseWeight += Weights[I];
4762 }
4763 while (TrueWeight > UINT32_MAX(4294967295U) || FalseWeight > UINT32_MAX(4294967295U)) {
4764 TrueWeight /= 2;
4765 FalseWeight /= 2;
4766 }
4767 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4768 }
4769 }
4770
4771 // Prune obsolete incoming values off the successors' PHI nodes.
4772 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4773 unsigned PreviousEdges = ContiguousCases->size();
4774 if (ContiguousDest == SI->getDefaultDest())
4775 ++PreviousEdges;
4776 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4777 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4778 }
4779 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4780 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4781 if (OtherDest == SI->getDefaultDest())
4782 ++PreviousEdges;
4783 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4784 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4785 }
4786
4787 // Clean up the default block - it may have phis or other instructions before
4788 // the unreachable terminator.
4789 if (!HasDefault)
4790 createUnreachableSwitchDefault(SI, DTU);
4791
4792 auto *UnreachableDefault = SI->getDefaultDest();
4793
4794 // Drop the switch.
4795 SI->eraseFromParent();
4796
4797 if (!HasDefault && DTU)
4798 DTU->applyUpdates({{DominatorTree::Delete, BB, UnreachableDefault}});
4799
4800 return true;
4801}
4802
4803/// Compute masked bits for the condition of a switch
4804/// and use it to remove dead cases.
4805static bool eliminateDeadSwitchCases(SwitchInst *SI, DomTreeUpdater *DTU,
4806 AssumptionCache *AC,
4807 const DataLayout &DL) {
4808 Value *Cond = SI->getCondition();
4809 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4810 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4811
4812 // We can also eliminate cases by determining that their values are outside of
4813 // the limited range of the condition based on how many significant (non-sign)
4814 // bits are in the condition value.
4815 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4816 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4817
4818 // Gather dead cases.
4819 SmallVector<ConstantInt *, 8> DeadCases;
4820 SmallMapVector<BasicBlock *, int, 8> NumPerSuccessorCases;
4821 for (auto &Case : SI->cases()) {
4822 auto *Successor = Case.getCaseSuccessor();
4823 ++NumPerSuccessorCases[Successor];
4824 const APInt &CaseVal = Case.getCaseValue()->getValue();
4825 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4826 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4827 DeadCases.push_back(Case.getCaseValue());
4828 --NumPerSuccessorCases[Successor];
4829 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
4830 << " is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
;
4831 }
4832 }
4833
4834 // If we can prove that the cases must cover all possible values, the
4835 // default destination becomes dead and we can remove it. If we know some
4836 // of the bits in the value, we can use that to more precisely compute the
4837 // number of possible unique case values.
4838 bool HasDefault =
4839 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4840 const unsigned NumUnknownBits =
4841 Bits - (Known.Zero | Known.One).countPopulation();
4842 assert(NumUnknownBits <= Bits)((NumUnknownBits <= Bits) ? static_cast<void> (0) : __assert_fail
("NumUnknownBits <= Bits", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4842, __PRETTY_FUNCTION__))
;
4843 if (HasDefault && DeadCases.empty() &&
4844 NumUnknownBits < 64 /* avoid overflow */ &&
4845 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4846 createUnreachableSwitchDefault(SI, DTU);
4847 return true;
4848 }
4849
4850 if (DeadCases.empty())
4851 return false;
4852
4853 SwitchInstProfUpdateWrapper SIW(*SI);
4854 for (ConstantInt *DeadCase : DeadCases) {
4855 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4856 assert(CaseI != SI->case_default() &&((CaseI != SI->case_default() && "Case was not found. Probably mistake in DeadCases forming."
) ? static_cast<void> (0) : __assert_fail ("CaseI != SI->case_default() && \"Case was not found. Probably mistake in DeadCases forming.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4857, __PRETTY_FUNCTION__))
4857 "Case was not found. Probably mistake in DeadCases forming.")((CaseI != SI->case_default() && "Case was not found. Probably mistake in DeadCases forming."
) ? static_cast<void> (0) : __assert_fail ("CaseI != SI->case_default() && \"Case was not found. Probably mistake in DeadCases forming.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4857, __PRETTY_FUNCTION__))
;
4858 // Prune unused values from PHI nodes.
4859 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4860 SIW.removeCase(CaseI);
4861 }
4862
4863 std::vector<DominatorTree::UpdateType> Updates;
4864 for (const std::pair<BasicBlock *, int> &I : NumPerSuccessorCases)
4865 if (I.second == 0)
4866 Updates.push_back({DominatorTree::Delete, SI->getParent(), I.first});
4867 if (DTU)
4868 DTU->applyUpdates(Updates);
4869
4870 return true;
4871}
4872
4873/// If BB would be eligible for simplification by
4874/// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4875/// by an unconditional branch), look at the phi node for BB in the successor
4876/// block and see if the incoming value is equal to CaseValue. If so, return
4877/// the phi node, and set PhiIndex to BB's index in the phi node.
4878static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4879 BasicBlock *BB, int *PhiIndex) {
4880 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4881 return nullptr; // BB must be empty to be a candidate for simplification.
4882 if (!BB->getSinglePredecessor())
4883 return nullptr; // BB must be dominated by the switch.
4884
4885 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4886 if (!Branch || !Branch->isUnconditional())
4887 return nullptr; // Terminator must be unconditional branch.
4888
4889 BasicBlock *Succ = Branch->getSuccessor(0);
4890
4891 for (PHINode &PHI : Succ->phis()) {
4892 int Idx = PHI.getBasicBlockIndex(BB);
4893 assert(Idx >= 0 && "PHI has no entry for predecessor?")((Idx >= 0 && "PHI has no entry for predecessor?")
? static_cast<void> (0) : __assert_fail ("Idx >= 0 && \"PHI has no entry for predecessor?\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4893, __PRETTY_FUNCTION__))
;
4894
4895 Value *InValue = PHI.getIncomingValue(Idx);
4896 if (InValue != CaseValue)
4897 continue;
4898
4899 *PhiIndex = Idx;
4900 return &PHI;
4901 }
4902
4903 return nullptr;
4904}
4905
4906/// Try to forward the condition of a switch instruction to a phi node
4907/// dominated by the switch, if that would mean that some of the destination
4908/// blocks of the switch can be folded away. Return true if a change is made.
4909static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4910 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4911
4912 ForwardingNodesMap ForwardingNodes;
4913 BasicBlock *SwitchBlock = SI->getParent();
4914 bool Changed = false;
4915 for (auto &Case : SI->cases()) {
4916 ConstantInt *CaseValue = Case.getCaseValue();
4917 BasicBlock *CaseDest = Case.getCaseSuccessor();
4918
4919 // Replace phi operands in successor blocks that are using the constant case
4920 // value rather than the switch condition variable:
4921 // switchbb:
4922 // switch i32 %x, label %default [
4923 // i32 17, label %succ
4924 // ...
4925 // succ:
4926 // %r = phi i32 ... [ 17, %switchbb ] ...
4927 // -->
4928 // %r = phi i32 ... [ %x, %switchbb ] ...
4929
4930 for (PHINode &Phi : CaseDest->phis()) {
4931 // This only works if there is exactly 1 incoming edge from the switch to
4932 // a phi. If there is >1, that means multiple cases of the switch map to 1
4933 // value in the phi, and that phi value is not the switch condition. Thus,
4934 // this transform would not make sense (the phi would be invalid because
4935 // a phi can't have different incoming values from the same block).
4936 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4937 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4938 count(Phi.blocks(), SwitchBlock) == 1) {
4939 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4940 Changed = true;
4941 }
4942 }
4943
4944 // Collect phi nodes that are indirectly using this switch's case constants.
4945 int PhiIdx;
4946 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4947 ForwardingNodes[Phi].push_back(PhiIdx);
4948 }
4949
4950 for (auto &ForwardingNode : ForwardingNodes) {
4951 PHINode *Phi = ForwardingNode.first;
4952 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4953 if (Indexes.size() < 2)
4954 continue;
4955
4956 for (int Index : Indexes)
4957 Phi->setIncomingValue(Index, SI->getCondition());
4958 Changed = true;
4959 }
4960
4961 return Changed;
4962}
4963
4964/// Return true if the backend will be able to handle
4965/// initializing an array of constants like C.
4966static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4967 if (C->isThreadDependent())
4968 return false;
4969 if (C->isDLLImportDependent())
4970 return false;
4971
4972 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4973 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4974 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4975 return false;
4976
4977 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4978 if (!CE->isGEPWithNoNotionalOverIndexing())
4979 return false;
4980 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4981 return false;
4982 }
4983
4984 if (!TTI.shouldBuildLookupTablesForConstant(C))
4985 return false;
4986
4987 return true;
4988}
4989
4990/// If V is a Constant, return it. Otherwise, try to look up
4991/// its constant value in ConstantPool, returning 0 if it's not there.
4992static Constant *
4993LookupConstant(Value *V,
4994 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4995 if (Constant *C = dyn_cast<Constant>(V))
4996 return C;
4997 return ConstantPool.lookup(V);
4998}
4999
5000/// Try to fold instruction I into a constant. This works for
5001/// simple instructions such as binary operations where both operands are
5002/// constant or can be replaced by constants from the ConstantPool. Returns the
5003/// resulting constant on success, 0 otherwise.
5004static Constant *
5005ConstantFold(Instruction *I, const DataLayout &DL,
5006 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
5007 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
5008 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
5009 if (!A)
5010 return nullptr;
5011 if (A->isAllOnesValue())
5012 return LookupConstant(Select->getTrueValue(), ConstantPool);
5013 if (A->isNullValue())
5014 return LookupConstant(Select->getFalseValue(), ConstantPool);
5015 return nullptr;
5016 }
5017
5018 SmallVector<Constant *, 4> COps;
5019 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
5020 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
5021 COps.push_back(A);
5022 else
5023 return nullptr;
5024 }
5025
5026 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
5027 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
5028 COps[1], DL);
5029 }
5030
5031 return ConstantFoldInstOperands(I, COps, DL);
5032}
5033
5034/// Try to determine the resulting constant values in phi nodes
5035/// at the common destination basic block, *CommonDest, for one of the case
5036/// destionations CaseDest corresponding to value CaseVal (0 for the default
5037/// case), of a switch instruction SI.
5038static bool
5039GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
5040 BasicBlock **CommonDest,
5041 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
5042 const DataLayout &DL, const TargetTransformInfo &TTI) {
5043 // The block from which we enter the common destination.
5044 BasicBlock *Pred = SI->getParent();
5045
5046 // If CaseDest is empty except for some side-effect free instructions through
5047 // which we can constant-propagate the CaseVal, continue to its successor.
5048 SmallDenseMap<Value *, Constant *> ConstantPool;
5049 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
5050 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
5051 if (I.isTerminator()) {
5052 // If the terminator is a simple branch, continue to the next block.
5053 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
5054 return false;
5055 Pred = CaseDest;
5056 CaseDest = I.getSuccessor(0);
5057 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
5058 // Instruction is side-effect free and constant.
5059
5060 // If the instruction has uses outside this block or a phi node slot for
5061 // the block, it is not safe to bypass the instruction since it would then
5062 // no longer dominate all its uses.
5063 for (auto &Use : I.uses()) {
5064 User *User = Use.getUser();
5065 if (Instruction *I = dyn_cast<Instruction>(User))
5066 if (I->getParent() == CaseDest)
5067 continue;
5068 if (PHINode *Phi = dyn_cast<PHINode>(User))
5069 if (Phi->getIncomingBlock(Use) == CaseDest)
5070 continue;
5071 return false;
5072 }
5073
5074 ConstantPool.insert(std::make_pair(&I, C));
5075 } else {
5076 break;
5077 }
5078 }
5079
5080 // If we did not have a CommonDest before, use the current one.
5081 if (!*CommonDest)
5082 *CommonDest = CaseDest;
5083 // If the destination isn't the common one, abort.
5084 if (CaseDest != *CommonDest)
5085 return false;
5086
5087 // Get the values for this case from phi nodes in the destination block.
5088 for (PHINode &PHI : (*CommonDest)->phis()) {
5089 int Idx = PHI.getBasicBlockIndex(Pred);
5090 if (Idx == -1)
5091 continue;
5092
5093 Constant *ConstVal =
5094 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
5095 if (!ConstVal)
5096 return false;
5097
5098 // Be conservative about which kinds of constants we support.
5099 if (!ValidLookupTableConstant(ConstVal, TTI))
5100 return false;
5101
5102 Res.push_back(std::make_pair(&PHI, ConstVal));
5103 }
5104
5105 return Res.size() > 0;
5106}
5107
5108// Helper function used to add CaseVal to the list of cases that generate
5109// Result. Returns the updated number of cases that generate this result.
5110static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
5111 SwitchCaseResultVectorTy &UniqueResults,
5112 Constant *Result) {
5113 for (auto &I : UniqueResults) {
5114 if (I.first == Result) {
5115 I.second.push_back(CaseVal);
5116 return I.second.size();
5117 }
5118 }
5119 UniqueResults.push_back(
5120 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
5121 return 1;
5122}
5123
5124// Helper function that initializes a map containing
5125// results for the PHI node of the common destination block for a switch
5126// instruction. Returns false if multiple PHI nodes have been found or if
5127// there is not a common destination block for the switch.
5128static bool
5129InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
5130 SwitchCaseResultVectorTy &UniqueResults,
5131 Constant *&DefaultResult, const DataLayout &DL,
5132 const TargetTransformInfo &TTI,
5133 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
5134 for (auto &I : SI->cases()) {
5135 ConstantInt *CaseVal = I.getCaseValue();
5136
5137 // Resulting value at phi nodes for this case value.
5138 SwitchCaseResultsTy Results;
5139 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
5140 DL, TTI))
5141 return false;
5142
5143 // Only one value per case is permitted.
5144 if (Results.size() > 1)
5145 return false;
5146
5147 // Add the case->result mapping to UniqueResults.
5148 const uintptr_t NumCasesForResult =
5149 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
5150
5151 // Early out if there are too many cases for this result.
5152 if (NumCasesForResult > MaxCasesPerResult)
5153 return false;
5154
5155 // Early out if there are too many unique results.
5156 if (UniqueResults.size() > MaxUniqueResults)
5157 return false;
5158
5159 // Check the PHI consistency.
5160 if (!PHI)
5161 PHI = Results[0].first;
5162 else if (PHI != Results[0].first)
5163 return false;
5164 }
5165 // Find the default result value.
5166 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
5167 BasicBlock *DefaultDest = SI->getDefaultDest();
5168 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
5169 DL, TTI);
5170 // If the default value is not found abort unless the default destination
5171 // is unreachable.
5172 DefaultResult =
5173 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
5174 if ((!DefaultResult &&
5175 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
5176 return false;
5177
5178 return true;
5179}
5180
5181// Helper function that checks if it is possible to transform a switch with only
5182// two cases (or two cases + default) that produces a result into a select.
5183// Example:
5184// switch (a) {
5185// case 10: %0 = icmp eq i32 %a, 10
5186// return 10; %1 = select i1 %0, i32 10, i32 4
5187// case 20: ----> %2 = icmp eq i32 %a, 20
5188// return 2; %3 = select i1 %2, i32 2, i32 %1
5189// default:
5190// return 4;
5191// }
5192static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
5193 Constant *DefaultResult, Value *Condition,
5194 IRBuilder<> &Builder) {
5195 assert(ResultVector.size() == 2 &&((ResultVector.size() == 2 && "We should have exactly two unique results at this point"
) ? static_cast<void> (0) : __assert_fail ("ResultVector.size() == 2 && \"We should have exactly two unique results at this point\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5196, __PRETTY_FUNCTION__))
5196 "We should have exactly two unique results at this point")((ResultVector.size() == 2 && "We should have exactly two unique results at this point"
) ? static_cast<void> (0) : __assert_fail ("ResultVector.size() == 2 && \"We should have exactly two unique results at this point\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5196, __PRETTY_FUNCTION__))
;
5197 // If we are selecting between only two cases transform into a simple
5198 // select or a two-way select if default is possible.
5199 if (ResultVector[0].second.size() == 1 &&
5200 ResultVector[1].second.size() == 1) {
5201 ConstantInt *const FirstCase = ResultVector[0].second[0];
5202 ConstantInt *const SecondCase = ResultVector[1].second[0];
5203
5204 bool DefaultCanTrigger = DefaultResult;
5205 Value *SelectValue = ResultVector[1].first;
5206 if (DefaultCanTrigger) {
5207 Value *const ValueCompare =
5208 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
5209 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
5210 DefaultResult, "switch.select");
5211 }
5212 Value *const ValueCompare =
5213 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
5214 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
5215 SelectValue, "switch.select");
5216 }
5217
5218 return nullptr;
5219}
5220
5221// Helper function to cleanup a switch instruction that has been converted into
5222// a select, fixing up PHI nodes and basic blocks.
5223static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
5224 Value *SelectValue,
5225 IRBuilder<> &Builder,
5226 DomTreeUpdater *DTU) {
5227 std::vector<DominatorTree::UpdateType> Updates;
5228
5229 BasicBlock *SelectBB = SI->getParent();
5230 BasicBlock *DestBB = PHI->getParent();
5231
5232 if (!is_contained(predecessors(DestBB), SelectBB))
5233 Updates.push_back({DominatorTree::Insert, SelectBB, DestBB});
5234 Builder.CreateBr(DestBB);
5235
5236 // Remove the switch.
5237
5238 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
5239 PHI->removeIncomingValue(SelectBB);
5240 PHI->addIncoming(SelectValue, SelectBB);
5241
5242 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5243 BasicBlock *Succ = SI->getSuccessor(i);
5244
5245 if (Succ == DestBB)
5246 continue;
5247 Succ->removePredecessor(SelectBB);
5248 Updates.push_back({DominatorTree::Delete, SelectBB, Succ});
5249 }
5250 SI->eraseFromParent();
5251 if (DTU)
5252 DTU->applyUpdates(Updates);
5253}
5254
5255/// If the switch is only used to initialize one or more
5256/// phi nodes in a common successor block with only two different
5257/// constant values, replace the switch with select.
5258static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
5259 DomTreeUpdater *DTU, const DataLayout &DL,
5260 const TargetTransformInfo &TTI) {
5261 Value *const Cond = SI->getCondition();
5262 PHINode *PHI = nullptr;
5263 BasicBlock *CommonDest = nullptr;
5264 Constant *DefaultResult;
5265 SwitchCaseResultVectorTy UniqueResults;
5266 // Collect all the cases that will deliver the same value from the switch.
5267 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
5268 DL, TTI, 2, 1))
5269 return false;
5270 // Selects choose between maximum two values.
5271 if (UniqueResults.size() != 2)
5272 return false;
5273 assert(PHI != nullptr && "PHI for value select not found")((PHI != nullptr && "PHI for value select not found")
? static_cast<void> (0) : __assert_fail ("PHI != nullptr && \"PHI for value select not found\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5273, __PRETTY_FUNCTION__))
;
5274
5275 Builder.SetInsertPoint(SI);
5276 Value *SelectValue =
5277 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
5278 if (SelectValue) {
5279 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder, DTU);
5280 return true;
5281 }
5282 // The switch couldn't be converted into a select.
5283 return false;
5284}
5285
5286namespace {
5287
5288/// This class represents a lookup table that can be used to replace a switch.
5289class SwitchLookupTable {
5290public:
5291 /// Create a lookup table to use as a switch replacement with the contents
5292 /// of Values, using DefaultValue to fill any holes in the table.
5293 SwitchLookupTable(
5294 Module &M, uint64_t TableSize, ConstantInt *Offset,
5295 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5296 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
5297
5298 /// Build instructions with Builder to retrieve the value at
5299 /// the position given by Index in the lookup table.
5300 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
5301
5302 /// Return true if a table with TableSize elements of
5303 /// type ElementType would fit in a target-legal register.
5304 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
5305 Type *ElementType);
5306
5307private:
5308 // Depending on the contents of the table, it can be represented in
5309 // different ways.
5310 enum {
5311 // For tables where each element contains the same value, we just have to
5312 // store that single value and return it for each lookup.
5313 SingleValueKind,
5314
5315 // For tables where there is a linear relationship between table index
5316 // and values. We calculate the result with a simple multiplication
5317 // and addition instead of a table lookup.
5318 LinearMapKind,
5319
5320 // For small tables with integer elements, we can pack them into a bitmap
5321 // that fits into a target-legal register. Values are retrieved by
5322 // shift and mask operations.
5323 BitMapKind,
5324
5325 // The table is stored as an array of values. Values are retrieved by load
5326 // instructions from the table.
5327 ArrayKind
5328 } Kind;
5329
5330 // For SingleValueKind, this is the single value.
5331 Constant *SingleValue = nullptr;
5332
5333 // For BitMapKind, this is the bitmap.
5334 ConstantInt *BitMap = nullptr;
5335 IntegerType *BitMapElementTy = nullptr;
5336
5337 // For LinearMapKind, these are the constants used to derive the value.
5338 ConstantInt *LinearOffset = nullptr;
5339 ConstantInt *LinearMultiplier = nullptr;
5340
5341 // For ArrayKind, this is the array.
5342 GlobalVariable *Array = nullptr;
5343};
5344
5345} // end anonymous namespace
5346
5347SwitchLookupTable::SwitchLookupTable(
5348 Module &M, uint64_t TableSize, ConstantInt *Offset,
5349 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
5350 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
5351 assert(Values.size() && "Can't build lookup table without values!")((Values.size() && "Can't build lookup table without values!"
) ? static_cast<void> (0) : __assert_fail ("Values.size() && \"Can't build lookup table without values!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5351, __PRETTY_FUNCTION__))
;
5352 assert(TableSize >= Values.size() && "Can't fit values in table!")((TableSize >= Values.size() && "Can't fit values in table!"
) ? static_cast<void> (0) : __assert_fail ("TableSize >= Values.size() && \"Can't fit values in table!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5352, __PRETTY_FUNCTION__))
;
5353
5354 // If all values in the table are equal, this is that value.
5355 SingleValue = Values.begin()->second;
5356
5357 Type *ValueType = Values.begin()->second->getType();
5358
5359 // Build up the table contents.
5360 SmallVector<Constant *, 64> TableContents(TableSize);
5361 for (size_t I = 0, E = Values.size(); I != E; ++I) {
5362 ConstantInt *CaseVal = Values[I].first;
5363 Constant *CaseRes = Values[I].second;
5364 assert(CaseRes->getType() == ValueType)((CaseRes->getType() == ValueType) ? static_cast<void>
(0) : __assert_fail ("CaseRes->getType() == ValueType", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5364, __PRETTY_FUNCTION__))
;
5365
5366 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
5367 TableContents[Idx] = CaseRes;
5368
5369 if (CaseRes != SingleValue)
5370 SingleValue = nullptr;
5371 }
5372
5373 // Fill in any holes in the table with the default result.
5374 if (Values.size() < TableSize) {
5375 assert(DefaultValue &&((DefaultValue && "Need a default value to fill the lookup table holes."
) ? static_cast<void> (0) : __assert_fail ("DefaultValue && \"Need a default value to fill the lookup table holes.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5376, __PRETTY_FUNCTION__))
5376 "Need a default value to fill the lookup table holes.")((DefaultValue && "Need a default value to fill the lookup table holes."
) ? static_cast<void> (0) : __assert_fail ("DefaultValue && \"Need a default value to fill the lookup table holes.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5376, __PRETTY_FUNCTION__))
;
5377 assert(DefaultValue->getType() == ValueType)((DefaultValue->getType() == ValueType) ? static_cast<void
> (0) : __assert_fail ("DefaultValue->getType() == ValueType"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5377, __PRETTY_FUNCTION__))
;
5378 for (uint64_t I = 0; I < TableSize; ++I) {
5379 if (!TableContents[I])
5380 TableContents[I] = DefaultValue;
5381 }
5382
5383 if (DefaultValue != SingleValue)
5384 SingleValue = nullptr;
5385 }
5386
5387 // If each element in the table contains the same value, we only need to store
5388 // that single value.
5389 if (SingleValue) {
5390 Kind = SingleValueKind;
5391 return;
5392 }
5393
5394 // Check if we can derive the value with a linear transformation from the
5395 // table index.
5396 if (isa<IntegerType>(ValueType)) {
5397 bool LinearMappingPossible = true;
5398 APInt PrevVal;
5399 APInt DistToPrev;
5400 assert(TableSize >= 2 && "Should be a SingleValue table.")((TableSize >= 2 && "Should be a SingleValue table."
) ? static_cast<void> (0) : __assert_fail ("TableSize >= 2 && \"Should be a SingleValue table.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5400, __PRETTY_FUNCTION__))
;
5401 // Check if there is the same distance between two consecutive values.
5402 for (uint64_t I = 0; I < TableSize; ++I) {
5403 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5404 if (!ConstVal) {
5405 // This is an undef. We could deal with it, but undefs in lookup tables
5406 // are very seldom. It's probably not worth the additional complexity.
5407 LinearMappingPossible = false;
5408 break;
5409 }
5410 const APInt &Val = ConstVal->getValue();
5411 if (I != 0) {
5412 APInt Dist = Val - PrevVal;
5413 if (I == 1) {
5414 DistToPrev = Dist;
5415 } else if (Dist != DistToPrev) {
5416 LinearMappingPossible = false;
5417 break;
5418 }
5419 }
5420 PrevVal = Val;
5421 }
5422 if (LinearMappingPossible) {
5423 LinearOffset = cast<ConstantInt>(TableContents[0]);
5424 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5425 Kind = LinearMapKind;
5426 ++NumLinearMaps;
5427 return;
5428 }
5429 }
5430
5431 // If the type is integer and the table fits in a register, build a bitmap.
5432 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5433 IntegerType *IT = cast<IntegerType>(ValueType);
5434 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5435 for (uint64_t I = TableSize; I > 0; --I) {
5436 TableInt <<= IT->getBitWidth();
5437 // Insert values into the bitmap. Undef values are set to zero.
5438 if (!isa<UndefValue>(TableContents[I - 1])) {
5439 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5440 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5441 }
5442 }
5443 BitMap = ConstantInt::get(M.getContext(), TableInt);
5444 BitMapElementTy = IT;
5445 Kind = BitMapKind;
5446 ++NumBitMaps;
5447 return;
5448 }
5449
5450 // Store the table in an array.
5451 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5452 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5453
5454 Array = new GlobalVariable(M, ArrayTy, /*isConstant=*/true,
5455 GlobalVariable::PrivateLinkage, Initializer,
5456 "switch.table." + FuncName);
5457 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5458 // Set the alignment to that of an array items. We will be only loading one
5459 // value out of it.
5460 Array->setAlignment(Align(DL.getPrefTypeAlignment(ValueType)));
5461 Kind = ArrayKind;
5462}
5463
5464Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5465 switch (Kind) {
5466 case SingleValueKind:
5467 return SingleValue;
5468 case LinearMapKind: {
5469 // Derive the result value from the input value.
5470 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5471 false, "switch.idx.cast");
5472 if (!LinearMultiplier->isOne())
5473 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5474 if (!LinearOffset->isZero())
5475 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5476 return Result;
5477 }
5478 case BitMapKind: {
5479 // Type of the bitmap (e.g. i59).
5480 IntegerType *MapTy = BitMap->getType();
5481
5482 // Cast Index to the same type as the bitmap.
5483 // Note: The Index is <= the number of elements in the table, so
5484 // truncating it to the width of the bitmask is safe.
5485 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5486
5487 // Multiply the shift amount by the element width.
5488 ShiftAmt = Builder.CreateMul(
5489 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5490 "switch.shiftamt");
5491
5492 // Shift down.
5493 Value *DownShifted =
5494 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5495 // Mask off.
5496 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5497 }
5498 case ArrayKind: {
5499 // Make sure the table index will not overflow when treated as signed.
5500 IntegerType *IT = cast<IntegerType>(Index->getType());
5501 uint64_t TableSize =
5502 Array->getInitializer()->getType()->getArrayNumElements();
5503 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5504 Index = Builder.CreateZExt(
5505 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5506 "switch.tableidx.zext");
5507
5508 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5509 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5510 GEPIndices, "switch.gep");
5511 return Builder.CreateLoad(
5512 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5513 "switch.load");
5514 }
5515 }
5516 llvm_unreachable("Unknown lookup table kind!")::llvm::llvm_unreachable_internal("Unknown lookup table kind!"
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5516)
;
5517}
5518
5519bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5520 uint64_t TableSize,
5521 Type *ElementType) {
5522 auto *IT = dyn_cast<IntegerType>(ElementType);
5523 if (!IT)
5524 return false;
5525 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5526 // are <= 15, we could try to narrow the type.
5527
5528 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5529 if (TableSize >= UINT_MAX(2147483647 *2U +1U) / IT->getBitWidth())
5530 return false;
5531 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5532}
5533
5534/// Determine whether a lookup table should be built for this switch, based on
5535/// the number of cases, size of the table, and the types of the results.
5536static bool
5537ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5538 const TargetTransformInfo &TTI, const DataLayout &DL,
5539 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5540 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX(18446744073709551615UL) / 10)
5541 return false; // TableSize overflowed, or mul below might overflow.
5542
5543 bool AllTablesFitInRegister = true;
5544 bool HasIllegalType = false;
5545 for (const auto &I : ResultTypes) {
5546 Type *Ty = I.second;
5547
5548 // Saturate this flag to true.
5549 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5550
5551 // Saturate this flag to false.
5552 AllTablesFitInRegister =
5553 AllTablesFitInRegister &&
5554 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5555
5556 // If both flags saturate, we're done. NOTE: This *only* works with
5557 // saturating flags, and all flags have to saturate first due to the
5558 // non-deterministic behavior of iterating over a dense map.
5559 if (HasIllegalType && !AllTablesFitInRegister)
5560 break;
5561 }
5562
5563 // If each table would fit in a register, we should build it anyway.
5564 if (AllTablesFitInRegister)
5565 return true;
5566
5567 // Don't build a table that doesn't fit in-register if it has illegal types.
5568 if (HasIllegalType)
5569 return false;
5570
5571 // The table density should be at least 40%. This is the same criterion as for
5572 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5573 // FIXME: Find the best cut-off.
5574 return SI->getNumCases() * 10 >= TableSize * 4;
5575}
5576
5577/// Try to reuse the switch table index compare. Following pattern:
5578/// \code
5579/// if (idx < tablesize)
5580/// r = table[idx]; // table does not contain default_value
5581/// else
5582/// r = default_value;
5583/// if (r != default_value)
5584/// ...
5585/// \endcode
5586/// Is optimized to:
5587/// \code
5588/// cond = idx < tablesize;
5589/// if (cond)
5590/// r = table[idx];
5591/// else
5592/// r = default_value;
5593/// if (cond)
5594/// ...
5595/// \endcode
5596/// Jump threading will then eliminate the second if(cond).
5597static void reuseTableCompare(
5598 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5599 Constant *DefaultValue,
5600 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5601 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5602 if (!CmpInst)
5603 return;
5604
5605 // We require that the compare is in the same block as the phi so that jump
5606 // threading can do its work afterwards.
5607 if (CmpInst->getParent() != PhiBlock)
5608 return;
5609
5610 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5611 if (!CmpOp1)
5612 return;
5613
5614 Value *RangeCmp = RangeCheckBranch->getCondition();
5615 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5616 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5617
5618 // Check if the compare with the default value is constant true or false.
5619 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5620 DefaultValue, CmpOp1, true);
5621 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5622 return;
5623
5624 // Check if the compare with the case values is distinct from the default
5625 // compare result.
5626 for (auto ValuePair : Values) {
5627 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5628 ValuePair.second, CmpOp1, true);
5629 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5630 return;
5631 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&(((CaseConst == TrueConst || CaseConst == FalseConst) &&
"Expect true or false as compare result.") ? static_cast<
void> (0) : __assert_fail ("(CaseConst == TrueConst || CaseConst == FalseConst) && \"Expect true or false as compare result.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5632, __PRETTY_FUNCTION__))
5632 "Expect true or false as compare result.")(((CaseConst == TrueConst || CaseConst == FalseConst) &&
"Expect true or false as compare result.") ? static_cast<
void> (0) : __assert_fail ("(CaseConst == TrueConst || CaseConst == FalseConst) && \"Expect true or false as compare result.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5632, __PRETTY_FUNCTION__))
;
5633 }
5634
5635 // Check if the branch instruction dominates the phi node. It's a simple
5636 // dominance check, but sufficient for our needs.
5637 // Although this check is invariant in the calling loops, it's better to do it
5638 // at this late stage. Practically we do it at most once for a switch.
5639 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5640 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5641 BasicBlock *Pred = *PI;
5642 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5643 return;
5644 }
5645
5646 if (DefaultConst == FalseConst) {
5647 // The compare yields the same result. We can replace it.
5648 CmpInst->replaceAllUsesWith(RangeCmp);
5649 ++NumTableCmpReuses;
5650 } else {
5651 // The compare yields the same result, just inverted. We can replace it.
5652 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5653 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5654 RangeCheckBranch);
5655 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5656 ++NumTableCmpReuses;
5657 }
5658}
5659
5660/// If the switch is only used to initialize one or more phi nodes in a common
5661/// successor block with different constant values, replace the switch with
5662/// lookup tables.
5663static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5664 DomTreeUpdater *DTU, const DataLayout &DL,
5665 const TargetTransformInfo &TTI) {
5666 assert(SI->getNumCases() > 1 && "Degenerate switch?")((SI->getNumCases() > 1 && "Degenerate switch?"
) ? static_cast<void> (0) : __assert_fail ("SI->getNumCases() > 1 && \"Degenerate switch?\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5666, __PRETTY_FUNCTION__))
;
5667
5668 BasicBlock *BB = SI->getParent();
5669 Function *Fn = BB->getParent();
5670 // Only build lookup table when we have a target that supports it or the
5671 // attribute is not set.
5672 if (!TTI.shouldBuildLookupTables() ||
5673 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5674 return false;
5675
5676 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5677 // split off a dense part and build a lookup table for that.
5678
5679 // FIXME: This creates arrays of GEPs to constant strings, which means each
5680 // GEP needs a runtime relocation in PIC code. We should just build one big
5681 // string and lookup indices into that.
5682
5683 // Ignore switches with less than three cases. Lookup tables will not make
5684 // them faster, so we don't analyze them.
5685 if (SI->getNumCases() < 3)
5686 return false;
5687
5688 // Figure out the corresponding result for each case value and phi node in the
5689 // common destination, as well as the min and max case values.
5690 assert(!SI->cases().empty())((!SI->cases().empty()) ? static_cast<void> (0) : __assert_fail
("!SI->cases().empty()", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5690, __PRETTY_FUNCTION__))
;
5691 SwitchInst::CaseIt CI = SI->case_begin();
5692 ConstantInt *MinCaseVal = CI->getCaseValue();
5693 ConstantInt *MaxCaseVal = CI->getCaseValue();
5694
5695 BasicBlock *CommonDest = nullptr;
5696
5697 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5698 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5699
5700 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5701 SmallDenseMap<PHINode *, Type *> ResultTypes;
5702 SmallVector<PHINode *, 4> PHIs;
5703
5704 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5705 ConstantInt *CaseVal = CI->getCaseValue();
5706 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5707 MinCaseVal = CaseVal;
5708 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5709 MaxCaseVal = CaseVal;
5710
5711 // Resulting value at phi nodes for this case value.
5712 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5713 ResultsTy Results;
5714 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5715 Results, DL, TTI))
5716 return false;
5717
5718 // Append the result from this case to the list for each phi.
5719 for (const auto &I : Results) {
5720 PHINode *PHI = I.first;
5721 Constant *Value = I.second;
5722 if (!ResultLists.count(PHI))
5723 PHIs.push_back(PHI);
5724 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5725 }
5726 }
5727
5728 // Keep track of the result types.
5729 for (PHINode *PHI : PHIs) {
5730 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5731 }
5732
5733 uint64_t NumResults = ResultLists[PHIs[0]].size();
5734 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5735 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5736 bool TableHasHoles = (NumResults < TableSize);
5737
5738 // If the table has holes, we need a constant result for the default case
5739 // or a bitmask that fits in a register.
5740 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5741 bool HasDefaultResults =
5742 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5743 DefaultResultsList, DL, TTI);
5744
5745 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5746 if (NeedMask) {
5747 // As an extra penalty for the validity test we require more cases.
5748 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5749 return false;
5750 if (!DL.fitsInLegalInteger(TableSize))
5751 return false;
5752 }
5753
5754 for (const auto &I : DefaultResultsList) {
5755 PHINode *PHI = I.first;
5756 Constant *Result = I.second;
5757 DefaultResults[PHI] = Result;
5758 }
5759
5760 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5761 return false;
5762
5763 std::vector<DominatorTree::UpdateType> Updates;
5764
5765 // Create the BB that does the lookups.
5766 Module &Mod = *CommonDest->getParent()->getParent();
5767 BasicBlock *LookupBB = BasicBlock::Create(
5768 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5769
5770 // Compute the table index value.
5771 Builder.SetInsertPoint(SI);
5772 Value *TableIndex;
5773 if (MinCaseVal->isNullValue())
5774 TableIndex = SI->getCondition();
5775 else
5776 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5777 "switch.tableidx");
5778
5779 // Compute the maximum table size representable by the integer type we are
5780 // switching upon.
5781 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5782 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX(18446744073709551615UL) : 1ULL << CaseSize;
5783 assert(MaxTableSize >= TableSize &&((MaxTableSize >= TableSize && "It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? static_cast
<void> (0) : __assert_fail ("MaxTableSize >= TableSize && \"It is impossible for a switch to have more entries than the max \" \"representable value of its input integer type's size.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5785, __PRETTY_FUNCTION__))
5784 "It is impossible for a switch to have more entries than the max "((MaxTableSize >= TableSize && "It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? static_cast
<void> (0) : __assert_fail ("MaxTableSize >= TableSize && \"It is impossible for a switch to have more entries than the max \" \"representable value of its input integer type's size.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5785, __PRETTY_FUNCTION__))
5785 "representable value of its input integer type's size.")((MaxTableSize >= TableSize && "It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? static_cast
<void> (0) : __assert_fail ("MaxTableSize >= TableSize && \"It is impossible for a switch to have more entries than the max \" \"representable value of its input integer type's size.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5785, __PRETTY_FUNCTION__))
;
5786
5787 // If the default destination is unreachable, or if the lookup table covers
5788 // all values of the conditional variable, branch directly to the lookup table
5789 // BB. Otherwise, check that the condition is within the case range.
5790 const bool DefaultIsReachable =
5791 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5792 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5793 BranchInst *RangeCheckBranch = nullptr;
5794
5795 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5796 Builder.CreateBr(LookupBB);
5797 Updates.push_back({DominatorTree::Insert, BB, LookupBB});
5798 // Note: We call removeProdecessor later since we need to be able to get the
5799 // PHI value for the default case in case we're using a bit mask.
5800 } else {
5801 Value *Cmp = Builder.CreateICmpULT(
5802 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5803 RangeCheckBranch =
5804 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5805 Updates.push_back({DominatorTree::Insert, BB, LookupBB});
5806 }
5807
5808 // Populate the BB that does the lookups.
5809 Builder.SetInsertPoint(LookupBB);
5810
5811 if (NeedMask) {
5812 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5813 // re-purposed to do the hole check, and we create a new LookupBB.
5814 BasicBlock *MaskBB = LookupBB;
5815 MaskBB->setName("switch.hole_check");
5816 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5817 CommonDest->getParent(), CommonDest);
5818
5819 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5820 // unnecessary illegal types.
5821 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5822 APInt MaskInt(TableSizePowOf2, 0);
5823 APInt One(TableSizePowOf2, 1);
5824 // Build bitmask; fill in a 1 bit for every case.
5825 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5826 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5827 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5828 .getLimitedValue();
5829 MaskInt |= One << Idx;
5830 }
5831 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5832
5833 // Get the TableIndex'th bit of the bitmask.
5834 // If this bit is 0 (meaning hole) jump to the default destination,
5835 // else continue with table lookup.
5836 IntegerType *MapTy = TableMask->getType();
5837 Value *MaskIndex =
5838 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5839 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5840 Value *LoBit = Builder.CreateTrunc(
5841 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5842 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5843 Updates.push_back({DominatorTree::Insert, MaskBB, LookupBB});
5844 Updates.push_back({DominatorTree::Insert, MaskBB, SI->getDefaultDest()});
5845 Builder.SetInsertPoint(LookupBB);
5846 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, BB);
5847 }
5848
5849 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5850 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5851 // do not delete PHINodes here.
5852 SI->getDefaultDest()->removePredecessor(BB,
5853 /*KeepOneInputPHIs=*/true);
5854 Updates.push_back({DominatorTree::Delete, BB, SI->getDefaultDest()});
5855 }
5856
5857 bool ReturnedEarly = false;
5858 for (PHINode *PHI : PHIs) {
5859 const ResultListTy &ResultList = ResultLists[PHI];
5860