Bug Summary

File:lib/Transforms/Utils/SimplifyCFG.cpp
Warning:line 2395, column 27
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name 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 -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/include -I /build/llvm-toolchain-snapshot-9~svn362543/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/9.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-9/lib/clang/9.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-9~svn362543/build-llvm/lib/Transforms/Utils -fdebug-prefix-map=/build/llvm-toolchain-snapshot-9~svn362543=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2019-06-05-060531-1271-1 -x c++ /build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp -faddrsig
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/Optional.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SetOperations.h"
19#include "llvm/ADT/SetVector.h"
20#include "llvm/ADT/SmallPtrSet.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/ADT/StringRef.h"
24#include "llvm/Analysis/AssumptionCache.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Analysis/EHPersonalities.h"
27#include "llvm/Analysis/InstructionSimplify.h"
28#include "llvm/Analysis/MemorySSA.h"
29#include "llvm/Analysis/MemorySSAUpdater.h"
30#include "llvm/Analysis/TargetTransformInfo.h"
31#include "llvm/Analysis/ValueTracking.h"
32#include "llvm/IR/Attributes.h"
33#include "llvm/IR/BasicBlock.h"
34#include "llvm/IR/CFG.h"
35#include "llvm/IR/CallSite.h"
36#include "llvm/IR/Constant.h"
37#include "llvm/IR/ConstantRange.h"
38#include "llvm/IR/Constants.h"
39#include "llvm/IR/DataLayout.h"
40#include "llvm/IR/DerivedTypes.h"
41#include "llvm/IR/Function.h"
42#include "llvm/IR/GlobalValue.h"
43#include "llvm/IR/GlobalVariable.h"
44#include "llvm/IR/IRBuilder.h"
45#include "llvm/IR/InstrTypes.h"
46#include "llvm/IR/Instruction.h"
47#include "llvm/IR/Instructions.h"
48#include "llvm/IR/IntrinsicInst.h"
49#include "llvm/IR/Intrinsics.h"
50#include "llvm/IR/LLVMContext.h"
51#include "llvm/IR/MDBuilder.h"
52#include "llvm/IR/Metadata.h"
53#include "llvm/IR/Module.h"
54#include "llvm/IR/NoFolder.h"
55#include "llvm/IR/Operator.h"
56#include "llvm/IR/PatternMatch.h"
57#include "llvm/IR/Type.h"
58#include "llvm/IR/Use.h"
59#include "llvm/IR/User.h"
60#include "llvm/IR/Value.h"
61#include "llvm/Support/Casting.h"
62#include "llvm/Support/CommandLine.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/ErrorHandling.h"
65#include "llvm/Support/KnownBits.h"
66#include "llvm/Support/MathExtras.h"
67#include "llvm/Support/raw_ostream.h"
68#include "llvm/Transforms/Utils/BasicBlockUtils.h"
69#include "llvm/Transforms/Utils/Local.h"
70#include "llvm/Transforms/Utils/ValueMapper.h"
71#include <algorithm>
72#include <cassert>
73#include <climits>
74#include <cstddef>
75#include <cstdint>
76#include <iterator>
77#include <map>
78#include <set>
79#include <tuple>
80#include <utility>
81#include <vector>
82
83using namespace llvm;
84using namespace PatternMatch;
85
86#define DEBUG_TYPE"simplifycfg" "simplifycfg"
87
88// Chosen as 2 so as to be cheap, but still to have enough power to fold
89// a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90// To catch this, we need to fold a compare and a select, hence '2' being the
91// minimum reasonable default.
92static cl::opt<unsigned> PHINodeFoldingThreshold(
93 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94 cl::desc(
95 "Control the amount of phi node folding to perform (default = 2)"));
96
97static cl::opt<bool> DupRet(
98 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99 cl::desc("Duplicate return instructions into unconditional branches"));
100
101static cl::opt<bool>
102 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103 cl::desc("Sink common instructions down to the end block"));
104
105static cl::opt<bool> HoistCondStores(
106 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107 cl::desc("Hoist conditional stores if an unconditional store precedes"));
108
109static cl::opt<bool> MergeCondStores(
110 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111 cl::desc("Hoist conditional stores even if an unconditional store does not "
112 "precede - hoist multiple conditional stores into a single "
113 "predicated store"));
114
115static cl::opt<bool> MergeCondStoresAggressively(
116 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117 cl::desc("When merging conditional stores, do so even if the resultant "
118 "basic blocks are unlikely to be if-converted as a result"));
119
120static cl::opt<bool> SpeculateOneExpensiveInst(
121 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122 cl::desc("Allow exactly one expensive instruction to be speculatively "
123 "executed"));
124
125static cl::opt<unsigned> MaxSpeculationDepth(
126 "max-speculation-depth", cl::Hidden, cl::init(10),
127 cl::desc("Limit maximum recursion depth when calculating costs of "
128 "speculatively executed instructions"));
129
130STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps")static llvm::Statistic NumBitMaps = {"simplifycfg", "NumBitMaps"
, "Number of switch instructions turned into bitmaps", {0}, {
false}}
;
131STATISTIC(NumLinearMaps,static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
132 "Number of switch instructions turned into linear mapping")static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
;
133STATISTIC(NumLookupTables,static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
134 "Number of switch instructions turned into lookup tables")static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
;
135STATISTIC(static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
136 NumLookupTablesHoles,static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
137 "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)"
, {0}, {false}}
;
138STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares")static llvm::Statistic NumTableCmpReuses = {"simplifycfg", "NumTableCmpReuses"
, "Number of reused switch table lookup compares", {0}, {false
}}
;
139STATISTIC(NumSinkCommons,static llvm::Statistic NumSinkCommons = {"simplifycfg", "NumSinkCommons"
, "Number of common instructions sunk down to the end block",
{0}, {false}}
140 "Number of common instructions sunk down to the end block")static llvm::Statistic NumSinkCommons = {"simplifycfg", "NumSinkCommons"
, "Number of common instructions sunk down to the end block",
{0}, {false}}
;
141STATISTIC(NumSpeculations, "Number of speculative executed instructions")static llvm::Statistic NumSpeculations = {"simplifycfg", "NumSpeculations"
, "Number of speculative executed instructions", {0}, {false}
}
;
142
143namespace {
144
145// The first field contains the value that the switch produces when a certain
146// case group is selected, and the second field is a vector containing the
147// cases composing the case group.
148using SwitchCaseResultVectorTy =
149 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
150
151// The first field contains the phi node that generates a result of the switch
152// and the second field contains the value generated for a certain case in the
153// switch for that PHI.
154using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155
156/// ValueEqualityComparisonCase - Represents a case of a switch.
157struct ValueEqualityComparisonCase {
158 ConstantInt *Value;
159 BasicBlock *Dest;
160
161 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162 : Value(Value), Dest(Dest) {}
163
164 bool operator<(ValueEqualityComparisonCase RHS) const {
165 // Comparing pointers is ok as we only rely on the order for uniquing.
166 return Value < RHS.Value;
167 }
168
169 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170};
171
172class SimplifyCFGOpt {
173 const TargetTransformInfo &TTI;
174 const DataLayout &DL;
175 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176 const SimplifyCFGOptions &Options;
177 bool Resimplify;
178
179 Value *isValueEqualityComparison(Instruction *TI);
180 BasicBlock *GetValueEqualityComparisonCases(
181 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
182 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
183 BasicBlock *Pred,
184 IRBuilder<> &Builder);
185 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
186 IRBuilder<> &Builder);
187
188 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
189 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
190 bool SimplifySingleResume(ResumeInst *RI);
191 bool SimplifyCommonResume(ResumeInst *RI);
192 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
193 bool SimplifyUnreachable(UnreachableInst *UI);
194 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
195 bool SimplifyIndirectBr(IndirectBrInst *IBI);
196 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
198
199 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
200 IRBuilder<> &Builder);
201
202public:
203 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
204 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
205 const SimplifyCFGOptions &Opts)
206 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
207
208 bool run(BasicBlock *BB);
209 bool simplifyOnce(BasicBlock *BB);
210
211 // Helper to set Resimplify and return change indication.
212 bool requestResimplify() {
213 Resimplify = true;
214 return true;
215 }
216};
217
218} // end anonymous namespace
219
220/// Return true if it is safe to merge these two
221/// terminator instructions together.
222static bool
223SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
224 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
225 if (SI1 == SI2)
226 return false; // Can't merge with self!
227
228 // It is not safe to merge these two switch instructions if they have a common
229 // successor, and if that successor has a PHI node, and if *that* PHI node has
230 // conflicting incoming values from the two switch blocks.
231 BasicBlock *SI1BB = SI1->getParent();
232 BasicBlock *SI2BB = SI2->getParent();
233
234 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
235 bool Fail = false;
236 for (BasicBlock *Succ : successors(SI2BB))
237 if (SI1Succs.count(Succ))
238 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
239 PHINode *PN = cast<PHINode>(BBI);
240 if (PN->getIncomingValueForBlock(SI1BB) !=
241 PN->getIncomingValueForBlock(SI2BB)) {
242 if (FailBlocks)
243 FailBlocks->insert(Succ);
244 Fail = true;
245 }
246 }
247
248 return !Fail;
249}
250
251/// Return true if it is safe and profitable to merge these two terminator
252/// instructions together, where SI1 is an unconditional branch. PhiNodes will
253/// store all PHI nodes in common successors.
254static bool
255isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
256 Instruction *Cond,
257 SmallVectorImpl<PHINode *> &PhiNodes) {
258 if (SI1 == SI2)
259 return false; // Can't merge with self!
260 assert(SI1->isUnconditional() && SI2->isConditional())((SI1->isUnconditional() && SI2->isConditional(
)) ? static_cast<void> (0) : __assert_fail ("SI1->isUnconditional() && SI2->isConditional()"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 260, __PRETTY_FUNCTION__))
;
261
262 // We fold the unconditional branch if we can easily update all PHI nodes in
263 // common successors:
264 // 1> We have a constant incoming value for the conditional branch;
265 // 2> We have "Cond" as the incoming value for the unconditional branch;
266 // 3> SI2->getCondition() and Cond have same operands.
267 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
268 if (!Ci2)
269 return false;
270 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
271 Cond->getOperand(1) == Ci2->getOperand(1)) &&
272 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
273 Cond->getOperand(1) == Ci2->getOperand(0)))
274 return false;
275
276 BasicBlock *SI1BB = SI1->getParent();
277 BasicBlock *SI2BB = SI2->getParent();
278 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
279 for (BasicBlock *Succ : successors(SI2BB))
280 if (SI1Succs.count(Succ))
281 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
282 PHINode *PN = cast<PHINode>(BBI);
283 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
284 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
285 return false;
286 PhiNodes.push_back(PN);
287 }
288 return true;
289}
290
291/// Update PHI nodes in Succ to indicate that there will now be entries in it
292/// from the 'NewPred' block. The values that will be flowing into the PHI nodes
293/// will be the same as those coming in from ExistPred, an existing predecessor
294/// of Succ.
295static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
296 BasicBlock *ExistPred,
297 MemorySSAUpdater *MSSAU = nullptr) {
298 for (PHINode &PN : Succ->phis())
299 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
300 if (MSSAU)
301 if (auto *MPhi = MSSAU->getMemorySSA()->getMemoryAccess(Succ))
302 MPhi->addIncoming(MPhi->getIncomingValueForBlock(ExistPred), NewPred);
303}
304
305/// Compute an abstract "cost" of speculating the given instruction,
306/// which is assumed to be safe to speculate. TCC_Free means cheap,
307/// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
308/// expensive.
309static unsigned ComputeSpeculationCost(const User *I,
310 const TargetTransformInfo &TTI) {
311 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 312, __PRETTY_FUNCTION__))
312 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 312, __PRETTY_FUNCTION__))
;
313 return TTI.getUserCost(I);
314}
315
316/// If we have a merge point of an "if condition" as accepted above,
317/// return true if the specified value dominates the block. We
318/// don't handle the true generality of domination here, just a special case
319/// which works well enough for us.
320///
321/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
322/// see if V (which must be an instruction) and its recursive operands
323/// that do not dominate BB have a combined cost lower than CostRemaining and
324/// are non-trapping. If both are true, the instruction is inserted into the
325/// set and true is returned.
326///
327/// The cost for most non-trapping instructions is defined as 1 except for
328/// Select whose cost is 2.
329///
330/// After this function returns, CostRemaining is decreased by the cost of
331/// V plus its non-dominating operands. If that cost is greater than
332/// CostRemaining, false is returned and CostRemaining is undefined.
333static bool DominatesMergePoint(Value *V, BasicBlock *BB,
334 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
335 unsigned &CostRemaining,
336 const TargetTransformInfo &TTI,
337 unsigned Depth = 0) {
338 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
339 // so limit the recursion depth.
340 // TODO: While this recursion limit does prevent pathological behavior, it
341 // would be better to track visited instructions to avoid cycles.
342 if (Depth == MaxSpeculationDepth)
343 return false;
344
345 Instruction *I = dyn_cast<Instruction>(V);
346 if (!I) {
347 // Non-instructions all dominate instructions, but not all constantexprs
348 // can be executed unconditionally.
349 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
350 if (C->canTrap())
351 return false;
352 return true;
353 }
354 BasicBlock *PBB = I->getParent();
355
356 // We don't want to allow weird loops that might have the "if condition" in
357 // the bottom of this block.
358 if (PBB == BB)
359 return false;
360
361 // If this instruction is defined in a block that contains an unconditional
362 // branch to BB, then it must be in the 'conditional' part of the "if
363 // statement". If not, it definitely dominates the region.
364 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
365 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
366 return true;
367
368 // If we have seen this instruction before, don't count it again.
369 if (AggressiveInsts.count(I))
370 return true;
371
372 // Okay, it looks like the instruction IS in the "condition". Check to
373 // see if it's a cheap instruction to unconditionally compute, and if it
374 // only uses stuff defined outside of the condition. If so, hoist it out.
375 if (!isSafeToSpeculativelyExecute(I))
376 return false;
377
378 unsigned Cost = ComputeSpeculationCost(I, TTI);
379
380 // Allow exactly one instruction to be speculated regardless of its cost
381 // (as long as it is safe to do so).
382 // This is intended to flatten the CFG even if the instruction is a division
383 // or other expensive operation. The speculation of an expensive instruction
384 // is expected to be undone in CodeGenPrepare if the speculation has not
385 // enabled further IR optimizations.
386 if (Cost > CostRemaining &&
387 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
388 return false;
389
390 // Avoid unsigned wrap.
391 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
392
393 // Okay, we can only really hoist these out if their operands do
394 // not take us over the cost threshold.
395 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
396 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
397 Depth + 1))
398 return false;
399 // Okay, it's safe to do this! Remember this instruction.
400 AggressiveInsts.insert(I);
401 return true;
402}
403
404/// Extract ConstantInt from value, looking through IntToPtr
405/// and PointerNullValue. Return NULL if value is not a constant int.
406static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
407 // Normal constant int.
408 ConstantInt *CI = dyn_cast<ConstantInt>(V);
409 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
410 return CI;
411
412 // This is some kind of pointer constant. Turn it into a pointer-sized
413 // ConstantInt if possible.
414 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
415
416 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
417 if (isa<ConstantPointerNull>(V))
418 return ConstantInt::get(PtrTy, 0);
419
420 // IntToPtr const int.
421 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
422 if (CE->getOpcode() == Instruction::IntToPtr)
423 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
424 // The constant is very likely to have the right type already.
425 if (CI->getType() == PtrTy)
426 return CI;
427 else
428 return cast<ConstantInt>(
429 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
430 }
431 return nullptr;
432}
433
434namespace {
435
436/// Given a chain of or (||) or and (&&) comparison of a value against a
437/// constant, this will try to recover the information required for a switch
438/// structure.
439/// It will depth-first traverse the chain of comparison, seeking for patterns
440/// like %a == 12 or %a < 4 and combine them to produce a set of integer
441/// representing the different cases for the switch.
442/// Note that if the chain is composed of '||' it will build the set of elements
443/// that matches the comparisons (i.e. any of this value validate the chain)
444/// while for a chain of '&&' it will build the set elements that make the test
445/// fail.
446struct ConstantComparesGatherer {
447 const DataLayout &DL;
448
449 /// Value found for the switch comparison
450 Value *CompValue = nullptr;
451
452 /// Extra clause to be checked before the switch
453 Value *Extra = nullptr;
454
455 /// Set of integers to match in switch
456 SmallVector<ConstantInt *, 8> Vals;
457
458 /// Number of comparisons matched in the and/or chain
459 unsigned UsedICmps = 0;
460
461 /// Construct and compute the result for the comparison instruction Cond
462 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
463 gather(Cond);
464 }
465
466 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
467 ConstantComparesGatherer &
468 operator=(const ConstantComparesGatherer &) = delete;
469
470private:
471 /// Try to set the current value used for the comparison, it succeeds only if
472 /// it wasn't set before or if the new value is the same as the old one
473 bool setValueOnce(Value *NewVal) {
474 if (CompValue && CompValue != NewVal)
475 return false;
476 CompValue = NewVal;
477 return (CompValue != nullptr);
478 }
479
480 /// Try to match Instruction "I" as a comparison against a constant and
481 /// populates the array Vals with the set of values that match (or do not
482 /// match depending on isEQ).
483 /// Return false on failure. On success, the Value the comparison matched
484 /// against is placed in CompValue.
485 /// If CompValue is already set, the function is expected to fail if a match
486 /// is found but the value compared to is different.
487 bool matchInstruction(Instruction *I, bool isEQ) {
488 // If this is an icmp against a constant, handle this as one of the cases.
489 ICmpInst *ICI;
490 ConstantInt *C;
491 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
492 (C = GetConstantInt(I->getOperand(1), DL)))) {
493 return false;
494 }
495
496 Value *RHSVal;
497 const APInt *RHSC;
498
499 // Pattern match a special case
500 // (x & ~2^z) == y --> x == y || x == y|2^z
501 // This undoes a transformation done by instcombine to fuse 2 compares.
502 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
503 // It's a little bit hard to see why the following transformations are
504 // correct. Here is a CVC3 program to verify them for 64-bit values:
505
506 /*
507 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
508 x : BITVECTOR(64);
509 y : BITVECTOR(64);
510 z : BITVECTOR(64);
511 mask : BITVECTOR(64) = BVSHL(ONE, z);
512 QUERY( (y & ~mask = y) =>
513 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
514 );
515 QUERY( (y | mask = y) =>
516 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
517 );
518 */
519
520 // Please note that each pattern must be a dual implication (<--> or
521 // iff). One directional implication can create spurious matches. If the
522 // implication is only one-way, an unsatisfiable condition on the left
523 // side can imply a satisfiable condition on the right side. Dual
524 // implication ensures that satisfiable conditions are transformed to
525 // other satisfiable conditions and unsatisfiable conditions are
526 // transformed to other unsatisfiable conditions.
527
528 // Here is a concrete example of a unsatisfiable condition on the left
529 // implying a satisfiable condition on the right:
530 //
531 // mask = (1 << z)
532 // (x & ~mask) == y --> (x == y || x == (y | mask))
533 //
534 // Substituting y = 3, z = 0 yields:
535 // (x & -2) == 3 --> (x == 3 || x == 2)
536
537 // Pattern match a special case:
538 /*
539 QUERY( (y & ~mask = y) =>
540 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
541 );
542 */
543 if (match(ICI->getOperand(0),
544 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
545 APInt Mask = ~*RHSC;
546 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
547 // If we already have a value for the switch, it has to match!
548 if (!setValueOnce(RHSVal))
549 return false;
550
551 Vals.push_back(C);
552 Vals.push_back(
553 ConstantInt::get(C->getContext(),
554 C->getValue() | Mask));
555 UsedICmps++;
556 return true;
557 }
558 }
559
560 // Pattern match a special case:
561 /*
562 QUERY( (y | mask = y) =>
563 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
564 );
565 */
566 if (match(ICI->getOperand(0),
567 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
568 APInt Mask = *RHSC;
569 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
570 // If we already have a value for the switch, it has to match!
571 if (!setValueOnce(RHSVal))
572 return false;
573
574 Vals.push_back(C);
575 Vals.push_back(ConstantInt::get(C->getContext(),
576 C->getValue() & ~Mask));
577 UsedICmps++;
578 return true;
579 }
580 }
581
582 // If we already have a value for the switch, it has to match!
583 if (!setValueOnce(ICI->getOperand(0)))
584 return false;
585
586 UsedICmps++;
587 Vals.push_back(C);
588 return ICI->getOperand(0);
589 }
590
591 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
592 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
593 ICI->getPredicate(), C->getValue());
594
595 // Shift the range if the compare is fed by an add. This is the range
596 // compare idiom as emitted by instcombine.
597 Value *CandidateVal = I->getOperand(0);
598 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
599 Span = Span.subtract(*RHSC);
600 CandidateVal = RHSVal;
601 }
602
603 // If this is an and/!= check, then we are looking to build the set of
604 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
605 // x != 0 && x != 1.
606 if (!isEQ)
607 Span = Span.inverse();
608
609 // If there are a ton of values, we don't want to make a ginormous switch.
610 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
611 return false;
612 }
613
614 // If we already have a value for the switch, it has to match!
615 if (!setValueOnce(CandidateVal))
616 return false;
617
618 // Add all values from the range to the set
619 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
620 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
621
622 UsedICmps++;
623 return true;
624 }
625
626 /// Given a potentially 'or'd or 'and'd together collection of icmp
627 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
628 /// the value being compared, and stick the list constants into the Vals
629 /// vector.
630 /// One "Extra" case is allowed to differ from the other.
631 void gather(Value *V) {
632 Instruction *I = dyn_cast<Instruction>(V);
633 bool isEQ = (I->getOpcode() == Instruction::Or);
634
635 // Keep a stack (SmallVector for efficiency) for depth-first traversal
636 SmallVector<Value *, 8> DFT;
637 SmallPtrSet<Value *, 8> Visited;
638
639 // Initialize
640 Visited.insert(V);
641 DFT.push_back(V);
642
643 while (!DFT.empty()) {
644 V = DFT.pop_back_val();
645
646 if (Instruction *I = dyn_cast<Instruction>(V)) {
647 // If it is a || (or && depending on isEQ), process the operands.
648 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
649 if (Visited.insert(I->getOperand(1)).second)
650 DFT.push_back(I->getOperand(1));
651 if (Visited.insert(I->getOperand(0)).second)
652 DFT.push_back(I->getOperand(0));
653 continue;
654 }
655
656 // Try to match the current instruction
657 if (matchInstruction(I, isEQ))
658 // Match succeed, continue the loop
659 continue;
660 }
661
662 // One element of the sequence of || (or &&) could not be match as a
663 // comparison against the same value as the others.
664 // We allow only one "Extra" case to be checked before the switch
665 if (!Extra) {
666 Extra = V;
667 continue;
668 }
669 // Failed to parse a proper sequence, abort now
670 CompValue = nullptr;
671 break;
672 }
673 }
674};
675
676} // end anonymous namespace
677
678static void EraseTerminatorAndDCECond(Instruction *TI,
679 MemorySSAUpdater *MSSAU = nullptr) {
680 Instruction *Cond = nullptr;
681 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
682 Cond = dyn_cast<Instruction>(SI->getCondition());
683 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
684 if (BI->isConditional())
685 Cond = dyn_cast<Instruction>(BI->getCondition());
686 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
687 Cond = dyn_cast<Instruction>(IBI->getAddress());
688 }
689
690 TI->eraseFromParent();
691 if (Cond)
692 RecursivelyDeleteTriviallyDeadInstructions(Cond, nullptr, MSSAU);
693}
694
695/// Return true if the specified terminator checks
696/// to see if a value is equal to constant integer value.
697Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
698 Value *CV = nullptr;
699 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
700 // Do not permit merging of large switch instructions into their
701 // predecessors unless there is only one predecessor.
702 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
703 CV = SI->getCondition();
704 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
705 if (BI->isConditional() && BI->getCondition()->hasOneUse())
706 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
707 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
708 CV = ICI->getOperand(0);
709 }
710
711 // Unwrap any lossless ptrtoint cast.
712 if (CV) {
713 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
714 Value *Ptr = PTII->getPointerOperand();
715 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
716 CV = Ptr;
717 }
718 }
719 return CV;
720}
721
722/// Given a value comparison instruction,
723/// decode all of the 'cases' that it represents and return the 'default' block.
724BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
725 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
726 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
727 Cases.reserve(SI->getNumCases());
728 for (auto Case : SI->cases())
729 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
730 Case.getCaseSuccessor()));
731 return SI->getDefaultDest();
732 }
733
734 BranchInst *BI = cast<BranchInst>(TI);
735 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
736 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
737 Cases.push_back(ValueEqualityComparisonCase(
738 GetConstantInt(ICI->getOperand(1), DL), Succ));
739 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
740}
741
742/// Given a vector of bb/value pairs, remove any entries
743/// in the list that match the specified block.
744static void
745EliminateBlockCases(BasicBlock *BB,
746 std::vector<ValueEqualityComparisonCase> &Cases) {
747 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
748}
749
750/// Return true if there are any keys in C1 that exist in C2 as well.
751static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
752 std::vector<ValueEqualityComparisonCase> &C2) {
753 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
754
755 // Make V1 be smaller than V2.
756 if (V1->size() > V2->size())
757 std::swap(V1, V2);
758
759 if (V1->empty())
760 return false;
761 if (V1->size() == 1) {
762 // Just scan V2.
763 ConstantInt *TheVal = (*V1)[0].Value;
764 for (unsigned i = 0, e = V2->size(); i != e; ++i)
765 if (TheVal == (*V2)[i].Value)
766 return true;
767 }
768
769 // Otherwise, just sort both lists and compare element by element.
770 array_pod_sort(V1->begin(), V1->end());
771 array_pod_sort(V2->begin(), V2->end());
772 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
773 while (i1 != e1 && i2 != e2) {
774 if ((*V1)[i1].Value == (*V2)[i2].Value)
775 return true;
776 if ((*V1)[i1].Value < (*V2)[i2].Value)
777 ++i1;
778 else
779 ++i2;
780 }
781 return false;
782}
783
784// Set branch weights on SwitchInst. This sets the metadata if there is at
785// least one non-zero weight.
786static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
787 // Check that there is at least one non-zero weight. Otherwise, pass
788 // nullptr to setMetadata which will erase the existing metadata.
789 MDNode *N = nullptr;
790 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
791 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
792 SI->setMetadata(LLVMContext::MD_prof, N);
793}
794
795// Similar to the above, but for branch and select instructions that take
796// exactly 2 weights.
797static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
798 uint32_t FalseWeight) {
799 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 799, __PRETTY_FUNCTION__))
;
800 // Check that there is at least one non-zero weight. Otherwise, pass
801 // nullptr to setMetadata which will erase the existing metadata.
802 MDNode *N = nullptr;
803 if (TrueWeight || FalseWeight)
804 N = MDBuilder(I->getParent()->getContext())
805 .createBranchWeights(TrueWeight, FalseWeight);
806 I->setMetadata(LLVMContext::MD_prof, N);
807}
808
809/// If TI is known to be a terminator instruction and its block is known to
810/// only have a single predecessor block, check to see if that predecessor is
811/// also a value comparison with the same value, and if that comparison
812/// determines the outcome of this comparison. If so, simplify TI. This does a
813/// very limited form of jump threading.
814bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
815 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
816 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
817 if (!PredVal)
818 return false; // Not a value comparison in predecessor.
819
820 Value *ThisVal = isValueEqualityComparison(TI);
821 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 821, __PRETTY_FUNCTION__))
;
822 if (ThisVal != PredVal)
823 return false; // Different predicates.
824
825 // TODO: Preserve branch weight metadata, similarly to how
826 // FoldValueComparisonIntoPredecessors preserves it.
827
828 // Find out information about when control will move from Pred to TI's block.
829 std::vector<ValueEqualityComparisonCase> PredCases;
830 BasicBlock *PredDef =
831 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
832 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
833
834 // Find information about how control leaves this block.
835 std::vector<ValueEqualityComparisonCase> ThisCases;
836 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
837 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
838
839 // If TI's block is the default block from Pred's comparison, potentially
840 // simplify TI based on this knowledge.
841 if (PredDef == TI->getParent()) {
842 // If we are here, we know that the value is none of those cases listed in
843 // PredCases. If there are any cases in ThisCases that are in PredCases, we
844 // can simplify TI.
845 if (!ValuesOverlap(PredCases, ThisCases))
846 return false;
847
848 if (isa<BranchInst>(TI)) {
849 // Okay, one of the successors of this condbr is dead. Convert it to a
850 // uncond br.
851 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 851, __PRETTY_FUNCTION__))
;
852 // Insert the new branch.
853 Instruction *NI = Builder.CreateBr(ThisDef);
854 (void)NI;
855
856 // Remove PHI node entries for the dead edge.
857 ThisCases[0].Dest->removePredecessor(TI->getParent());
858
859 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)
860 << "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)
861 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
862
863 EraseTerminatorAndDCECond(TI);
864 return true;
865 }
866
867 SwitchInst *SI = cast<SwitchInst>(TI);
868 // Okay, TI has cases that are statically dead, prune them away.
869 SmallPtrSet<Constant *, 16> DeadCases;
870 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
871 DeadCases.insert(PredCases[i].Value);
872
873 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)
874 << "Through successor TI: " << *TI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI; } } while (false)
;
875
876 // Collect branch weights into a vector.
877 SmallVector<uint32_t, 8> Weights;
878 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
879 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
880 if (HasWeight)
881 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
882 ++MD_i) {
883 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
884 Weights.push_back(CI->getValue().getZExtValue());
885 }
886 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
887 --i;
888 if (DeadCases.count(i->getCaseValue())) {
889 if (HasWeight) {
890 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
891 Weights.pop_back();
892 }
893 i->getCaseSuccessor()->removePredecessor(TI->getParent());
894 SI->removeCase(i);
895 }
896 }
897 if (HasWeight && Weights.size() >= 2)
898 setBranchWeights(SI, Weights);
899
900 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Leaving: " << *TI <<
"\n"; } } while (false)
;
901 return true;
902 }
903
904 // Otherwise, TI's block must correspond to some matched value. Find out
905 // which value (or set of values) this is.
906 ConstantInt *TIV = nullptr;
907 BasicBlock *TIBB = TI->getParent();
908 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
909 if (PredCases[i].Dest == TIBB) {
910 if (TIV)
911 return false; // Cannot handle multiple values coming to this block.
912 TIV = PredCases[i].Value;
913 }
914 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 914, __PRETTY_FUNCTION__))
;
915
916 // Okay, we found the one constant that our value can be if we get into TI's
917 // BB. Find out which successor will unconditionally be branched to.
918 BasicBlock *TheRealDest = nullptr;
919 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
920 if (ThisCases[i].Value == TIV) {
921 TheRealDest = ThisCases[i].Dest;
922 break;
923 }
924
925 // If not handled by any explicit cases, it is handled by the default case.
926 if (!TheRealDest)
927 TheRealDest = ThisDef;
928
929 // Remove PHI node entries for dead edges.
930 BasicBlock *CheckEdge = TheRealDest;
931 for (BasicBlock *Succ : successors(TIBB))
932 if (Succ != CheckEdge)
933 Succ->removePredecessor(TIBB);
934 else
935 CheckEdge = nullptr;
936
937 // Insert the new branch.
938 Instruction *NI = Builder.CreateBr(TheRealDest);
939 (void)NI;
940
941 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)
942 << "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)
943 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
944
945 EraseTerminatorAndDCECond(TI);
946 return true;
947}
948
949namespace {
950
951/// This class implements a stable ordering of constant
952/// integers that does not depend on their address. This is important for
953/// applications that sort ConstantInt's to ensure uniqueness.
954struct ConstantIntOrdering {
955 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
956 return LHS->getValue().ult(RHS->getValue());
957 }
958};
959
960} // end anonymous namespace
961
962static int ConstantIntSortPredicate(ConstantInt *const *P1,
963 ConstantInt *const *P2) {
964 const ConstantInt *LHS = *P1;
965 const ConstantInt *RHS = *P2;
966 if (LHS == RHS)
967 return 0;
968 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
969}
970
971static inline bool HasBranchWeights(const Instruction *I) {
972 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
973 if (ProfMD && ProfMD->getOperand(0))
974 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
975 return MDS->getString().equals("branch_weights");
976
977 return false;
978}
979
980/// Get Weights of a given terminator, the default weight is at the front
981/// of the vector. If TI is a conditional eq, we need to swap the branch-weight
982/// metadata.
983static void GetBranchWeights(Instruction *TI,
984 SmallVectorImpl<uint64_t> &Weights) {
985 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
986 assert(MD)((MD) ? static_cast<void> (0) : __assert_fail ("MD", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 986, __PRETTY_FUNCTION__))
;
987 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
988 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
989 Weights.push_back(CI->getValue().getZExtValue());
990 }
991
992 // If TI is a conditional eq, the default case is the false case,
993 // and the corresponding branch-weight data is at index 2. We swap the
994 // default weight to be the first entry.
995 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
996 assert(Weights.size() == 2)((Weights.size() == 2) ? static_cast<void> (0) : __assert_fail
("Weights.size() == 2", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 996, __PRETTY_FUNCTION__))
;
997 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
998 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
999 std::swap(Weights.front(), Weights.back());
1000 }
1001}
1002
1003/// Keep halving the weights until all can fit in uint32_t.
1004static void FitWeights(MutableArrayRef<uint64_t> Weights) {
1005 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1006 if (Max > UINT_MAX(2147483647 *2U +1U)) {
1007 unsigned Offset = 32 - countLeadingZeros(Max);
1008 for (uint64_t &I : Weights)
1009 I >>= Offset;
1010 }
1011}
1012
1013/// The specified terminator is a value equality comparison instruction
1014/// (either a switch or a branch on "X == c").
1015/// See if any of the predecessors of the terminator block are value comparisons
1016/// on the same value. If so, and if safe to do so, fold them together.
1017bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1018 IRBuilder<> &Builder) {
1019 BasicBlock *BB = TI->getParent();
1020 Value *CV = isValueEqualityComparison(TI); // CondVal
1021 assert(CV && "Not a comparison?")((CV && "Not a comparison?") ? static_cast<void>
(0) : __assert_fail ("CV && \"Not a comparison?\"", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1021, __PRETTY_FUNCTION__))
;
1022 bool Changed = false;
1023
1024 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1025 while (!Preds.empty()) {
1026 BasicBlock *Pred = Preds.pop_back_val();
1027
1028 // See if the predecessor is a comparison with the same value.
1029 Instruction *PTI = Pred->getTerminator();
1030 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1031
1032 if (PCV == CV && TI != PTI) {
1033 SmallSetVector<BasicBlock*, 4> FailBlocks;
1034 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1035 for (auto *Succ : FailBlocks) {
1036 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1037 return false;
1038 }
1039 }
1040
1041 // Figure out which 'cases' to copy from SI to PSI.
1042 std::vector<ValueEqualityComparisonCase> BBCases;
1043 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1044
1045 std::vector<ValueEqualityComparisonCase> PredCases;
1046 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1047
1048 // Based on whether the default edge from PTI goes to BB or not, fill in
1049 // PredCases and PredDefault with the new switch cases we would like to
1050 // build.
1051 SmallVector<BasicBlock *, 8> NewSuccessors;
1052
1053 // Update the branch weight metadata along the way
1054 SmallVector<uint64_t, 8> Weights;
1055 bool PredHasWeights = HasBranchWeights(PTI);
1056 bool SuccHasWeights = HasBranchWeights(TI);
1057
1058 if (PredHasWeights) {
1059 GetBranchWeights(PTI, Weights);
1060 // branch-weight metadata is inconsistent here.
1061 if (Weights.size() != 1 + PredCases.size())
1062 PredHasWeights = SuccHasWeights = false;
1063 } else if (SuccHasWeights)
1064 // If there are no predecessor weights but there are successor weights,
1065 // populate Weights with 1, which will later be scaled to the sum of
1066 // successor's weights
1067 Weights.assign(1 + PredCases.size(), 1);
1068
1069 SmallVector<uint64_t, 8> SuccWeights;
1070 if (SuccHasWeights) {
1071 GetBranchWeights(TI, SuccWeights);
1072 // branch-weight metadata is inconsistent here.
1073 if (SuccWeights.size() != 1 + BBCases.size())
1074 PredHasWeights = SuccHasWeights = false;
1075 } else if (PredHasWeights)
1076 SuccWeights.assign(1 + BBCases.size(), 1);
1077
1078 if (PredDefault == BB) {
1079 // If this is the default destination from PTI, only the edges in TI
1080 // that don't occur in PTI, or that branch to BB will be activated.
1081 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1082 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1083 if (PredCases[i].Dest != BB)
1084 PTIHandled.insert(PredCases[i].Value);
1085 else {
1086 // The default destination is BB, we don't need explicit targets.
1087 std::swap(PredCases[i], PredCases.back());
1088
1089 if (PredHasWeights || SuccHasWeights) {
1090 // Increase weight for the default case.
1091 Weights[0] += Weights[i + 1];
1092 std::swap(Weights[i + 1], Weights.back());
1093 Weights.pop_back();
1094 }
1095
1096 PredCases.pop_back();
1097 --i;
1098 --e;
1099 }
1100
1101 // Reconstruct the new switch statement we will be building.
1102 if (PredDefault != BBDefault) {
1103 PredDefault->removePredecessor(Pred);
1104 PredDefault = BBDefault;
1105 NewSuccessors.push_back(BBDefault);
1106 }
1107
1108 unsigned CasesFromPred = Weights.size();
1109 uint64_t ValidTotalSuccWeight = 0;
1110 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1111 if (!PTIHandled.count(BBCases[i].Value) &&
1112 BBCases[i].Dest != BBDefault) {
1113 PredCases.push_back(BBCases[i]);
1114 NewSuccessors.push_back(BBCases[i].Dest);
1115 if (SuccHasWeights || PredHasWeights) {
1116 // The default weight is at index 0, so weight for the ith case
1117 // should be at index i+1. Scale the cases from successor by
1118 // PredDefaultWeight (Weights[0]).
1119 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1120 ValidTotalSuccWeight += SuccWeights[i + 1];
1121 }
1122 }
1123
1124 if (SuccHasWeights || PredHasWeights) {
1125 ValidTotalSuccWeight += SuccWeights[0];
1126 // Scale the cases from predecessor by ValidTotalSuccWeight.
1127 for (unsigned i = 1; i < CasesFromPred; ++i)
1128 Weights[i] *= ValidTotalSuccWeight;
1129 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1130 Weights[0] *= SuccWeights[0];
1131 }
1132 } else {
1133 // If this is not the default destination from PSI, only the edges
1134 // in SI that occur in PSI with a destination of BB will be
1135 // activated.
1136 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1137 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1138 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1139 if (PredCases[i].Dest == BB) {
1140 PTIHandled.insert(PredCases[i].Value);
1141
1142 if (PredHasWeights || SuccHasWeights) {
1143 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1144 std::swap(Weights[i + 1], Weights.back());
1145 Weights.pop_back();
1146 }
1147
1148 std::swap(PredCases[i], PredCases.back());
1149 PredCases.pop_back();
1150 --i;
1151 --e;
1152 }
1153
1154 // Okay, now we know which constants were sent to BB from the
1155 // predecessor. Figure out where they will all go now.
1156 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1157 if (PTIHandled.count(BBCases[i].Value)) {
1158 // If this is one we are capable of getting...
1159 if (PredHasWeights || SuccHasWeights)
1160 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1161 PredCases.push_back(BBCases[i]);
1162 NewSuccessors.push_back(BBCases[i].Dest);
1163 PTIHandled.erase(
1164 BBCases[i].Value); // This constant is taken care of
1165 }
1166
1167 // If there are any constants vectored to BB that TI doesn't handle,
1168 // they must go to the default destination of TI.
1169 for (ConstantInt *I : PTIHandled) {
1170 if (PredHasWeights || SuccHasWeights)
1171 Weights.push_back(WeightsForHandled[I]);
1172 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1173 NewSuccessors.push_back(BBDefault);
1174 }
1175 }
1176
1177 // Okay, at this point, we know which new successor Pred will get. Make
1178 // sure we update the number of entries in the PHI nodes for these
1179 // successors.
1180 for (BasicBlock *NewSuccessor : NewSuccessors)
1181 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1182
1183 Builder.SetInsertPoint(PTI);
1184 // Convert pointer to int before we switch.
1185 if (CV->getType()->isPointerTy()) {
1186 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1187 "magicptr");
1188 }
1189
1190 // Now that the successors are updated, create the new Switch instruction.
1191 SwitchInst *NewSI =
1192 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1193 NewSI->setDebugLoc(PTI->getDebugLoc());
1194 for (ValueEqualityComparisonCase &V : PredCases)
1195 NewSI->addCase(V.Value, V.Dest);
1196
1197 if (PredHasWeights || SuccHasWeights) {
1198 // Halve the weights if any of them cannot fit in an uint32_t
1199 FitWeights(Weights);
1200
1201 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1202
1203 setBranchWeights(NewSI, MDWeights);
1204 }
1205
1206 EraseTerminatorAndDCECond(PTI);
1207
1208 // Okay, last check. If BB is still a successor of PSI, then we must
1209 // have an infinite loop case. If so, add an infinitely looping block
1210 // to handle the case to preserve the behavior of the code.
1211 BasicBlock *InfLoopBlock = nullptr;
1212 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1213 if (NewSI->getSuccessor(i) == BB) {
1214 if (!InfLoopBlock) {
1215 // Insert it at the end of the function, because it's either code,
1216 // or it won't matter if it's hot. :)
1217 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1218 BB->getParent());
1219 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1220 }
1221 NewSI->setSuccessor(i, InfLoopBlock);
1222 }
1223
1224 Changed = true;
1225 }
1226 }
1227 return Changed;
1228}
1229
1230// If we would need to insert a select that uses the value of this invoke
1231// (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1232// can't hoist the invoke, as there is nowhere to put the select in this case.
1233static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1234 Instruction *I1, Instruction *I2) {
1235 for (BasicBlock *Succ : successors(BB1)) {
1236 for (const PHINode &PN : Succ->phis()) {
1237 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1238 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1239 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1240 return false;
1241 }
1242 }
1243 }
1244 return true;
1245}
1246
1247static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1248
1249/// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1250/// in the two blocks up into the branch block. The caller of this function
1251/// guarantees that BI's block dominates BB1 and BB2.
1252static bool HoistThenElseCodeToIf(BranchInst *BI,
1253 const TargetTransformInfo &TTI) {
1254 // This does very trivial matching, with limited scanning, to find identical
1255 // instructions in the two blocks. In particular, we don't want to get into
1256 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1257 // such, we currently just scan for obviously identical instructions in an
1258 // identical order.
1259 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1260 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1261
1262 BasicBlock::iterator BB1_Itr = BB1->begin();
1263 BasicBlock::iterator BB2_Itr = BB2->begin();
1264
1265 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1266 // Skip debug info if it is not identical.
1267 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1268 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1269 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1270 while (isa<DbgInfoIntrinsic>(I1))
1271 I1 = &*BB1_Itr++;
1272 while (isa<DbgInfoIntrinsic>(I2))
1273 I2 = &*BB2_Itr++;
1274 }
1275 // FIXME: Can we define a safety predicate for CallBr?
1276 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1277 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1278 isa<CallBrInst>(I1))
1279 return false;
1280
1281 BasicBlock *BIParent = BI->getParent();
1282
1283 bool Changed = false;
1284 do {
1285 // If we are hoisting the terminator instruction, don't move one (making a
1286 // broken BB), instead clone it, and remove BI.
1287 if (I1->isTerminator())
1288 goto HoistTerminator;
1289
1290 // If we're going to hoist a call, make sure that the two instructions we're
1291 // commoning/hoisting are both marked with musttail, or neither of them is
1292 // marked as such. Otherwise, we might end up in a situation where we hoist
1293 // from a block where the terminator is a `ret` to a block where the terminator
1294 // is a `br`, and `musttail` calls expect to be followed by a return.
1295 auto *C1 = dyn_cast<CallInst>(I1);
1296 auto *C2 = dyn_cast<CallInst>(I2);
1297 if (C1 && C2)
1298 if (C1->isMustTailCall() != C2->isMustTailCall())
1299 return Changed;
1300
1301 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1302 return Changed;
1303
1304 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1305 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1305, __PRETTY_FUNCTION__))
;
1306 // The debug location is an integral part of a debug info intrinsic
1307 // and can't be separated from it or replaced. Instead of attempting
1308 // to merge locations, simply hoist both copies of the intrinsic.
1309 BIParent->getInstList().splice(BI->getIterator(),
1310 BB1->getInstList(), I1);
1311 BIParent->getInstList().splice(BI->getIterator(),
1312 BB2->getInstList(), I2);
1313 Changed = true;
1314 } else {
1315 // For a normal instruction, we just move one to right before the branch,
1316 // then replace all uses of the other with the first. Finally, we remove
1317 // the now redundant second instruction.
1318 BIParent->getInstList().splice(BI->getIterator(),
1319 BB1->getInstList(), I1);
1320 if (!I2->use_empty())
1321 I2->replaceAllUsesWith(I1);
1322 I1->andIRFlags(I2);
1323 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1324 LLVMContext::MD_range,
1325 LLVMContext::MD_fpmath,
1326 LLVMContext::MD_invariant_load,
1327 LLVMContext::MD_nonnull,
1328 LLVMContext::MD_invariant_group,
1329 LLVMContext::MD_align,
1330 LLVMContext::MD_dereferenceable,
1331 LLVMContext::MD_dereferenceable_or_null,
1332 LLVMContext::MD_mem_parallel_loop_access,
1333 LLVMContext::MD_access_group};
1334 combineMetadata(I1, I2, KnownIDs, true);
1335
1336 // I1 and I2 are being combined into a single instruction. Its debug
1337 // location is the merged locations of the original instructions.
1338 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1339
1340 I2->eraseFromParent();
1341 Changed = true;
1342 }
1343
1344 I1 = &*BB1_Itr++;
1345 I2 = &*BB2_Itr++;
1346 // Skip debug info if it is not identical.
1347 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1348 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1349 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1350 while (isa<DbgInfoIntrinsic>(I1))
1351 I1 = &*BB1_Itr++;
1352 while (isa<DbgInfoIntrinsic>(I2))
1353 I2 = &*BB2_Itr++;
1354 }
1355 } while (I1->isIdenticalToWhenDefined(I2));
1356
1357 return true;
1358
1359HoistTerminator:
1360 // It may not be possible to hoist an invoke.
1361 // FIXME: Can we define a safety predicate for CallBr?
1362 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1363 return Changed;
1364
1365 // TODO: callbr hoisting currently disabled pending further study.
1366 if (isa<CallBrInst>(I1))
1367 return Changed;
1368
1369 for (BasicBlock *Succ : successors(BB1)) {
1370 for (PHINode &PN : Succ->phis()) {
1371 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1372 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1373 if (BB1V == BB2V)
1374 continue;
1375
1376 // Check for passingValueIsAlwaysUndefined here because we would rather
1377 // eliminate undefined control flow then converting it to a select.
1378 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1379 passingValueIsAlwaysUndefined(BB2V, &PN))
1380 return Changed;
1381
1382 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1383 return Changed;
1384 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1385 return Changed;
1386 }
1387 }
1388
1389 // Okay, it is safe to hoist the terminator.
1390 Instruction *NT = I1->clone();
1391 BIParent->getInstList().insert(BI->getIterator(), NT);
1392 if (!NT->getType()->isVoidTy()) {
1393 I1->replaceAllUsesWith(NT);
1394 I2->replaceAllUsesWith(NT);
1395 NT->takeName(I1);
1396 }
1397
1398 // Ensure terminator gets a debug location, even an unknown one, in case
1399 // it involves inlinable calls.
1400 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1401
1402 // PHIs created below will adopt NT's merged DebugLoc.
1403 IRBuilder<NoFolder> Builder(NT);
1404
1405 // Hoisting one of the terminators from our successor is a great thing.
1406 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1407 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1408 // nodes, so we insert select instruction to compute the final result.
1409 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1410 for (BasicBlock *Succ : successors(BB1)) {
1411 for (PHINode &PN : Succ->phis()) {
1412 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1413 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1414 if (BB1V == BB2V)
1415 continue;
1416
1417 // These values do not agree. Insert a select instruction before NT
1418 // that determines the right value.
1419 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1420 if (!SI)
1421 SI = cast<SelectInst>(
1422 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1423 BB1V->getName() + "." + BB2V->getName(), BI));
1424
1425 // Make the PHI node use the select for all incoming values for BB1/BB2
1426 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1427 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1428 PN.setIncomingValue(i, SI);
1429 }
1430 }
1431
1432 // Update any PHI nodes in our new successors.
1433 for (BasicBlock *Succ : successors(BB1))
1434 AddPredecessorToBlock(Succ, BIParent, BB1);
1435
1436 EraseTerminatorAndDCECond(BI);
1437 return true;
1438}
1439
1440// All instructions in Insts belong to different blocks that all unconditionally
1441// branch to a common successor. Analyze each instruction and return true if it
1442// would be possible to sink them into their successor, creating one common
1443// instruction instead. For every value that would be required to be provided by
1444// PHI node (because an operand varies in each input block), add to PHIOperands.
1445static bool canSinkInstructions(
1446 ArrayRef<Instruction *> Insts,
1447 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1448 // Prune out obviously bad instructions to move. Each instruction must have
1449 // exactly zero or one use, and we check later that use is by a single, common
1450 // PHI instruction in the successor.
1451 bool HasUse = !Insts.front()->user_empty();
1452 for (auto *I : Insts) {
1453 // These instructions may change or break semantics if moved.
1454 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1455 I->getType()->isTokenTy())
1456 return false;
1457
1458 // Conservatively return false if I is an inline-asm instruction. Sinking
1459 // and merging inline-asm instructions can potentially create arguments
1460 // that cannot satisfy the inline-asm constraints.
1461 if (const auto *C = dyn_cast<CallBase>(I))
1462 if (C->isInlineAsm())
1463 return false;
1464
1465 // Each instruction must have zero or one use.
1466 if (HasUse && !I->hasOneUse())
1467 return false;
1468 if (!HasUse && !I->user_empty())
1469 return false;
1470 }
1471
1472 const Instruction *I0 = Insts.front();
1473 for (auto *I : Insts)
1474 if (!I->isSameOperationAs(I0))
1475 return false;
1476
1477 // All instructions in Insts are known to be the same opcode. If they have a
1478 // use, check that the only user is a PHI or in the same block as the
1479 // instruction, because if a user is in the same block as an instruction we're
1480 // contemplating sinking, it must already be determined to be sinkable.
1481 if (HasUse) {
1482 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1483 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1484 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1485 auto *U = cast<Instruction>(*I->user_begin());
1486 return (PNUse &&
1487 PNUse->getParent() == Succ &&
1488 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1489 U->getParent() == I->getParent();
1490 }))
1491 return false;
1492 }
1493
1494 // Because SROA can't handle speculating stores of selects, try not
1495 // to sink loads or stores of allocas when we'd have to create a PHI for
1496 // the address operand. Also, because it is likely that loads or stores
1497 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1498 // This can cause code churn which can have unintended consequences down
1499 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1500 // FIXME: This is a workaround for a deficiency in SROA - see
1501 // https://llvm.org/bugs/show_bug.cgi?id=30188
1502 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1503 return isa<AllocaInst>(I->getOperand(1));
1504 }))
1505 return false;
1506 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1507 return isa<AllocaInst>(I->getOperand(0));
1508 }))
1509 return false;
1510
1511 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1512 if (I0->getOperand(OI)->getType()->isTokenTy())
1513 // Don't touch any operand of token type.
1514 return false;
1515
1516 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1517 assert(I->getNumOperands() == I0->getNumOperands())((I->getNumOperands() == I0->getNumOperands()) ? static_cast
<void> (0) : __assert_fail ("I->getNumOperands() == I0->getNumOperands()"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1517, __PRETTY_FUNCTION__))
;
1518 return I->getOperand(OI) == I0->getOperand(OI);
1519 };
1520 if (!all_of(Insts, SameAsI0)) {
1521 if (!canReplaceOperandWithVariable(I0, OI))
1522 // We can't create a PHI from this GEP.
1523 return false;
1524 // Don't create indirect calls! The called value is the final operand.
1525 if (isa<CallBase>(I0) && OI == OE - 1) {
1526 // FIXME: if the call was *already* indirect, we should do this.
1527 return false;
1528 }
1529 for (auto *I : Insts)
1530 PHIOperands[I].push_back(I->getOperand(OI));
1531 }
1532 }
1533 return true;
1534}
1535
1536// Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1537// instruction of every block in Blocks to their common successor, commoning
1538// into one instruction.
1539static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1540 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1541
1542 // canSinkLastInstruction returning true guarantees that every block has at
1543 // least one non-terminator instruction.
1544 SmallVector<Instruction*,4> Insts;
1545 for (auto *BB : Blocks) {
1546 Instruction *I = BB->getTerminator();
1547 do {
1548 I = I->getPrevNode();
1549 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1550 if (!isa<DbgInfoIntrinsic>(I))
1551 Insts.push_back(I);
1552 }
1553
1554 // The only checking we need to do now is that all users of all instructions
1555 // are the same PHI node. canSinkLastInstruction should have checked this but
1556 // it is slightly over-aggressive - it gets confused by commutative instructions
1557 // so double-check it here.
1558 Instruction *I0 = Insts.front();
1559 if (!I0->user_empty()) {
1560 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1561 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1562 auto *U = cast<Instruction>(*I->user_begin());
1563 return U == PNUse;
1564 }))
1565 return false;
1566 }
1567
1568 // We don't need to do any more checking here; canSinkLastInstruction should
1569 // have done it all for us.
1570 SmallVector<Value*, 4> NewOperands;
1571 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1572 // This check is different to that in canSinkLastInstruction. There, we
1573 // cared about the global view once simplifycfg (and instcombine) have
1574 // completed - it takes into account PHIs that become trivially
1575 // simplifiable. However here we need a more local view; if an operand
1576 // differs we create a PHI and rely on instcombine to clean up the very
1577 // small mess we may make.
1578 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1579 return I->getOperand(O) != I0->getOperand(O);
1580 });
1581 if (!NeedPHI) {
1582 NewOperands.push_back(I0->getOperand(O));
1583 continue;
1584 }
1585
1586 // Create a new PHI in the successor block and populate it.
1587 auto *Op = I0->getOperand(O);
1588 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1588, __PRETTY_FUNCTION__))
;
1589 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1590 Op->getName() + ".sink", &BBEnd->front());
1591 for (auto *I : Insts)
1592 PN->addIncoming(I->getOperand(O), I->getParent());
1593 NewOperands.push_back(PN);
1594 }
1595
1596 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1597 // and move it to the start of the successor block.
1598 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1599 I0->getOperandUse(O).set(NewOperands[O]);
1600 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1601
1602 // Update metadata and IR flags, and merge debug locations.
1603 for (auto *I : Insts)
1604 if (I != I0) {
1605 // The debug location for the "common" instruction is the merged locations
1606 // of all the commoned instructions. We start with the original location
1607 // of the "common" instruction and iteratively merge each location in the
1608 // loop below.
1609 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1610 // However, as N-way merge for CallInst is rare, so we use simplified API
1611 // instead of using complex API for N-way merge.
1612 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1613 combineMetadataForCSE(I0, I, true);
1614 I0->andIRFlags(I);
1615 }
1616
1617 if (!I0->user_empty()) {
1618 // canSinkLastInstruction checked that all instructions were used by
1619 // one and only one PHI node. Find that now, RAUW it to our common
1620 // instruction and nuke it.
1621 auto *PN = cast<PHINode>(*I0->user_begin());
1622 PN->replaceAllUsesWith(I0);
1623 PN->eraseFromParent();
1624 }
1625
1626 // Finally nuke all instructions apart from the common instruction.
1627 for (auto *I : Insts)
1628 if (I != I0)
1629 I->eraseFromParent();
1630
1631 return true;
1632}
1633
1634namespace {
1635
1636 // LockstepReverseIterator - Iterates through instructions
1637 // in a set of blocks in reverse order from the first non-terminator.
1638 // For example (assume all blocks have size n):
1639 // LockstepReverseIterator I([B1, B2, B3]);
1640 // *I-- = [B1[n], B2[n], B3[n]];
1641 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1642 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1643 // ...
1644 class LockstepReverseIterator {
1645 ArrayRef<BasicBlock*> Blocks;
1646 SmallVector<Instruction*,4> Insts;
1647 bool Fail;
1648
1649 public:
1650 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1651 reset();
1652 }
1653
1654 void reset() {
1655 Fail = false;
1656 Insts.clear();
1657 for (auto *BB : Blocks) {
1658 Instruction *Inst = BB->getTerminator();
1659 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1660 Inst = Inst->getPrevNode();
1661 if (!Inst) {
1662 // Block wasn't big enough.
1663 Fail = true;
1664 return;
1665 }
1666 Insts.push_back(Inst);
1667 }
1668 }
1669
1670 bool isValid() const {
1671 return !Fail;
1672 }
1673
1674 void operator--() {
1675 if (Fail)
1676 return;
1677 for (auto *&Inst : Insts) {
1678 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1679 Inst = Inst->getPrevNode();
1680 // Already at beginning of block.
1681 if (!Inst) {
1682 Fail = true;
1683 return;
1684 }
1685 }
1686 }
1687
1688 ArrayRef<Instruction*> operator * () const {
1689 return Insts;
1690 }
1691 };
1692
1693} // end anonymous namespace
1694
1695/// Check whether BB's predecessors end with unconditional branches. If it is
1696/// true, sink any common code from the predecessors to BB.
1697/// We also allow one predecessor to end with conditional branch (but no more
1698/// than one).
1699static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1700 // We support two situations:
1701 // (1) all incoming arcs are unconditional
1702 // (2) one incoming arc is conditional
1703 //
1704 // (2) is very common in switch defaults and
1705 // else-if patterns;
1706 //
1707 // if (a) f(1);
1708 // else if (b) f(2);
1709 //
1710 // produces:
1711 //
1712 // [if]
1713 // / \
1714 // [f(1)] [if]
1715 // | | \
1716 // | | |
1717 // | [f(2)]|
1718 // \ | /
1719 // [ end ]
1720 //
1721 // [end] has two unconditional predecessor arcs and one conditional. The
1722 // conditional refers to the implicit empty 'else' arc. This conditional
1723 // arc can also be caused by an empty default block in a switch.
1724 //
1725 // In this case, we attempt to sink code from all *unconditional* arcs.
1726 // If we can sink instructions from these arcs (determined during the scan
1727 // phase below) we insert a common successor for all unconditional arcs and
1728 // connect that to [end], to enable sinking:
1729 //
1730 // [if]
1731 // / \
1732 // [x(1)] [if]
1733 // | | \
1734 // | | \
1735 // | [x(2)] |
1736 // \ / |
1737 // [sink.split] |
1738 // \ /
1739 // [ end ]
1740 //
1741 SmallVector<BasicBlock*,4> UnconditionalPreds;
1742 Instruction *Cond = nullptr;
1743 for (auto *B : predecessors(BB)) {
1744 auto *T = B->getTerminator();
1745 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1746 UnconditionalPreds.push_back(B);
1747 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1748 Cond = T;
1749 else
1750 return false;
1751 }
1752 if (UnconditionalPreds.size() < 2)
1753 return false;
1754
1755 bool Changed = false;
1756 // We take a two-step approach to tail sinking. First we scan from the end of
1757 // each block upwards in lockstep. If the n'th instruction from the end of each
1758 // block can be sunk, those instructions are added to ValuesToSink and we
1759 // carry on. If we can sink an instruction but need to PHI-merge some operands
1760 // (because they're not identical in each instruction) we add these to
1761 // PHIOperands.
1762 unsigned ScanIdx = 0;
1763 SmallPtrSet<Value*,4> InstructionsToSink;
1764 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1765 LockstepReverseIterator LRI(UnconditionalPreds);
1766 while (LRI.isValid() &&
1767 canSinkInstructions(*LRI, PHIOperands)) {
1768 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)
1769 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: instruction can be sunk: "
<< *(*LRI)[0] << "\n"; } } while (false)
;
1770 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1771 ++ScanIdx;
1772 --LRI;
1773 }
1774
1775 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1776 unsigned NumPHIdValues = 0;
1777 for (auto *I : *LRI)
1778 for (auto *V : PHIOperands[I])
1779 if (InstructionsToSink.count(V) == 0)
1780 ++NumPHIdValues;
1781 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: #phid values: " <<
NumPHIdValues << "\n"; } } while (false)
;
1782 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1783 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1784 NumPHIInsts++;
1785
1786 return NumPHIInsts <= 1;
1787 };
1788
1789 if (ScanIdx > 0 && Cond) {
1790 // Check if we would actually sink anything first! This mutates the CFG and
1791 // adds an extra block. The goal in doing this is to allow instructions that
1792 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1793 // (such as trunc, add) can be sunk and predicated already. So we check that
1794 // we're going to sink at least one non-speculatable instruction.
1795 LRI.reset();
1796 unsigned Idx = 0;
1797 bool Profitable = false;
1798 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1799 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1800 Profitable = true;
1801 break;
1802 }
1803 --LRI;
1804 ++Idx;
1805 }
1806 if (!Profitable)
1807 return false;
1808
1809 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Splitting edge\n"; }
} while (false)
;
1810 // We have a conditional edge and we're going to sink some instructions.
1811 // Insert a new block postdominating all blocks we're going to sink from.
1812 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1813 // Edges couldn't be split.
1814 return false;
1815 Changed = true;
1816 }
1817
1818 // Now that we've analyzed all potential sinking candidates, perform the
1819 // actual sink. We iteratively sink the last non-terminator of the source
1820 // blocks into their common successor unless doing so would require too
1821 // many PHI instructions to be generated (currently only one PHI is allowed
1822 // per sunk instruction).
1823 //
1824 // We can use InstructionsToSink to discount values needing PHI-merging that will
1825 // actually be sunk in a later iteration. This allows us to be more
1826 // aggressive in what we sink. This does allow a false positive where we
1827 // sink presuming a later value will also be sunk, but stop half way through
1828 // and never actually sink it which means we produce more PHIs than intended.
1829 // This is unlikely in practice though.
1830 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1831 LLVM_DEBUG(dbgs() << "SINK: Sink: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1832 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1833 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
;
1834
1835 // Because we've sunk every instruction in turn, the current instruction to
1836 // sink is always at index 0.
1837 LRI.reset();
1838 if (!ProfitableToSinkInstruction(LRI)) {
1839 // Too many PHIs would be created.
1840 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, too many PHIs would be created!\n"
; } } while (false)
1841 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)
;
1842 break;
1843 }
1844
1845 if (!sinkLastInstruction(UnconditionalPreds))
1846 return Changed;
1847 NumSinkCommons++;
1848 Changed = true;
1849 }
1850 return Changed;
1851}
1852
1853/// Determine if we can hoist sink a sole store instruction out of a
1854/// conditional block.
1855///
1856/// We are looking for code like the following:
1857/// BrBB:
1858/// store i32 %add, i32* %arrayidx2
1859/// ... // No other stores or function calls (we could be calling a memory
1860/// ... // function).
1861/// %cmp = icmp ult %x, %y
1862/// br i1 %cmp, label %EndBB, label %ThenBB
1863/// ThenBB:
1864/// store i32 %add5, i32* %arrayidx2
1865/// br label EndBB
1866/// EndBB:
1867/// ...
1868/// We are going to transform this into:
1869/// BrBB:
1870/// store i32 %add, i32* %arrayidx2
1871/// ... //
1872/// %cmp = icmp ult %x, %y
1873/// %add.add5 = select i1 %cmp, i32 %add, %add5
1874/// store i32 %add.add5, i32* %arrayidx2
1875/// ...
1876///
1877/// \return The pointer to the value of the previous store if the store can be
1878/// hoisted into the predecessor block. 0 otherwise.
1879static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1880 BasicBlock *StoreBB, BasicBlock *EndBB) {
1881 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1882 if (!StoreToHoist)
1883 return nullptr;
1884
1885 // Volatile or atomic.
1886 if (!StoreToHoist->isSimple())
1887 return nullptr;
1888
1889 Value *StorePtr = StoreToHoist->getPointerOperand();
1890
1891 // Look for a store to the same pointer in BrBB.
1892 unsigned MaxNumInstToLookAt = 9;
1893 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1894 if (!MaxNumInstToLookAt)
1895 break;
1896 --MaxNumInstToLookAt;
1897
1898 // Could be calling an instruction that affects memory like free().
1899 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1900 return nullptr;
1901
1902 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1903 // Found the previous store make sure it stores to the same location.
1904 if (SI->getPointerOperand() == StorePtr)
1905 // Found the previous store, return its value operand.
1906 return SI->getValueOperand();
1907 return nullptr; // Unknown store.
1908 }
1909 }
1910
1911 return nullptr;
1912}
1913
1914/// Speculate a conditional basic block flattening the CFG.
1915///
1916/// Note that this is a very risky transform currently. Speculating
1917/// instructions like this is most often not desirable. Instead, there is an MI
1918/// pass which can do it with full awareness of the resource constraints.
1919/// However, some cases are "obvious" and we should do directly. An example of
1920/// this is speculating a single, reasonably cheap instruction.
1921///
1922/// There is only one distinct advantage to flattening the CFG at the IR level:
1923/// it makes very common but simplistic optimizations such as are common in
1924/// instcombine and the DAG combiner more powerful by removing CFG edges and
1925/// modeling their effects with easier to reason about SSA value graphs.
1926///
1927///
1928/// An illustration of this transform is turning this IR:
1929/// \code
1930/// BB:
1931/// %cmp = icmp ult %x, %y
1932/// br i1 %cmp, label %EndBB, label %ThenBB
1933/// ThenBB:
1934/// %sub = sub %x, %y
1935/// br label BB2
1936/// EndBB:
1937/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1938/// ...
1939/// \endcode
1940///
1941/// Into this IR:
1942/// \code
1943/// BB:
1944/// %cmp = icmp ult %x, %y
1945/// %sub = sub %x, %y
1946/// %cond = select i1 %cmp, 0, %sub
1947/// ...
1948/// \endcode
1949///
1950/// \returns true if the conditional block is removed.
1951static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1952 const TargetTransformInfo &TTI) {
1953 // Be conservative for now. FP select instruction can often be expensive.
1954 Value *BrCond = BI->getCondition();
1955 if (isa<FCmpInst>(BrCond))
1956 return false;
1957
1958 BasicBlock *BB = BI->getParent();
1959 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1960
1961 // If ThenBB is actually on the false edge of the conditional branch, remember
1962 // to swap the select operands later.
1963 bool Invert = false;
1964 if (ThenBB != BI->getSuccessor(0)) {
1965 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1965, __PRETTY_FUNCTION__))
;
1966 Invert = true;
1967 }
1968 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1968, __PRETTY_FUNCTION__))
;
1969
1970 // Keep a count of how many times instructions are used within ThenBB when
1971 // they are candidates for sinking into ThenBB. Specifically:
1972 // - They are defined in BB, and
1973 // - They have no side effects, and
1974 // - All of their uses are in ThenBB.
1975 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1976
1977 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1978
1979 unsigned SpeculationCost = 0;
1980 Value *SpeculatedStoreValue = nullptr;
1981 StoreInst *SpeculatedStore = nullptr;
1982 for (BasicBlock::iterator BBI = ThenBB->begin(),
1983 BBE = std::prev(ThenBB->end());
1984 BBI != BBE; ++BBI) {
1985 Instruction *I = &*BBI;
1986 // Skip debug info.
1987 if (isa<DbgInfoIntrinsic>(I)) {
1988 SpeculatedDbgIntrinsics.push_back(I);
1989 continue;
1990 }
1991
1992 // Only speculatively execute a single instruction (not counting the
1993 // terminator) for now.
1994 ++SpeculationCost;
1995 if (SpeculationCost > 1)
1996 return false;
1997
1998 // Don't hoist the instruction if it's unsafe or expensive.
1999 if (!isSafeToSpeculativelyExecute(I) &&
2000 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
2001 I, BB, ThenBB, EndBB))))
2002 return false;
2003 if (!SpeculatedStoreValue &&
2004 ComputeSpeculationCost(I, TTI) >
2005 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2006 return false;
2007
2008 // Store the store speculation candidate.
2009 if (SpeculatedStoreValue)
2010 SpeculatedStore = cast<StoreInst>(I);
2011
2012 // Do not hoist the instruction if any of its operands are defined but not
2013 // used in BB. The transformation will prevent the operand from
2014 // being sunk into the use block.
2015 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2016 Instruction *OpI = dyn_cast<Instruction>(*i);
2017 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2018 continue; // Not a candidate for sinking.
2019
2020 ++SinkCandidateUseCounts[OpI];
2021 }
2022 }
2023
2024 // Consider any sink candidates which are only used in ThenBB as costs for
2025 // speculation. Note, while we iterate over a DenseMap here, we are summing
2026 // and so iteration order isn't significant.
2027 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2028 I = SinkCandidateUseCounts.begin(),
2029 E = SinkCandidateUseCounts.end();
2030 I != E; ++I)
2031 if (I->first->hasNUses(I->second)) {
2032 ++SpeculationCost;
2033 if (SpeculationCost > 1)
2034 return false;
2035 }
2036
2037 // Check that the PHI nodes can be converted to selects.
2038 bool HaveRewritablePHIs = false;
2039 for (PHINode &PN : EndBB->phis()) {
2040 Value *OrigV = PN.getIncomingValueForBlock(BB);
2041 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2042
2043 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2044 // Skip PHIs which are trivial.
2045 if (ThenV == OrigV)
2046 continue;
2047
2048 // Don't convert to selects if we could remove undefined behavior instead.
2049 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2050 passingValueIsAlwaysUndefined(ThenV, &PN))
2051 return false;
2052
2053 HaveRewritablePHIs = true;
2054 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2055 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2056 if (!OrigCE && !ThenCE)
2057 continue; // Known safe and cheap.
2058
2059 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2060 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2061 return false;
2062 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2063 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2064 unsigned MaxCost =
2065 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2066 if (OrigCost + ThenCost > MaxCost)
2067 return false;
2068
2069 // Account for the cost of an unfolded ConstantExpr which could end up
2070 // getting expanded into Instructions.
2071 // FIXME: This doesn't account for how many operations are combined in the
2072 // constant expression.
2073 ++SpeculationCost;
2074 if (SpeculationCost > 1)
2075 return false;
2076 }
2077
2078 // If there are no PHIs to process, bail early. This helps ensure idempotence
2079 // as well.
2080 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2081 return false;
2082
2083 // If we get here, we can hoist the instruction and if-convert.
2084 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SPECULATIVELY EXECUTING BB"
<< *ThenBB << "\n";; } } while (false)
;
2085
2086 // Insert a select of the value of the speculated store.
2087 if (SpeculatedStoreValue) {
2088 IRBuilder<NoFolder> Builder(BI);
2089 Value *TrueV = SpeculatedStore->getValueOperand();
2090 Value *FalseV = SpeculatedStoreValue;
2091 if (Invert)
2092 std::swap(TrueV, FalseV);
2093 Value *S = Builder.CreateSelect(
2094 BrCond, TrueV, FalseV, "spec.store.select", BI);
2095 SpeculatedStore->setOperand(0, S);
2096 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2097 SpeculatedStore->getDebugLoc());
2098 }
2099
2100 // Metadata can be dependent on the condition we are hoisting above.
2101 // Conservatively strip all metadata on the instruction.
2102 for (auto &I : *ThenBB)
2103 I.dropUnknownNonDebugMetadata();
2104
2105 // Hoist the instructions.
2106 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2107 ThenBB->begin(), std::prev(ThenBB->end()));
2108
2109 // Insert selects and rewrite the PHI operands.
2110 IRBuilder<NoFolder> Builder(BI);
2111 for (PHINode &PN : EndBB->phis()) {
2112 unsigned OrigI = PN.getBasicBlockIndex(BB);
2113 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2114 Value *OrigV = PN.getIncomingValue(OrigI);
2115 Value *ThenV = PN.getIncomingValue(ThenI);
2116
2117 // Skip PHIs which are trivial.
2118 if (OrigV == ThenV)
2119 continue;
2120
2121 // Create a select whose true value is the speculatively executed value and
2122 // false value is the preexisting value. Swap them if the branch
2123 // destinations were inverted.
2124 Value *TrueV = ThenV, *FalseV = OrigV;
2125 if (Invert)
2126 std::swap(TrueV, FalseV);
2127 Value *V = Builder.CreateSelect(
2128 BrCond, TrueV, FalseV, "spec.select", BI);
2129 PN.setIncomingValue(OrigI, V);
2130 PN.setIncomingValue(ThenI, V);
2131 }
2132
2133 // Remove speculated dbg intrinsics.
2134 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2135 // dbg value for the different flows and inserting it after the select.
2136 for (Instruction *I : SpeculatedDbgIntrinsics)
2137 I->eraseFromParent();
2138
2139 ++NumSpeculations;
2140 return true;
2141}
2142
2143/// Return true if we can thread a branch across this block.
2144static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2145 unsigned Size = 0;
2146
2147 for (Instruction &I : BB->instructionsWithoutDebug()) {
2148 if (Size > 10)
2149 return false; // Don't clone large BB's.
2150 ++Size;
2151
2152 // We can only support instructions that do not define values that are
2153 // live outside of the current basic block.
2154 for (User *U : I.users()) {
2155 Instruction *UI = cast<Instruction>(U);
2156 if (UI->getParent() != BB || isa<PHINode>(UI))
2157 return false;
2158 }
2159
2160 // Looks ok, continue checking.
2161 }
2162
2163 return true;
2164}
2165
2166/// If we have a conditional branch on a PHI node value that is defined in the
2167/// same block as the branch and if any PHI entries are constants, thread edges
2168/// corresponding to that entry to be branches to their ultimate destination.
2169static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2170 AssumptionCache *AC) {
2171 BasicBlock *BB = BI->getParent();
2172 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2173 // NOTE: we currently cannot transform this case if the PHI node is used
2174 // outside of the block.
2175 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2176 return false;
2177
2178 // Degenerate case of a single entry PHI.
2179 if (PN->getNumIncomingValues() == 1) {
2180 FoldSingleEntryPHINodes(PN->getParent());
2181 return true;
2182 }
2183
2184 // Now we know that this block has multiple preds and two succs.
2185 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2186 return false;
2187
2188 // Can't fold blocks that contain noduplicate or convergent calls.
2189 if (any_of(*BB, [](const Instruction &I) {
2190 const CallInst *CI = dyn_cast<CallInst>(&I);
2191 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2192 }))
2193 return false;
2194
2195 // Okay, this is a simple enough basic block. See if any phi values are
2196 // constants.
2197 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2198 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2199 if (!CB || !CB->getType()->isIntegerTy(1))
2200 continue;
2201
2202 // Okay, we now know that all edges from PredBB should be revectored to
2203 // branch to RealDest.
2204 BasicBlock *PredBB = PN->getIncomingBlock(i);
2205 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2206
2207 if (RealDest == BB)
2208 continue; // Skip self loops.
2209 // Skip if the predecessor's terminator is an indirect branch.
2210 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2211 continue;
2212
2213 // The dest block might have PHI nodes, other predecessors and other
2214 // difficult cases. Instead of being smart about this, just insert a new
2215 // block that jumps to the destination block, effectively splitting
2216 // the edge we are about to create.
2217 BasicBlock *EdgeBB =
2218 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2219 RealDest->getParent(), RealDest);
2220 BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2221 CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2222
2223 // Update PHI nodes.
2224 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2225
2226 // BB may have instructions that are being threaded over. Clone these
2227 // instructions into EdgeBB. We know that there will be no uses of the
2228 // cloned instructions outside of EdgeBB.
2229 BasicBlock::iterator InsertPt = EdgeBB->begin();
2230 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2231 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2232 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2233 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2234 continue;
2235 }
2236 // Clone the instruction.
2237 Instruction *N = BBI->clone();
2238 if (BBI->hasName())
2239 N->setName(BBI->getName() + ".c");
2240
2241 // Update operands due to translation.
2242 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2243 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2244 if (PI != TranslateMap.end())
2245 *i = PI->second;
2246 }
2247
2248 // Check for trivial simplification.
2249 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2250 if (!BBI->use_empty())
2251 TranslateMap[&*BBI] = V;
2252 if (!N->mayHaveSideEffects()) {
2253 N->deleteValue(); // Instruction folded away, don't need actual inst
2254 N = nullptr;
2255 }
2256 } else {
2257 if (!BBI->use_empty())
2258 TranslateMap[&*BBI] = N;
2259 }
2260 // Insert the new instruction into its new home.
2261 if (N)
2262 EdgeBB->getInstList().insert(InsertPt, N);
2263
2264 // Register the new instruction with the assumption cache if necessary.
2265 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2266 if (II->getIntrinsicID() == Intrinsic::assume)
2267 AC->registerAssumption(II);
2268 }
2269
2270 // Loop over all of the edges from PredBB to BB, changing them to branch
2271 // to EdgeBB instead.
2272 Instruction *PredBBTI = PredBB->getTerminator();
2273 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2274 if (PredBBTI->getSuccessor(i) == BB) {
2275 BB->removePredecessor(PredBB);
2276 PredBBTI->setSuccessor(i, EdgeBB);
2277 }
2278
2279 // Recurse, simplifying any other constants.
2280 return FoldCondBranchOnPHI(BI, DL, AC) || true;
2281 }
2282
2283 return false;
2284}
2285
2286/// Given a BB that starts with the specified two-entry PHI node,
2287/// see if we can eliminate it.
2288static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2289 const DataLayout &DL) {
2290 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2291 // statement", which has a very simple dominance structure. Basically, we
2292 // are trying to find the condition that is being branched on, which
2293 // subsequently causes this merge to happen. We really want control
2294 // dependence information for this check, but simplifycfg can't keep it up
2295 // to date, and this catches most of the cases we care about anyway.
2296 BasicBlock *BB = PN->getParent();
2297 const Function *Fn = BB->getParent();
2298 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
18
Assuming 'Fn' is null
2299 return false;
2300
2301 BasicBlock *IfTrue, *IfFalse;
2302 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2303 if (!IfCond ||
19
Assuming 'IfCond' is non-null
20
Taking false branch
2304 // Don't bother if the branch will be constant folded trivially.
2305 isa<ConstantInt>(IfCond))
2306 return false;
2307
2308 // Okay, we found that we can merge this two-entry phi node into a select.
2309 // Doing so would require us to fold *all* two entry phi nodes in this block.
2310 // At some point this becomes non-profitable (particularly if the target
2311 // doesn't support cmov's). Only do this transformation if there are two or
2312 // fewer PHI nodes in this block.
2313 unsigned NumPhis = 0;
2314 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
21
Loop condition is false. Execution continues on line 2321
2315 if (NumPhis > 2)
2316 return false;
2317
2318 // Loop over the PHI's seeing if we can promote them all to select
2319 // instructions. While we are at it, keep track of the instructions
2320 // that need to be moved to the dominating block.
2321 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2322 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2323 MaxCostVal1 = PHINodeFoldingThreshold;
2324 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2325 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2326
2327 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
22
Loop condition is false. Execution continues on line 2344
2328 PHINode *PN = cast<PHINode>(II++);
2329 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2330 PN->replaceAllUsesWith(V);
2331 PN->eraseFromParent();
2332 continue;
2333 }
2334
2335 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2336 MaxCostVal0, TTI) ||
2337 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2338 MaxCostVal1, TTI))
2339 return false;
2340 }
2341
2342 // If we folded the first phi, PN dangles at this point. Refresh it. If
2343 // we ran out of PHIs then we simplified them all.
2344 PN = dyn_cast<PHINode>(BB->begin());
2345 if (!PN)
23
Assuming 'PN' is non-null
24
Taking false branch
2346 return true;
2347
2348 // Don't fold i1 branches on PHIs which contain binary operators. These can
2349 // often be turned into switches and other things.
2350 if (PN->getType()->isIntegerTy(1) &&
25
Assuming the condition is false
26
Taking false branch
2351 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2352 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2353 isa<BinaryOperator>(IfCond)))
2354 return false;
2355
2356 // If all PHI nodes are promotable, check to make sure that all instructions
2357 // in the predecessor blocks can be promoted as well. If not, we won't be able
2358 // to get rid of the control flow, so it's not worth promoting to select
2359 // instructions.
2360 BasicBlock *DomBlock = nullptr;
27
'DomBlock' initialized to a null pointer value
2361 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2362 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2363 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
28
Taking true branch
2364 IfBlock1 = nullptr;
2365 } else {
2366 DomBlock = *pred_begin(IfBlock1);
2367 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2368 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2369 // This is not an aggressive instruction that we can promote.
2370 // Because of this, we won't be able to get rid of the control flow, so
2371 // the xform is not worth it.
2372 return false;
2373 }
2374 }
2375
2376 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
29
Taking true branch
2377 IfBlock2 = nullptr;
2378 } else {
2379 DomBlock = *pred_begin(IfBlock2);
2380 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2381 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2382 // This is not an aggressive instruction that we can promote.
2383 // Because of this, we won't be able to get rid of the control flow, so
2384 // the xform is not worth it.
2385 return false;
2386 }
2387 }
2388
2389 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)
30
Assuming 'DebugFlag' is 0
31
Loop condition is false. Exiting loop
2390 << " T: " << IfTrue->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
2391 << " 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)
;
2392
2393 // If we can still promote the PHI nodes after this gauntlet of tests,
2394 // do all of the PHI's now.
2395 Instruction *InsertPt = DomBlock->getTerminator();
32
Called C++ object pointer is null
2396 IRBuilder<NoFolder> Builder(InsertPt);
2397
2398 // Move all 'aggressive' instructions, which are defined in the
2399 // conditional parts of the if's up to the dominating block.
2400 if (IfBlock1)
2401 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2402 if (IfBlock2)
2403 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2404
2405 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2406 // Change the PHI node into a select instruction.
2407 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2408 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2409
2410 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2411 PN->replaceAllUsesWith(Sel);
2412 Sel->takeName(PN);
2413 PN->eraseFromParent();
2414 }
2415
2416 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2417 // has been flattened. Change DomBlock to jump directly to our new block to
2418 // avoid other simplifycfg's kicking in on the diamond.
2419 Instruction *OldTI = DomBlock->getTerminator();
2420 Builder.SetInsertPoint(OldTI);
2421 Builder.CreateBr(BB);
2422 OldTI->eraseFromParent();
2423 return true;
2424}
2425
2426/// If we found a conditional branch that goes to two returning blocks,
2427/// try to merge them together into one return,
2428/// introducing a select if the return values disagree.
2429static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2430 IRBuilder<> &Builder) {
2431 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2431, __PRETTY_FUNCTION__))
;
2432 BasicBlock *TrueSucc = BI->getSuccessor(0);
2433 BasicBlock *FalseSucc = BI->getSuccessor(1);
2434 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2435 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2436
2437 // Check to ensure both blocks are empty (just a return) or optionally empty
2438 // with PHI nodes. If there are other instructions, merging would cause extra
2439 // computation on one path or the other.
2440 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2441 return false;
2442 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2443 return false;
2444
2445 Builder.SetInsertPoint(BI);
2446 // Okay, we found a branch that is going to two return nodes. If
2447 // there is no return value for this function, just change the
2448 // branch into a return.
2449 if (FalseRet->getNumOperands() == 0) {
2450 TrueSucc->removePredecessor(BI->getParent());
2451 FalseSucc->removePredecessor(BI->getParent());
2452 Builder.CreateRetVoid();
2453 EraseTerminatorAndDCECond(BI);
2454 return true;
2455 }
2456
2457 // Otherwise, figure out what the true and false return values are
2458 // so we can insert a new select instruction.
2459 Value *TrueValue = TrueRet->getReturnValue();
2460 Value *FalseValue = FalseRet->getReturnValue();
2461
2462 // Unwrap any PHI nodes in the return blocks.
2463 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2464 if (TVPN->getParent() == TrueSucc)
2465 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2466 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2467 if (FVPN->getParent() == FalseSucc)
2468 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2469
2470 // In order for this transformation to be safe, we must be able to
2471 // unconditionally execute both operands to the return. This is
2472 // normally the case, but we could have a potentially-trapping
2473 // constant expression that prevents this transformation from being
2474 // safe.
2475 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2476 if (TCV->canTrap())
2477 return false;
2478 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2479 if (FCV->canTrap())
2480 return false;
2481
2482 // Okay, we collected all the mapped values and checked them for sanity, and
2483 // defined to really do this transformation. First, update the CFG.
2484 TrueSucc->removePredecessor(BI->getParent());
2485 FalseSucc->removePredecessor(BI->getParent());
2486
2487 // Insert select instructions where needed.
2488 Value *BrCond = BI->getCondition();
2489 if (TrueValue) {
2490 // Insert a select if the results differ.
2491 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2492 } else if (isa<UndefValue>(TrueValue)) {
2493 TrueValue = FalseValue;
2494 } else {
2495 TrueValue =
2496 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2497 }
2498 }
2499
2500 Value *RI =
2501 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2502
2503 (void)RI;
2504
2505 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 << "NewRet = " << *
RI << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "
<< *FalseSucc; } } while (false)
2506 << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "NewRet = " << *
RI << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "
<< *FalseSucc; } } while (false)
2507 << *TrueSucc << "FALSEBLOCK: " << *FalseSucc)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "NewRet = " << *
RI << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "
<< *FalseSucc; } } while (false)
;
2508
2509 EraseTerminatorAndDCECond(BI);
2510
2511 return true;
2512}
2513
2514/// Return true if the given instruction is available
2515/// in its predecessor block. If yes, the instruction will be removed.
2516static bool tryCSEWithPredecessor(Instruction *Inst, BasicBlock *PB) {
2517 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2518 return false;
2519 for (Instruction &I : *PB) {
2520 Instruction *PBI = &I;
2521 // Check whether Inst and PBI generate the same value.
2522 if (Inst->isIdenticalTo(PBI)) {
2523 Inst->replaceAllUsesWith(PBI);
2524 Inst->eraseFromParent();
2525 return true;
2526 }
2527 }
2528 return false;
2529}
2530
2531/// Return true if either PBI or BI has branch weight available, and store
2532/// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2533/// not have branch weight, use 1:1 as its weight.
2534static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2535 uint64_t &PredTrueWeight,
2536 uint64_t &PredFalseWeight,
2537 uint64_t &SuccTrueWeight,
2538 uint64_t &SuccFalseWeight) {
2539 bool PredHasWeights =
2540 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2541 bool SuccHasWeights =
2542 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2543 if (PredHasWeights || SuccHasWeights) {
2544 if (!PredHasWeights)
2545 PredTrueWeight = PredFalseWeight = 1;
2546 if (!SuccHasWeights)
2547 SuccTrueWeight = SuccFalseWeight = 1;
2548 return true;
2549 } else {
2550 return false;
2551 }
2552}
2553
2554/// If this basic block is simple enough, and if a predecessor branches to us
2555/// and one of our successors, fold the block into the predecessor and use
2556/// logical operations to pick the right destination.
2557bool llvm::FoldBranchToCommonDest(BranchInst *BI, MemorySSAUpdater *MSSAU,
2558 unsigned BonusInstThreshold) {
2559 BasicBlock *BB = BI->getParent();
2560
2561 const unsigned PredCount = pred_size(BB);
2562
2563 Instruction *Cond = nullptr;
2564 if (BI->isConditional())
2565 Cond = dyn_cast<Instruction>(BI->getCondition());
2566 else {
2567 // For unconditional branch, check for a simple CFG pattern, where
2568 // BB has a single predecessor and BB's successor is also its predecessor's
2569 // successor. If such pattern exists, check for CSE between BB and its
2570 // predecessor.
2571 if (BasicBlock *PB = BB->getSinglePredecessor())
2572 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2573 if (PBI->isConditional() &&
2574 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2575 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2576 for (auto I = BB->instructionsWithoutDebug().begin(),
2577 E = BB->instructionsWithoutDebug().end();
2578 I != E;) {
2579 Instruction *Curr = &*I++;
2580 if (isa<CmpInst>(Curr)) {
2581 Cond = Curr;
2582 break;
2583 }
2584 // Quit if we can't remove this instruction.
2585 if (!tryCSEWithPredecessor(Curr, PB))
2586 return false;
2587 }
2588 }
2589
2590 if (!Cond)
2591 return false;
2592 }
2593
2594 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2595 Cond->getParent() != BB || !Cond->hasOneUse())
2596 return false;
2597
2598 // Make sure the instruction after the condition is the cond branch.
2599 BasicBlock::iterator CondIt = ++Cond->getIterator();
2600
2601 // Ignore dbg intrinsics.
2602 while (isa<DbgInfoIntrinsic>(CondIt))
2603 ++CondIt;
2604
2605 if (&*CondIt != BI)
2606 return false;
2607
2608 // Only allow this transformation if computing the condition doesn't involve
2609 // too many instructions and these involved instructions can be executed
2610 // unconditionally. We denote all involved instructions except the condition
2611 // as "bonus instructions", and only allow this transformation when the
2612 // number of the bonus instructions we'll need to create when cloning into
2613 // each predecessor does not exceed a certain threshold.
2614 unsigned NumBonusInsts = 0;
2615 for (auto I = BB->begin(); Cond != &*I; ++I) {
2616 // Ignore dbg intrinsics.
2617 if (isa<DbgInfoIntrinsic>(I))
2618 continue;
2619 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2620 return false;
2621 // I has only one use and can be executed unconditionally.
2622 Instruction *User = dyn_cast<Instruction>(I->user_back());
2623 if (User == nullptr || User->getParent() != BB)
2624 return false;
2625 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2626 // to use any other instruction, User must be an instruction between next(I)
2627 // and Cond.
2628
2629 // Account for the cost of duplicating this instruction into each
2630 // predecessor.
2631 NumBonusInsts += PredCount;
2632 // Early exits once we reach the limit.
2633 if (NumBonusInsts > BonusInstThreshold)
2634 return false;
2635 }
2636
2637 // Cond is known to be a compare or binary operator. Check to make sure that
2638 // neither operand is a potentially-trapping constant expression.
2639 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2640 if (CE->canTrap())
2641 return false;
2642 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2643 if (CE->canTrap())
2644 return false;
2645
2646 // Finally, don't infinitely unroll conditional loops.
2647 BasicBlock *TrueDest = BI->getSuccessor(0);
2648 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2649 if (TrueDest == BB || FalseDest == BB)
2650 return false;
2651
2652 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2653 BasicBlock *PredBlock = *PI;
2654 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2655
2656 // Check that we have two conditional branches. If there is a PHI node in
2657 // the common successor, verify that the same value flows in from both
2658 // blocks.
2659 SmallVector<PHINode *, 4> PHIs;
2660 if (!PBI || PBI->isUnconditional() ||
2661 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2662 (!BI->isConditional() &&
2663 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2664 continue;
2665
2666 // Determine if the two branches share a common destination.
2667 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2668 bool InvertPredCond = false;
2669
2670 if (BI->isConditional()) {
2671 if (PBI->getSuccessor(0) == TrueDest) {
2672 Opc = Instruction::Or;
2673 } else if (PBI->getSuccessor(1) == FalseDest) {
2674 Opc = Instruction::And;
2675 } else if (PBI->getSuccessor(0) == FalseDest) {
2676 Opc = Instruction::And;
2677 InvertPredCond = true;
2678 } else if (PBI->getSuccessor(1) == TrueDest) {
2679 Opc = Instruction::Or;
2680 InvertPredCond = true;
2681 } else {
2682 continue;
2683 }
2684 } else {
2685 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2686 continue;
2687 }
2688
2689 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)
;
2690 IRBuilder<> Builder(PBI);
2691
2692 // If we need to invert the condition in the pred block to match, do so now.
2693 if (InvertPredCond) {
2694 Value *NewCond = PBI->getCondition();
2695
2696 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2697 CmpInst *CI = cast<CmpInst>(NewCond);
2698 CI->setPredicate(CI->getInversePredicate());
2699 } else {
2700 NewCond =
2701 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2702 }
2703
2704 PBI->setCondition(NewCond);
2705 PBI->swapSuccessors();
2706 }
2707
2708 // If we have bonus instructions, clone them into the predecessor block.
2709 // Note that there may be multiple predecessor blocks, so we cannot move
2710 // bonus instructions to a predecessor block.
2711 ValueToValueMapTy VMap; // maps original values to cloned values
2712 // We already make sure Cond is the last instruction before BI. Therefore,
2713 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2714 // instructions.
2715 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2716 if (isa<DbgInfoIntrinsic>(BonusInst))
2717 continue;
2718 Instruction *NewBonusInst = BonusInst->clone();
2719 RemapInstruction(NewBonusInst, VMap,
2720 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2721 VMap[&*BonusInst] = NewBonusInst;
2722
2723 // If we moved a load, we cannot any longer claim any knowledge about
2724 // its potential value. The previous information might have been valid
2725 // only given the branch precondition.
2726 // For an analogous reason, we must also drop all the metadata whose
2727 // semantics we don't understand.
2728 NewBonusInst->dropUnknownNonDebugMetadata();
2729
2730 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2731 NewBonusInst->takeName(&*BonusInst);
2732 BonusInst->setName(BonusInst->getName() + ".old");
2733 }
2734
2735 // Clone Cond into the predecessor basic block, and or/and the
2736 // two conditions together.
2737 Instruction *CondInPred = Cond->clone();
2738 RemapInstruction(CondInPred, VMap,
2739 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2740 PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2741 CondInPred->takeName(Cond);
2742 Cond->setName(CondInPred->getName() + ".old");
2743
2744 if (BI->isConditional()) {
2745 Instruction *NewCond = cast<Instruction>(
2746 Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2747 PBI->setCondition(NewCond);
2748
2749 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2750 bool HasWeights =
2751 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2752 SuccTrueWeight, SuccFalseWeight);
2753 SmallVector<uint64_t, 8> NewWeights;
2754
2755 if (PBI->getSuccessor(0) == BB) {
2756 if (HasWeights) {
2757 // PBI: br i1 %x, BB, FalseDest
2758 // BI: br i1 %y, TrueDest, FalseDest
2759 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2760 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2761 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2762 // TrueWeight for PBI * FalseWeight for BI.
2763 // We assume that total weights of a BranchInst can fit into 32 bits.
2764 // Therefore, we will not have overflow using 64-bit arithmetic.
2765 NewWeights.push_back(PredFalseWeight *
2766 (SuccFalseWeight + SuccTrueWeight) +
2767 PredTrueWeight * SuccFalseWeight);
2768 }
2769 AddPredecessorToBlock(TrueDest, PredBlock, BB, MSSAU);
2770 PBI->setSuccessor(0, TrueDest);
2771 }
2772 if (PBI->getSuccessor(1) == BB) {
2773 if (HasWeights) {
2774 // PBI: br i1 %x, TrueDest, BB
2775 // BI: br i1 %y, TrueDest, FalseDest
2776 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2777 // FalseWeight for PBI * TrueWeight for BI.
2778 NewWeights.push_back(PredTrueWeight *
2779 (SuccFalseWeight + SuccTrueWeight) +
2780 PredFalseWeight * SuccTrueWeight);
2781 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2782 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2783 }
2784 AddPredecessorToBlock(FalseDest, PredBlock, BB, MSSAU);
2785 PBI->setSuccessor(1, FalseDest);
2786 }
2787 if (NewWeights.size() == 2) {
2788 // Halve the weights if any of them cannot fit in an uint32_t
2789 FitWeights(NewWeights);
2790
2791 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2792 NewWeights.end());
2793 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2794 } else
2795 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2796 } else {
2797 // Update PHI nodes in the common successors.
2798 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2799 ConstantInt *PBI_C = cast<ConstantInt>(
2800 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2801 assert(PBI_C->getType()->isIntegerTy(1))((PBI_C->getType()->isIntegerTy(1)) ? static_cast<void
> (0) : __assert_fail ("PBI_C->getType()->isIntegerTy(1)"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2801, __PRETTY_FUNCTION__))
;
2802 Instruction *MergedCond = nullptr;
2803 if (PBI->getSuccessor(0) == TrueDest) {
2804 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2805 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2806 // is false: !PBI_Cond and BI_Value
2807 Instruction *NotCond = cast<Instruction>(
2808 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2809 MergedCond = cast<Instruction>(
2810 Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2811 "and.cond"));
2812 if (PBI_C->isOne())
2813 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2814 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2815 } else {
2816 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2817 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2818 // is false: PBI_Cond and BI_Value
2819 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2820 Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2821 if (PBI_C->isOne()) {
2822 Instruction *NotCond = cast<Instruction>(
2823 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2824 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2825 Instruction::Or, NotCond, MergedCond, "or.cond"));
2826 }
2827 }
2828 // Update PHI Node.
2829 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2830 MergedCond);
2831 }
2832
2833 // PBI is changed to branch to TrueDest below. Remove itself from
2834 // potential phis from all other successors.
2835 if (MSSAU)
2836 MSSAU->changeCondBranchToUnconditionalTo(PBI, TrueDest);
2837
2838 // Change PBI from Conditional to Unconditional.
2839 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2840 EraseTerminatorAndDCECond(PBI, MSSAU);
2841 PBI = New_PBI;
2842 }
2843
2844 // If BI was a loop latch, it may have had associated loop metadata.
2845 // We need to copy it to the new latch, that is, PBI.
2846 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2847 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2848
2849 // TODO: If BB is reachable from all paths through PredBlock, then we
2850 // could replace PBI's branch probabilities with BI's.
2851
2852 // Copy any debug value intrinsics into the end of PredBlock.
2853 for (Instruction &I : *BB)
2854 if (isa<DbgInfoIntrinsic>(I))
2855 I.clone()->insertBefore(PBI);
2856
2857 return true;
2858 }
2859 return false;
2860}
2861
2862// If there is only one store in BB1 and BB2, return it, otherwise return
2863// nullptr.
2864static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2865 StoreInst *S = nullptr;
2866 for (auto *BB : {BB1, BB2}) {
2867 if (!BB)
2868 continue;
2869 for (auto &I : *BB)
2870 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2871 if (S)
2872 // Multiple stores seen.
2873 return nullptr;
2874 else
2875 S = SI;
2876 }
2877 }
2878 return S;
2879}
2880
2881static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2882 Value *AlternativeV = nullptr) {
2883 // PHI is going to be a PHI node that allows the value V that is defined in
2884 // BB to be referenced in BB's only successor.
2885 //
2886 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2887 // doesn't matter to us what the other operand is (it'll never get used). We
2888 // could just create a new PHI with an undef incoming value, but that could
2889 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2890 // other PHI. So here we directly look for some PHI in BB's successor with V
2891 // as an incoming operand. If we find one, we use it, else we create a new
2892 // one.
2893 //
2894 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2895 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2896 // where OtherBB is the single other predecessor of BB's only successor.
2897 PHINode *PHI = nullptr;
2898 BasicBlock *Succ = BB->getSingleSuccessor();
2899
2900 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2901 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2902 PHI = cast<PHINode>(I);
2903 if (!AlternativeV)
2904 break;
2905
2906 assert(Succ->hasNPredecessors(2))((Succ->hasNPredecessors(2)) ? static_cast<void> (0)
: __assert_fail ("Succ->hasNPredecessors(2)", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2906, __PRETTY_FUNCTION__))
;
2907 auto PredI = pred_begin(Succ);
2908 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2909 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2910 break;
2911 PHI = nullptr;
2912 }
2913 if (PHI)
2914 return PHI;
2915
2916 // If V is not an instruction defined in BB, just return it.
2917 if (!AlternativeV &&
2918 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2919 return V;
2920
2921 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2922 PHI->addIncoming(V, BB);
2923 for (BasicBlock *PredBB : predecessors(Succ))
2924 if (PredBB != BB)
2925 PHI->addIncoming(
2926 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2927 return PHI;
2928}
2929
2930static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2931 BasicBlock *QTB, BasicBlock *QFB,
2932 BasicBlock *PostBB, Value *Address,
2933 bool InvertPCond, bool InvertQCond,
2934 const DataLayout &DL) {
2935 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2936 return Operator::getOpcode(&I) == Instruction::BitCast &&
2937 I.getType()->isPointerTy();
2938 };
2939
2940 // If we're not in aggressive mode, we only optimize if we have some
2941 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2942 auto IsWorthwhile = [&](BasicBlock *BB) {
2943 if (!BB)
2944 return true;
2945 // Heuristic: if the block can be if-converted/phi-folded and the
2946 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2947 // thread this store.
2948 unsigned N = 0;
2949 for (auto &I : BB->instructionsWithoutDebug()) {
2950 // Cheap instructions viable for folding.
2951 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2952 isa<StoreInst>(I))
2953 ++N;
2954 // Free instructions.
2955 else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2956 continue;
2957 else
2958 return false;
2959 }
2960 // The store we want to merge is counted in N, so add 1 to make sure
2961 // we're counting the instructions that would be left.
2962 return N <= (PHINodeFoldingThreshold + 1);
2963 };
2964
2965 if (!MergeCondStoresAggressively &&
2966 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2967 !IsWorthwhile(QFB)))
2968 return false;
2969
2970 // For every pointer, there must be exactly two stores, one coming from
2971 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2972 // store (to any address) in PTB,PFB or QTB,QFB.
2973 // FIXME: We could relax this restriction with a bit more work and performance
2974 // testing.
2975 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2976 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2977 if (!PStore || !QStore)
2978 return false;
2979
2980 // Now check the stores are compatible.
2981 if (!QStore->isUnordered() || !PStore->isUnordered())
2982 return false;
2983
2984 // Check that sinking the store won't cause program behavior changes. Sinking
2985 // the store out of the Q blocks won't change any behavior as we're sinking
2986 // from a block to its unconditional successor. But we're moving a store from
2987 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2988 // So we need to check that there are no aliasing loads or stores in
2989 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2990 // operations between PStore and the end of its parent block.
2991 //
2992 // The ideal way to do this is to query AliasAnalysis, but we don't
2993 // preserve AA currently so that is dangerous. Be super safe and just
2994 // check there are no other memory operations at all.
2995 for (auto &I : *QFB->getSinglePredecessor())
2996 if (I.mayReadOrWriteMemory())
2997 return false;
2998 for (auto &I : *QFB)
2999 if (&I != QStore && I.mayReadOrWriteMemory())
3000 return false;
3001 if (QTB)
3002 for (auto &I : *QTB)
3003 if (&I != QStore && I.mayReadOrWriteMemory())
3004 return false;
3005 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
3006 I != E; ++I)
3007 if (&*I != PStore && I->mayReadOrWriteMemory())
3008 return false;
3009
3010 // If PostBB has more than two predecessors, we need to split it so we can
3011 // sink the store.
3012 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3013 // We know that QFB's only successor is PostBB. And QFB has a single
3014 // predecessor. If QTB exists, then its only successor is also PostBB.
3015 // If QTB does not exist, then QFB's only predecessor has a conditional
3016 // branch to QFB and PostBB.
3017 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3018 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3019 "condstore.split");
3020 if (!NewBB)
3021 return false;
3022 PostBB = NewBB;
3023 }
3024
3025 // OK, we're going to sink the stores to PostBB. The store has to be
3026 // conditional though, so first create the predicate.
3027 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3028 ->getCondition();
3029 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3030 ->getCondition();
3031
3032 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3033 PStore->getParent());
3034 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3035 QStore->getParent(), PPHI);
3036
3037 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3038
3039 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3040 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3041
3042 if (InvertPCond)
3043 PPred = QB.CreateNot(PPred);
3044 if (InvertQCond)
3045 QPred = QB.CreateNot(QPred);
3046 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3047
3048 auto *T =
3049 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3050 QB.SetInsertPoint(T);
3051 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3052 AAMDNodes AAMD;
3053 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3054 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3055 SI->setAAMetadata(AAMD);
3056 unsigned PAlignment = PStore->getAlignment();
3057 unsigned QAlignment = QStore->getAlignment();
3058 unsigned TypeAlignment =
3059 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3060 unsigned MinAlignment;
3061 unsigned MaxAlignment;
3062 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3063 // Choose the minimum alignment. If we could prove both stores execute, we
3064 // could use biggest one. In this case, though, we only know that one of the
3065 // stores executes. And we don't know it's safe to take the alignment from a
3066 // store that doesn't execute.
3067 if (MinAlignment != 0) {
3068 // Choose the minimum of all non-zero alignments.
3069 SI->setAlignment(MinAlignment);
3070 } else if (MaxAlignment != 0) {
3071 // Choose the minimal alignment between the non-zero alignment and the ABI
3072 // default alignment for the type of the stored value.
3073 SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3074 } else {
3075 // If both alignments are zero, use ABI default alignment for the type of
3076 // the stored value.
3077 SI->setAlignment(TypeAlignment);
3078 }
3079
3080 QStore->eraseFromParent();
3081 PStore->eraseFromParent();
3082
3083 return true;
3084}
3085
3086static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3087 const DataLayout &DL) {
3088 // The intention here is to find diamonds or triangles (see below) where each
3089 // conditional block contains a store to the same address. Both of these
3090 // stores are conditional, so they can't be unconditionally sunk. But it may
3091 // be profitable to speculatively sink the stores into one merged store at the
3092 // end, and predicate the merged store on the union of the two conditions of
3093 // PBI and QBI.
3094 //
3095 // This can reduce the number of stores executed if both of the conditions are
3096 // true, and can allow the blocks to become small enough to be if-converted.
3097 // This optimization will also chain, so that ladders of test-and-set
3098 // sequences can be if-converted away.
3099 //
3100 // We only deal with simple diamonds or triangles:
3101 //
3102 // PBI or PBI or a combination of the two
3103 // / \ | \
3104 // PTB PFB | PFB
3105 // \ / | /
3106 // QBI QBI
3107 // / \ | \
3108 // QTB QFB | QFB
3109 // \ / | /
3110 // PostBB PostBB
3111 //
3112 // We model triangles as a type of diamond with a nullptr "true" block.
3113 // Triangles are canonicalized so that the fallthrough edge is represented by
3114 // a true condition, as in the diagram above.
3115 BasicBlock *PTB = PBI->getSuccessor(0);
3116 BasicBlock *PFB = PBI->getSuccessor(1);
3117 BasicBlock *QTB = QBI->getSuccessor(0);
3118 BasicBlock *QFB = QBI->getSuccessor(1);
3119 BasicBlock *PostBB = QFB->getSingleSuccessor();
3120
3121 // Make sure we have a good guess for PostBB. If QTB's only successor is
3122 // QFB, then QFB is a better PostBB.
3123 if (QTB->getSingleSuccessor() == QFB)
3124 PostBB = QFB;
3125
3126 // If we couldn't find a good PostBB, stop.
3127 if (!PostBB)
3128 return false;
3129
3130 bool InvertPCond = false, InvertQCond = false;
3131 // Canonicalize fallthroughs to the true branches.
3132 if (PFB == QBI->getParent()) {
3133 std::swap(PFB, PTB);
3134 InvertPCond = true;
3135 }
3136 if (QFB == PostBB) {
3137 std::swap(QFB, QTB);
3138 InvertQCond = true;
3139 }
3140
3141 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3142 // and QFB may not. Model fallthroughs as a nullptr block.
3143 if (PTB == QBI->getParent())
3144 PTB = nullptr;
3145 if (QTB == PostBB)
3146 QTB = nullptr;
3147
3148 // Legality bailouts. We must have at least the non-fallthrough blocks and
3149 // the post-dominating block, and the non-fallthroughs must only have one
3150 // predecessor.
3151 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3152 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3153 };
3154 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3155 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3156 return false;
3157 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3158 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3159 return false;
3160 if (!QBI->getParent()->hasNUses(2))
3161 return false;
3162
3163 // OK, this is a sequence of two diamonds or triangles.
3164 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3165 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3166 for (auto *BB : {PTB, PFB}) {
3167 if (!BB)
3168 continue;
3169 for (auto &I : *BB)
3170 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3171 PStoreAddresses.insert(SI->getPointerOperand());
3172 }
3173 for (auto *BB : {QTB, QFB}) {
3174 if (!BB)
3175 continue;
3176 for (auto &I : *BB)
3177 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3178 QStoreAddresses.insert(SI->getPointerOperand());
3179 }
3180
3181 set_intersect(PStoreAddresses, QStoreAddresses);
3182 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3183 // clear what it contains.
3184 auto &CommonAddresses = PStoreAddresses;
3185
3186 bool Changed = false;
3187 for (auto *Address : CommonAddresses)
3188 Changed |= mergeConditionalStoreToAddress(
3189 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3190 return Changed;
3191}
3192
3193/// If we have a conditional branch as a predecessor of another block,
3194/// this function tries to simplify it. We know
3195/// that PBI and BI are both conditional branches, and BI is in one of the
3196/// successor blocks of PBI - PBI branches to BI.
3197static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3198 const DataLayout &DL) {
3199 assert(PBI->isConditional() && BI->isConditional())((PBI->isConditional() && BI->isConditional()) ?
static_cast<void> (0) : __assert_fail ("PBI->isConditional() && BI->isConditional()"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3199, __PRETTY_FUNCTION__))
;
3200 BasicBlock *BB = BI->getParent();
3201
3202 // If this block ends with a branch instruction, and if there is a
3203 // predecessor that ends on a branch of the same condition, make
3204 // this conditional branch redundant.
3205 if (PBI->getCondition() == BI->getCondition() &&
3206 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3207 // Okay, the outcome of this conditional branch is statically
3208 // knowable. If this block had a single pred, handle specially.
3209 if (BB->getSinglePredecessor()) {
3210 // Turn this into a branch on constant.
3211 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3212 BI->setCondition(
3213 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3214 return true; // Nuke the branch on constant.
3215 }
3216
3217 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3218 // in the constant and simplify the block result. Subsequent passes of
3219 // simplifycfg will thread the block.
3220 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3221 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3222 PHINode *NewPN = PHINode::Create(
3223 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3224 BI->getCondition()->getName() + ".pr", &BB->front());
3225 // Okay, we're going to insert the PHI node. Since PBI is not the only
3226 // predecessor, compute the PHI'd conditional value for all of the preds.
3227 // Any predecessor where the condition is not computable we keep symbolic.
3228 for (pred_iterator PI = PB; PI != PE; ++PI) {
3229 BasicBlock *P = *PI;
3230 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3231 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3232 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3233 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3234 NewPN->addIncoming(
3235 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3236 P);
3237 } else {
3238 NewPN->addIncoming(BI->getCondition(), P);
3239 }
3240 }
3241
3242 BI->setCondition(NewPN);
3243 return true;
3244 }
3245 }
3246
3247 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3248 if (CE->canTrap())
3249 return false;
3250
3251 // If both branches are conditional and both contain stores to the same
3252 // address, remove the stores from the conditionals and create a conditional
3253 // merged store at the end.
3254 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3255 return true;
3256
3257 // If this is a conditional branch in an empty block, and if any
3258 // predecessors are a conditional branch to one of our destinations,
3259 // fold the conditions into logical ops and one cond br.
3260
3261 // Ignore dbg intrinsics.
3262 if (&*BB->instructionsWithoutDebug().begin() != BI)
3263 return false;
3264
3265 int PBIOp, BIOp;
3266 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3267 PBIOp = 0;
3268 BIOp = 0;
3269 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3270 PBIOp = 0;
3271 BIOp = 1;
3272 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3273 PBIOp = 1;
3274 BIOp = 0;
3275 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3276 PBIOp = 1;
3277 BIOp = 1;
3278 } else {
3279 return false;
3280 }
3281
3282 // Check to make sure that the other destination of this branch
3283 // isn't BB itself. If so, this is an infinite loop that will
3284 // keep getting unwound.
3285 if (PBI->getSuccessor(PBIOp) == BB)
3286 return false;
3287
3288 // Do not perform this transformation if it would require
3289 // insertion of a large number of select instructions. For targets
3290 // without predication/cmovs, this is a big pessimization.
3291
3292 // Also do not perform this transformation if any phi node in the common
3293 // destination block can trap when reached by BB or PBB (PR17073). In that
3294 // case, it would be unsafe to hoist the operation into a select instruction.
3295
3296 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3297 unsigned NumPhis = 0;
3298 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3299 ++II, ++NumPhis) {
3300 if (NumPhis > 2) // Disable this xform.
3301 return false;
3302
3303 PHINode *PN = cast<PHINode>(II);
3304 Value *BIV = PN->getIncomingValueForBlock(BB);
3305 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3306 if (CE->canTrap())
3307 return false;
3308
3309 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3310 Value *PBIV = PN->getIncomingValue(PBBIdx);
3311 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3312 if (CE->canTrap())
3313 return false;
3314 }
3315
3316 // Finally, if everything is ok, fold the branches to logical ops.
3317 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3318
3319 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
3320 << "AND: " << *BI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
;
3321
3322 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3323 // branch in it, where one edge (OtherDest) goes back to itself but the other
3324 // exits. We don't *know* that the program avoids the infinite loop
3325 // (even though that seems likely). If we do this xform naively, we'll end up
3326 // recursively unpeeling the loop. Since we know that (after the xform is
3327 // done) that the block *is* infinite if reached, we just make it an obviously
3328 // infinite loop with no cond branch.
3329 if (OtherDest == BB) {
3330 // Insert it at the end of the function, because it's either code,
3331 // or it won't matter if it's hot. :)
3332 BasicBlock *InfLoopBlock =
3333 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3334 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3335 OtherDest = InfLoopBlock;
3336 }
3337
3338 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3339
3340 // BI may have other predecessors. Because of this, we leave
3341 // it alone, but modify PBI.
3342
3343 // Make sure we get to CommonDest on True&True directions.
3344 Value *PBICond = PBI->getCondition();
3345 IRBuilder<NoFolder> Builder(PBI);
3346 if (PBIOp)
3347 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3348
3349 Value *BICond = BI->getCondition();
3350 if (BIOp)
3351 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3352
3353 // Merge the conditions.
3354 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3355
3356 // Modify PBI to branch on the new condition to the new dests.
3357 PBI->setCondition(Cond);
3358 PBI->setSuccessor(0, CommonDest);
3359 PBI->setSuccessor(1, OtherDest);
3360
3361 // Update branch weight for PBI.
3362 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3363 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3364 bool HasWeights =
3365 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3366 SuccTrueWeight, SuccFalseWeight);
3367 if (HasWeights) {
3368 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3369 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3370 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3371 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3372 // The weight to CommonDest should be PredCommon * SuccTotal +
3373 // PredOther * SuccCommon.
3374 // The weight to OtherDest should be PredOther * SuccOther.
3375 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3376 PredOther * SuccCommon,
3377 PredOther * SuccOther};
3378 // Halve the weights if any of them cannot fit in an uint32_t
3379 FitWeights(NewWeights);
3380
3381 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3382 }
3383
3384 // OtherDest may have phi nodes. If so, add an entry from PBI's
3385 // block that are identical to the entries for BI's block.
3386 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3387
3388 // We know that the CommonDest already had an edge from PBI to
3389 // it. If it has PHIs though, the PHIs may have different
3390 // entries for BB and PBI's BB. If so, insert a select to make
3391 // them agree.
3392 for (PHINode &PN : CommonDest->phis()) {
3393 Value *BIV = PN.getIncomingValueForBlock(BB);
3394 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3395 Value *PBIV = PN.getIncomingValue(PBBIdx);
3396 if (BIV != PBIV) {
3397 // Insert a select in PBI to pick the right value.
3398 SelectInst *NV = cast<SelectInst>(
3399 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3400 PN.setIncomingValue(PBBIdx, NV);
3401 // Although the select has the same condition as PBI, the original branch
3402 // weights for PBI do not apply to the new select because the select's
3403 // 'logical' edges are incoming edges of the phi that is eliminated, not
3404 // the outgoing edges of PBI.
3405 if (HasWeights) {
3406 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3407 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3408 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3409 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3410 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3411 // The weight to PredOtherDest should be PredOther * SuccCommon.
3412 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3413 PredOther * SuccCommon};
3414
3415 FitWeights(NewWeights);
3416
3417 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3418 }
3419 }
3420 }
3421
3422 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "INTO: " << *PBI->
getParent(); } } while (false)
;
3423 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3424
3425 // This basic block is probably dead. We know it has at least
3426 // one fewer predecessor.
3427 return true;
3428}
3429
3430// Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3431// true or to FalseBB if Cond is false.
3432// Takes care of updating the successors and removing the old terminator.
3433// Also makes sure not to introduce new successors by assuming that edges to
3434// non-successor TrueBBs and FalseBBs aren't reachable.
3435static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3436 BasicBlock *TrueBB, BasicBlock *FalseBB,
3437 uint32_t TrueWeight,
3438 uint32_t FalseWeight) {
3439 // Remove any superfluous successor edges from the CFG.
3440 // First, figure out which successors to preserve.
3441 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3442 // successor.
3443 BasicBlock *KeepEdge1 = TrueBB;
3444 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3445
3446 // Then remove the rest.
3447 for (BasicBlock *Succ : successors(OldTerm)) {
3448 // Make sure only to keep exactly one copy of each edge.
3449 if (Succ == KeepEdge1)
3450 KeepEdge1 = nullptr;
3451 else if (Succ == KeepEdge2)
3452 KeepEdge2 = nullptr;
3453 else
3454 Succ->removePredecessor(OldTerm->getParent(),
3455 /*KeepOneInputPHIs=*/true);
3456 }
3457
3458 IRBuilder<> Builder(OldTerm);
3459 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3460
3461 // Insert an appropriate new terminator.
3462 if (!KeepEdge1 && !KeepEdge2) {
3463 if (TrueBB == FalseBB)
3464 // We were only looking for one successor, and it was present.
3465 // Create an unconditional branch to it.
3466 Builder.CreateBr(TrueBB);
3467 else {
3468 // We found both of the successors we were looking for.
3469 // Create a conditional branch sharing the condition of the select.
3470 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3471 if (TrueWeight != FalseWeight)
3472 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3473 }
3474 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3475 // Neither of the selected blocks were successors, so this
3476 // terminator must be unreachable.
3477 new UnreachableInst(OldTerm->getContext(), OldTerm);
3478 } else {
3479 // One of the selected values was a successor, but the other wasn't.
3480 // Insert an unconditional branch to the one that was found;
3481 // the edge to the one that wasn't must be unreachable.
3482 if (!KeepEdge1)
3483 // Only TrueBB was found.
3484 Builder.CreateBr(TrueBB);
3485 else
3486 // Only FalseBB was found.
3487 Builder.CreateBr(FalseBB);
3488 }
3489
3490 EraseTerminatorAndDCECond(OldTerm);
3491 return true;
3492}
3493
3494// Replaces
3495// (switch (select cond, X, Y)) on constant X, Y
3496// with a branch - conditional if X and Y lead to distinct BBs,
3497// unconditional otherwise.
3498static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3499 // Check for constant integer values in the select.
3500 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3501 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3502 if (!TrueVal || !FalseVal)
3503 return false;
3504
3505 // Find the relevant condition and destinations.
3506 Value *Condition = Select->getCondition();
3507 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3508 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3509
3510 // Get weight for TrueBB and FalseBB.
3511 uint32_t TrueWeight = 0, FalseWeight = 0;
3512 SmallVector<uint64_t, 8> Weights;
3513 bool HasWeights = HasBranchWeights(SI);
3514 if (HasWeights) {
3515 GetBranchWeights(SI, Weights);
3516 if (Weights.size() == 1 + SI->getNumCases()) {
3517 TrueWeight =
3518 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3519 FalseWeight =
3520 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3521 }
3522 }
3523
3524 // Perform the actual simplification.
3525 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3526 FalseWeight);
3527}
3528
3529// Replaces
3530// (indirectbr (select cond, blockaddress(@fn, BlockA),
3531// blockaddress(@fn, BlockB)))
3532// with
3533// (br cond, BlockA, BlockB).
3534static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3535 // Check that both operands of the select are block addresses.
3536 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3537 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3538 if (!TBA || !FBA)
3539 return false;
3540
3541 // Extract the actual blocks.
3542 BasicBlock *TrueBB = TBA->getBasicBlock();
3543 BasicBlock *FalseBB = FBA->getBasicBlock();
3544
3545 // Perform the actual simplification.
3546 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3547 0);
3548}
3549
3550/// This is called when we find an icmp instruction
3551/// (a seteq/setne with a constant) as the only instruction in a
3552/// block that ends with an uncond branch. We are looking for a very specific
3553/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3554/// this case, we merge the first two "or's of icmp" into a switch, but then the
3555/// default value goes to an uncond block with a seteq in it, we get something
3556/// like:
3557///
3558/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3559/// DEFAULT:
3560/// %tmp = icmp eq i8 %A, 92
3561/// br label %end
3562/// end:
3563/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3564///
3565/// We prefer to split the edge to 'end' so that there is a true/false entry to
3566/// the PHI, merging the third icmp into the switch.
3567bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3568 ICmpInst *ICI, IRBuilder<> &Builder) {
3569 BasicBlock *BB = ICI->getParent();
3570
3571 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3572 // complex.
3573 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3574 return false;
3575
3576 Value *V = ICI->getOperand(0);
3577 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3578
3579 // The pattern we're looking for is where our only predecessor is a switch on
3580 // 'V' and this block is the default case for the switch. In this case we can
3581 // fold the compared value into the switch to simplify things.
3582 BasicBlock *Pred = BB->getSinglePredecessor();
3583 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3584 return false;
3585
3586 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3587 if (SI->getCondition() != V)
3588 return false;
3589
3590 // If BB is reachable on a non-default case, then we simply know the value of
3591 // V in this block. Substitute it and constant fold the icmp instruction
3592 // away.
3593 if (SI->getDefaultDest() != BB) {
3594 ConstantInt *VVal = SI->findCaseDest(BB);
3595 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3595, __PRETTY_FUNCTION__))
;
3596 ICI->setOperand(0, VVal);
3597
3598 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3599 ICI->replaceAllUsesWith(V);
3600 ICI->eraseFromParent();
3601 }
3602 // BB is now empty, so it is likely to simplify away.
3603 return requestResimplify();
3604 }
3605
3606 // Ok, the block is reachable from the default dest. If the constant we're
3607 // comparing exists in one of the other edges, then we can constant fold ICI
3608 // and zap it.
3609 if (SI->findCaseValue(Cst) != SI->case_default()) {
3610 Value *V;
3611 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3612 V = ConstantInt::getFalse(BB->getContext());
3613 else
3614 V = ConstantInt::getTrue(BB->getContext());
3615
3616 ICI->replaceAllUsesWith(V);
3617 ICI->eraseFromParent();
3618 // BB is now empty, so it is likely to simplify away.
3619 return requestResimplify();
3620 }
3621
3622 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3623 // the block.
3624 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3625 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3626 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3627 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3628 return false;
3629
3630 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3631 // true in the PHI.
3632 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3633 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3634
3635 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3636 std::swap(DefaultCst, NewCst);
3637
3638 // Replace ICI (which is used by the PHI for the default value) with true or
3639 // false depending on if it is EQ or NE.
3640 ICI->replaceAllUsesWith(DefaultCst);
3641 ICI->eraseFromParent();
3642
3643 // Okay, the switch goes to this block on a default value. Add an edge from
3644 // the switch to the merge point on the compared value.
3645 BasicBlock *NewBB =
3646 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3647 SmallVector<uint64_t, 8> Weights;
3648 bool HasWeights = HasBranchWeights(SI);
3649 if (HasWeights) {
3650 GetBranchWeights(SI, Weights);
3651 if (Weights.size() == 1 + SI->getNumCases()) {
3652 // Split weight for default case to case for "Cst".
3653 Weights[0] = (Weights[0] + 1) >> 1;
3654 Weights.push_back(Weights[0]);
3655
3656 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3657 setBranchWeights(SI, MDWeights);
3658 }
3659 }
3660 SI->addCase(Cst, NewBB);
3661
3662 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3663 Builder.SetInsertPoint(NewBB);
3664 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3665 Builder.CreateBr(SuccBlock);
3666 PHIUse->addIncoming(NewCst, NewBB);
3667 return true;
3668}
3669
3670/// The specified branch is a conditional branch.
3671/// Check to see if it is branching on an or/and chain of icmp instructions, and
3672/// fold it into a switch instruction if so.
3673static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3674 const DataLayout &DL) {
3675 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3676 if (!Cond)
3677 return false;
3678
3679 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3680 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3681 // 'setne's and'ed together, collect them.
3682
3683 // Try to gather values from a chain of and/or to be turned into a switch
3684 ConstantComparesGatherer ConstantCompare(Cond, DL);
3685 // Unpack the result
3686 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3687 Value *CompVal = ConstantCompare.CompValue;
3688 unsigned UsedICmps = ConstantCompare.UsedICmps;
3689 Value *ExtraCase = ConstantCompare.Extra;
3690
3691 // If we didn't have a multiply compared value, fail.
3692 if (!CompVal)
3693 return false;
3694
3695 // Avoid turning single icmps into a switch.
3696 if (UsedICmps <= 1)
3697 return false;
3698
3699 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3700
3701 // There might be duplicate constants in the list, which the switch
3702 // instruction can't handle, remove them now.
3703 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3704 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3705
3706 // If Extra was used, we require at least two switch values to do the
3707 // transformation. A switch with one value is just a conditional branch.
3708 if (ExtraCase && Values.size() < 2)
3709 return false;
3710
3711 // TODO: Preserve branch weight metadata, similarly to how
3712 // FoldValueComparisonIntoPredecessors preserves it.
3713
3714 // Figure out which block is which destination.
3715 BasicBlock *DefaultBB = BI->getSuccessor(1);
3716 BasicBlock *EdgeBB = BI->getSuccessor(0);
3717 if (!TrueWhenEqual)
3718 std::swap(DefaultBB, EdgeBB);
3719
3720 BasicBlock *BB = BI->getParent();
3721
3722 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)
3723 << " 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)
3724 << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
;
3725
3726 // If there are any extra values that couldn't be folded into the switch
3727 // then we evaluate them with an explicit branch first. Split the block
3728 // right before the condbr to handle it.
3729 if (ExtraCase) {
3730 BasicBlock *NewBB =
3731 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3732 // Remove the uncond branch added to the old block.
3733 Instruction *OldTI = BB->getTerminator();
3734 Builder.SetInsertPoint(OldTI);
3735
3736 if (TrueWhenEqual)
3737 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3738 else
3739 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3740
3741 OldTI->eraseFromParent();
3742
3743 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3744 // for the edge we just added.
3745 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3746
3747 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCasedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
3748 << "\nEXTRABB = " << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
;
3749 BB = NewBB;
3750 }
3751
3752 Builder.SetInsertPoint(BI);
3753 // Convert pointer to int before we switch.
3754 if (CompVal->getType()->isPointerTy()) {
3755 CompVal = Builder.CreatePtrToInt(
3756 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3757 }
3758
3759 // Create the new switch instruction now.
3760 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3761
3762 // Add all of the 'cases' to the switch instruction.
3763 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3764 New->addCase(Values[i], EdgeBB);
3765
3766 // We added edges from PI to the EdgeBB. As such, if there were any
3767 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3768 // the number of edges added.
3769 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3770 PHINode *PN = cast<PHINode>(BBI);
3771 Value *InVal = PN->getIncomingValueForBlock(BB);
3772 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3773 PN->addIncoming(InVal, BB);
3774 }
3775
3776 // Erase the old branch instruction.
3777 EraseTerminatorAndDCECond(BI);
3778
3779 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)
;
3780 return true;
3781}
3782
3783bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3784 if (isa<PHINode>(RI->getValue()))
3785 return SimplifyCommonResume(RI);
3786 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3787 RI->getValue() == RI->getParent()->getFirstNonPHI())
3788 // The resume must unwind the exception that caused control to branch here.
3789 return SimplifySingleResume(RI);
3790
3791 return false;
3792}
3793
3794// Simplify resume that is shared by several landing pads (phi of landing pad).
3795bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3796 BasicBlock *BB = RI->getParent();
3797
3798 // Check that there are no other instructions except for debug intrinsics
3799 // between the phi of landing pads (RI->getValue()) and resume instruction.
3800 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3801 E = RI->getIterator();
3802 while (++I != E)
3803 if (!isa<DbgInfoIntrinsic>(I))
3804 return false;
3805
3806 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3807 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3808
3809 // Check incoming blocks to see if any of them are trivial.
3810 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3811 Idx++) {
3812 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3813 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3814
3815 // If the block has other successors, we can not delete it because
3816 // it has other dependents.
3817 if (IncomingBB->getUniqueSuccessor() != BB)
3818 continue;
3819
3820 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3821 // Not the landing pad that caused the control to branch here.
3822 if (IncomingValue != LandingPad)
3823 continue;
3824
3825 bool isTrivial = true;
3826
3827 I = IncomingBB->getFirstNonPHI()->getIterator();
3828 E = IncomingBB->getTerminator()->getIterator();
3829 while (++I != E)
3830 if (!isa<DbgInfoIntrinsic>(I)) {
3831 isTrivial = false;
3832 break;
3833 }
3834
3835 if (isTrivial)
3836 TrivialUnwindBlocks.insert(IncomingBB);
3837 }
3838
3839 // If no trivial unwind blocks, don't do any simplifications.
3840 if (TrivialUnwindBlocks.empty())
3841 return false;
3842
3843 // Turn all invokes that unwind here into calls.
3844 for (auto *TrivialBB : TrivialUnwindBlocks) {
3845 // Blocks that will be simplified should be removed from the phi node.
3846 // Note there could be multiple edges to the resume block, and we need
3847 // to remove them all.
3848 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3849 BB->removePredecessor(TrivialBB, true);
3850
3851 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3852 PI != PE;) {
3853 BasicBlock *Pred = *PI++;
3854 removeUnwindEdge(Pred);
3855 }
3856
3857 // In each SimplifyCFG run, only the current processed block can be erased.
3858 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3859 // of erasing TrivialBB, we only remove the branch to the common resume
3860 // block so that we can later erase the resume block since it has no
3861 // predecessors.
3862 TrivialBB->getTerminator()->eraseFromParent();
3863 new UnreachableInst(RI->getContext(), TrivialBB);
3864 }
3865
3866 // Delete the resume block if all its predecessors have been removed.
3867 if (pred_empty(BB))
3868 BB->eraseFromParent();
3869
3870 return !TrivialUnwindBlocks.empty();
3871}
3872
3873// Simplify resume that is only used by a single (non-phi) landing pad.
3874bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3875 BasicBlock *BB = RI->getParent();
3876 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3877 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3878, __PRETTY_FUNCTION__))
3878 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3878, __PRETTY_FUNCTION__))
;
3879
3880 // Check that there are no other instructions except for debug intrinsics.
3881 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3882 while (++I != E)
3883 if (!isa<DbgInfoIntrinsic>(I))
3884 return false;
3885
3886 // Turn all invokes that unwind here into calls and delete the basic block.
3887 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3888 BasicBlock *Pred = *PI++;
3889 removeUnwindEdge(Pred);
3890 }
3891
3892 // The landingpad is now unreachable. Zap it.
3893 if (LoopHeaders)
3894 LoopHeaders->erase(BB);
3895 BB->eraseFromParent();
3896 return true;
3897}
3898
3899static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3900 // If this is a trivial cleanup pad that executes no instructions, it can be
3901 // eliminated. If the cleanup pad continues to the caller, any predecessor
3902 // that is an EH pad will be updated to continue to the caller and any
3903 // predecessor that terminates with an invoke instruction will have its invoke
3904 // instruction converted to a call instruction. If the cleanup pad being
3905 // simplified does not continue to the caller, each predecessor will be
3906 // updated to continue to the unwind destination of the cleanup pad being
3907 // simplified.
3908 BasicBlock *BB = RI->getParent();
3909 CleanupPadInst *CPInst = RI->getCleanupPad();
3910 if (CPInst->getParent() != BB)
3911 // This isn't an empty cleanup.
3912 return false;
3913
3914 // We cannot kill the pad if it has multiple uses. This typically arises
3915 // from unreachable basic blocks.
3916 if (!CPInst->hasOneUse())
3917 return false;
3918
3919 // Check that there are no other instructions except for benign intrinsics.
3920 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3921 while (++I != E) {
3922 auto *II = dyn_cast<IntrinsicInst>(I);
3923 if (!II)
3924 return false;
3925
3926 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3927 switch (IntrinsicID) {
3928 case Intrinsic::dbg_declare:
3929 case Intrinsic::dbg_value:
3930 case Intrinsic::dbg_label:
3931 case Intrinsic::lifetime_end:
3932 break;
3933 default:
3934 return false;
3935 }
3936 }
3937
3938 // If the cleanup return we are simplifying unwinds to the caller, this will
3939 // set UnwindDest to nullptr.
3940 BasicBlock *UnwindDest = RI->getUnwindDest();
3941 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3942
3943 // We're about to remove BB from the control flow. Before we do, sink any
3944 // PHINodes into the unwind destination. Doing this before changing the
3945 // control flow avoids some potentially slow checks, since we can currently
3946 // be certain that UnwindDest and BB have no common predecessors (since they
3947 // are both EH pads).
3948 if (UnwindDest) {
3949 // First, go through the PHI nodes in UnwindDest and update any nodes that
3950 // reference the block we are removing
3951 for (BasicBlock::iterator I = UnwindDest->begin(),
3952 IE = DestEHPad->getIterator();
3953 I != IE; ++I) {
3954 PHINode *DestPN = cast<PHINode>(I);
3955
3956 int Idx = DestPN->getBasicBlockIndex(BB);
3957 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3958 assert(Idx != -1)((Idx != -1) ? static_cast<void> (0) : __assert_fail ("Idx != -1"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3958, __PRETTY_FUNCTION__))
;
3959 // This PHI node has an incoming value that corresponds to a control
3960 // path through the cleanup pad we are removing. If the incoming
3961 // value is in the cleanup pad, it must be a PHINode (because we
3962 // verified above that the block is otherwise empty). Otherwise, the
3963 // value is either a constant or a value that dominates the cleanup
3964 // pad being removed.
3965 //
3966 // Because BB and UnwindDest are both EH pads, all of their
3967 // predecessors must unwind to these blocks, and since no instruction
3968 // can have multiple unwind destinations, there will be no overlap in
3969 // incoming blocks between SrcPN and DestPN.
3970 Value *SrcVal = DestPN->getIncomingValue(Idx);
3971 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3972
3973 // Remove the entry for the block we are deleting.
3974 DestPN->removeIncomingValue(Idx, false);
3975
3976 if (SrcPN && SrcPN->getParent() == BB) {
3977 // If the incoming value was a PHI node in the cleanup pad we are
3978 // removing, we need to merge that PHI node's incoming values into
3979 // DestPN.
3980 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3981 SrcIdx != SrcE; ++SrcIdx) {
3982 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3983 SrcPN->getIncomingBlock(SrcIdx));
3984 }
3985 } else {
3986 // Otherwise, the incoming value came from above BB and
3987 // so we can just reuse it. We must associate all of BB's
3988 // predecessors with this value.
3989 for (auto *pred : predecessors(BB)) {
3990 DestPN->addIncoming(SrcVal, pred);
3991 }
3992 }
3993 }
3994
3995 // Sink any remaining PHI nodes directly into UnwindDest.
3996 Instruction *InsertPt = DestEHPad;
3997 for (BasicBlock::iterator I = BB->begin(),
3998 IE = BB->getFirstNonPHI()->getIterator();
3999 I != IE;) {
4000 // The iterator must be incremented here because the instructions are
4001 // being moved to another block.
4002 PHINode *PN = cast<PHINode>(I++);
4003 if (PN->use_empty())
4004 // If the PHI node has no uses, just leave it. It will be erased
4005 // when we erase BB below.
4006 continue;
4007
4008 // Otherwise, sink this PHI node into UnwindDest.
4009 // Any predecessors to UnwindDest which are not already represented
4010 // must be back edges which inherit the value from the path through
4011 // BB. In this case, the PHI value must reference itself.
4012 for (auto *pred : predecessors(UnwindDest))
4013 if (pred != BB)
4014 PN->addIncoming(PN, pred);
4015 PN->moveBefore(InsertPt);
4016 }
4017 }
4018
4019 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4020 // The iterator must be updated here because we are removing this pred.
4021 BasicBlock *PredBB = *PI++;
4022 if (UnwindDest == nullptr) {
4023 removeUnwindEdge(PredBB);
4024 } else {
4025 Instruction *TI = PredBB->getTerminator();
4026 TI->replaceUsesOfWith(BB, UnwindDest);
4027 }
4028 }
4029
4030 // The cleanup pad is now unreachable. Zap it.
4031 BB->eraseFromParent();
4032 return true;
4033}
4034
4035// Try to merge two cleanuppads together.
4036static bool mergeCleanupPad(CleanupReturnInst *RI) {
4037 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4038 // with.
4039 BasicBlock *UnwindDest = RI->getUnwindDest();
4040 if (!UnwindDest)
4041 return false;
4042
4043 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4044 // be safe to merge without code duplication.
4045 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4046 return false;
4047
4048 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4049 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4050 if (!SuccessorCleanupPad)
4051 return false;
4052
4053 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4054 // Replace any uses of the successor cleanupad with the predecessor pad
4055 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4056 // funclet bundle operands.
4057 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4058 // Remove the old cleanuppad.
4059 SuccessorCleanupPad->eraseFromParent();
4060 // Now, we simply replace the cleanupret with a branch to the unwind
4061 // destination.
4062 BranchInst::Create(UnwindDest, RI->getParent());
4063 RI->eraseFromParent();
4064
4065 return true;
4066}
4067
4068bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4069 // It is possible to transiantly have an undef cleanuppad operand because we
4070 // have deleted some, but not all, dead blocks.
4071 // Eventually, this block will be deleted.
4072 if (isa<UndefValue>(RI->getOperand(0)))
4073 return false;
4074
4075 if (mergeCleanupPad(RI))
4076 return true;
4077
4078 if (removeEmptyCleanup(RI))
4079 return true;
4080
4081 return false;
4082}
4083
4084bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4085 BasicBlock *BB = RI->getParent();
4086 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4087 return false;
4088
4089 // Find predecessors that end with branches.
4090 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4091 SmallVector<BranchInst *, 8> CondBranchPreds;
4092 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4093 BasicBlock *P = *PI;
4094 Instruction *PTI = P->getTerminator();
4095 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4096 if (BI->isUnconditional())
4097 UncondBranchPreds.push_back(P);
4098 else
4099 CondBranchPreds.push_back(BI);
4100 }
4101 }
4102
4103 // If we found some, do the transformation!
4104 if (!UncondBranchPreds.empty() && DupRet) {
4105 while (!UncondBranchPreds.empty()) {
4106 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4107 LLVM_DEBUG(dbgs() << "FOLDING: " << *BBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
4108 << "INTO UNCOND BRANCH PRED: " << *Pred)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
;
4109 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4110 }
4111
4112 // If we eliminated all predecessors of the block, delete the block now.
4113 if (pred_empty(BB)) {
4114 // We know there are no successors, so just nuke the block.
4115 if (LoopHeaders)
4116 LoopHeaders->erase(BB);
4117 BB->eraseFromParent();
4118 }
4119
4120 return true;
4121 }
4122
4123 // Check out all of the conditional branches going to this return
4124 // instruction. If any of them just select between returns, change the
4125 // branch itself into a select/return pair.
4126 while (!CondBranchPreds.empty()) {
4127 BranchInst *BI = CondBranchPreds.pop_back_val();
4128
4129 // Check to see if the non-BB successor is also a return block.
4130 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4131 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4132 SimplifyCondBranchToTwoReturns(BI, Builder))
4133 return true;
4134 }
4135 return false;
4136}
4137
4138bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4139 BasicBlock *BB = UI->getParent();
4140
4141 bool Changed = false;
4142
4143 // If there are any instructions immediately before the unreachable that can
4144 // be removed, do so.
4145 while (UI->getIterator() != BB->begin()) {
4146 BasicBlock::iterator BBI = UI->getIterator();
4147 --BBI;
4148 // Do not delete instructions that can have side effects which might cause
4149 // the unreachable to not be reachable; specifically, calls and volatile
4150 // operations may have this effect.
4151 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4152 break;
4153
4154 if (BBI->mayHaveSideEffects()) {
4155 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4156 if (SI->isVolatile())
4157 break;
4158 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4159 if (LI->isVolatile())
4160 break;
4161 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4162 if (RMWI->isVolatile())
4163 break;
4164 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4165 if (CXI->isVolatile())
4166 break;
4167 } else if (isa<CatchPadInst>(BBI)) {
4168 // A catchpad may invoke exception object constructors and such, which
4169 // in some languages can be arbitrary code, so be conservative by
4170 // default.
4171 // For CoreCLR, it just involves a type test, so can be removed.
4172 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4173 EHPersonality::CoreCLR)
4174 break;
4175 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4176 !isa<LandingPadInst>(BBI)) {
4177 break;
4178 }
4179 // Note that deleting LandingPad's here is in fact okay, although it
4180 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4181 // all the predecessors of this block will be the unwind edges of Invokes,
4182 // and we can therefore guarantee this block will be erased.
4183 }
4184
4185 // Delete this instruction (any uses are guaranteed to be dead)
4186 if (!BBI->use_empty())
4187 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4188 BBI->eraseFromParent();
4189 Changed = true;
4190 }
4191
4192 // If the unreachable instruction is the first in the block, take a gander
4193 // at all of the predecessors of this instruction, and simplify them.
4194 if (&BB->front() != UI)
4195 return Changed;
4196
4197 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4198 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4199 Instruction *TI = Preds[i]->getTerminator();
4200 IRBuilder<> Builder(TI);
4201 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4202 if (BI->isUnconditional()) {
4203 if (BI->getSuccessor(0) == BB) {
4204 new UnreachableInst(TI->getContext(), TI);
4205 TI->eraseFromParent();
4206 Changed = true;
4207 }
4208 } else {
4209 Value* Cond = BI->getCondition();
4210 if (BI->getSuccessor(0) == BB) {
4211 Builder.CreateAssumption(Builder.CreateNot(Cond));
4212 Builder.CreateBr(BI->getSuccessor(1));
4213 EraseTerminatorAndDCECond(BI);
4214 } else if (BI->getSuccessor(1) == BB) {
4215 Builder.CreateAssumption(Cond);
4216 Builder.CreateBr(BI->getSuccessor(0));
4217 EraseTerminatorAndDCECond(BI);
4218 Changed = true;
4219 }
4220 }
4221 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4222 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4223 if (i->getCaseSuccessor() != BB) {
4224 ++i;
4225 continue;
4226 }
4227 BB->removePredecessor(SI->getParent());
4228 i = SI->removeCase(i);
4229 e = SI->case_end();
4230 Changed = true;
4231 }
4232 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4233 if (II->getUnwindDest() == BB) {
4234 removeUnwindEdge(TI->getParent());
4235 Changed = true;
4236 }
4237 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4238 if (CSI->getUnwindDest() == BB) {
4239 removeUnwindEdge(TI->getParent());
4240 Changed = true;
4241 continue;
4242 }
4243
4244 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4245 E = CSI->handler_end();
4246 I != E; ++I) {
4247 if (*I == BB) {
4248 CSI->removeHandler(I);
4249 --I;
4250 --E;
4251 Changed = true;
4252 }
4253 }
4254 if (CSI->getNumHandlers() == 0) {
4255 BasicBlock *CatchSwitchBB = CSI->getParent();
4256 if (CSI->hasUnwindDest()) {
4257 // Redirect preds to the unwind dest
4258 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4259 } else {
4260 // Rewrite all preds to unwind to caller (or from invoke to call).
4261 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4262 for (BasicBlock *EHPred : EHPreds)
4263 removeUnwindEdge(EHPred);
4264 }
4265 // The catchswitch is no longer reachable.
4266 new UnreachableInst(CSI->getContext(), CSI);
4267 CSI->eraseFromParent();
4268 Changed = true;
4269 }
4270 } else if (isa<CleanupReturnInst>(TI)) {
4271 new UnreachableInst(TI->getContext(), TI);
4272 TI->eraseFromParent();
4273 Changed = true;
4274 }
4275 }
4276
4277 // If this block is now dead, remove it.
4278 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4279 // We know there are no successors, so just nuke the block.
4280 if (LoopHeaders)
4281 LoopHeaders->erase(BB);
4282 BB->eraseFromParent();
4283 return true;
4284 }
4285
4286 return Changed;
4287}
4288
4289static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4290 assert(Cases.size() >= 1)((Cases.size() >= 1) ? static_cast<void> (0) : __assert_fail
("Cases.size() >= 1", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4290, __PRETTY_FUNCTION__))
;
4291
4292 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4293 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4294 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4295 return false;
4296 }
4297 return true;
4298}
4299
4300/// Turn a switch with two reachable destinations into an integer range
4301/// comparison and branch.
4302static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4303 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4303, __PRETTY_FUNCTION__))
;
4304
4305 bool HasDefault =
4306 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4307
4308 // Partition the cases into two sets with different destinations.
4309 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4310 BasicBlock *DestB = nullptr;
4311 SmallVector<ConstantInt *, 16> CasesA;
4312 SmallVector<ConstantInt *, 16> CasesB;
4313
4314 for (auto Case : SI->cases()) {
4315 BasicBlock *Dest = Case.getCaseSuccessor();
4316 if (!DestA)
4317 DestA = Dest;
4318 if (Dest == DestA) {
4319 CasesA.push_back(Case.getCaseValue());
4320 continue;
4321 }
4322 if (!DestB)
4323 DestB = Dest;
4324 if (Dest == DestB) {
4325 CasesB.push_back(Case.getCaseValue());
4326 continue;
4327 }
4328 return false; // More than two destinations.
4329 }
4330
4331 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4332, __PRETTY_FUNCTION__))
4332 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4332, __PRETTY_FUNCTION__))
;
4333 assert(DestA != DestB)((DestA != DestB) ? static_cast<void> (0) : __assert_fail
("DestA != DestB", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4333, __PRETTY_FUNCTION__))
;
4334 assert(DestB != SI->getDefaultDest())((DestB != SI->getDefaultDest()) ? static_cast<void>
(0) : __assert_fail ("DestB != SI->getDefaultDest()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4334, __PRETTY_FUNCTION__))
;
4335 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4335, __PRETTY_FUNCTION__))
;
4336 assert(!CasesA.empty() || HasDefault)((!CasesA.empty() || HasDefault) ? static_cast<void> (0
) : __assert_fail ("!CasesA.empty() || HasDefault", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4336, __PRETTY_FUNCTION__))
;
4337
4338 // Figure out if one of the sets of cases form a contiguous range.
4339 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4340 BasicBlock *ContiguousDest = nullptr;
4341 BasicBlock *OtherDest = nullptr;
4342 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4343 ContiguousCases = &CasesA;
4344 ContiguousDest = DestA;
4345 OtherDest = DestB;
4346 } else if (CasesAreContiguous(CasesB)) {
4347 ContiguousCases = &CasesB;
4348 ContiguousDest = DestB;
4349 OtherDest = DestA;
4350 } else
4351 return false;
4352
4353 // Start building the compare and branch.
4354
4355 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4356 Constant *NumCases =
4357 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4358
4359 Value *Sub = SI->getCondition();
4360 if (!Offset->isNullValue())
4361 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4362
4363 Value *Cmp;
4364 // If NumCases overflowed, then all possible values jump to the successor.
4365 if (NumCases->isNullValue() && !ContiguousCases->empty())
4366 Cmp = ConstantInt::getTrue(SI->getContext());
4367 else
4368 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4369 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4370
4371 // Update weight for the newly-created conditional branch.
4372 if (HasBranchWeights(SI)) {
4373 SmallVector<uint64_t, 8> Weights;
4374 GetBranchWeights(SI, Weights);
4375 if (Weights.size() == 1 + SI->getNumCases()) {
4376 uint64_t TrueWeight = 0;
4377 uint64_t FalseWeight = 0;
4378 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4379 if (SI->getSuccessor(I) == ContiguousDest)
4380 TrueWeight += Weights[I];
4381 else
4382 FalseWeight += Weights[I];
4383 }
4384 while (TrueWeight > UINT32_MAX(4294967295U) || FalseWeight > UINT32_MAX(4294967295U)) {
4385 TrueWeight /= 2;
4386 FalseWeight /= 2;
4387 }
4388 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4389 }
4390 }
4391
4392 // Prune obsolete incoming values off the successors' PHI nodes.
4393 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4394 unsigned PreviousEdges = ContiguousCases->size();
4395 if (ContiguousDest == SI->getDefaultDest())
4396 ++PreviousEdges;
4397 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4398 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4399 }
4400 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4401 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4402 if (OtherDest == SI->getDefaultDest())
4403 ++PreviousEdges;
4404 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4405 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4406 }
4407
4408 // Drop the switch.
4409 SI->eraseFromParent();
4410
4411 return true;
4412}
4413
4414/// Compute masked bits for the condition of a switch
4415/// and use it to remove dead cases.
4416static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4417 const DataLayout &DL) {
4418 Value *Cond = SI->getCondition();
4419 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4420 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4421
4422 // We can also eliminate cases by determining that their values are outside of
4423 // the limited range of the condition based on how many significant (non-sign)
4424 // bits are in the condition value.
4425 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4426 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4427
4428 // Gather dead cases.
4429 SmallVector<ConstantInt *, 8> DeadCases;
4430 for (auto &Case : SI->cases()) {
4431 const APInt &CaseVal = Case.getCaseValue()->getValue();
4432 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4433 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4434 DeadCases.push_back(Case.getCaseValue());
4435 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
4436 << " is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
;
4437 }
4438 }
4439
4440 // If we can prove that the cases must cover all possible values, the
4441 // default destination becomes dead and we can remove it. If we know some
4442 // of the bits in the value, we can use that to more precisely compute the
4443 // number of possible unique case values.
4444 bool HasDefault =
4445 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4446 const unsigned NumUnknownBits =
4447 Bits - (Known.Zero | Known.One).countPopulation();
4448 assert(NumUnknownBits <= Bits)((NumUnknownBits <= Bits) ? static_cast<void> (0) : __assert_fail
("NumUnknownBits <= Bits", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4448, __PRETTY_FUNCTION__))
;
4449 if (HasDefault && DeadCases.empty() &&
4450 NumUnknownBits < 64 /* avoid overflow */ &&
4451 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4452 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)
;
4453 BasicBlock *NewDefault =
4454 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4455 SI->setDefaultDest(&*NewDefault);
4456 SplitBlock(&*NewDefault, &NewDefault->front());
4457 auto *OldTI = NewDefault->getTerminator();
4458 new UnreachableInst(SI->getContext(), OldTI);
4459 EraseTerminatorAndDCECond(OldTI);
4460 return true;
4461 }
4462
4463 SmallVector<uint64_t, 8> Weights;
4464 bool HasWeight = HasBranchWeights(SI);
4465 if (HasWeight) {
4466 GetBranchWeights(SI, Weights);
4467 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4468 }
4469
4470 // Remove dead cases from the switch.
4471 for (ConstantInt *DeadCase : DeadCases) {
4472 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4473 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4474, __PRETTY_FUNCTION__))
4474 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4474, __PRETTY_FUNCTION__))
;
4475 if (HasWeight) {
4476 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4477 Weights.pop_back();
4478 }
4479
4480 // Prune unused values from PHI nodes.
4481 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4482 SI->removeCase(CaseI);
4483 }
4484 if (HasWeight && Weights.size() >= 2) {
4485 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4486 setBranchWeights(SI, MDWeights);
4487 }
4488
4489 return !DeadCases.empty();
4490}
4491
4492/// If BB would be eligible for simplification by
4493/// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4494/// by an unconditional branch), look at the phi node for BB in the successor
4495/// block and see if the incoming value is equal to CaseValue. If so, return
4496/// the phi node, and set PhiIndex to BB's index in the phi node.
4497static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4498 BasicBlock *BB, int *PhiIndex) {
4499 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4500 return nullptr; // BB must be empty to be a candidate for simplification.
4501 if (!BB->getSinglePredecessor())
4502 return nullptr; // BB must be dominated by the switch.
4503
4504 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4505 if (!Branch || !Branch->isUnconditional())
4506 return nullptr; // Terminator must be unconditional branch.
4507
4508 BasicBlock *Succ = Branch->getSuccessor(0);
4509
4510 for (PHINode &PHI : Succ->phis()) {
4511 int Idx = PHI.getBasicBlockIndex(BB);
4512 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4512, __PRETTY_FUNCTION__))
;
4513
4514 Value *InValue = PHI.getIncomingValue(Idx);
4515 if (InValue != CaseValue)
4516 continue;
4517
4518 *PhiIndex = Idx;
4519 return &PHI;
4520 }
4521
4522 return nullptr;
4523}
4524
4525/// Try to forward the condition of a switch instruction to a phi node
4526/// dominated by the switch, if that would mean that some of the destination
4527/// blocks of the switch can be folded away. Return true if a change is made.
4528static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4529 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4530
4531 ForwardingNodesMap ForwardingNodes;
4532 BasicBlock *SwitchBlock = SI->getParent();
4533 bool Changed = false;
4534 for (auto &Case : SI->cases()) {
4535 ConstantInt *CaseValue = Case.getCaseValue();
4536 BasicBlock *CaseDest = Case.getCaseSuccessor();
4537
4538 // Replace phi operands in successor blocks that are using the constant case
4539 // value rather than the switch condition variable:
4540 // switchbb:
4541 // switch i32 %x, label %default [
4542 // i32 17, label %succ
4543 // ...
4544 // succ:
4545 // %r = phi i32 ... [ 17, %switchbb ] ...
4546 // -->
4547 // %r = phi i32 ... [ %x, %switchbb ] ...
4548
4549 for (PHINode &Phi : CaseDest->phis()) {
4550 // This only works if there is exactly 1 incoming edge from the switch to
4551 // a phi. If there is >1, that means multiple cases of the switch map to 1
4552 // value in the phi, and that phi value is not the switch condition. Thus,
4553 // this transform would not make sense (the phi would be invalid because
4554 // a phi can't have different incoming values from the same block).
4555 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4556 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4557 count(Phi.blocks(), SwitchBlock) == 1) {
4558 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4559 Changed = true;
4560 }
4561 }
4562
4563 // Collect phi nodes that are indirectly using this switch's case constants.
4564 int PhiIdx;
4565 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4566 ForwardingNodes[Phi].push_back(PhiIdx);
4567 }
4568
4569 for (auto &ForwardingNode : ForwardingNodes) {
4570 PHINode *Phi = ForwardingNode.first;
4571 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4572 if (Indexes.size() < 2)
4573 continue;
4574
4575 for (int Index : Indexes)
4576 Phi->setIncomingValue(Index, SI->getCondition());
4577 Changed = true;
4578 }
4579
4580 return Changed;
4581}
4582
4583/// Return true if the backend will be able to handle
4584/// initializing an array of constants like C.
4585static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4586 if (C->isThreadDependent())
4587 return false;
4588 if (C->isDLLImportDependent())
4589 return false;
4590
4591 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4592 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4593 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4594 return false;
4595
4596 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4597 if (!CE->isGEPWithNoNotionalOverIndexing())
4598 return false;
4599 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4600 return false;
4601 }
4602
4603 if (!TTI.shouldBuildLookupTablesForConstant(C))
4604 return false;
4605
4606 return true;
4607}
4608
4609/// If V is a Constant, return it. Otherwise, try to look up
4610/// its constant value in ConstantPool, returning 0 if it's not there.
4611static Constant *
4612LookupConstant(Value *V,
4613 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4614 if (Constant *C = dyn_cast<Constant>(V))
4615 return C;
4616 return ConstantPool.lookup(V);
4617}
4618
4619/// Try to fold instruction I into a constant. This works for
4620/// simple instructions such as binary operations where both operands are
4621/// constant or can be replaced by constants from the ConstantPool. Returns the
4622/// resulting constant on success, 0 otherwise.
4623static Constant *
4624ConstantFold(Instruction *I, const DataLayout &DL,
4625 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4626 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4627 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4628 if (!A)
4629 return nullptr;
4630 if (A->isAllOnesValue())
4631 return LookupConstant(Select->getTrueValue(), ConstantPool);
4632 if (A->isNullValue())
4633 return LookupConstant(Select->getFalseValue(), ConstantPool);
4634 return nullptr;
4635 }
4636
4637 SmallVector<Constant *, 4> COps;
4638 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4639 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4640 COps.push_back(A);
4641 else
4642 return nullptr;
4643 }
4644
4645 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4646 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4647 COps[1], DL);
4648 }
4649
4650 return ConstantFoldInstOperands(I, COps, DL);
4651}
4652
4653/// Try to determine the resulting constant values in phi nodes
4654/// at the common destination basic block, *CommonDest, for one of the case
4655/// destionations CaseDest corresponding to value CaseVal (0 for the default
4656/// case), of a switch instruction SI.
4657static bool
4658GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4659 BasicBlock **CommonDest,
4660 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4661 const DataLayout &DL, const TargetTransformInfo &TTI) {
4662 // The block from which we enter the common destination.
4663 BasicBlock *Pred = SI->getParent();
4664
4665 // If CaseDest is empty except for some side-effect free instructions through
4666 // which we can constant-propagate the CaseVal, continue to its successor.
4667 SmallDenseMap<Value *, Constant *> ConstantPool;
4668 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4669 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4670 if (I.isTerminator()) {
4671 // If the terminator is a simple branch, continue to the next block.
4672 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4673 return false;
4674 Pred = CaseDest;
4675 CaseDest = I.getSuccessor(0);
4676 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4677 // Instruction is side-effect free and constant.
4678
4679 // If the instruction has uses outside this block or a phi node slot for
4680 // the block, it is not safe to bypass the instruction since it would then
4681 // no longer dominate all its uses.
4682 for (auto &Use : I.uses()) {
4683 User *User = Use.getUser();
4684 if (Instruction *I = dyn_cast<Instruction>(User))
4685 if (I->getParent() == CaseDest)
4686 continue;
4687 if (PHINode *Phi = dyn_cast<PHINode>(User))
4688 if (Phi->getIncomingBlock(Use) == CaseDest)
4689 continue;
4690 return false;
4691 }
4692
4693 ConstantPool.insert(std::make_pair(&I, C));
4694 } else {
4695 break;
4696 }
4697 }
4698
4699 // If we did not have a CommonDest before, use the current one.
4700 if (!*CommonDest)
4701 *CommonDest = CaseDest;
4702 // If the destination isn't the common one, abort.
4703 if (CaseDest != *CommonDest)
4704 return false;
4705
4706 // Get the values for this case from phi nodes in the destination block.
4707 for (PHINode &PHI : (*CommonDest)->phis()) {
4708 int Idx = PHI.getBasicBlockIndex(Pred);
4709 if (Idx == -1)
4710 continue;
4711
4712 Constant *ConstVal =
4713 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4714 if (!ConstVal)
4715 return false;
4716
4717 // Be conservative about which kinds of constants we support.
4718 if (!ValidLookupTableConstant(ConstVal, TTI))
4719 return false;
4720
4721 Res.push_back(std::make_pair(&PHI, ConstVal));
4722 }
4723
4724 return Res.size() > 0;
4725}
4726
4727// Helper function used to add CaseVal to the list of cases that generate
4728// Result. Returns the updated number of cases that generate this result.
4729static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4730 SwitchCaseResultVectorTy &UniqueResults,
4731 Constant *Result) {
4732 for (auto &I : UniqueResults) {
4733 if (I.first == Result) {
4734 I.second.push_back(CaseVal);
4735 return I.second.size();
4736 }
4737 }
4738 UniqueResults.push_back(
4739 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4740 return 1;
4741}
4742
4743// Helper function that initializes a map containing
4744// results for the PHI node of the common destination block for a switch
4745// instruction. Returns false if multiple PHI nodes have been found or if
4746// there is not a common destination block for the switch.
4747static bool
4748InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4749 SwitchCaseResultVectorTy &UniqueResults,
4750 Constant *&DefaultResult, const DataLayout &DL,
4751 const TargetTransformInfo &TTI,
4752 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4753 for (auto &I : SI->cases()) {
4754 ConstantInt *CaseVal = I.getCaseValue();
4755
4756 // Resulting value at phi nodes for this case value.
4757 SwitchCaseResultsTy Results;
4758 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4759 DL, TTI))
4760 return false;
4761
4762 // Only one value per case is permitted.
4763 if (Results.size() > 1)
4764 return false;
4765
4766 // Add the case->result mapping to UniqueResults.
4767 const uintptr_t NumCasesForResult =
4768 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4769
4770 // Early out if there are too many cases for this result.
4771 if (NumCasesForResult > MaxCasesPerResult)
4772 return false;
4773
4774 // Early out if there are too many unique results.
4775 if (UniqueResults.size() > MaxUniqueResults)
4776 return false;
4777
4778 // Check the PHI consistency.
4779 if (!PHI)
4780 PHI = Results[0].first;
4781 else if (PHI != Results[0].first)
4782 return false;
4783 }
4784 // Find the default result value.
4785 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4786 BasicBlock *DefaultDest = SI->getDefaultDest();
4787 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4788 DL, TTI);
4789 // If the default value is not found abort unless the default destination
4790 // is unreachable.
4791 DefaultResult =
4792 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4793 if ((!DefaultResult &&
4794 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4795 return false;
4796
4797 return true;
4798}
4799
4800// Helper function that checks if it is possible to transform a switch with only
4801// two cases (or two cases + default) that produces a result into a select.
4802// Example:
4803// switch (a) {
4804// case 10: %0 = icmp eq i32 %a, 10
4805// return 10; %1 = select i1 %0, i32 10, i32 4
4806// case 20: ----> %2 = icmp eq i32 %a, 20
4807// return 2; %3 = select i1 %2, i32 2, i32 %1
4808// default:
4809// return 4;
4810// }
4811static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4812 Constant *DefaultResult, Value *Condition,
4813 IRBuilder<> &Builder) {
4814 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4815, __PRETTY_FUNCTION__))
4815 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4815, __PRETTY_FUNCTION__))
;
4816 // If we are selecting between only two cases transform into a simple
4817 // select or a two-way select if default is possible.
4818 if (ResultVector[0].second.size() == 1 &&
4819 ResultVector[1].second.size() == 1) {
4820 ConstantInt *const FirstCase = ResultVector[0].second[0];
4821 ConstantInt *const SecondCase = ResultVector[1].second[0];
4822
4823 bool DefaultCanTrigger = DefaultResult;
4824 Value *SelectValue = ResultVector[1].first;
4825 if (DefaultCanTrigger) {
4826 Value *const ValueCompare =
4827 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4828 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4829 DefaultResult, "switch.select");
4830 }
4831 Value *const ValueCompare =
4832 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4833 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4834 SelectValue, "switch.select");
4835 }
4836
4837 return nullptr;
4838}
4839
4840// Helper function to cleanup a switch instruction that has been converted into
4841// a select, fixing up PHI nodes and basic blocks.
4842static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4843 Value *SelectValue,
4844 IRBuilder<> &Builder) {
4845 BasicBlock *SelectBB = SI->getParent();
4846 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4847 PHI->removeIncomingValue(SelectBB);
4848 PHI->addIncoming(SelectValue, SelectBB);
4849
4850 Builder.CreateBr(PHI->getParent());
4851
4852 // Remove the switch.
4853 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4854 BasicBlock *Succ = SI->getSuccessor(i);
4855
4856 if (Succ == PHI->getParent())
4857 continue;
4858 Succ->removePredecessor(SelectBB);
4859 }
4860 SI->eraseFromParent();
4861}
4862
4863/// If the switch is only used to initialize one or more
4864/// phi nodes in a common successor block with only two different
4865/// constant values, replace the switch with select.
4866static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4867 const DataLayout &DL,
4868 const TargetTransformInfo &TTI) {
4869 Value *const Cond = SI->getCondition();
4870 PHINode *PHI = nullptr;
4871 BasicBlock *CommonDest = nullptr;
4872 Constant *DefaultResult;
4873 SwitchCaseResultVectorTy UniqueResults;
4874 // Collect all the cases that will deliver the same value from the switch.
4875 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4876 DL, TTI, 2, 1))
4877 return false;
4878 // Selects choose between maximum two values.
4879 if (UniqueResults.size() != 2)
4880 return false;
4881 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4881, __PRETTY_FUNCTION__))
;
4882
4883 Builder.SetInsertPoint(SI);
4884 Value *SelectValue =
4885 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4886 if (SelectValue) {
4887 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4888 return true;
4889 }
4890 // The switch couldn't be converted into a select.
4891 return false;
4892}
4893
4894namespace {
4895
4896/// This class represents a lookup table that can be used to replace a switch.
4897class SwitchLookupTable {
4898public:
4899 /// Create a lookup table to use as a switch replacement with the contents
4900 /// of Values, using DefaultValue to fill any holes in the table.
4901 SwitchLookupTable(
4902 Module &M, uint64_t TableSize, ConstantInt *Offset,
4903 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4904 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4905
4906 /// Build instructions with Builder to retrieve the value at
4907 /// the position given by Index in the lookup table.
4908 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4909
4910 /// Return true if a table with TableSize elements of
4911 /// type ElementType would fit in a target-legal register.
4912 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4913 Type *ElementType);
4914
4915private:
4916 // Depending on the contents of the table, it can be represented in
4917 // different ways.
4918 enum {
4919 // For tables where each element contains the same value, we just have to
4920 // store that single value and return it for each lookup.
4921 SingleValueKind,
4922
4923 // For tables where there is a linear relationship between table index
4924 // and values. We calculate the result with a simple multiplication
4925 // and addition instead of a table lookup.
4926 LinearMapKind,
4927
4928 // For small tables with integer elements, we can pack them into a bitmap
4929 // that fits into a target-legal register. Values are retrieved by
4930 // shift and mask operations.
4931 BitMapKind,
4932
4933 // The table is stored as an array of values. Values are retrieved by load
4934 // instructions from the table.
4935 ArrayKind
4936 } Kind;
4937
4938 // For SingleValueKind, this is the single value.
4939 Constant *SingleValue = nullptr;
4940
4941 // For BitMapKind, this is the bitmap.
4942 ConstantInt *BitMap = nullptr;
4943 IntegerType *BitMapElementTy = nullptr;
4944
4945 // For LinearMapKind, these are the constants used to derive the value.
4946 ConstantInt *LinearOffset = nullptr;
4947 ConstantInt *LinearMultiplier = nullptr;
4948
4949 // For ArrayKind, this is the array.
4950 GlobalVariable *Array = nullptr;
4951};
4952
4953} // end anonymous namespace
4954
4955SwitchLookupTable::SwitchLookupTable(
4956 Module &M, uint64_t TableSize, ConstantInt *Offset,
4957 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4958 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4959 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4959, __PRETTY_FUNCTION__))
;
4960 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4960, __PRETTY_FUNCTION__))
;
4961
4962 // If all values in the table are equal, this is that value.
4963 SingleValue = Values.begin()->second;
4964
4965 Type *ValueType = Values.begin()->second->getType();
4966
4967 // Build up the table contents.
4968 SmallVector<Constant *, 64> TableContents(TableSize);
4969 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4970 ConstantInt *CaseVal = Values[I].first;
4971 Constant *CaseRes = Values[I].second;
4972 assert(CaseRes->getType() == ValueType)((CaseRes->getType() == ValueType) ? static_cast<void>
(0) : __assert_fail ("CaseRes->getType() == ValueType", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4972, __PRETTY_FUNCTION__))
;
4973
4974 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4975 TableContents[Idx] = CaseRes;
4976
4977 if (CaseRes != SingleValue)
4978 SingleValue = nullptr;
4979 }
4980
4981 // Fill in any holes in the table with the default result.
4982 if (Values.size() < TableSize) {
4983 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4984, __PRETTY_FUNCTION__))
4984 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4984, __PRETTY_FUNCTION__))
;
4985 assert(DefaultValue->getType() == ValueType)((DefaultValue->getType() == ValueType) ? static_cast<void
> (0) : __assert_fail ("DefaultValue->getType() == ValueType"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4985, __PRETTY_FUNCTION__))
;
4986 for (uint64_t I = 0; I < TableSize; ++I) {
4987 if (!TableContents[I])
4988 TableContents[I] = DefaultValue;
4989 }
4990
4991 if (DefaultValue != SingleValue)
4992 SingleValue = nullptr;
4993 }
4994
4995 // If each element in the table contains the same value, we only need to store
4996 // that single value.
4997 if (SingleValue) {
4998 Kind = SingleValueKind;
4999 return;
5000 }
5001
5002 // Check if we can derive the value with a linear transformation from the
5003 // table index.
5004 if (isa<IntegerType>(ValueType)) {
5005 bool LinearMappingPossible = true;
5006 APInt PrevVal;
5007 APInt DistToPrev;
5008 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5008, __PRETTY_FUNCTION__))
;
5009 // Check if there is the same distance between two consecutive values.
5010 for (uint64_t I = 0; I < TableSize; ++I) {
5011 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
5012 if (!ConstVal) {
5013 // This is an undef. We could deal with it, but undefs in lookup tables
5014 // are very seldom. It's probably not worth the additional complexity.
5015 LinearMappingPossible = false;
5016 break;
5017 }
5018 const APInt &Val = ConstVal->getValue();
5019 if (I != 0) {
5020 APInt Dist = Val - PrevVal;
5021 if (I == 1) {
5022 DistToPrev = Dist;
5023 } else if (Dist != DistToPrev) {
5024 LinearMappingPossible = false;
5025 break;
5026 }
5027 }
5028 PrevVal = Val;
5029 }
5030 if (LinearMappingPossible) {
5031 LinearOffset = cast<ConstantInt>(TableContents[0]);
5032 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5033 Kind = LinearMapKind;
5034 ++NumLinearMaps;
5035 return;
5036 }
5037 }
5038
5039 // If the type is integer and the table fits in a register, build a bitmap.
5040 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5041 IntegerType *IT = cast<IntegerType>(ValueType);
5042 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5043 for (uint64_t I = TableSize; I > 0; --I) {
5044 TableInt <<= IT->getBitWidth();
5045 // Insert values into the bitmap. Undef values are set to zero.
5046 if (!isa<UndefValue>(TableContents[I - 1])) {
5047 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5048 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5049 }
5050 }
5051 BitMap = ConstantInt::get(M.getContext(), TableInt);
5052 BitMapElementTy = IT;
5053 Kind = BitMapKind;
5054 ++NumBitMaps;
5055 return;
5056 }
5057
5058 // Store the table in an array.
5059 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5060 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5061
5062 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5063 GlobalVariable::PrivateLinkage, Initializer,
5064 "switch.table." + FuncName);
5065 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5066 // Set the alignment to that of an array items. We will be only loading one
5067 // value out of it.
5068 Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5069 Kind = ArrayKind;
5070}
5071
5072Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5073 switch (Kind) {
5074 case SingleValueKind:
5075 return SingleValue;
5076 case LinearMapKind: {
5077 // Derive the result value from the input value.
5078 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5079 false, "switch.idx.cast");
5080 if (!LinearMultiplier->isOne())
5081 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5082 if (!LinearOffset->isZero())
5083 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5084 return Result;
5085 }
5086 case BitMapKind: {
5087 // Type of the bitmap (e.g. i59).
5088 IntegerType *MapTy = BitMap->getType();
5089
5090 // Cast Index to the same type as the bitmap.
5091 // Note: The Index is <= the number of elements in the table, so
5092 // truncating it to the width of the bitmask is safe.
5093 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5094
5095 // Multiply the shift amount by the element width.
5096 ShiftAmt = Builder.CreateMul(
5097 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5098 "switch.shiftamt");
5099
5100 // Shift down.
5101 Value *DownShifted =
5102 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5103 // Mask off.
5104 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5105 }
5106 case ArrayKind: {
5107 // Make sure the table index will not overflow when treated as signed.
5108 IntegerType *IT = cast<IntegerType>(Index->getType());
5109 uint64_t TableSize =
5110 Array->getInitializer()->getType()->getArrayNumElements();
5111 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5112 Index = Builder.CreateZExt(
5113 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5114 "switch.tableidx.zext");
5115
5116 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5117 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5118 GEPIndices, "switch.gep");
5119 return Builder.CreateLoad(
5120 cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5121 "switch.load");
5122 }
5123 }
5124 llvm_unreachable("Unknown lookup table kind!")::llvm::llvm_unreachable_internal("Unknown lookup table kind!"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5124)
;
5125}
5126
5127bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5128 uint64_t TableSize,
5129 Type *ElementType) {
5130 auto *IT = dyn_cast<IntegerType>(ElementType);
5131 if (!IT)
5132 return false;
5133 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5134 // are <= 15, we could try to narrow the type.
5135
5136 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5137 if (TableSize >= UINT_MAX(2147483647 *2U +1U) / IT->getBitWidth())
5138 return false;
5139 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5140}
5141
5142/// Determine whether a lookup table should be built for this switch, based on
5143/// the number of cases, size of the table, and the types of the results.
5144static bool
5145ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5146 const TargetTransformInfo &TTI, const DataLayout &DL,
5147 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5148 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX(18446744073709551615UL) / 10)
5149 return false; // TableSize overflowed, or mul below might overflow.
5150
5151 bool AllTablesFitInRegister = true;
5152 bool HasIllegalType = false;
5153 for (const auto &I : ResultTypes) {
5154 Type *Ty = I.second;
5155
5156 // Saturate this flag to true.
5157 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5158
5159 // Saturate this flag to false.
5160 AllTablesFitInRegister =
5161 AllTablesFitInRegister &&
5162 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5163
5164 // If both flags saturate, we're done. NOTE: This *only* works with
5165 // saturating flags, and all flags have to saturate first due to the
5166 // non-deterministic behavior of iterating over a dense map.
5167 if (HasIllegalType && !AllTablesFitInRegister)
5168 break;
5169 }
5170
5171 // If each table would fit in a register, we should build it anyway.
5172 if (AllTablesFitInRegister)
5173 return true;
5174
5175 // Don't build a table that doesn't fit in-register if it has illegal types.
5176 if (HasIllegalType)
5177 return false;
5178
5179 // The table density should be at least 40%. This is the same criterion as for
5180 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5181 // FIXME: Find the best cut-off.
5182 return SI->getNumCases() * 10 >= TableSize * 4;
5183}
5184
5185/// Try to reuse the switch table index compare. Following pattern:
5186/// \code
5187/// if (idx < tablesize)
5188/// r = table[idx]; // table does not contain default_value
5189/// else
5190/// r = default_value;
5191/// if (r != default_value)
5192/// ...
5193/// \endcode
5194/// Is optimized to:
5195/// \code
5196/// cond = idx < tablesize;
5197/// if (cond)
5198/// r = table[idx];
5199/// else
5200/// r = default_value;
5201/// if (cond)
5202/// ...
5203/// \endcode
5204/// Jump threading will then eliminate the second if(cond).
5205static void reuseTableCompare(
5206 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5207 Constant *DefaultValue,
5208 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5209 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5210 if (!CmpInst)
5211 return;
5212
5213 // We require that the compare is in the same block as the phi so that jump
5214 // threading can do its work afterwards.
5215 if (CmpInst->getParent() != PhiBlock)
5216 return;
5217
5218 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5219 if (!CmpOp1)
5220 return;
5221
5222 Value *RangeCmp = RangeCheckBranch->getCondition();
5223 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5224 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5225
5226 // Check if the compare with the default value is constant true or false.
5227 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5228 DefaultValue, CmpOp1, true);
5229 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5230 return;
5231
5232 // Check if the compare with the case values is distinct from the default
5233 // compare result.
5234 for (auto ValuePair : Values) {
5235 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5236 ValuePair.second, CmpOp1, true);
5237 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5238 return;
5239 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5240, __PRETTY_FUNCTION__))
5240 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5240, __PRETTY_FUNCTION__))
;
5241 }
5242
5243 // Check if the branch instruction dominates the phi node. It's a simple
5244 // dominance check, but sufficient for our needs.
5245 // Although this check is invariant in the calling loops, it's better to do it
5246 // at this late stage. Practically we do it at most once for a switch.
5247 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5248 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5249 BasicBlock *Pred = *PI;
5250 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5251 return;
5252 }
5253
5254 if (DefaultConst == FalseConst) {
5255 // The compare yields the same result. We can replace it.
5256 CmpInst->replaceAllUsesWith(RangeCmp);
5257 ++NumTableCmpReuses;
5258 } else {
5259 // The compare yields the same result, just inverted. We can replace it.
5260 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5261 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5262 RangeCheckBranch);
5263 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5264 ++NumTableCmpReuses;
5265 }
5266}
5267
5268/// If the switch is only used to initialize one or more phi nodes in a common
5269/// successor block with different constant values, replace the switch with
5270/// lookup tables.
5271static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5272 const DataLayout &DL,
5273 const TargetTransformInfo &TTI) {
5274 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5274, __PRETTY_FUNCTION__))
;
5275
5276 Function *Fn = SI->getParent()->getParent();
5277 // Only build lookup table when we have a target that supports it or the
5278 // attribute is not set.
5279 if (!TTI.shouldBuildLookupTables() ||
5280 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5281 return false;
5282
5283 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5284 // split off a dense part and build a lookup table for that.
5285
5286 // FIXME: This creates arrays of GEPs to constant strings, which means each
5287 // GEP needs a runtime relocation in PIC code. We should just build one big
5288 // string and lookup indices into that.
5289
5290 // Ignore switches with less than three cases. Lookup tables will not make
5291 // them faster, so we don't analyze them.
5292 if (SI->getNumCases() < 3)
5293 return false;
5294
5295 // Figure out the corresponding result for each case value and phi node in the
5296 // common destination, as well as the min and max case values.
5297 assert(!empty(SI->cases()))((!empty(SI->cases())) ? static_cast<void> (0) : __assert_fail
("!empty(SI->cases())", "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5297, __PRETTY_FUNCTION__))
;
5298 SwitchInst::CaseIt CI = SI->case_begin();
5299 ConstantInt *MinCaseVal = CI->getCaseValue();
5300 ConstantInt *MaxCaseVal = CI->getCaseValue();
5301
5302 BasicBlock *CommonDest = nullptr;
5303
5304 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5305 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5306
5307 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5308 SmallDenseMap<PHINode *, Type *> ResultTypes;
5309 SmallVector<PHINode *, 4> PHIs;
5310
5311 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5312 ConstantInt *CaseVal = CI->getCaseValue();
5313 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5314 MinCaseVal = CaseVal;
5315 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5316 MaxCaseVal = CaseVal;
5317
5318 // Resulting value at phi nodes for this case value.
5319 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5320 ResultsTy Results;
5321 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5322 Results, DL, TTI))
5323 return false;
5324
5325 // Append the result from this case to the list for each phi.
5326 for (const auto &I : Results) {
5327 PHINode *PHI = I.first;
5328 Constant *Value = I.second;
5329 if (!ResultLists.count(PHI))
5330 PHIs.push_back(PHI);
5331 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5332 }
5333 }
5334
5335 // Keep track of the result types.
5336 for (PHINode *PHI : PHIs) {
5337 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5338 }
5339
5340 uint64_t NumResults = ResultLists[PHIs[0]].size();
5341 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5342 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5343 bool TableHasHoles = (NumResults < TableSize);
5344
5345 // If the table has holes, we need a constant result for the default case
5346 // or a bitmask that fits in a register.
5347 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5348 bool HasDefaultResults =
5349 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5350 DefaultResultsList, DL, TTI);
5351
5352 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5353 if (NeedMask) {
5354 // As an extra penalty for the validity test we require more cases.
5355 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5356 return false;
5357 if (!DL.fitsInLegalInteger(TableSize))
5358 return false;
5359 }
5360
5361 for (const auto &I : DefaultResultsList) {
5362 PHINode *PHI = I.first;
5363 Constant *Result = I.second;
5364 DefaultResults[PHI] = Result;
5365 }
5366
5367 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5368 return false;
5369
5370 // Create the BB that does the lookups.
5371 Module &Mod = *CommonDest->getParent()->getParent();
5372 BasicBlock *LookupBB = BasicBlock::Create(
5373 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5374
5375 // Compute the table index value.
5376 Builder.SetInsertPoint(SI);
5377 Value *TableIndex;
5378 if (MinCaseVal->isNullValue())
5379 TableIndex = SI->getCondition();
5380 else
5381 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5382 "switch.tableidx");
5383
5384 // Compute the maximum table size representable by the integer type we are
5385 // switching upon.
5386 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5387 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX(18446744073709551615UL) : 1ULL << CaseSize;
5388 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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5390, __PRETTY_FUNCTION__))
5389 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5390, __PRETTY_FUNCTION__))
5390 "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-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5390, __PRETTY_FUNCTION__))
;
5391
5392 // If the default destination is unreachable, or if the lookup table covers
5393 // all values of the conditional variable, branch directly to the lookup table
5394 // BB. Otherwise, check that the condition is within the case range.
5395 const bool DefaultIsReachable =
5396 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5397 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5398 BranchInst *RangeCheckBranch = nullptr;
5399
5400 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5401 Builder.CreateBr(LookupBB);
5402 // Note: We call removeProdecessor later since we need to be able to get the
5403 // PHI value for the default case in case we're using a bit mask.
5404 } else {
5405 Value *Cmp = Builder.CreateICmpULT(
5406 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5407 RangeCheckBranch =
5408 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5409 }
5410
5411 // Populate the BB that does the lookups.
5412 Builder.SetInsertPoint(LookupBB);
5413
5414 if (NeedMask) {
5415 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5416 // re-purposed to do the hole check, and we create a new LookupBB.
5417 BasicBlock *MaskBB = LookupBB;
5418 MaskBB->setName("switch.hole_check");
5419 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5420 CommonDest->getParent(), CommonDest);
5421
5422 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5423 // unnecessary illegal types.
5424 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5425 APInt MaskInt(TableSizePowOf2, 0);
5426 APInt One(TableSizePowOf2, 1);
5427 // Build bitmask; fill in a 1 bit for every case.
5428 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5429 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5430 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5431 .getLimitedValue();
5432 MaskInt |= One << Idx;
5433 }
5434 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5435
5436 // Get the TableIndex'th bit of the bitmask.
5437 // If this bit is 0 (meaning hole) jump to the default destination,
5438 // else continue with table lookup.
5439 IntegerType *MapTy = TableMask->getType();
5440 Value *MaskIndex =
5441 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5442 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5443 Value *LoBit = Builder.CreateTrunc(
5444 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5445 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5446
5447 Builder.SetInsertPoint(LookupBB);
5448 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5449 }
5450
5451 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5452 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5453 // do not delete PHINodes here.
5454 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5455 /*KeepOneInputPHIs=*/true);
5456 }
5457
5458 bool ReturnedEarly = false;
5459 for (PHINode *PHI : PHIs) {
5460 const ResultListTy &ResultList = ResultLists[PHI];
5461
5462 // If using a bitmask, use any value to fill the lookup table holes.
5463 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5464 StringRef FuncName = Fn->getName();
5465 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5466 FuncName);
5467
5468 Value *Result = Table.BuildLookup(TableIndex, Builder);
5469
5470 // If the result is used to return immediately from the function, we want to
5471 // do that right here.
5472 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5473 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5474 Builder.CreateRet(Result);
5475 ReturnedEarly = true;
5476 break;
5477 }
5478
5479 // Do a small peephole optimization: re-use the switch table compare if
5480 // possible.
5481 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5482 BasicBlock *PhiBlock = PHI->getParent();
5483 // Search for compare instructions which use the phi.
5484 for (auto *User : PHI->users()) {
5485 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5486 }
5487 }
5488
5489 PHI->addIncoming(Result, LookupBB);
5490 }
5491
5492 if (!ReturnedEarly)
5493 Builder.CreateBr(CommonDest);
5494
5495 // Remove the switch.
5496 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5497 BasicBlock *Succ = SI->getSuccessor(i);
5498
5499 if (Succ == SI->getDefaultDest())
5500 continue;
5501 Succ->removePredecessor(SI->getParent());
5502 }
5503 SI->eraseFromParent();
5504
5505 ++NumLookupTables;
5506 if (NeedMask)
5507 ++NumLookupTablesHoles;
5508 return true;
5509}
5510
5511static bool isSwitchDense(ArrayRef<int64_t> Values) {
5512 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5513 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5514 uint64_t Range = Diff + 1;
5515 uint64_t NumCases = Values.size();
5516 // 40% is the default density for building a jump table in optsize/minsize mode.
5517 uint64_t MinDensity = 40;
5518
5519 return NumCases * 100 >= Range * MinDensity;
5520}
5521
5522/// Try to transform a switch that has "holes" in it to a contiguous sequence
5523/// of cases.
5524///
5525/// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5526/// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5527///
5528/// This converts a sparse switch into a dense switch which allows better
5529/// lowering and could also allow transforming into a lookup table.
5530static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5531 const DataLayout &DL,
5532 const TargetTransformInfo &TTI) {
5533 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5534 if (CondTy->getIntegerBitWidth() > 64 ||
5535 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5536 return false;
5537 // Only bother with this optimization if there are more than 3 switch cases;
5538 // SDAG will only bother creating jump tables for 4 or more cases.
5539 if (SI->getNumCases() < 4)
5540 return false;
5541
5542 // This transform is agnostic to the signedness of the input or case values. We
5543 // can treat the case values as signed or unsigned. We can optimize more common
5544 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5545 // as signed.
5546 SmallVector<int64_t,4> Values;
5547 for (auto &C : SI->cases())
5548 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5549 llvm::sort(Values);
5550
5551 // If the switch is already dense, there's nothing useful to do here.
5552 if (isSwitchDense(Values))
5553 return false;
5554
5555 // First, transform the values such that they start at zero and ascend.
5556 int64_t Base = Values[0];
5557 for (auto &V : Values)
5558 V -= (uint64_t)(Base);
5559
5560 // Now we have signed numbers that have been shifted so that, given enough
5561 // precision, there are no negative values. Since the rest of the transform
5562 // is bitwise only, we switch now to an unsigned representation.
5563 uint64_t GCD = 0;
5564 for (auto &V : Values)
5565 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5566
5567 // This transform can be done speculatively because it is so cheap - it results
5568 // in a single rotate operation being inserted. This can only happen if the
5569 // factor extracted is a power of 2.
5570 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5571 // inverse of GCD and then perform this transform.
5572 // FIXME: It's possible that optimizing a switch on powers of two might also
5573 // be beneficial - flag values are often powers of two and we could use a CLZ
5574 // as the key function.
5575 if (GCD <= 1 || !isPowerOf2_64(GCD))
5576 // No common divisor found or too expensive to compute key function.
5577 return false;
5578
5579 unsigned Shift = Log2_64(GCD);
5580 for (auto &V : Values)
5581 V = (int64_t)((uint64_t)V >> Shift);
5582
5583 if (!isSwitchDense(Values))
5584 // Transform didn't create a dense switch.
5585 return false;
5586
5587 // The obvious transform is to shift the switch condition right and emit a
5588 // check that the condition actually cleanly divided by GCD, i.e.
5589 // C & (1 << Shift - 1) == 0
5590 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5591 //
5592 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5593 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5594 // are nonzero then the switch condition will be very large and will hit the
5595 // default case.
5596
5597 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5598 Builder.SetInsertPoint(SI);
5599 auto *ShiftC = ConstantInt::get(Ty, Shift);
5600 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5601 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5602 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5603 auto *Rot = Builder.CreateOr(LShr, Shl);
5604 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5605
5606 for (auto Case : SI->cases()) {
5607 auto *Orig = Case.getCaseValue();
5608 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5609 Case.setValue(
5610 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5611 }
5612 return true;
5613}
5614
5615bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5616 BasicBlock *BB = SI->getParent();
5617
5618 if (isValueEqualityComparison(SI)) {
5619 // If we only have one predecessor, and if it is a branch on this value,
5620 // see if that predecessor totally determines the outcome of this switch.
5621 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5622 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5623 return requestResimplify();
5624
5625 Value *Cond = SI->getCondition();
5626 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5627 if (SimplifySwitchOnSelect(SI, Select))
5628 return requestResimplify();
5629
5630 // If the block only contains the switch, see if we can fold the block
5631 // away into any preds.
5632 if (SI == &*BB->instructionsWithoutDebug().begin())
5633 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5634 return requestResimplify();
5635 }
5636
5637 // Try to transform the switch into an icmp and a branch.
5638 if (TurnSwitchRangeIntoICmp(SI, Builder))
5639 return requestResimplify();
5640
5641 // Remove unreachable cases.
5642 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5643 return requestResimplify();
5644
5645 if (switchToSelect(SI, Builder, DL, TTI))
5646 return requestResimplify();
5647
5648 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5649 return requestResimplify();
5650
5651 // The conversion from switch to lookup tables results in difficult-to-analyze
5652 // code and makes pruning branches much harder. This is a problem if the
5653 // switch expression itself can still be restricted as a result of inlining or
5654 // CVP. Therefore, only apply this transformation during late stages of the
5655 // optimisation pipeline.
5656 if (Options.ConvertSwitchToLookupTable &&
5657 SwitchToLookupTable(SI, Builder, DL, TTI))
5658 return requestResimplify();
5659
5660 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5661 return requestResimplify();
5662
5663 return false;
5664}
5665
5666bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5667 BasicBlock *BB = IBI->getParent();
5668 bool Changed = false;
5669
5670 // Eliminate redundant destinations.
5671 SmallPtrSet<Value *, 8> Succs;
5672 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5673 BasicBlock *Dest = IBI->getDestination(i);
5674 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5675 Dest->removePredecessor(BB);
5676 IBI->removeDestination(i);
5677 --i;
5678 --e;
5679 Changed = true;
5680 }
5681 }
5682
5683 if (IBI->getNumDestinations() == 0) {
5684 // If the indirectbr has no successors, change it to unreachable.
5685 new UnreachableInst(IBI->getContext(), IBI);
5686 EraseTerminatorAndDCECond(IBI);
5687 return true;
5688 }
5689
5690 if (IBI->getNumDestinations() == 1) {
5691 // If the indirectbr has one successor, change it to a direct branch.
5692 BranchInst::Create(IBI->getDestination(0), IBI);
5693 EraseTerminatorAndDCECond(IBI);
5694 return true;
5695 }
5696
5697 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5698 if (SimplifyIndirectBrOnSelect(IBI, SI))
5699 return requestResimplify();
5700 }
5701 return Changed;
5702}
5703
5704/// Given an block with only a single landing pad and a unconditional branch
5705/// try to find another basic block which this one can be merged with. This
5706/// handles cases where we have multiple invokes with unique landing pads, but
5707/// a shared handler.
5708///
5709/// We specifically choose to not worry about merging non-empty blocks
5710/// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5711/// practice, the optimizer produces empty landing pad blocks quite frequently
5712/// when dealing with exception dense code. (see: instcombine, gvn, if-else
5713/// sinking in this file)
5714///
5715/// This is primarily a code size optimization. We need to avoid performing
5716/// any transform which might inhibit optimization (such as our ability to
5717/// specialize a particular handler via tail commoning). We do this by not
5718/// merging any blocks which require us to introduce a phi. Since the same
5719/// values are flowing through both blocks, we don't lose any ability to
5720/// specialize. If anything, we make such specialization more likely.
5721///
5722/// TODO - This transformation could remove entries from a phi in the target
5723/// block when the inputs in the phi are the same for the two blocks being
5724/// merged. In some cases, this could result in removal of the PHI entirely.
5725static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5726 BasicBlock *BB) {
5727 auto Succ = BB->getUniqueSuccessor();
5728 assert(Succ)((Succ) ? static_cast<void> (0) : __assert_fail ("Succ"
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5728, __PRETTY_FUNCTION__))
;
5729 // If there's a phi in the successor block, we'd likely have to introduce
5730 // a phi into the merged landing pad block.
5731 if (isa<PHINode>(*Succ->begin()))
5732 return false;
5733
5734 for (BasicBlock *OtherPred : predecessors(Succ)) {
5735 if (BB == OtherPred)
5736 continue;
5737 BasicBlock::iterator I = OtherPred->begin();
5738 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5739 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5740 continue;
5741 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5742 ;
5743 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5744 if (!BI2 || !BI2->isIdenticalTo(BI))
5745 continue;
5746
5747 // We've found an identical block. Update our predecessors to take that
5748 // path instead and make ourselves dead.
5749 SmallPtrSet<BasicBlock *, 16> Preds;
5750 Preds.insert(pred_begin(BB), pred_end(BB));
5751 for (BasicBlock *Pred : Preds) {
5752 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5753 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&((II->getNormalDest() != BB && II->getUnwindDest
() == BB && "unexpected successor") ? static_cast<
void> (0) : __assert_fail ("II->getNormalDest() != BB && II->getUnwindDest() == BB && \"unexpected successor\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5754, __PRETTY_FUNCTION__))
5754 "unexpected successor")((II->getNormalDest() != BB && II->getUnwindDest
() == BB && "unexpected successor") ? static_cast<
void> (0) : __assert_fail ("II->getNormalDest() != BB && II->getUnwindDest() == BB && \"unexpected successor\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5754, __PRETTY_FUNCTION__))
;
5755 II->setUnwindDest(OtherPred);
5756 }
5757
5758 // The debug info in OtherPred doesn't cover the merged control flow that
5759 // used to go through BB. We need to delete it or update it.
5760 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5761 Instruction &Inst = *I;
5762 I++;
5763 if (isa<DbgInfoIntrinsic>(Inst))
5764 Inst.eraseFromParent();
5765 }
5766
5767 SmallPtrSet<BasicBlock *, 16> Succs;
5768 Succs.insert(succ_begin(BB), succ_end(BB));
5769 for (BasicBlock *Succ : Succs) {
5770 Succ->removePredecessor(BB);
5771 }
5772
5773 IRBuilder<> Builder(BI);
5774 Builder.CreateUnreachable();
5775 BI->eraseFromParent();
5776 return true;
5777 }
5778 return false;
5779}
5780
5781bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5782 IRBuilder<> &Builder) {
5783 BasicBlock *BB = BI->getParent();
5784 BasicBlock *Succ = BI->getSuccessor(0);
5785
5786 // If the Terminator is the only non-phi instruction, simplify the block.
5787 // If LoopHeader is provided, check if the block or its successor is a loop
5788 // header. (This is for early invocations before loop simplify and
5789 // vectorization to keep canonical loop forms for nested loops. These blocks
5790 // can be eliminated when the pass is invoked later in the back-end.)
5791 // Note that if BB has only one predecessor then we do not introduce new
5792 // backedge, so we can eliminate BB.
5793 bool NeedCanonicalLoop =
5794 Options.NeedCanonicalLoop &&
5795 (LoopHeaders && BB->hasNPredecessorsOrMore(2) &&
5796 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5797 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5798 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5799 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5800 return true;
5801
5802 // If the only instruction in the block is a seteq/setne comparison against a
5803 // constant, try to simplify the block.
5804 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5805 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5806 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5807 ;
5808 if (I->isTerminator() &&
5809 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
5810 return true;
5811 }
5812
5813 // See if we can merge an empty landing pad block with another which is
5814 // equivalent.
5815 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5816 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5817 ;
5818 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5819 return true;
5820 }
5821
5822 // If this basic block is ONLY a compare and a branch, and if a predecessor
5823 // branches to us and our successor, fold the comparison into the
5824 // predecessor and use logical operations to update the incoming value
5825 // for PHI nodes in common successor.
5826 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5827 return requestResimplify();
5828 return false;
5829}
5830
5831static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5832 BasicBlock *PredPred = nullptr;
5833 for (auto *P : predecessors(BB)) {
5834 BasicBlock *PPred = P->getSinglePredecessor();
5835 if (!PPred || (PredPred && PredPred != PPred))
5836 return nullptr;
5837 PredPred = PPred;
5838 }
5839 return PredPred;
5840}
5841
5842bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5843 BasicBlock *BB = BI->getParent();
5844 const Function *Fn = BB->getParent();
5845 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
5846 return false;
5847
5848 // Conditional branch
5849 if (isValueEqualityComparison(BI)) {
5850 // If we only have one predecessor, and if it is a branch on this value,
5851 // see if that predecessor totally determines the outcome of this
5852 // switch.
5853 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5854 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5855 return requestResimplify();
5856
5857 // This block must be empty, except for the setcond inst, if it exists.
5858 // Ignore dbg intrinsics.
5859 auto I = BB->instructionsWithoutDebug().begin();
5860 if (&*I == BI) {
5861 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5862 return requestResimplify();
5863 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5864 ++I;
5865 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5866 return requestResimplify();
5867 }
5868 }
5869
5870 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5871 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5872 return true;
5873
5874 // If this basic block has dominating predecessor blocks and the dominating
5875 // blocks' conditions imply BI's condition, we know the direction of BI.
5876 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
5877 if (Imp) {
5878 // Turn this into a branch on constant.
5879 auto *OldCond = BI->getCondition();
5880 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
5881 : ConstantInt::getFalse(BB->getContext());
5882 BI->setCondition(TorF);
5883 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5884 return requestResimplify();
5885 }
5886
5887 // If this basic block is ONLY a compare and a branch, and if a predecessor
5888 // branches to us and one of our successors, fold the comparison into the
5889 // predecessor and use logical operations to pick the right destination.
5890 if (FoldBranchToCommonDest(BI, nullptr, Options.BonusInstThreshold))
5891 return requestResimplify();
5892
5893 // We have a conditional branch to two blocks that are only reachable
5894 // from BI. We know that the condbr dominates the two blocks, so see if
5895 // there is any identical code in the "then" and "else" blocks. If so, we
5896 // can hoist it up to the branching block.
5897 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5898 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5899 if (HoistThenElseCodeToIf(BI, TTI))
5900 return requestResimplify();
5901 } else {
5902 // If Successor #1 has multiple preds, we may be able to conditionally
5903 // execute Successor #0 if it branches to Successor #1.
5904 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
5905 if (Succ0TI->getNumSuccessors() == 1 &&
5906 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5907 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5908 return requestResimplify();
5909 }
5910 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5911 // If Successor #0 has multiple preds, we may be able to conditionally
5912 // execute Successor #1 if it branches to Successor #0.
5913 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
5914 if (Succ1TI->getNumSuccessors() == 1 &&
5915 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5916 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5917 return requestResimplify();
5918 }
5919
5920 // If this is a branch on a phi node in the current block, thread control
5921 // through this block if any PHI node entries are constants.
5922 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5923 if (PN->getParent() == BI->getParent())
5924 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
5925 return requestResimplify();
5926
5927 // Scan predecessor blocks for conditional branches.
5928 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5929 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5930 if (PBI != BI && PBI->isConditional())
5931 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5932 return requestResimplify();
5933
5934 // Look for diamond patterns.
5935 if (MergeCondStores)
5936 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5937 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5938 if (PBI != BI && PBI->isConditional())
5939 if (mergeConditionalStores(PBI, BI, DL))
5940 return requestResimplify();
5941
5942 return false;
5943}
5944
5945/// Check if passing a value to an instruction will cause undefined behavior.
5946static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5947 Constant *C = dyn_cast<Constant>(V);
5948 if (!C)
5949 return false;
5950
5951 if (I->use_empty())
5952 return false;
5953
5954 if (C->isNullValue() || isa<UndefValue>(C)) {
5955 // Only look at the first use, avoid hurting compile time with long uselists
5956 User *Use = *I->user_begin();
5957
5958 // Now make sure that there are no instructions in between that can alter
5959 // control flow (eg. calls)
5960 for (BasicBlock::iterator
5961 i = ++BasicBlock::iterator(I),
5962 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5963 i != UI; ++i)
5964 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5965 return false;
5966
5967 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5968 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5969 if (GEP->getPointerOperand() == I)
5970 return passingValueIsAlwaysUndefined(V, GEP);
5971
5972 // Look through bitcasts.
5973 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5974 return passingValueIsAlwaysUndefined(V, BC);
5975
5976 // Load from null is undefined.
5977 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5978 if (!LI->isVolatile())
5979 return !NullPointerIsDefined(LI->getFunction(),
5980 LI->getPointerAddressSpace());
5981
5982 // Store to null is undefined.
5983 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5984 if (!SI->isVolatile())
5985 return (!NullPointerIsDefined(SI->getFunction(),
5986 SI->getPointerAddressSpace())) &&
5987 SI->getPointerOperand() == I;
5988
5989 // A call to null is undefined.
5990 if (auto CS = CallSite(Use))
5991 return !NullPointerIsDefined(CS->getFunction()) &&
5992 CS.getCalledValue() == I;
5993 }
5994 return false;
5995}
5996
5997/// If BB has an incoming value that will always trigger undefined behavior
5998/// (eg. null pointer dereference), remove the branch leading here.
5999static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
6000 for (PHINode &PHI : BB->phis())
6001 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
6002 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
6003 Instruction *T = PHI.getIncomingBlock(i)->getTerminator();
6004 IRBuilder<> Builder(T);
6005 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
6006 BB->removePredecessor(PHI.getIncomingBlock(i));
6007 // Turn uncoditional branches into unreachables and remove the dead
6008 // destination from conditional branches.
6009 if (BI->isUnconditional())
6010 Builder.CreateUnreachable();
6011 else
6012 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
6013 : BI->getSuccessor(0));
6014 BI->eraseFromParent();
6015 return true;
6016 }
6017 // TODO: SwitchInst.
6018 }
6019
6020 return false;
6021}
6022
6023bool SimplifyCFGOpt::simplifyOnce(BasicBlock *BB) {
6024 bool Changed = false;
6025
6026 assert(BB && BB->getParent() && "Block not embedded in function!")((BB && BB->getParent() && "Block not embedded in function!"
) ? static_cast<void> (0) : __assert_fail ("BB && BB->getParent() && \"Block not embedded in function!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 6026, __PRETTY_FUNCTION__))
;
3
Assuming the condition is true
4
'?' condition is true
6027 assert(BB->getTerminator() && "Degenerate basic block encountered!")((BB->getTerminator() && "Degenerate basic block encountered!"
) ? static_cast<void> (0) : __assert_fail ("BB->getTerminator() && \"Degenerate basic block encountered!\""
, "/build/llvm-toolchain-snapshot-9~svn362543/lib/Transforms/Utils/SimplifyCFG.cpp"
, 6027, __PRETTY_FUNCTION__))
;
5
Assuming the condition is true
6
'?' condition is true
6028
6029 // Remove basic blocks that have no predecessors (except the entry block)...
6030 // or that just have themself as a predecessor. These are unreachable.
6031 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
7
Assuming the condition is false
9
Taking false branch
6032 BB->getSinglePredecessor() == BB) {
8
Assuming the condition is false
6033 LLVM_DEBUG(dbgs() << "Removing BB: \n" << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Removing BB: \n" <<
*BB; } } while (false)
;
6034 DeleteDeadBlock(BB);
6035 return true;
6036 }
6037
6038 // Check to see if we can constant propagate this terminator instruction
6039 // away...
6040 Changed |= ConstantFoldTerminator(BB, true);
6041
6042 // Check for and eliminate duplicate PHI nodes in this block.
6043 Changed |= EliminateDuplicatePHINodes(BB);
6044
6045 // Check for and remove branches that will always cause undefined behavior.
6046 Changed |= removeUndefIntroducingPredecessor(BB);
6047
6048 // Merge basic blocks into their predecessor if there is only one distinct
6049 // pred, and if there is only one distinct successor of the predecessor, and
6050 // if there are no PHI nodes.
6051 if (MergeBlockIntoPredecessor(BB))
10
Assuming the condition is false
11
Taking false branch
6052 return true;
6053
6054 if (SinkCommon && Options.SinkCommonInsts)
12
Assuming the condition is false
6055 Changed |= SinkCommonCodeFromPredecessors(BB);
6056
6057 IRBuilder<> Builder(BB);
6058
6059 // If there is a trivial two-entry PHI node in this basic block, and we can
6060 // eliminate it, do so now.
6061 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
13
Assuming 'PN' is non-null
14
Taking true branch
6062 if (PN->getNumIncomingValues() == 2)
15
Assuming the condition is true
16
Taking true branch
6063 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
17
Calling 'FoldTwoEntryPHINode'
6064
6065 Builder.SetInsertPoint(BB->getTerminator());
6066 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
6067 if (BI->isUnconditional()) {
6068 if (SimplifyUncondBranch(BI, Builder))
6069 return true;
6070 } else {
6071 if (SimplifyCondBranch(BI, Builder))
6072 return true;
6073 }
6074 } else if (auto *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6075 if (SimplifyReturn(RI, Builder))
6076 return true;
6077 } else if (auto *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6078 if (SimplifyResume(RI, Builder))
6079 return true;
6080 } else if (auto *RI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6081 if (SimplifyCleanupReturn(RI))
6082 return true;
6083 } else if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6084 if (SimplifySwitch(SI, Builder))
6085 return true;
6086 } else if (auto *UI = dyn_cast<UnreachableInst>(BB->getTerminator())) {
6087 if (SimplifyUnreachable(UI))
6088 return true;
6089 } else if (auto *IBI = dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6090 if (SimplifyIndirectBr(IBI))
6091 return true;
6092 }
6093
6094 return Changed;
6095}
6096
6097bool SimplifyCFGOpt::run(BasicBlock *BB) {
6098 bool Changed = false;
6099
6100 // Repeated simplify BB as long as resimplification is requested.
6101 do {
6102 Resimplify = false;
6103
6104 // Perform one round of simplifcation. Resimplify flag will be set if
6105 // another iteration is requested.
6106 Changed |= simplifyOnce(BB);
2
Calling 'SimplifyCFGOpt::simplifyOnce'
6107 } while (Resimplify);
6108
6109 return Changed;
6110}
6111
6112bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6113 const SimplifyCFGOptions &Options,
6114 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6115 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,
1
Calling 'SimplifyCFGOpt::run'
6116 Options)
6117 .run(BB);
6118}