Bug Summary

File:lib/Transforms/Utils/SimplifyCFG.cpp
Warning:line 2391, 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-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Transforms/Utils -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp -faddrsig
1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// Peephole optimize the CFG.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/ADT/APInt.h"
15#include "llvm/ADT/ArrayRef.h"
16#include "llvm/ADT/DenseMap.h"
17#include "llvm/ADT/Optional.h"
18#include "llvm/ADT/STLExtras.h"
19#include "llvm/ADT/SetOperations.h"
20#include "llvm/ADT/SetVector.h"
21#include "llvm/ADT/SmallPtrSet.h"
22#include "llvm/ADT/SmallVector.h"
23#include "llvm/ADT/Statistic.h"
24#include "llvm/ADT/StringRef.h"
25#include "llvm/Analysis/AssumptionCache.h"
26#include "llvm/Analysis/ConstantFolding.h"
27#include "llvm/Analysis/EHPersonalities.h"
28#include "llvm/Analysis/InstructionSimplify.h"
29#include "llvm/Analysis/TargetTransformInfo.h"
30#include "llvm/Transforms/Utils/Local.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/ValueMapper.h"
70#include <algorithm>
71#include <cassert>
72#include <climits>
73#include <cstddef>
74#include <cstdint>
75#include <iterator>
76#include <map>
77#include <set>
78#include <tuple>
79#include <utility>
80#include <vector>
81
82using namespace llvm;
83using namespace PatternMatch;
84
85#define DEBUG_TYPE"simplifycfg" "simplifycfg"
86
87// Chosen as 2 so as to be cheap, but still to have enough power to fold
88// a select, so the "clamp" idiom (of a min followed by a max) will be caught.
89// To catch this, we need to fold a compare and a select, hence '2' being the
90// minimum reasonable default.
91static cl::opt<unsigned> PHINodeFoldingThreshold(
92 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
93 cl::desc(
94 "Control the amount of phi node folding to perform (default = 2)"));
95
96static cl::opt<bool> DupRet(
97 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
98 cl::desc("Duplicate return instructions into unconditional branches"));
99
100static cl::opt<bool>
101 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
102 cl::desc("Sink common instructions down to the end block"));
103
104static cl::opt<bool> HoistCondStores(
105 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
106 cl::desc("Hoist conditional stores if an unconditional store precedes"));
107
108static cl::opt<bool> MergeCondStores(
109 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
110 cl::desc("Hoist conditional stores even if an unconditional store does not "
111 "precede - hoist multiple conditional stores into a single "
112 "predicated store"));
113
114static cl::opt<bool> MergeCondStoresAggressively(
115 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
116 cl::desc("When merging conditional stores, do so even if the resultant "
117 "basic blocks are unlikely to be if-converted as a result"));
118
119static cl::opt<bool> SpeculateOneExpensiveInst(
120 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
121 cl::desc("Allow exactly one expensive instruction to be speculatively "
122 "executed"));
123
124static cl::opt<unsigned> MaxSpeculationDepth(
125 "max-speculation-depth", cl::Hidden, cl::init(10),
126 cl::desc("Limit maximum recursion depth when calculating costs of "
127 "speculatively executed instructions"));
128
129STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps")static llvm::Statistic NumBitMaps = {"simplifycfg", "NumBitMaps"
, "Number of switch instructions turned into bitmaps", {0}, {
false}}
;
130STATISTIC(NumLinearMaps,static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
131 "Number of switch instructions turned into linear mapping")static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
;
132STATISTIC(NumLookupTables,static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
133 "Number of switch instructions turned into lookup tables")static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
;
134STATISTIC(static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
135 NumLookupTablesHoles,static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
136 "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}}
;
137STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares")static llvm::Statistic NumTableCmpReuses = {"simplifycfg", "NumTableCmpReuses"
, "Number of reused switch table lookup compares", {0}, {false
}}
;
138STATISTIC(NumSinkCommons,static llvm::Statistic NumSinkCommons = {"simplifycfg", "NumSinkCommons"
, "Number of common instructions sunk down to the end block",
{0}, {false}}
139 "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}}
;
140STATISTIC(NumSpeculations, "Number of speculative executed instructions")static llvm::Statistic NumSpeculations = {"simplifycfg", "NumSpeculations"
, "Number of speculative executed instructions", {0}, {false}
}
;
141
142namespace {
143
144// The first field contains the value that the switch produces when a certain
145// case group is selected, and the second field is a vector containing the
146// cases composing the case group.
147using SwitchCaseResultVectorTy =
148 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
149
150// The first field contains the phi node that generates a result of the switch
151// and the second field contains the value generated for a certain case in the
152// switch for that PHI.
153using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
154
155/// ValueEqualityComparisonCase - Represents a case of a switch.
156struct ValueEqualityComparisonCase {
157 ConstantInt *Value;
158 BasicBlock *Dest;
159
160 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
161 : Value(Value), Dest(Dest) {}
162
163 bool operator<(ValueEqualityComparisonCase RHS) const {
164 // Comparing pointers is ok as we only rely on the order for uniquing.
165 return Value < RHS.Value;
166 }
167
168 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
169};
170
171class SimplifyCFGOpt {
172 const TargetTransformInfo &TTI;
173 const DataLayout &DL;
174 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
175 const SimplifyCFGOptions &Options;
176 bool Resimplify;
177
178 Value *isValueEqualityComparison(Instruction *TI);
179 BasicBlock *GetValueEqualityComparisonCases(
180 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181 bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
182 BasicBlock *Pred,
183 IRBuilder<> &Builder);
184 bool FoldValueComparisonIntoPredecessors(Instruction *TI,
185 IRBuilder<> &Builder);
186
187 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189 bool SimplifySingleResume(ResumeInst *RI);
190 bool SimplifyCommonResume(ResumeInst *RI);
191 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192 bool SimplifyUnreachable(UnreachableInst *UI);
193 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194 bool SimplifyIndirectBr(IndirectBrInst *IBI);
195 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197
198 bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
199 IRBuilder<> &Builder);
200
201public:
202 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
203 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
204 const SimplifyCFGOptions &Opts)
205 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
206
207 bool run(BasicBlock *BB);
208 bool simplifyOnce(BasicBlock *BB);
209
210 // Helper to set Resimplify and return change indication.
211 bool requestResimplify() {
212 Resimplify = true;
213 return true;
214 }
215};
216
217} // end anonymous namespace
218
219/// Return true if it is safe to merge these two
220/// terminator instructions together.
221static bool
222SafeToMergeTerminators(Instruction *SI1, Instruction *SI2,
223 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
224 if (SI1 == SI2)
225 return false; // Can't merge with self!
226
227 // It is not safe to merge these two switch instructions if they have a common
228 // successor, and if that successor has a PHI node, and if *that* PHI node has
229 // conflicting incoming values from the two switch blocks.
230 BasicBlock *SI1BB = SI1->getParent();
231 BasicBlock *SI2BB = SI2->getParent();
232
233 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
234 bool Fail = false;
235 for (BasicBlock *Succ : successors(SI2BB))
236 if (SI1Succs.count(Succ))
237 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
238 PHINode *PN = cast<PHINode>(BBI);
239 if (PN->getIncomingValueForBlock(SI1BB) !=
240 PN->getIncomingValueForBlock(SI2BB)) {
241 if (FailBlocks)
242 FailBlocks->insert(Succ);
243 Fail = true;
244 }
245 }
246
247 return !Fail;
248}
249
250/// Return true if it is safe and profitable to merge these two terminator
251/// instructions together, where SI1 is an unconditional branch. PhiNodes will
252/// store all PHI nodes in common successors.
253static bool
254isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
255 Instruction *Cond,
256 SmallVectorImpl<PHINode *> &PhiNodes) {
257 if (SI1 == SI2)
258 return false; // Can't merge with self!
259 assert(SI1->isUnconditional() && SI2->isConditional())((SI1->isUnconditional() && SI2->isConditional(
)) ? static_cast<void> (0) : __assert_fail ("SI1->isUnconditional() && SI2->isConditional()"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 259, __PRETTY_FUNCTION__))
;
260
261 // We fold the unconditional branch if we can easily update all PHI nodes in
262 // common successors:
263 // 1> We have a constant incoming value for the conditional branch;
264 // 2> We have "Cond" as the incoming value for the unconditional branch;
265 // 3> SI2->getCondition() and Cond have same operands.
266 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
267 if (!Ci2)
268 return false;
269 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
270 Cond->getOperand(1) == Ci2->getOperand(1)) &&
271 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
272 Cond->getOperand(1) == Ci2->getOperand(0)))
273 return false;
274
275 BasicBlock *SI1BB = SI1->getParent();
276 BasicBlock *SI2BB = SI2->getParent();
277 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
278 for (BasicBlock *Succ : successors(SI2BB))
279 if (SI1Succs.count(Succ))
280 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
281 PHINode *PN = cast<PHINode>(BBI);
282 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
283 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
284 return false;
285 PhiNodes.push_back(PN);
286 }
287 return true;
288}
289
290/// Update PHI nodes in Succ to indicate that there will now be entries in it
291/// from the 'NewPred' block. The values that will be flowing into the PHI nodes
292/// will be the same as those coming in from ExistPred, an existing predecessor
293/// of Succ.
294static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
295 BasicBlock *ExistPred) {
296 for (PHINode &PN : Succ->phis())
297 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
298}
299
300/// Compute an abstract "cost" of speculating the given instruction,
301/// which is assumed to be safe to speculate. TCC_Free means cheap,
302/// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
303/// expensive.
304static unsigned ComputeSpeculationCost(const User *I,
305 const TargetTransformInfo &TTI) {
306 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 307, __PRETTY_FUNCTION__))
307 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 307, __PRETTY_FUNCTION__))
;
308 return TTI.getUserCost(I);
309}
310
311/// If we have a merge point of an "if condition" as accepted above,
312/// return true if the specified value dominates the block. We
313/// don't handle the true generality of domination here, just a special case
314/// which works well enough for us.
315///
316/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
317/// see if V (which must be an instruction) and its recursive operands
318/// that do not dominate BB have a combined cost lower than CostRemaining and
319/// are non-trapping. If both are true, the instruction is inserted into the
320/// set and true is returned.
321///
322/// The cost for most non-trapping instructions is defined as 1 except for
323/// Select whose cost is 2.
324///
325/// After this function returns, CostRemaining is decreased by the cost of
326/// V plus its non-dominating operands. If that cost is greater than
327/// CostRemaining, false is returned and CostRemaining is undefined.
328static bool DominatesMergePoint(Value *V, BasicBlock *BB,
329 SmallPtrSetImpl<Instruction *> &AggressiveInsts,
330 unsigned &CostRemaining,
331 const TargetTransformInfo &TTI,
332 unsigned Depth = 0) {
333 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
334 // so limit the recursion depth.
335 // TODO: While this recursion limit does prevent pathological behavior, it
336 // would be better to track visited instructions to avoid cycles.
337 if (Depth == MaxSpeculationDepth)
338 return false;
339
340 Instruction *I = dyn_cast<Instruction>(V);
341 if (!I) {
342 // Non-instructions all dominate instructions, but not all constantexprs
343 // can be executed unconditionally.
344 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
345 if (C->canTrap())
346 return false;
347 return true;
348 }
349 BasicBlock *PBB = I->getParent();
350
351 // We don't want to allow weird loops that might have the "if condition" in
352 // the bottom of this block.
353 if (PBB == BB)
354 return false;
355
356 // If this instruction is defined in a block that contains an unconditional
357 // branch to BB, then it must be in the 'conditional' part of the "if
358 // statement". If not, it definitely dominates the region.
359 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
360 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
361 return true;
362
363 // If we have seen this instruction before, don't count it again.
364 if (AggressiveInsts.count(I))
365 return true;
366
367 // Okay, it looks like the instruction IS in the "condition". Check to
368 // see if it's a cheap instruction to unconditionally compute, and if it
369 // only uses stuff defined outside of the condition. If so, hoist it out.
370 if (!isSafeToSpeculativelyExecute(I))
371 return false;
372
373 unsigned Cost = ComputeSpeculationCost(I, TTI);
374
375 // Allow exactly one instruction to be speculated regardless of its cost
376 // (as long as it is safe to do so).
377 // This is intended to flatten the CFG even if the instruction is a division
378 // or other expensive operation. The speculation of an expensive instruction
379 // is expected to be undone in CodeGenPrepare if the speculation has not
380 // enabled further IR optimizations.
381 if (Cost > CostRemaining &&
382 (!SpeculateOneExpensiveInst || !AggressiveInsts.empty() || Depth > 0))
383 return false;
384
385 // Avoid unsigned wrap.
386 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
387
388 // Okay, we can only really hoist these out if their operands do
389 // not take us over the cost threshold.
390 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
391 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
392 Depth + 1))
393 return false;
394 // Okay, it's safe to do this! Remember this instruction.
395 AggressiveInsts.insert(I);
396 return true;
397}
398
399/// Extract ConstantInt from value, looking through IntToPtr
400/// and PointerNullValue. Return NULL if value is not a constant int.
401static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
402 // Normal constant int.
403 ConstantInt *CI = dyn_cast<ConstantInt>(V);
404 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
405 return CI;
406
407 // This is some kind of pointer constant. Turn it into a pointer-sized
408 // ConstantInt if possible.
409 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
410
411 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
412 if (isa<ConstantPointerNull>(V))
413 return ConstantInt::get(PtrTy, 0);
414
415 // IntToPtr const int.
416 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
417 if (CE->getOpcode() == Instruction::IntToPtr)
418 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
419 // The constant is very likely to have the right type already.
420 if (CI->getType() == PtrTy)
421 return CI;
422 else
423 return cast<ConstantInt>(
424 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
425 }
426 return nullptr;
427}
428
429namespace {
430
431/// Given a chain of or (||) or and (&&) comparison of a value against a
432/// constant, this will try to recover the information required for a switch
433/// structure.
434/// It will depth-first traverse the chain of comparison, seeking for patterns
435/// like %a == 12 or %a < 4 and combine them to produce a set of integer
436/// representing the different cases for the switch.
437/// Note that if the chain is composed of '||' it will build the set of elements
438/// that matches the comparisons (i.e. any of this value validate the chain)
439/// while for a chain of '&&' it will build the set elements that make the test
440/// fail.
441struct ConstantComparesGatherer {
442 const DataLayout &DL;
443
444 /// Value found for the switch comparison
445 Value *CompValue = nullptr;
446
447 /// Extra clause to be checked before the switch
448 Value *Extra = nullptr;
449
450 /// Set of integers to match in switch
451 SmallVector<ConstantInt *, 8> Vals;
452
453 /// Number of comparisons matched in the and/or chain
454 unsigned UsedICmps = 0;
455
456 /// Construct and compute the result for the comparison instruction Cond
457 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
458 gather(Cond);
459 }
460
461 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
462 ConstantComparesGatherer &
463 operator=(const ConstantComparesGatherer &) = delete;
464
465private:
466 /// Try to set the current value used for the comparison, it succeeds only if
467 /// it wasn't set before or if the new value is the same as the old one
468 bool setValueOnce(Value *NewVal) {
469 if (CompValue && CompValue != NewVal)
470 return false;
471 CompValue = NewVal;
472 return (CompValue != nullptr);
473 }
474
475 /// Try to match Instruction "I" as a comparison against a constant and
476 /// populates the array Vals with the set of values that match (or do not
477 /// match depending on isEQ).
478 /// Return false on failure. On success, the Value the comparison matched
479 /// against is placed in CompValue.
480 /// If CompValue is already set, the function is expected to fail if a match
481 /// is found but the value compared to is different.
482 bool matchInstruction(Instruction *I, bool isEQ) {
483 // If this is an icmp against a constant, handle this as one of the cases.
484 ICmpInst *ICI;
485 ConstantInt *C;
486 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
487 (C = GetConstantInt(I->getOperand(1), DL)))) {
488 return false;
489 }
490
491 Value *RHSVal;
492 const APInt *RHSC;
493
494 // Pattern match a special case
495 // (x & ~2^z) == y --> x == y || x == y|2^z
496 // This undoes a transformation done by instcombine to fuse 2 compares.
497 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
498 // It's a little bit hard to see why the following transformations are
499 // correct. Here is a CVC3 program to verify them for 64-bit values:
500
501 /*
502 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
503 x : BITVECTOR(64);
504 y : BITVECTOR(64);
505 z : BITVECTOR(64);
506 mask : BITVECTOR(64) = BVSHL(ONE, z);
507 QUERY( (y & ~mask = y) =>
508 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
509 );
510 QUERY( (y | mask = y) =>
511 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
512 );
513 */
514
515 // Please note that each pattern must be a dual implication (<--> or
516 // iff). One directional implication can create spurious matches. If the
517 // implication is only one-way, an unsatisfiable condition on the left
518 // side can imply a satisfiable condition on the right side. Dual
519 // implication ensures that satisfiable conditions are transformed to
520 // other satisfiable conditions and unsatisfiable conditions are
521 // transformed to other unsatisfiable conditions.
522
523 // Here is a concrete example of a unsatisfiable condition on the left
524 // implying a satisfiable condition on the right:
525 //
526 // mask = (1 << z)
527 // (x & ~mask) == y --> (x == y || x == (y | mask))
528 //
529 // Substituting y = 3, z = 0 yields:
530 // (x & -2) == 3 --> (x == 3 || x == 2)
531
532 // Pattern match a special case:
533 /*
534 QUERY( (y & ~mask = y) =>
535 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
536 );
537 */
538 if (match(ICI->getOperand(0),
539 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
540 APInt Mask = ~*RHSC;
541 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
542 // If we already have a value for the switch, it has to match!
543 if (!setValueOnce(RHSVal))
544 return false;
545
546 Vals.push_back(C);
547 Vals.push_back(
548 ConstantInt::get(C->getContext(),
549 C->getValue() | Mask));
550 UsedICmps++;
551 return true;
552 }
553 }
554
555 // Pattern match a special case:
556 /*
557 QUERY( (y | mask = y) =>
558 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
559 );
560 */
561 if (match(ICI->getOperand(0),
562 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
563 APInt Mask = *RHSC;
564 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
565 // If we already have a value for the switch, it has to match!
566 if (!setValueOnce(RHSVal))
567 return false;
568
569 Vals.push_back(C);
570 Vals.push_back(ConstantInt::get(C->getContext(),
571 C->getValue() & ~Mask));
572 UsedICmps++;
573 return true;
574 }
575 }
576
577 // If we already have a value for the switch, it has to match!
578 if (!setValueOnce(ICI->getOperand(0)))
579 return false;
580
581 UsedICmps++;
582 Vals.push_back(C);
583 return ICI->getOperand(0);
584 }
585
586 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
587 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
588 ICI->getPredicate(), C->getValue());
589
590 // Shift the range if the compare is fed by an add. This is the range
591 // compare idiom as emitted by instcombine.
592 Value *CandidateVal = I->getOperand(0);
593 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
594 Span = Span.subtract(*RHSC);
595 CandidateVal = RHSVal;
596 }
597
598 // If this is an and/!= check, then we are looking to build the set of
599 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
600 // x != 0 && x != 1.
601 if (!isEQ)
602 Span = Span.inverse();
603
604 // If there are a ton of values, we don't want to make a ginormous switch.
605 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
606 return false;
607 }
608
609 // If we already have a value for the switch, it has to match!
610 if (!setValueOnce(CandidateVal))
611 return false;
612
613 // Add all values from the range to the set
614 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
615 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
616
617 UsedICmps++;
618 return true;
619 }
620
621 /// Given a potentially 'or'd or 'and'd together collection of icmp
622 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
623 /// the value being compared, and stick the list constants into the Vals
624 /// vector.
625 /// One "Extra" case is allowed to differ from the other.
626 void gather(Value *V) {
627 Instruction *I = dyn_cast<Instruction>(V);
628 bool isEQ = (I->getOpcode() == Instruction::Or);
629
630 // Keep a stack (SmallVector for efficiency) for depth-first traversal
631 SmallVector<Value *, 8> DFT;
632 SmallPtrSet<Value *, 8> Visited;
633
634 // Initialize
635 Visited.insert(V);
636 DFT.push_back(V);
637
638 while (!DFT.empty()) {
639 V = DFT.pop_back_val();
640
641 if (Instruction *I = dyn_cast<Instruction>(V)) {
642 // If it is a || (or && depending on isEQ), process the operands.
643 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
644 if (Visited.insert(I->getOperand(1)).second)
645 DFT.push_back(I->getOperand(1));
646 if (Visited.insert(I->getOperand(0)).second)
647 DFT.push_back(I->getOperand(0));
648 continue;
649 }
650
651 // Try to match the current instruction
652 if (matchInstruction(I, isEQ))
653 // Match succeed, continue the loop
654 continue;
655 }
656
657 // One element of the sequence of || (or &&) could not be match as a
658 // comparison against the same value as the others.
659 // We allow only one "Extra" case to be checked before the switch
660 if (!Extra) {
661 Extra = V;
662 continue;
663 }
664 // Failed to parse a proper sequence, abort now
665 CompValue = nullptr;
666 break;
667 }
668 }
669};
670
671} // end anonymous namespace
672
673static void EraseTerminatorAndDCECond(Instruction *TI) {
674 Instruction *Cond = nullptr;
675 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
676 Cond = dyn_cast<Instruction>(SI->getCondition());
677 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
678 if (BI->isConditional())
679 Cond = dyn_cast<Instruction>(BI->getCondition());
680 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
681 Cond = dyn_cast<Instruction>(IBI->getAddress());
682 }
683
684 TI->eraseFromParent();
685 if (Cond)
686 RecursivelyDeleteTriviallyDeadInstructions(Cond);
687}
688
689/// Return true if the specified terminator checks
690/// to see if a value is equal to constant integer value.
691Value *SimplifyCFGOpt::isValueEqualityComparison(Instruction *TI) {
692 Value *CV = nullptr;
693 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
694 // Do not permit merging of large switch instructions into their
695 // predecessors unless there is only one predecessor.
696 if (!SI->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
697 CV = SI->getCondition();
698 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
699 if (BI->isConditional() && BI->getCondition()->hasOneUse())
700 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
701 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
702 CV = ICI->getOperand(0);
703 }
704
705 // Unwrap any lossless ptrtoint cast.
706 if (CV) {
707 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
708 Value *Ptr = PTII->getPointerOperand();
709 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
710 CV = Ptr;
711 }
712 }
713 return CV;
714}
715
716/// Given a value comparison instruction,
717/// decode all of the 'cases' that it represents and return the 'default' block.
718BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
719 Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
720 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
721 Cases.reserve(SI->getNumCases());
722 for (auto Case : SI->cases())
723 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
724 Case.getCaseSuccessor()));
725 return SI->getDefaultDest();
726 }
727
728 BranchInst *BI = cast<BranchInst>(TI);
729 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
730 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
731 Cases.push_back(ValueEqualityComparisonCase(
732 GetConstantInt(ICI->getOperand(1), DL), Succ));
733 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
734}
735
736/// Given a vector of bb/value pairs, remove any entries
737/// in the list that match the specified block.
738static void
739EliminateBlockCases(BasicBlock *BB,
740 std::vector<ValueEqualityComparisonCase> &Cases) {
741 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
742}
743
744/// Return true if there are any keys in C1 that exist in C2 as well.
745static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
746 std::vector<ValueEqualityComparisonCase> &C2) {
747 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
748
749 // Make V1 be smaller than V2.
750 if (V1->size() > V2->size())
751 std::swap(V1, V2);
752
753 if (V1->empty())
754 return false;
755 if (V1->size() == 1) {
756 // Just scan V2.
757 ConstantInt *TheVal = (*V1)[0].Value;
758 for (unsigned i = 0, e = V2->size(); i != e; ++i)
759 if (TheVal == (*V2)[i].Value)
760 return true;
761 }
762
763 // Otherwise, just sort both lists and compare element by element.
764 array_pod_sort(V1->begin(), V1->end());
765 array_pod_sort(V2->begin(), V2->end());
766 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
767 while (i1 != e1 && i2 != e2) {
768 if ((*V1)[i1].Value == (*V2)[i2].Value)
769 return true;
770 if ((*V1)[i1].Value < (*V2)[i2].Value)
771 ++i1;
772 else
773 ++i2;
774 }
775 return false;
776}
777
778// Set branch weights on SwitchInst. This sets the metadata if there is at
779// least one non-zero weight.
780static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
781 // Check that there is at least one non-zero weight. Otherwise, pass
782 // nullptr to setMetadata which will erase the existing metadata.
783 MDNode *N = nullptr;
784 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
785 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
786 SI->setMetadata(LLVMContext::MD_prof, N);
787}
788
789// Similar to the above, but for branch and select instructions that take
790// exactly 2 weights.
791static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
792 uint32_t FalseWeight) {
793 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 793, __PRETTY_FUNCTION__))
;
794 // Check that there is at least one non-zero weight. Otherwise, pass
795 // nullptr to setMetadata which will erase the existing metadata.
796 MDNode *N = nullptr;
797 if (TrueWeight || FalseWeight)
798 N = MDBuilder(I->getParent()->getContext())
799 .createBranchWeights(TrueWeight, FalseWeight);
800 I->setMetadata(LLVMContext::MD_prof, N);
801}
802
803/// If TI is known to be a terminator instruction and its block is known to
804/// only have a single predecessor block, check to see if that predecessor is
805/// also a value comparison with the same value, and if that comparison
806/// determines the outcome of this comparison. If so, simplify TI. This does a
807/// very limited form of jump threading.
808bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
809 Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
810 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
811 if (!PredVal)
812 return false; // Not a value comparison in predecessor.
813
814 Value *ThisVal = isValueEqualityComparison(TI);
815 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 815, __PRETTY_FUNCTION__))
;
816 if (ThisVal != PredVal)
817 return false; // Different predicates.
818
819 // TODO: Preserve branch weight metadata, similarly to how
820 // FoldValueComparisonIntoPredecessors preserves it.
821
822 // Find out information about when control will move from Pred to TI's block.
823 std::vector<ValueEqualityComparisonCase> PredCases;
824 BasicBlock *PredDef =
825 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
826 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
827
828 // Find information about how control leaves this block.
829 std::vector<ValueEqualityComparisonCase> ThisCases;
830 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
831 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
832
833 // If TI's block is the default block from Pred's comparison, potentially
834 // simplify TI based on this knowledge.
835 if (PredDef == TI->getParent()) {
836 // If we are here, we know that the value is none of those cases listed in
837 // PredCases. If there are any cases in ThisCases that are in PredCases, we
838 // can simplify TI.
839 if (!ValuesOverlap(PredCases, ThisCases))
840 return false;
841
842 if (isa<BranchInst>(TI)) {
843 // Okay, one of the successors of this condbr is dead. Convert it to a
844 // uncond br.
845 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 845, __PRETTY_FUNCTION__))
;
846 // Insert the new branch.
847 Instruction *NI = Builder.CreateBr(ThisDef);
848 (void)NI;
849
850 // Remove PHI node entries for the dead edge.
851 ThisCases[0].Dest->removePredecessor(TI->getParent());
852
853 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)
854 << "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)
855 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
856
857 EraseTerminatorAndDCECond(TI);
858 return true;
859 }
860
861 SwitchInst *SI = cast<SwitchInst>(TI);
862 // Okay, TI has cases that are statically dead, prune them away.
863 SmallPtrSet<Constant *, 16> DeadCases;
864 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
865 DeadCases.insert(PredCases[i].Value);
866
867 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)
868 << "Through successor TI: " << *TI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI; } } while (false)
;
869
870 // Collect branch weights into a vector.
871 SmallVector<uint32_t, 8> Weights;
872 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
873 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
874 if (HasWeight)
875 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
876 ++MD_i) {
877 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
878 Weights.push_back(CI->getValue().getZExtValue());
879 }
880 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
881 --i;
882 if (DeadCases.count(i->getCaseValue())) {
883 if (HasWeight) {
884 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
885 Weights.pop_back();
886 }
887 i->getCaseSuccessor()->removePredecessor(TI->getParent());
888 SI->removeCase(i);
889 }
890 }
891 if (HasWeight && Weights.size() >= 2)
892 setBranchWeights(SI, Weights);
893
894 LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Leaving: " << *TI <<
"\n"; } } while (false)
;
895 return true;
896 }
897
898 // Otherwise, TI's block must correspond to some matched value. Find out
899 // which value (or set of values) this is.
900 ConstantInt *TIV = nullptr;
901 BasicBlock *TIBB = TI->getParent();
902 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
903 if (PredCases[i].Dest == TIBB) {
904 if (TIV)
905 return false; // Cannot handle multiple values coming to this block.
906 TIV = PredCases[i].Value;
907 }
908 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 908, __PRETTY_FUNCTION__))
;
909
910 // Okay, we found the one constant that our value can be if we get into TI's
911 // BB. Find out which successor will unconditionally be branched to.
912 BasicBlock *TheRealDest = nullptr;
913 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
914 if (ThisCases[i].Value == TIV) {
915 TheRealDest = ThisCases[i].Dest;
916 break;
917 }
918
919 // If not handled by any explicit cases, it is handled by the default case.
920 if (!TheRealDest)
921 TheRealDest = ThisDef;
922
923 // Remove PHI node entries for dead edges.
924 BasicBlock *CheckEdge = TheRealDest;
925 for (BasicBlock *Succ : successors(TIBB))
926 if (Succ != CheckEdge)
927 Succ->removePredecessor(TIBB);
928 else
929 CheckEdge = nullptr;
930
931 // Insert the new branch.
932 Instruction *NI = Builder.CreateBr(TheRealDest);
933 (void)NI;
934
935 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)
936 << "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)
937 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
938
939 EraseTerminatorAndDCECond(TI);
940 return true;
941}
942
943namespace {
944
945/// This class implements a stable ordering of constant
946/// integers that does not depend on their address. This is important for
947/// applications that sort ConstantInt's to ensure uniqueness.
948struct ConstantIntOrdering {
949 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
950 return LHS->getValue().ult(RHS->getValue());
951 }
952};
953
954} // end anonymous namespace
955
956static int ConstantIntSortPredicate(ConstantInt *const *P1,
957 ConstantInt *const *P2) {
958 const ConstantInt *LHS = *P1;
959 const ConstantInt *RHS = *P2;
960 if (LHS == RHS)
961 return 0;
962 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
963}
964
965static inline bool HasBranchWeights(const Instruction *I) {
966 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
967 if (ProfMD && ProfMD->getOperand(0))
968 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
969 return MDS->getString().equals("branch_weights");
970
971 return false;
972}
973
974/// Get Weights of a given terminator, the default weight is at the front
975/// of the vector. If TI is a conditional eq, we need to swap the branch-weight
976/// metadata.
977static void GetBranchWeights(Instruction *TI,
978 SmallVectorImpl<uint64_t> &Weights) {
979 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
980 assert(MD)((MD) ? static_cast<void> (0) : __assert_fail ("MD", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 980, __PRETTY_FUNCTION__))
;
981 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
982 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
983 Weights.push_back(CI->getValue().getZExtValue());
984 }
985
986 // If TI is a conditional eq, the default case is the false case,
987 // and the corresponding branch-weight data is at index 2. We swap the
988 // default weight to be the first entry.
989 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
990 assert(Weights.size() == 2)((Weights.size() == 2) ? static_cast<void> (0) : __assert_fail
("Weights.size() == 2", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 990, __PRETTY_FUNCTION__))
;
991 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
992 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
993 std::swap(Weights.front(), Weights.back());
994 }
995}
996
997/// Keep halving the weights until all can fit in uint32_t.
998static void FitWeights(MutableArrayRef<uint64_t> Weights) {
999 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1000 if (Max > UINT_MAX(2147483647 *2U +1U)) {
1001 unsigned Offset = 32 - countLeadingZeros(Max);
1002 for (uint64_t &I : Weights)
1003 I >>= Offset;
1004 }
1005}
1006
1007/// The specified terminator is a value equality comparison instruction
1008/// (either a switch or a branch on "X == c").
1009/// See if any of the predecessors of the terminator block are value comparisons
1010/// on the same value. If so, and if safe to do so, fold them together.
1011bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1012 IRBuilder<> &Builder) {
1013 BasicBlock *BB = TI->getParent();
1014 Value *CV = isValueEqualityComparison(TI); // CondVal
1015 assert(CV && "Not a comparison?")((CV && "Not a comparison?") ? static_cast<void>
(0) : __assert_fail ("CV && \"Not a comparison?\"", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1015, __PRETTY_FUNCTION__))
;
1016 bool Changed = false;
1017
1018 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1019 while (!Preds.empty()) {
1020 BasicBlock *Pred = Preds.pop_back_val();
1021
1022 // See if the predecessor is a comparison with the same value.
1023 Instruction *PTI = Pred->getTerminator();
1024 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1025
1026 if (PCV == CV && TI != PTI) {
1027 SmallSetVector<BasicBlock*, 4> FailBlocks;
1028 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1029 for (auto *Succ : FailBlocks) {
1030 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1031 return false;
1032 }
1033 }
1034
1035 // Figure out which 'cases' to copy from SI to PSI.
1036 std::vector<ValueEqualityComparisonCase> BBCases;
1037 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1038
1039 std::vector<ValueEqualityComparisonCase> PredCases;
1040 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1041
1042 // Based on whether the default edge from PTI goes to BB or not, fill in
1043 // PredCases and PredDefault with the new switch cases we would like to
1044 // build.
1045 SmallVector<BasicBlock *, 8> NewSuccessors;
1046
1047 // Update the branch weight metadata along the way
1048 SmallVector<uint64_t, 8> Weights;
1049 bool PredHasWeights = HasBranchWeights(PTI);
1050 bool SuccHasWeights = HasBranchWeights(TI);
1051
1052 if (PredHasWeights) {
1053 GetBranchWeights(PTI, Weights);
1054 // branch-weight metadata is inconsistent here.
1055 if (Weights.size() != 1 + PredCases.size())
1056 PredHasWeights = SuccHasWeights = false;
1057 } else if (SuccHasWeights)
1058 // If there are no predecessor weights but there are successor weights,
1059 // populate Weights with 1, which will later be scaled to the sum of
1060 // successor's weights
1061 Weights.assign(1 + PredCases.size(), 1);
1062
1063 SmallVector<uint64_t, 8> SuccWeights;
1064 if (SuccHasWeights) {
1065 GetBranchWeights(TI, SuccWeights);
1066 // branch-weight metadata is inconsistent here.
1067 if (SuccWeights.size() != 1 + BBCases.size())
1068 PredHasWeights = SuccHasWeights = false;
1069 } else if (PredHasWeights)
1070 SuccWeights.assign(1 + BBCases.size(), 1);
1071
1072 if (PredDefault == BB) {
1073 // If this is the default destination from PTI, only the edges in TI
1074 // that don't occur in PTI, or that branch to BB will be activated.
1075 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1076 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1077 if (PredCases[i].Dest != BB)
1078 PTIHandled.insert(PredCases[i].Value);
1079 else {
1080 // The default destination is BB, we don't need explicit targets.
1081 std::swap(PredCases[i], PredCases.back());
1082
1083 if (PredHasWeights || SuccHasWeights) {
1084 // Increase weight for the default case.
1085 Weights[0] += Weights[i + 1];
1086 std::swap(Weights[i + 1], Weights.back());
1087 Weights.pop_back();
1088 }
1089
1090 PredCases.pop_back();
1091 --i;
1092 --e;
1093 }
1094
1095 // Reconstruct the new switch statement we will be building.
1096 if (PredDefault != BBDefault) {
1097 PredDefault->removePredecessor(Pred);
1098 PredDefault = BBDefault;
1099 NewSuccessors.push_back(BBDefault);
1100 }
1101
1102 unsigned CasesFromPred = Weights.size();
1103 uint64_t ValidTotalSuccWeight = 0;
1104 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1105 if (!PTIHandled.count(BBCases[i].Value) &&
1106 BBCases[i].Dest != BBDefault) {
1107 PredCases.push_back(BBCases[i]);
1108 NewSuccessors.push_back(BBCases[i].Dest);
1109 if (SuccHasWeights || PredHasWeights) {
1110 // The default weight is at index 0, so weight for the ith case
1111 // should be at index i+1. Scale the cases from successor by
1112 // PredDefaultWeight (Weights[0]).
1113 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1114 ValidTotalSuccWeight += SuccWeights[i + 1];
1115 }
1116 }
1117
1118 if (SuccHasWeights || PredHasWeights) {
1119 ValidTotalSuccWeight += SuccWeights[0];
1120 // Scale the cases from predecessor by ValidTotalSuccWeight.
1121 for (unsigned i = 1; i < CasesFromPred; ++i)
1122 Weights[i] *= ValidTotalSuccWeight;
1123 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1124 Weights[0] *= SuccWeights[0];
1125 }
1126 } else {
1127 // If this is not the default destination from PSI, only the edges
1128 // in SI that occur in PSI with a destination of BB will be
1129 // activated.
1130 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1131 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1132 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1133 if (PredCases[i].Dest == BB) {
1134 PTIHandled.insert(PredCases[i].Value);
1135
1136 if (PredHasWeights || SuccHasWeights) {
1137 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1138 std::swap(Weights[i + 1], Weights.back());
1139 Weights.pop_back();
1140 }
1141
1142 std::swap(PredCases[i], PredCases.back());
1143 PredCases.pop_back();
1144 --i;
1145 --e;
1146 }
1147
1148 // Okay, now we know which constants were sent to BB from the
1149 // predecessor. Figure out where they will all go now.
1150 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1151 if (PTIHandled.count(BBCases[i].Value)) {
1152 // If this is one we are capable of getting...
1153 if (PredHasWeights || SuccHasWeights)
1154 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1155 PredCases.push_back(BBCases[i]);
1156 NewSuccessors.push_back(BBCases[i].Dest);
1157 PTIHandled.erase(
1158 BBCases[i].Value); // This constant is taken care of
1159 }
1160
1161 // If there are any constants vectored to BB that TI doesn't handle,
1162 // they must go to the default destination of TI.
1163 for (ConstantInt *I : PTIHandled) {
1164 if (PredHasWeights || SuccHasWeights)
1165 Weights.push_back(WeightsForHandled[I]);
1166 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1167 NewSuccessors.push_back(BBDefault);
1168 }
1169 }
1170
1171 // Okay, at this point, we know which new successor Pred will get. Make
1172 // sure we update the number of entries in the PHI nodes for these
1173 // successors.
1174 for (BasicBlock *NewSuccessor : NewSuccessors)
1175 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1176
1177 Builder.SetInsertPoint(PTI);
1178 // Convert pointer to int before we switch.
1179 if (CV->getType()->isPointerTy()) {
1180 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1181 "magicptr");
1182 }
1183
1184 // Now that the successors are updated, create the new Switch instruction.
1185 SwitchInst *NewSI =
1186 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1187 NewSI->setDebugLoc(PTI->getDebugLoc());
1188 for (ValueEqualityComparisonCase &V : PredCases)
1189 NewSI->addCase(V.Value, V.Dest);
1190
1191 if (PredHasWeights || SuccHasWeights) {
1192 // Halve the weights if any of them cannot fit in an uint32_t
1193 FitWeights(Weights);
1194
1195 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1196
1197 setBranchWeights(NewSI, MDWeights);
1198 }
1199
1200 EraseTerminatorAndDCECond(PTI);
1201
1202 // Okay, last check. If BB is still a successor of PSI, then we must
1203 // have an infinite loop case. If so, add an infinitely looping block
1204 // to handle the case to preserve the behavior of the code.
1205 BasicBlock *InfLoopBlock = nullptr;
1206 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1207 if (NewSI->getSuccessor(i) == BB) {
1208 if (!InfLoopBlock) {
1209 // Insert it at the end of the function, because it's either code,
1210 // or it won't matter if it's hot. :)
1211 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1212 BB->getParent());
1213 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1214 }
1215 NewSI->setSuccessor(i, InfLoopBlock);
1216 }
1217
1218 Changed = true;
1219 }
1220 }
1221 return Changed;
1222}
1223
1224// If we would need to insert a select that uses the value of this invoke
1225// (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1226// can't hoist the invoke, as there is nowhere to put the select in this case.
1227static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1228 Instruction *I1, Instruction *I2) {
1229 for (BasicBlock *Succ : successors(BB1)) {
1230 for (const PHINode &PN : Succ->phis()) {
1231 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1232 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1233 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1234 return false;
1235 }
1236 }
1237 }
1238 return true;
1239}
1240
1241static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1242
1243/// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1244/// in the two blocks up into the branch block. The caller of this function
1245/// guarantees that BI's block dominates BB1 and BB2.
1246static bool HoistThenElseCodeToIf(BranchInst *BI,
1247 const TargetTransformInfo &TTI) {
1248 // This does very trivial matching, with limited scanning, to find identical
1249 // instructions in the two blocks. In particular, we don't want to get into
1250 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1251 // such, we currently just scan for obviously identical instructions in an
1252 // identical order.
1253 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1254 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1255
1256 BasicBlock::iterator BB1_Itr = BB1->begin();
1257 BasicBlock::iterator BB2_Itr = BB2->begin();
1258
1259 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1260 // Skip debug info if it is not identical.
1261 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1262 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1263 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1264 while (isa<DbgInfoIntrinsic>(I1))
1265 I1 = &*BB1_Itr++;
1266 while (isa<DbgInfoIntrinsic>(I2))
1267 I2 = &*BB2_Itr++;
1268 }
1269 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1270 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1271 return false;
1272
1273 BasicBlock *BIParent = BI->getParent();
1274
1275 bool Changed = false;
1276 do {
1277 // If we are hoisting the terminator instruction, don't move one (making a
1278 // broken BB), instead clone it, and remove BI.
1279 if (I1->isTerminator())
1280 goto HoistTerminator;
1281
1282 // If we're going to hoist a call, make sure that the two instructions we're
1283 // commoning/hoisting are both marked with musttail, or neither of them is
1284 // marked as such. Otherwise, we might end up in a situation where we hoist
1285 // from a block where the terminator is a `ret` to a block where the terminator
1286 // is a `br`, and `musttail` calls expect to be followed by a return.
1287 auto *C1 = dyn_cast<CallInst>(I1);
1288 auto *C2 = dyn_cast<CallInst>(I2);
1289 if (C1 && C2)
1290 if (C1->isMustTailCall() != C2->isMustTailCall())
1291 return Changed;
1292
1293 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1294 return Changed;
1295
1296 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1297 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1297, __PRETTY_FUNCTION__))
;
1298 // The debug location is an integral part of a debug info intrinsic
1299 // and can't be separated from it or replaced. Instead of attempting
1300 // to merge locations, simply hoist both copies of the intrinsic.
1301 BIParent->getInstList().splice(BI->getIterator(),
1302 BB1->getInstList(), I1);
1303 BIParent->getInstList().splice(BI->getIterator(),
1304 BB2->getInstList(), I2);
1305 Changed = true;
1306 } else {
1307 // For a normal instruction, we just move one to right before the branch,
1308 // then replace all uses of the other with the first. Finally, we remove
1309 // the now redundant second instruction.
1310 BIParent->getInstList().splice(BI->getIterator(),
1311 BB1->getInstList(), I1);
1312 if (!I2->use_empty())
1313 I2->replaceAllUsesWith(I1);
1314 I1->andIRFlags(I2);
1315 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1316 LLVMContext::MD_range,
1317 LLVMContext::MD_fpmath,
1318 LLVMContext::MD_invariant_load,
1319 LLVMContext::MD_nonnull,
1320 LLVMContext::MD_invariant_group,
1321 LLVMContext::MD_align,
1322 LLVMContext::MD_dereferenceable,
1323 LLVMContext::MD_dereferenceable_or_null,
1324 LLVMContext::MD_mem_parallel_loop_access,
1325 LLVMContext::MD_access_group};
1326 combineMetadata(I1, I2, KnownIDs, true);
1327
1328 // I1 and I2 are being combined into a single instruction. Its debug
1329 // location is the merged locations of the original instructions.
1330 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1331
1332 I2->eraseFromParent();
1333 Changed = true;
1334 }
1335
1336 I1 = &*BB1_Itr++;
1337 I2 = &*BB2_Itr++;
1338 // Skip debug info if it is not identical.
1339 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1340 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1341 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1342 while (isa<DbgInfoIntrinsic>(I1))
1343 I1 = &*BB1_Itr++;
1344 while (isa<DbgInfoIntrinsic>(I2))
1345 I2 = &*BB2_Itr++;
1346 }
1347 } while (I1->isIdenticalToWhenDefined(I2));
1348
1349 return true;
1350
1351HoistTerminator:
1352 // It may not be possible to hoist an invoke.
1353 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1354 return Changed;
1355
1356 for (BasicBlock *Succ : successors(BB1)) {
1357 for (PHINode &PN : Succ->phis()) {
1358 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1359 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1360 if (BB1V == BB2V)
1361 continue;
1362
1363 // Check for passingValueIsAlwaysUndefined here because we would rather
1364 // eliminate undefined control flow then converting it to a select.
1365 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1366 passingValueIsAlwaysUndefined(BB2V, &PN))
1367 return Changed;
1368
1369 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1370 return Changed;
1371 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1372 return Changed;
1373 }
1374 }
1375
1376 // As the parent basic block terminator is a branch instruction which is
1377 // removed at the end of the current transformation, use its previous
1378 // non-debug instruction, as the reference insertion point, which will
1379 // provide the debug location for generated select instructions. For BBs
1380 // with only debug instructions, use an empty debug location.
1381 Instruction *InsertPt =
1382 BIParent->getTerminator()->getPrevNonDebugInstruction();
1383
1384 // Okay, it is safe to hoist the terminator.
1385 Instruction *NT = I1->clone();
1386 BIParent->getInstList().insert(BI->getIterator(), NT);
1387 if (!NT->getType()->isVoidTy()) {
1388 I1->replaceAllUsesWith(NT);
1389 I2->replaceAllUsesWith(NT);
1390 NT->takeName(I1);
1391 }
1392
1393 // Ensure terminator gets a debug location, even an unknown one, in case
1394 // it involves inlinable calls.
1395 NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1396
1397 IRBuilder<NoFolder> Builder(NT);
1398 // If an earlier instruction in this BB had a location, adopt it, otherwise
1399 // clear debug locations.
1400 Builder.SetCurrentDebugLocation(InsertPt ? InsertPt->getDebugLoc()
1401 : DebugLoc());
1402
1403 // Hoisting one of the terminators from our successor is a great thing.
1404 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1405 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1406 // nodes, so we insert select instruction to compute the final result.
1407 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1408 for (BasicBlock *Succ : successors(BB1)) {
1409 for (PHINode &PN : Succ->phis()) {
1410 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1411 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1412 if (BB1V == BB2V)
1413 continue;
1414
1415 // These values do not agree. Insert a select instruction before NT
1416 // that determines the right value.
1417 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1418 if (!SI)
1419 SI = cast<SelectInst>(
1420 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1421 BB1V->getName() + "." + BB2V->getName(), BI));
1422
1423 // Make the PHI node use the select for all incoming values for BB1/BB2
1424 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1425 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1426 PN.setIncomingValue(i, SI);
1427 }
1428 }
1429
1430 // Update any PHI nodes in our new successors.
1431 for (BasicBlock *Succ : successors(BB1))
1432 AddPredecessorToBlock(Succ, BIParent, BB1);
1433
1434 EraseTerminatorAndDCECond(BI);
1435 return true;
1436}
1437
1438// All instructions in Insts belong to different blocks that all unconditionally
1439// branch to a common successor. Analyze each instruction and return true if it
1440// would be possible to sink them into their successor, creating one common
1441// instruction instead. For every value that would be required to be provided by
1442// PHI node (because an operand varies in each input block), add to PHIOperands.
1443static bool canSinkInstructions(
1444 ArrayRef<Instruction *> Insts,
1445 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1446 // Prune out obviously bad instructions to move. Any non-store instruction
1447 // must have exactly one use, and we check later that use is by a single,
1448 // common PHI instruction in the successor.
1449 for (auto *I : Insts) {
1450 // These instructions may change or break semantics if moved.
1451 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1452 I->getType()->isTokenTy())
1453 return false;
1454
1455 // Conservatively return false if I is an inline-asm instruction. Sinking
1456 // and merging inline-asm instructions can potentially create arguments
1457 // that cannot satisfy the inline-asm constraints.
1458 if (const auto *C = dyn_cast<CallInst>(I))
1459 if (C->isInlineAsm())
1460 return false;
1461
1462 // Everything must have only one use too, apart from stores which
1463 // have no uses.
1464 if (!isa<StoreInst>(I) && !I->hasOneUse())
1465 return false;
1466 }
1467
1468 const Instruction *I0 = Insts.front();
1469 for (auto *I : Insts)
1470 if (!I->isSameOperationAs(I0))
1471 return false;
1472
1473 // All instructions in Insts are known to be the same opcode. If they aren't
1474 // stores, check the only user of each is a PHI or in the same block as the
1475 // instruction, because if a user is in the same block as an instruction
1476 // we're contemplating sinking, it must already be determined to be sinkable.
1477 if (!isa<StoreInst>(I0)) {
1478 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1479 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1480 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1481 auto *U = cast<Instruction>(*I->user_begin());
1482 return (PNUse &&
1483 PNUse->getParent() == Succ &&
1484 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1485 U->getParent() == I->getParent();
1486 }))
1487 return false;
1488 }
1489
1490 // Because SROA can't handle speculating stores of selects, try not
1491 // to sink loads or stores of allocas when we'd have to create a PHI for
1492 // the address operand. Also, because it is likely that loads or stores
1493 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1494 // This can cause code churn which can have unintended consequences down
1495 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1496 // FIXME: This is a workaround for a deficiency in SROA - see
1497 // https://llvm.org/bugs/show_bug.cgi?id=30188
1498 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1499 return isa<AllocaInst>(I->getOperand(1));
1500 }))
1501 return false;
1502 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1503 return isa<AllocaInst>(I->getOperand(0));
1504 }))
1505 return false;
1506
1507 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1508 if (I0->getOperand(OI)->getType()->isTokenTy())
1509 // Don't touch any operand of token type.
1510 return false;
1511
1512 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1513 assert(I->getNumOperands() == I0->getNumOperands())((I->getNumOperands() == I0->getNumOperands()) ? static_cast
<void> (0) : __assert_fail ("I->getNumOperands() == I0->getNumOperands()"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1513, __PRETTY_FUNCTION__))
;
1514 return I->getOperand(OI) == I0->getOperand(OI);
1515 };
1516 if (!all_of(Insts, SameAsI0)) {
1517 if (!canReplaceOperandWithVariable(I0, OI))
1518 // We can't create a PHI from this GEP.
1519 return false;
1520 // Don't create indirect calls! The called value is the final operand.
1521 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1522 // FIXME: if the call was *already* indirect, we should do this.
1523 return false;
1524 }
1525 for (auto *I : Insts)
1526 PHIOperands[I].push_back(I->getOperand(OI));
1527 }
1528 }
1529 return true;
1530}
1531
1532// Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1533// instruction of every block in Blocks to their common successor, commoning
1534// into one instruction.
1535static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1536 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1537
1538 // canSinkLastInstruction returning true guarantees that every block has at
1539 // least one non-terminator instruction.
1540 SmallVector<Instruction*,4> Insts;
1541 for (auto *BB : Blocks) {
1542 Instruction *I = BB->getTerminator();
1543 do {
1544 I = I->getPrevNode();
1545 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1546 if (!isa<DbgInfoIntrinsic>(I))
1547 Insts.push_back(I);
1548 }
1549
1550 // The only checking we need to do now is that all users of all instructions
1551 // are the same PHI node. canSinkLastInstruction should have checked this but
1552 // it is slightly over-aggressive - it gets confused by commutative instructions
1553 // so double-check it here.
1554 Instruction *I0 = Insts.front();
1555 if (!isa<StoreInst>(I0)) {
1556 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1557 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1558 auto *U = cast<Instruction>(*I->user_begin());
1559 return U == PNUse;
1560 }))
1561 return false;
1562 }
1563
1564 // We don't need to do any more checking here; canSinkLastInstruction should
1565 // have done it all for us.
1566 SmallVector<Value*, 4> NewOperands;
1567 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1568 // This check is different to that in canSinkLastInstruction. There, we
1569 // cared about the global view once simplifycfg (and instcombine) have
1570 // completed - it takes into account PHIs that become trivially
1571 // simplifiable. However here we need a more local view; if an operand
1572 // differs we create a PHI and rely on instcombine to clean up the very
1573 // small mess we may make.
1574 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1575 return I->getOperand(O) != I0->getOperand(O);
1576 });
1577 if (!NeedPHI) {
1578 NewOperands.push_back(I0->getOperand(O));
1579 continue;
1580 }
1581
1582 // Create a new PHI in the successor block and populate it.
1583 auto *Op = I0->getOperand(O);
1584 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1584, __PRETTY_FUNCTION__))
;
1585 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1586 Op->getName() + ".sink", &BBEnd->front());
1587 for (auto *I : Insts)
1588 PN->addIncoming(I->getOperand(O), I->getParent());
1589 NewOperands.push_back(PN);
1590 }
1591
1592 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1593 // and move it to the start of the successor block.
1594 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1595 I0->getOperandUse(O).set(NewOperands[O]);
1596 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1597
1598 // Update metadata and IR flags, and merge debug locations.
1599 for (auto *I : Insts)
1600 if (I != I0) {
1601 // The debug location for the "common" instruction is the merged locations
1602 // of all the commoned instructions. We start with the original location
1603 // of the "common" instruction and iteratively merge each location in the
1604 // loop below.
1605 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1606 // However, as N-way merge for CallInst is rare, so we use simplified API
1607 // instead of using complex API for N-way merge.
1608 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1609 combineMetadataForCSE(I0, I, true);
1610 I0->andIRFlags(I);
1611 }
1612
1613 if (!isa<StoreInst>(I0)) {
1614 // canSinkLastInstruction checked that all instructions were used by
1615 // one and only one PHI node. Find that now, RAUW it to our common
1616 // instruction and nuke it.
1617 assert(I0->hasOneUse())((I0->hasOneUse()) ? static_cast<void> (0) : __assert_fail
("I0->hasOneUse()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1617, __PRETTY_FUNCTION__))
;
1618 auto *PN = cast<PHINode>(*I0->user_begin());
1619 PN->replaceAllUsesWith(I0);
1620 PN->eraseFromParent();
1621 }
1622
1623 // Finally nuke all instructions apart from the common instruction.
1624 for (auto *I : Insts)
1625 if (I != I0)
1626 I->eraseFromParent();
1627
1628 return true;
1629}
1630
1631namespace {
1632
1633 // LockstepReverseIterator - Iterates through instructions
1634 // in a set of blocks in reverse order from the first non-terminator.
1635 // For example (assume all blocks have size n):
1636 // LockstepReverseIterator I([B1, B2, B3]);
1637 // *I-- = [B1[n], B2[n], B3[n]];
1638 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1639 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1640 // ...
1641 class LockstepReverseIterator {
1642 ArrayRef<BasicBlock*> Blocks;
1643 SmallVector<Instruction*,4> Insts;
1644 bool Fail;
1645
1646 public:
1647 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1648 reset();
1649 }
1650
1651 void reset() {
1652 Fail = false;
1653 Insts.clear();
1654 for (auto *BB : Blocks) {
1655 Instruction *Inst = BB->getTerminator();
1656 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1657 Inst = Inst->getPrevNode();
1658 if (!Inst) {
1659 // Block wasn't big enough.
1660 Fail = true;
1661 return;
1662 }
1663 Insts.push_back(Inst);
1664 }
1665 }
1666
1667 bool isValid() const {
1668 return !Fail;
1669 }
1670
1671 void operator--() {
1672 if (Fail)
1673 return;
1674 for (auto *&Inst : Insts) {
1675 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1676 Inst = Inst->getPrevNode();
1677 // Already at beginning of block.
1678 if (!Inst) {
1679 Fail = true;
1680 return;
1681 }
1682 }
1683 }
1684
1685 ArrayRef<Instruction*> operator * () const {
1686 return Insts;
1687 }
1688 };
1689
1690} // end anonymous namespace
1691
1692/// Check whether BB's predecessors end with unconditional branches. If it is
1693/// true, sink any common code from the predecessors to BB.
1694/// We also allow one predecessor to end with conditional branch (but no more
1695/// than one).
1696static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1697 // We support two situations:
1698 // (1) all incoming arcs are unconditional
1699 // (2) one incoming arc is conditional
1700 //
1701 // (2) is very common in switch defaults and
1702 // else-if patterns;
1703 //
1704 // if (a) f(1);
1705 // else if (b) f(2);
1706 //
1707 // produces:
1708 //
1709 // [if]
1710 // / \
1711 // [f(1)] [if]
1712 // | | \
1713 // | | |
1714 // | [f(2)]|
1715 // \ | /
1716 // [ end ]
1717 //
1718 // [end] has two unconditional predecessor arcs and one conditional. The
1719 // conditional refers to the implicit empty 'else' arc. This conditional
1720 // arc can also be caused by an empty default block in a switch.
1721 //
1722 // In this case, we attempt to sink code from all *unconditional* arcs.
1723 // If we can sink instructions from these arcs (determined during the scan
1724 // phase below) we insert a common successor for all unconditional arcs and
1725 // connect that to [end], to enable sinking:
1726 //
1727 // [if]
1728 // / \
1729 // [x(1)] [if]
1730 // | | \
1731 // | | \
1732 // | [x(2)] |
1733 // \ / |
1734 // [sink.split] |
1735 // \ /
1736 // [ end ]
1737 //
1738 SmallVector<BasicBlock*,4> UnconditionalPreds;
1739 Instruction *Cond = nullptr;
1740 for (auto *B : predecessors(BB)) {
1741 auto *T = B->getTerminator();
1742 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1743 UnconditionalPreds.push_back(B);
1744 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1745 Cond = T;
1746 else
1747 return false;
1748 }
1749 if (UnconditionalPreds.size() < 2)
1750 return false;
1751
1752 bool Changed = false;
1753 // We take a two-step approach to tail sinking. First we scan from the end of
1754 // each block upwards in lockstep. If the n'th instruction from the end of each
1755 // block can be sunk, those instructions are added to ValuesToSink and we
1756 // carry on. If we can sink an instruction but need to PHI-merge some operands
1757 // (because they're not identical in each instruction) we add these to
1758 // PHIOperands.
1759 unsigned ScanIdx = 0;
1760 SmallPtrSet<Value*,4> InstructionsToSink;
1761 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1762 LockstepReverseIterator LRI(UnconditionalPreds);
1763 while (LRI.isValid() &&
1764 canSinkInstructions(*LRI, PHIOperands)) {
1765 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)
1766 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: instruction can be sunk: "
<< *(*LRI)[0] << "\n"; } } while (false)
;
1767 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1768 ++ScanIdx;
1769 --LRI;
1770 }
1771
1772 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1773 unsigned NumPHIdValues = 0;
1774 for (auto *I : *LRI)
1775 for (auto *V : PHIOperands[I])
1776 if (InstructionsToSink.count(V) == 0)
1777 ++NumPHIdValues;
1778 LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: #phid values: " <<
NumPHIdValues << "\n"; } } while (false)
;
1779 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1780 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1781 NumPHIInsts++;
1782
1783 return NumPHIInsts <= 1;
1784 };
1785
1786 if (ScanIdx > 0 && Cond) {
1787 // Check if we would actually sink anything first! This mutates the CFG and
1788 // adds an extra block. The goal in doing this is to allow instructions that
1789 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1790 // (such as trunc, add) can be sunk and predicated already. So we check that
1791 // we're going to sink at least one non-speculatable instruction.
1792 LRI.reset();
1793 unsigned Idx = 0;
1794 bool Profitable = false;
1795 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1796 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1797 Profitable = true;
1798 break;
1799 }
1800 --LRI;
1801 ++Idx;
1802 }
1803 if (!Profitable)
1804 return false;
1805
1806 LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Splitting edge\n"; }
} while (false)
;
1807 // We have a conditional edge and we're going to sink some instructions.
1808 // Insert a new block postdominating all blocks we're going to sink from.
1809 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1810 // Edges couldn't be split.
1811 return false;
1812 Changed = true;
1813 }
1814
1815 // Now that we've analyzed all potential sinking candidates, perform the
1816 // actual sink. We iteratively sink the last non-terminator of the source
1817 // blocks into their common successor unless doing so would require too
1818 // many PHI instructions to be generated (currently only one PHI is allowed
1819 // per sunk instruction).
1820 //
1821 // We can use InstructionsToSink to discount values needing PHI-merging that will
1822 // actually be sunk in a later iteration. This allows us to be more
1823 // aggressive in what we sink. This does allow a false positive where we
1824 // sink presuming a later value will also be sunk, but stop half way through
1825 // and never actually sink it which means we produce more PHIs than intended.
1826 // This is unlikely in practice though.
1827 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1828 LLVM_DEBUG(dbgs() << "SINK: Sink: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1829 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1830 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
;
1831
1832 // Because we've sunk every instruction in turn, the current instruction to
1833 // sink is always at index 0.
1834 LRI.reset();
1835 if (!ProfitableToSinkInstruction(LRI)) {
1836 // Too many PHIs would be created.
1837 LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: stopping here, too many PHIs would be created!\n"
; } } while (false)
1838 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)
;
1839 break;
1840 }
1841
1842 if (!sinkLastInstruction(UnconditionalPreds))
1843 return Changed;
1844 NumSinkCommons++;
1845 Changed = true;
1846 }
1847 return Changed;
1848}
1849
1850/// Determine if we can hoist sink a sole store instruction out of a
1851/// conditional block.
1852///
1853/// We are looking for code like the following:
1854/// BrBB:
1855/// store i32 %add, i32* %arrayidx2
1856/// ... // No other stores or function calls (we could be calling a memory
1857/// ... // function).
1858/// %cmp = icmp ult %x, %y
1859/// br i1 %cmp, label %EndBB, label %ThenBB
1860/// ThenBB:
1861/// store i32 %add5, i32* %arrayidx2
1862/// br label EndBB
1863/// EndBB:
1864/// ...
1865/// We are going to transform this into:
1866/// BrBB:
1867/// store i32 %add, i32* %arrayidx2
1868/// ... //
1869/// %cmp = icmp ult %x, %y
1870/// %add.add5 = select i1 %cmp, i32 %add, %add5
1871/// store i32 %add.add5, i32* %arrayidx2
1872/// ...
1873///
1874/// \return The pointer to the value of the previous store if the store can be
1875/// hoisted into the predecessor block. 0 otherwise.
1876static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1877 BasicBlock *StoreBB, BasicBlock *EndBB) {
1878 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1879 if (!StoreToHoist)
1880 return nullptr;
1881
1882 // Volatile or atomic.
1883 if (!StoreToHoist->isSimple())
1884 return nullptr;
1885
1886 Value *StorePtr = StoreToHoist->getPointerOperand();
1887
1888 // Look for a store to the same pointer in BrBB.
1889 unsigned MaxNumInstToLookAt = 9;
1890 for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1891 if (!MaxNumInstToLookAt)
1892 break;
1893 --MaxNumInstToLookAt;
1894
1895 // Could be calling an instruction that affects memory like free().
1896 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1897 return nullptr;
1898
1899 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1900 // Found the previous store make sure it stores to the same location.
1901 if (SI->getPointerOperand() == StorePtr)
1902 // Found the previous store, return its value operand.
1903 return SI->getValueOperand();
1904 return nullptr; // Unknown store.
1905 }
1906 }
1907
1908 return nullptr;
1909}
1910
1911/// Speculate a conditional basic block flattening the CFG.
1912///
1913/// Note that this is a very risky transform currently. Speculating
1914/// instructions like this is most often not desirable. Instead, there is an MI
1915/// pass which can do it with full awareness of the resource constraints.
1916/// However, some cases are "obvious" and we should do directly. An example of
1917/// this is speculating a single, reasonably cheap instruction.
1918///
1919/// There is only one distinct advantage to flattening the CFG at the IR level:
1920/// it makes very common but simplistic optimizations such as are common in
1921/// instcombine and the DAG combiner more powerful by removing CFG edges and
1922/// modeling their effects with easier to reason about SSA value graphs.
1923///
1924///
1925/// An illustration of this transform is turning this IR:
1926/// \code
1927/// BB:
1928/// %cmp = icmp ult %x, %y
1929/// br i1 %cmp, label %EndBB, label %ThenBB
1930/// ThenBB:
1931/// %sub = sub %x, %y
1932/// br label BB2
1933/// EndBB:
1934/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1935/// ...
1936/// \endcode
1937///
1938/// Into this IR:
1939/// \code
1940/// BB:
1941/// %cmp = icmp ult %x, %y
1942/// %sub = sub %x, %y
1943/// %cond = select i1 %cmp, 0, %sub
1944/// ...
1945/// \endcode
1946///
1947/// \returns true if the conditional block is removed.
1948static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1949 const TargetTransformInfo &TTI) {
1950 // Be conservative for now. FP select instruction can often be expensive.
1951 Value *BrCond = BI->getCondition();
1952 if (isa<FCmpInst>(BrCond))
1953 return false;
1954
1955 BasicBlock *BB = BI->getParent();
1956 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1957
1958 // If ThenBB is actually on the false edge of the conditional branch, remember
1959 // to swap the select operands later.
1960 bool Invert = false;
1961 if (ThenBB != BI->getSuccessor(0)) {
1962 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1962, __PRETTY_FUNCTION__))
;
1963 Invert = true;
1964 }
1965 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1965, __PRETTY_FUNCTION__))
;
1966
1967 // Keep a count of how many times instructions are used within ThenBB when
1968 // they are candidates for sinking into ThenBB. Specifically:
1969 // - They are defined in BB, and
1970 // - They have no side effects, and
1971 // - All of their uses are in ThenBB.
1972 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1973
1974 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1975
1976 unsigned SpeculationCost = 0;
1977 Value *SpeculatedStoreValue = nullptr;
1978 StoreInst *SpeculatedStore = nullptr;
1979 for (BasicBlock::iterator BBI = ThenBB->begin(),
1980 BBE = std::prev(ThenBB->end());
1981 BBI != BBE; ++BBI) {
1982 Instruction *I = &*BBI;
1983 // Skip debug info.
1984 if (isa<DbgInfoIntrinsic>(I)) {
1985 SpeculatedDbgIntrinsics.push_back(I);
1986 continue;
1987 }
1988
1989 // Only speculatively execute a single instruction (not counting the
1990 // terminator) for now.
1991 ++SpeculationCost;
1992 if (SpeculationCost > 1)
1993 return false;
1994
1995 // Don't hoist the instruction if it's unsafe or expensive.
1996 if (!isSafeToSpeculativelyExecute(I) &&
1997 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1998 I, BB, ThenBB, EndBB))))
1999 return false;
2000 if (!SpeculatedStoreValue &&
2001 ComputeSpeculationCost(I, TTI) >
2002 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
2003 return false;
2004
2005 // Store the store speculation candidate.
2006 if (SpeculatedStoreValue)
2007 SpeculatedStore = cast<StoreInst>(I);
2008
2009 // Do not hoist the instruction if any of its operands are defined but not
2010 // used in BB. The transformation will prevent the operand from
2011 // being sunk into the use block.
2012 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2013 Instruction *OpI = dyn_cast<Instruction>(*i);
2014 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2015 continue; // Not a candidate for sinking.
2016
2017 ++SinkCandidateUseCounts[OpI];
2018 }
2019 }
2020
2021 // Consider any sink candidates which are only used in ThenBB as costs for
2022 // speculation. Note, while we iterate over a DenseMap here, we are summing
2023 // and so iteration order isn't significant.
2024 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2025 I = SinkCandidateUseCounts.begin(),
2026 E = SinkCandidateUseCounts.end();
2027 I != E; ++I)
2028 if (I->first->hasNUses(I->second)) {
2029 ++SpeculationCost;
2030 if (SpeculationCost > 1)
2031 return false;
2032 }
2033
2034 // Check that the PHI nodes can be converted to selects.
2035 bool HaveRewritablePHIs = false;
2036 for (PHINode &PN : EndBB->phis()) {
2037 Value *OrigV = PN.getIncomingValueForBlock(BB);
2038 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2039
2040 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2041 // Skip PHIs which are trivial.
2042 if (ThenV == OrigV)
2043 continue;
2044
2045 // Don't convert to selects if we could remove undefined behavior instead.
2046 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2047 passingValueIsAlwaysUndefined(ThenV, &PN))
2048 return false;
2049
2050 HaveRewritablePHIs = true;
2051 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2052 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2053 if (!OrigCE && !ThenCE)
2054 continue; // Known safe and cheap.
2055
2056 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2057 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2058 return false;
2059 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2060 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2061 unsigned MaxCost =
2062 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2063 if (OrigCost + ThenCost > MaxCost)
2064 return false;
2065
2066 // Account for the cost of an unfolded ConstantExpr which could end up
2067 // getting expanded into Instructions.
2068 // FIXME: This doesn't account for how many operations are combined in the
2069 // constant expression.
2070 ++SpeculationCost;
2071 if (SpeculationCost > 1)
2072 return false;
2073 }
2074
2075 // If there are no PHIs to process, bail early. This helps ensure idempotence
2076 // as well.
2077 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2078 return false;
2079
2080 // If we get here, we can hoist the instruction and if-convert.
2081 LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SPECULATIVELY EXECUTING BB"
<< *ThenBB << "\n";; } } while (false)
;
2082
2083 // Insert a select of the value of the speculated store.
2084 if (SpeculatedStoreValue) {
2085 IRBuilder<NoFolder> Builder(BI);
2086 Value *TrueV = SpeculatedStore->getValueOperand();
2087 Value *FalseV = SpeculatedStoreValue;
2088 if (Invert)
2089 std::swap(TrueV, FalseV);
2090 Value *S = Builder.CreateSelect(
2091 BrCond, TrueV, FalseV, "spec.store.select", BI);
2092 SpeculatedStore->setOperand(0, S);
2093 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2094 SpeculatedStore->getDebugLoc());
2095 }
2096
2097 // Metadata can be dependent on the condition we are hoisting above.
2098 // Conservatively strip all metadata on the instruction.
2099 for (auto &I : *ThenBB)
2100 I.dropUnknownNonDebugMetadata();
2101
2102 // Hoist the instructions.
2103 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2104 ThenBB->begin(), std::prev(ThenBB->end()));
2105
2106 // Insert selects and rewrite the PHI operands.
2107 IRBuilder<NoFolder> Builder(BI);
2108 for (PHINode &PN : EndBB->phis()) {
2109 unsigned OrigI = PN.getBasicBlockIndex(BB);
2110 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2111 Value *OrigV = PN.getIncomingValue(OrigI);
2112 Value *ThenV = PN.getIncomingValue(ThenI);
2113
2114 // Skip PHIs which are trivial.
2115 if (OrigV == ThenV)
2116 continue;
2117
2118 // Create a select whose true value is the speculatively executed value and
2119 // false value is the preexisting value. Swap them if the branch
2120 // destinations were inverted.
2121 Value *TrueV = ThenV, *FalseV = OrigV;
2122 if (Invert)
2123 std::swap(TrueV, FalseV);
2124 Value *V = Builder.CreateSelect(
2125 BrCond, TrueV, FalseV, "spec.select", BI);
2126 PN.setIncomingValue(OrigI, V);
2127 PN.setIncomingValue(ThenI, V);
2128 }
2129
2130 // Remove speculated dbg intrinsics.
2131 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2132 // dbg value for the different flows and inserting it after the select.
2133 for (Instruction *I : SpeculatedDbgIntrinsics)
2134 I->eraseFromParent();
2135
2136 ++NumSpeculations;
2137 return true;
2138}
2139
2140/// Return true if we can thread a branch across this block.
2141static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2142 unsigned Size = 0;
2143
2144 for (Instruction &I : BB->instructionsWithoutDebug()) {
2145 if (Size > 10)
2146 return false; // Don't clone large BB's.
2147 ++Size;
2148
2149 // We can only support instructions that do not define values that are
2150 // live outside of the current basic block.
2151 for (User *U : I.users()) {
2152 Instruction *UI = cast<Instruction>(U);
2153 if (UI->getParent() != BB || isa<PHINode>(UI))
2154 return false;
2155 }
2156
2157 // Looks ok, continue checking.
2158 }
2159
2160 return true;
2161}
2162
2163/// If we have a conditional branch on a PHI node value that is defined in the
2164/// same block as the branch and if any PHI entries are constants, thread edges
2165/// corresponding to that entry to be branches to their ultimate destination.
2166static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2167 AssumptionCache *AC) {
2168 BasicBlock *BB = BI->getParent();
2169 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2170 // NOTE: we currently cannot transform this case if the PHI node is used
2171 // outside of the block.
2172 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2173 return false;
2174
2175 // Degenerate case of a single entry PHI.
2176 if (PN->getNumIncomingValues() == 1) {
2177 FoldSingleEntryPHINodes(PN->getParent());
2178 return true;
2179 }
2180
2181 // Now we know that this block has multiple preds and two succs.
2182 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2183 return false;
2184
2185 // Can't fold blocks that contain noduplicate or convergent calls.
2186 if (any_of(*BB, [](const Instruction &I) {
2187 const CallInst *CI = dyn_cast<CallInst>(&I);
2188 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2189 }))
2190 return false;
2191
2192 // Okay, this is a simple enough basic block. See if any phi values are
2193 // constants.
2194 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2195 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2196 if (!CB || !CB->getType()->isIntegerTy(1))
2197 continue;
2198
2199 // Okay, we now know that all edges from PredBB should be revectored to
2200 // branch to RealDest.
2201 BasicBlock *PredBB = PN->getIncomingBlock(i);
2202 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2203
2204 if (RealDest == BB)
2205 continue; // Skip self loops.
2206 // Skip if the predecessor's terminator is an indirect branch.
2207 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2208 continue;
2209
2210 // The dest block might have PHI nodes, other predecessors and other
2211 // difficult cases. Instead of being smart about this, just insert a new
2212 // block that jumps to the destination block, effectively splitting
2213 // the edge we are about to create.
2214 BasicBlock *EdgeBB =
2215 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2216 RealDest->getParent(), RealDest);
2217 BranchInst::Create(RealDest, EdgeBB);
2218
2219 // Update PHI nodes.
2220 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2221
2222 // BB may have instructions that are being threaded over. Clone these
2223 // instructions into EdgeBB. We know that there will be no uses of the
2224 // cloned instructions outside of EdgeBB.
2225 BasicBlock::iterator InsertPt = EdgeBB->begin();
2226 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2227 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2228 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2229 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2230 continue;
2231 }
2232 // Clone the instruction.
2233 Instruction *N = BBI->clone();
2234 if (BBI->hasName())
2235 N->setName(BBI->getName() + ".c");
2236
2237 // Update operands due to translation.
2238 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2239 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2240 if (PI != TranslateMap.end())
2241 *i = PI->second;
2242 }
2243
2244 // Check for trivial simplification.
2245 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2246 if (!BBI->use_empty())
2247 TranslateMap[&*BBI] = V;
2248 if (!N->mayHaveSideEffects()) {
2249 N->deleteValue(); // Instruction folded away, don't need actual inst
2250 N = nullptr;
2251 }
2252 } else {
2253 if (!BBI->use_empty())
2254 TranslateMap[&*BBI] = N;
2255 }
2256 // Insert the new instruction into its new home.
2257 if (N)
2258 EdgeBB->getInstList().insert(InsertPt, N);
2259
2260 // Register the new instruction with the assumption cache if necessary.
2261 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2262 if (II->getIntrinsicID() == Intrinsic::assume)
2263 AC->registerAssumption(II);
2264 }
2265
2266 // Loop over all of the edges from PredBB to BB, changing them to branch
2267 // to EdgeBB instead.
2268 Instruction *PredBBTI = PredBB->getTerminator();
2269 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2270 if (PredBBTI->getSuccessor(i) == BB) {
2271 BB->removePredecessor(PredBB);
2272 PredBBTI->setSuccessor(i, EdgeBB);
2273 }
2274
2275 // Recurse, simplifying any other constants.
2276 return FoldCondBranchOnPHI(BI, DL, AC) || true;
2277 }
2278
2279 return false;
2280}
2281
2282/// Given a BB that starts with the specified two-entry PHI node,
2283/// see if we can eliminate it.
2284static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2285 const DataLayout &DL) {
2286 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2287 // statement", which has a very simple dominance structure. Basically, we
2288 // are trying to find the condition that is being branched on, which
2289 // subsequently causes this merge to happen. We really want control
2290 // dependence information for this check, but simplifycfg can't keep it up
2291 // to date, and this catches most of the cases we care about anyway.
2292 BasicBlock *BB = PN->getParent();
2293 const Function *Fn = BB->getParent();
2294 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
18
Assuming 'Fn' is null
2295 return false;
2296
2297 BasicBlock *IfTrue, *IfFalse;
2298 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2299 if (!IfCond ||
19
Assuming 'IfCond' is non-null
20
Taking false branch
2300 // Don't bother if the branch will be constant folded trivially.
2301 isa<ConstantInt>(IfCond))
2302 return false;
2303
2304 // Okay, we found that we can merge this two-entry phi node into a select.
2305 // Doing so would require us to fold *all* two entry phi nodes in this block.
2306 // At some point this becomes non-profitable (particularly if the target
2307 // doesn't support cmov's). Only do this transformation if there are two or
2308 // fewer PHI nodes in this block.
2309 unsigned NumPhis = 0;
2310 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
21
Loop condition is false. Execution continues on line 2317
2311 if (NumPhis > 2)
2312 return false;
2313
2314 // Loop over the PHI's seeing if we can promote them all to select
2315 // instructions. While we are at it, keep track of the instructions
2316 // that need to be moved to the dominating block.
2317 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2318 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2319 MaxCostVal1 = PHINodeFoldingThreshold;
2320 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2321 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2322
2323 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
22
Loop condition is false. Execution continues on line 2340
2324 PHINode *PN = cast<PHINode>(II++);
2325 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2326 PN->replaceAllUsesWith(V);
2327 PN->eraseFromParent();
2328 continue;
2329 }
2330
2331 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2332 MaxCostVal0, TTI) ||
2333 !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2334 MaxCostVal1, TTI))
2335 return false;
2336 }
2337
2338 // If we folded the first phi, PN dangles at this point. Refresh it. If
2339 // we ran out of PHIs then we simplified them all.
2340 PN = dyn_cast<PHINode>(BB->begin());
2341 if (!PN)
23
Assuming 'PN' is non-null
24
Taking false branch
2342 return true;
2343
2344 // Don't fold i1 branches on PHIs which contain binary operators. These can
2345 // often be turned into switches and other things.
2346 if (PN->getType()->isIntegerTy(1) &&
25
Assuming the condition is false
26
Taking false branch
2347 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2348 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2349 isa<BinaryOperator>(IfCond)))
2350 return false;
2351
2352 // If all PHI nodes are promotable, check to make sure that all instructions
2353 // in the predecessor blocks can be promoted as well. If not, we won't be able
2354 // to get rid of the control flow, so it's not worth promoting to select
2355 // instructions.
2356 BasicBlock *DomBlock = nullptr;
27
'DomBlock' initialized to a null pointer value
2357 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2358 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2359 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
28
Taking true branch
2360 IfBlock1 = nullptr;
2361 } else {
2362 DomBlock = *pred_begin(IfBlock1);
2363 for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2364 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2365 // This is not an aggressive instruction that we can promote.
2366 // Because of this, we won't be able to get rid of the control flow, so
2367 // the xform is not worth it.
2368 return false;
2369 }
2370 }
2371
2372 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
29
Taking true branch
2373 IfBlock2 = nullptr;
2374 } else {
2375 DomBlock = *pred_begin(IfBlock2);
2376 for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2377 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2378 // This is not an aggressive instruction that we can promote.
2379 // Because of this, we won't be able to get rid of the control flow, so
2380 // the xform is not worth it.
2381 return false;
2382 }
2383 }
2384
2385 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
2386 << " T: " << IfTrue->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
2387 << " 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)
;
2388
2389 // If we can still promote the PHI nodes after this gauntlet of tests,
2390 // do all of the PHI's now.
2391 Instruction *InsertPt = DomBlock->getTerminator();
32
Called C++ object pointer is null
2392 IRBuilder<NoFolder> Builder(InsertPt);
2393
2394 // Move all 'aggressive' instructions, which are defined in the
2395 // conditional parts of the if's up to the dominating block.
2396 if (IfBlock1)
2397 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2398 if (IfBlock2)
2399 hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2400
2401 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2402 // Change the PHI node into a select instruction.
2403 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2404 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2405
2406 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2407 PN->replaceAllUsesWith(Sel);
2408 Sel->takeName(PN);
2409 PN->eraseFromParent();
2410 }
2411
2412 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2413 // has been flattened. Change DomBlock to jump directly to our new block to
2414 // avoid other simplifycfg's kicking in on the diamond.
2415 Instruction *OldTI = DomBlock->getTerminator();
2416 Builder.SetInsertPoint(OldTI);
2417 Builder.CreateBr(BB);
2418 OldTI->eraseFromParent();
2419 return true;
2420}
2421
2422/// If we found a conditional branch that goes to two returning blocks,
2423/// try to merge them together into one return,
2424/// introducing a select if the return values disagree.
2425static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2426 IRBuilder<> &Builder) {
2427 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2427, __PRETTY_FUNCTION__))
;
2428 BasicBlock *TrueSucc = BI->getSuccessor(0);
2429 BasicBlock *FalseSucc = BI->getSuccessor(1);
2430 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2431 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2432
2433 // Check to ensure both blocks are empty (just a return) or optionally empty
2434 // with PHI nodes. If there are other instructions, merging would cause extra
2435 // computation on one path or the other.
2436 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2437 return false;
2438 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2439 return false;
2440
2441 Builder.SetInsertPoint(BI);
2442 // Okay, we found a branch that is going to two return nodes. If
2443 // there is no return value for this function, just change the
2444 // branch into a return.
2445 if (FalseRet->getNumOperands() == 0) {
2446 TrueSucc->removePredecessor(BI->getParent());
2447 FalseSucc->removePredecessor(BI->getParent());
2448 Builder.CreateRetVoid();
2449 EraseTerminatorAndDCECond(BI);
2450 return true;
2451 }
2452
2453 // Otherwise, figure out what the true and false return values are
2454 // so we can insert a new select instruction.
2455 Value *TrueValue = TrueRet->getReturnValue();
2456 Value *FalseValue = FalseRet->getReturnValue();
2457
2458 // Unwrap any PHI nodes in the return blocks.
2459 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2460 if (TVPN->getParent() == TrueSucc)
2461 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2462 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2463 if (FVPN->getParent() == FalseSucc)
2464 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2465
2466 // In order for this transformation to be safe, we must be able to
2467 // unconditionally execute both operands to the return. This is
2468 // normally the case, but we could have a potentially-trapping
2469 // constant expression that prevents this transformation from being
2470 // safe.
2471 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2472 if (TCV->canTrap())
2473 return false;
2474 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2475 if (FCV->canTrap())
2476 return false;
2477
2478 // Okay, we collected all the mapped values and checked them for sanity, and
2479 // defined to really do this transformation. First, update the CFG.
2480 TrueSucc->removePredecessor(BI->getParent());
2481 FalseSucc->removePredecessor(BI->getParent());
2482
2483 // Insert select instructions where needed.
2484 Value *BrCond = BI->getCondition();
2485 if (TrueValue) {
2486 // Insert a select if the results differ.
2487 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2488 } else if (isa<UndefValue>(TrueValue)) {
2489 TrueValue = FalseValue;
2490 } else {
2491 TrueValue =
2492 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2493 }
2494 }
2495
2496 Value *RI =
2497 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2498
2499 (void)RI;
2500
2501 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)
2502 << "\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)
2503 << *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)
;
2504
2505 EraseTerminatorAndDCECond(BI);
2506
2507 return true;
2508}
2509
2510/// Return true if the given instruction is available
2511/// in its predecessor block. If yes, the instruction will be removed.
2512static bool tryCSEWithPredecessor(Instruction *Inst, BasicBlock *PB) {
2513 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2514 return false;
2515 for (Instruction &I : *PB) {
2516 Instruction *PBI = &I;
2517 // Check whether Inst and PBI generate the same value.
2518 if (Inst->isIdenticalTo(PBI)) {
2519 Inst->replaceAllUsesWith(PBI);
2520 Inst->eraseFromParent();
2521 return true;
2522 }
2523 }
2524 return false;
2525}
2526
2527/// Return true if either PBI or BI has branch weight available, and store
2528/// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2529/// not have branch weight, use 1:1 as its weight.
2530static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2531 uint64_t &PredTrueWeight,
2532 uint64_t &PredFalseWeight,
2533 uint64_t &SuccTrueWeight,
2534 uint64_t &SuccFalseWeight) {
2535 bool PredHasWeights =
2536 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2537 bool SuccHasWeights =
2538 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2539 if (PredHasWeights || SuccHasWeights) {
2540 if (!PredHasWeights)
2541 PredTrueWeight = PredFalseWeight = 1;
2542 if (!SuccHasWeights)
2543 SuccTrueWeight = SuccFalseWeight = 1;
2544 return true;
2545 } else {
2546 return false;
2547 }
2548}
2549
2550/// If this basic block is simple enough, and if a predecessor branches to us
2551/// and one of our successors, fold the block into the predecessor and use
2552/// logical operations to pick the right destination.
2553bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2554 BasicBlock *BB = BI->getParent();
2555
2556 const unsigned PredCount = pred_size(BB);
2557
2558 Instruction *Cond = nullptr;
2559 if (BI->isConditional())
2560 Cond = dyn_cast<Instruction>(BI->getCondition());
2561 else {
2562 // For unconditional branch, check for a simple CFG pattern, where
2563 // BB has a single predecessor and BB's successor is also its predecessor's
2564 // successor. If such pattern exists, check for CSE between BB and its
2565 // predecessor.
2566 if (BasicBlock *PB = BB->getSinglePredecessor())
2567 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2568 if (PBI->isConditional() &&
2569 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2570 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2571 for (auto I = BB->instructionsWithoutDebug().begin(),
2572 E = BB->instructionsWithoutDebug().end();
2573 I != E;) {
2574 Instruction *Curr = &*I++;
2575 if (isa<CmpInst>(Curr)) {
2576 Cond = Curr;
2577 break;
2578 }
2579 // Quit if we can't remove this instruction.
2580 if (!tryCSEWithPredecessor(Curr, PB))
2581 return false;
2582 }
2583 }
2584
2585 if (!Cond)
2586 return false;
2587 }
2588
2589 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2590 Cond->getParent() != BB || !Cond->hasOneUse())
2591 return false;
2592
2593 // Make sure the instruction after the condition is the cond branch.
2594 BasicBlock::iterator CondIt = ++Cond->getIterator();
2595
2596 // Ignore dbg intrinsics.
2597 while (isa<DbgInfoIntrinsic>(CondIt))
2598 ++CondIt;
2599
2600 if (&*CondIt != BI)
2601 return false;
2602
2603 // Only allow this transformation if computing the condition doesn't involve
2604 // too many instructions and these involved instructions can be executed
2605 // unconditionally. We denote all involved instructions except the condition
2606 // as "bonus instructions", and only allow this transformation when the
2607 // number of the bonus instructions we'll need to create when cloning into
2608 // each predecessor does not exceed a certain threshold.
2609 unsigned NumBonusInsts = 0;
2610 for (auto I = BB->begin(); Cond != &*I; ++I) {
2611 // Ignore dbg intrinsics.
2612 if (isa<DbgInfoIntrinsic>(I))
2613 continue;
2614 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2615 return false;
2616 // I has only one use and can be executed unconditionally.
2617 Instruction *User = dyn_cast<Instruction>(I->user_back());
2618 if (User == nullptr || User->getParent() != BB)
2619 return false;
2620 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2621 // to use any other instruction, User must be an instruction between next(I)
2622 // and Cond.
2623
2624 // Account for the cost of duplicating this instruction into each
2625 // predecessor.
2626 NumBonusInsts += PredCount;
2627 // Early exits once we reach the limit.
2628 if (NumBonusInsts > BonusInstThreshold)
2629 return false;
2630 }
2631
2632 // Cond is known to be a compare or binary operator. Check to make sure that
2633 // neither operand is a potentially-trapping constant expression.
2634 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2635 if (CE->canTrap())
2636 return false;
2637 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2638 if (CE->canTrap())
2639 return false;
2640
2641 // Finally, don't infinitely unroll conditional loops.
2642 BasicBlock *TrueDest = BI->getSuccessor(0);
2643 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2644 if (TrueDest == BB || FalseDest == BB)
2645 return false;
2646
2647 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2648 BasicBlock *PredBlock = *PI;
2649 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2650
2651 // Check that we have two conditional branches. If there is a PHI node in
2652 // the common successor, verify that the same value flows in from both
2653 // blocks.
2654 SmallVector<PHINode *, 4> PHIs;
2655 if (!PBI || PBI->isUnconditional() ||
2656 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2657 (!BI->isConditional() &&
2658 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2659 continue;
2660
2661 // Determine if the two branches share a common destination.
2662 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2663 bool InvertPredCond = false;
2664
2665 if (BI->isConditional()) {
2666 if (PBI->getSuccessor(0) == TrueDest) {
2667 Opc = Instruction::Or;
2668 } else if (PBI->getSuccessor(1) == FalseDest) {
2669 Opc = Instruction::And;
2670 } else if (PBI->getSuccessor(0) == FalseDest) {
2671 Opc = Instruction::And;
2672 InvertPredCond = true;
2673 } else if (PBI->getSuccessor(1) == TrueDest) {
2674 Opc = Instruction::Or;
2675 InvertPredCond = true;
2676 } else {
2677 continue;
2678 }
2679 } else {
2680 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2681 continue;
2682 }
2683
2684 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)
;
2685 IRBuilder<> Builder(PBI);
2686
2687 // If we need to invert the condition in the pred block to match, do so now.
2688 if (InvertPredCond) {
2689 Value *NewCond = PBI->getCondition();
2690
2691 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2692 CmpInst *CI = cast<CmpInst>(NewCond);
2693 CI->setPredicate(CI->getInversePredicate());
2694 } else {
2695 NewCond =
2696 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2697 }
2698
2699 PBI->setCondition(NewCond);
2700 PBI->swapSuccessors();
2701 }
2702
2703 // If we have bonus instructions, clone them into the predecessor block.
2704 // Note that there may be multiple predecessor blocks, so we cannot move
2705 // bonus instructions to a predecessor block.
2706 ValueToValueMapTy VMap; // maps original values to cloned values
2707 // We already make sure Cond is the last instruction before BI. Therefore,
2708 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2709 // instructions.
2710 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2711 if (isa<DbgInfoIntrinsic>(BonusInst))
2712 continue;
2713 Instruction *NewBonusInst = BonusInst->clone();
2714 RemapInstruction(NewBonusInst, VMap,
2715 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2716 VMap[&*BonusInst] = NewBonusInst;
2717
2718 // If we moved a load, we cannot any longer claim any knowledge about
2719 // its potential value. The previous information might have been valid
2720 // only given the branch precondition.
2721 // For an analogous reason, we must also drop all the metadata whose
2722 // semantics we don't understand.
2723 NewBonusInst->dropUnknownNonDebugMetadata();
2724
2725 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2726 NewBonusInst->takeName(&*BonusInst);
2727 BonusInst->setName(BonusInst->getName() + ".old");
2728 }
2729
2730 // Clone Cond into the predecessor basic block, and or/and the
2731 // two conditions together.
2732 Instruction *CondInPred = Cond->clone();
2733 RemapInstruction(CondInPred, VMap,
2734 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2735 PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2736 CondInPred->takeName(Cond);
2737 Cond->setName(CondInPred->getName() + ".old");
2738
2739 if (BI->isConditional()) {
2740 Instruction *NewCond = cast<Instruction>(
2741 Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2742 PBI->setCondition(NewCond);
2743
2744 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2745 bool HasWeights =
2746 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2747 SuccTrueWeight, SuccFalseWeight);
2748 SmallVector<uint64_t, 8> NewWeights;
2749
2750 if (PBI->getSuccessor(0) == BB) {
2751 if (HasWeights) {
2752 // PBI: br i1 %x, BB, FalseDest
2753 // BI: br i1 %y, TrueDest, FalseDest
2754 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2755 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2756 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2757 // TrueWeight for PBI * FalseWeight for BI.
2758 // We assume that total weights of a BranchInst can fit into 32 bits.
2759 // Therefore, we will not have overflow using 64-bit arithmetic.
2760 NewWeights.push_back(PredFalseWeight *
2761 (SuccFalseWeight + SuccTrueWeight) +
2762 PredTrueWeight * SuccFalseWeight);
2763 }
2764 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2765 PBI->setSuccessor(0, TrueDest);
2766 }
2767 if (PBI->getSuccessor(1) == BB) {
2768 if (HasWeights) {
2769 // PBI: br i1 %x, TrueDest, BB
2770 // BI: br i1 %y, TrueDest, FalseDest
2771 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2772 // FalseWeight for PBI * TrueWeight for BI.
2773 NewWeights.push_back(PredTrueWeight *
2774 (SuccFalseWeight + SuccTrueWeight) +
2775 PredFalseWeight * SuccTrueWeight);
2776 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2777 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2778 }
2779 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2780 PBI->setSuccessor(1, FalseDest);
2781 }
2782 if (NewWeights.size() == 2) {
2783 // Halve the weights if any of them cannot fit in an uint32_t
2784 FitWeights(NewWeights);
2785
2786 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2787 NewWeights.end());
2788 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2789 } else
2790 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2791 } else {
2792 // Update PHI nodes in the common successors.
2793 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2794 ConstantInt *PBI_C = cast<ConstantInt>(
2795 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2796 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2796, __PRETTY_FUNCTION__))
;
2797 Instruction *MergedCond = nullptr;
2798 if (PBI->getSuccessor(0) == TrueDest) {
2799 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2800 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2801 // is false: !PBI_Cond and BI_Value
2802 Instruction *NotCond = cast<Instruction>(
2803 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2804 MergedCond = cast<Instruction>(
2805 Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2806 "and.cond"));
2807 if (PBI_C->isOne())
2808 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2809 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2810 } else {
2811 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2812 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2813 // is false: PBI_Cond and BI_Value
2814 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2815 Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2816 if (PBI_C->isOne()) {
2817 Instruction *NotCond = cast<Instruction>(
2818 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2819 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2820 Instruction::Or, NotCond, MergedCond, "or.cond"));
2821 }
2822 }
2823 // Update PHI Node.
2824 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2825 MergedCond);
2826 }
2827 // Change PBI from Conditional to Unconditional.
2828 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2829 EraseTerminatorAndDCECond(PBI);
2830 PBI = New_PBI;
2831 }
2832
2833 // If BI was a loop latch, it may have had associated loop metadata.
2834 // We need to copy it to the new latch, that is, PBI.
2835 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2836 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2837
2838 // TODO: If BB is reachable from all paths through PredBlock, then we
2839 // could replace PBI's branch probabilities with BI's.
2840
2841 // Copy any debug value intrinsics into the end of PredBlock.
2842 for (Instruction &I : *BB)
2843 if (isa<DbgInfoIntrinsic>(I))
2844 I.clone()->insertBefore(PBI);
2845
2846 return true;
2847 }
2848 return false;
2849}
2850
2851// If there is only one store in BB1 and BB2, return it, otherwise return
2852// nullptr.
2853static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2854 StoreInst *S = nullptr;
2855 for (auto *BB : {BB1, BB2}) {
2856 if (!BB)
2857 continue;
2858 for (auto &I : *BB)
2859 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2860 if (S)
2861 // Multiple stores seen.
2862 return nullptr;
2863 else
2864 S = SI;
2865 }
2866 }
2867 return S;
2868}
2869
2870static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2871 Value *AlternativeV = nullptr) {
2872 // PHI is going to be a PHI node that allows the value V that is defined in
2873 // BB to be referenced in BB's only successor.
2874 //
2875 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2876 // doesn't matter to us what the other operand is (it'll never get used). We
2877 // could just create a new PHI with an undef incoming value, but that could
2878 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2879 // other PHI. So here we directly look for some PHI in BB's successor with V
2880 // as an incoming operand. If we find one, we use it, else we create a new
2881 // one.
2882 //
2883 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2884 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2885 // where OtherBB is the single other predecessor of BB's only successor.
2886 PHINode *PHI = nullptr;
2887 BasicBlock *Succ = BB->getSingleSuccessor();
2888
2889 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2890 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2891 PHI = cast<PHINode>(I);
2892 if (!AlternativeV)
2893 break;
2894
2895 assert(Succ->hasNPredecessors(2))((Succ->hasNPredecessors(2)) ? static_cast<void> (0)
: __assert_fail ("Succ->hasNPredecessors(2)", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2895, __PRETTY_FUNCTION__))
;
2896 auto PredI = pred_begin(Succ);
2897 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2898 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2899 break;
2900 PHI = nullptr;
2901 }
2902 if (PHI)
2903 return PHI;
2904
2905 // If V is not an instruction defined in BB, just return it.
2906 if (!AlternativeV &&
2907 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2908 return V;
2909
2910 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2911 PHI->addIncoming(V, BB);
2912 for (BasicBlock *PredBB : predecessors(Succ))
2913 if (PredBB != BB)
2914 PHI->addIncoming(
2915 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2916 return PHI;
2917}
2918
2919static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2920 BasicBlock *QTB, BasicBlock *QFB,
2921 BasicBlock *PostBB, Value *Address,
2922 bool InvertPCond, bool InvertQCond,
2923 const DataLayout &DL) {
2924 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2925 return Operator::getOpcode(&I) == Instruction::BitCast &&
2926 I.getType()->isPointerTy();
2927 };
2928
2929 // If we're not in aggressive mode, we only optimize if we have some
2930 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2931 auto IsWorthwhile = [&](BasicBlock *BB) {
2932 if (!BB)
2933 return true;
2934 // Heuristic: if the block can be if-converted/phi-folded and the
2935 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2936 // thread this store.
2937 unsigned N = 0;
2938 for (auto &I : BB->instructionsWithoutDebug()) {
2939 // Cheap instructions viable for folding.
2940 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2941 isa<StoreInst>(I))
2942 ++N;
2943 // Free instructions.
2944 else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2945 continue;
2946 else
2947 return false;
2948 }
2949 // The store we want to merge is counted in N, so add 1 to make sure
2950 // we're counting the instructions that would be left.
2951 return N <= (PHINodeFoldingThreshold + 1);
2952 };
2953
2954 if (!MergeCondStoresAggressively &&
2955 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2956 !IsWorthwhile(QFB)))
2957 return false;
2958
2959 // For every pointer, there must be exactly two stores, one coming from
2960 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2961 // store (to any address) in PTB,PFB or QTB,QFB.
2962 // FIXME: We could relax this restriction with a bit more work and performance
2963 // testing.
2964 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2965 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2966 if (!PStore || !QStore)
2967 return false;
2968
2969 // Now check the stores are compatible.
2970 if (!QStore->isUnordered() || !PStore->isUnordered())
2971 return false;
2972
2973 // Check that sinking the store won't cause program behavior changes. Sinking
2974 // the store out of the Q blocks won't change any behavior as we're sinking
2975 // from a block to its unconditional successor. But we're moving a store from
2976 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2977 // So we need to check that there are no aliasing loads or stores in
2978 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2979 // operations between PStore and the end of its parent block.
2980 //
2981 // The ideal way to do this is to query AliasAnalysis, but we don't
2982 // preserve AA currently so that is dangerous. Be super safe and just
2983 // check there are no other memory operations at all.
2984 for (auto &I : *QFB->getSinglePredecessor())
2985 if (I.mayReadOrWriteMemory())
2986 return false;
2987 for (auto &I : *QFB)
2988 if (&I != QStore && I.mayReadOrWriteMemory())
2989 return false;
2990 if (QTB)
2991 for (auto &I : *QTB)
2992 if (&I != QStore && I.mayReadOrWriteMemory())
2993 return false;
2994 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2995 I != E; ++I)
2996 if (&*I != PStore && I->mayReadOrWriteMemory())
2997 return false;
2998
2999 // If PostBB has more than two predecessors, we need to split it so we can
3000 // sink the store.
3001 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
3002 // We know that QFB's only successor is PostBB. And QFB has a single
3003 // predecessor. If QTB exists, then its only successor is also PostBB.
3004 // If QTB does not exist, then QFB's only predecessor has a conditional
3005 // branch to QFB and PostBB.
3006 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3007 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3008 "condstore.split");
3009 if (!NewBB)
3010 return false;
3011 PostBB = NewBB;
3012 }
3013
3014 // OK, we're going to sink the stores to PostBB. The store has to be
3015 // conditional though, so first create the predicate.
3016 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3017 ->getCondition();
3018 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3019 ->getCondition();
3020
3021 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3022 PStore->getParent());
3023 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3024 QStore->getParent(), PPHI);
3025
3026 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3027
3028 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3029 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3030
3031 if (InvertPCond)
3032 PPred = QB.CreateNot(PPred);
3033 if (InvertQCond)
3034 QPred = QB.CreateNot(QPred);
3035 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3036
3037 auto *T =
3038 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3039 QB.SetInsertPoint(T);
3040 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3041 AAMDNodes AAMD;
3042 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3043 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3044 SI->setAAMetadata(AAMD);
3045 unsigned PAlignment = PStore->getAlignment();
3046 unsigned QAlignment = QStore->getAlignment();
3047 unsigned TypeAlignment =
3048 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3049 unsigned MinAlignment;
3050 unsigned MaxAlignment;
3051 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3052 // Choose the minimum alignment. If we could prove both stores execute, we
3053 // could use biggest one. In this case, though, we only know that one of the
3054 // stores executes. And we don't know it's safe to take the alignment from a
3055 // store that doesn't execute.
3056 if (MinAlignment != 0) {
3057 // Choose the minimum of all non-zero alignments.
3058 SI->setAlignment(MinAlignment);
3059 } else if (MaxAlignment != 0) {
3060 // Choose the minimal alignment between the non-zero alignment and the ABI
3061 // default alignment for the type of the stored value.
3062 SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3063 } else {
3064 // If both alignments are zero, use ABI default alignment for the type of
3065 // the stored value.
3066 SI->setAlignment(TypeAlignment);
3067 }
3068
3069 QStore->eraseFromParent();
3070 PStore->eraseFromParent();
3071
3072 return true;
3073}
3074
3075static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3076 const DataLayout &DL) {
3077 // The intention here is to find diamonds or triangles (see below) where each
3078 // conditional block contains a store to the same address. Both of these
3079 // stores are conditional, so they can't be unconditionally sunk. But it may
3080 // be profitable to speculatively sink the stores into one merged store at the
3081 // end, and predicate the merged store on the union of the two conditions of
3082 // PBI and QBI.
3083 //
3084 // This can reduce the number of stores executed if both of the conditions are
3085 // true, and can allow the blocks to become small enough to be if-converted.
3086 // This optimization will also chain, so that ladders of test-and-set
3087 // sequences can be if-converted away.
3088 //
3089 // We only deal with simple diamonds or triangles:
3090 //
3091 // PBI or PBI or a combination of the two
3092 // / \ | \
3093 // PTB PFB | PFB
3094 // \ / | /
3095 // QBI QBI
3096 // / \ | \
3097 // QTB QFB | QFB
3098 // \ / | /
3099 // PostBB PostBB
3100 //
3101 // We model triangles as a type of diamond with a nullptr "true" block.
3102 // Triangles are canonicalized so that the fallthrough edge is represented by
3103 // a true condition, as in the diagram above.
3104 BasicBlock *PTB = PBI->getSuccessor(0);
3105 BasicBlock *PFB = PBI->getSuccessor(1);
3106 BasicBlock *QTB = QBI->getSuccessor(0);
3107 BasicBlock *QFB = QBI->getSuccessor(1);
3108 BasicBlock *PostBB = QFB->getSingleSuccessor();
3109
3110 // Make sure we have a good guess for PostBB. If QTB's only successor is
3111 // QFB, then QFB is a better PostBB.
3112 if (QTB->getSingleSuccessor() == QFB)
3113 PostBB = QFB;
3114
3115 // If we couldn't find a good PostBB, stop.
3116 if (!PostBB)
3117 return false;
3118
3119 bool InvertPCond = false, InvertQCond = false;
3120 // Canonicalize fallthroughs to the true branches.
3121 if (PFB == QBI->getParent()) {
3122 std::swap(PFB, PTB);
3123 InvertPCond = true;
3124 }
3125 if (QFB == PostBB) {
3126 std::swap(QFB, QTB);
3127 InvertQCond = true;
3128 }
3129
3130 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3131 // and QFB may not. Model fallthroughs as a nullptr block.
3132 if (PTB == QBI->getParent())
3133 PTB = nullptr;
3134 if (QTB == PostBB)
3135 QTB = nullptr;
3136
3137 // Legality bailouts. We must have at least the non-fallthrough blocks and
3138 // the post-dominating block, and the non-fallthroughs must only have one
3139 // predecessor.
3140 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3141 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3142 };
3143 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3144 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3145 return false;
3146 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3147 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3148 return false;
3149 if (!QBI->getParent()->hasNUses(2))
3150 return false;
3151
3152 // OK, this is a sequence of two diamonds or triangles.
3153 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3154 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3155 for (auto *BB : {PTB, PFB}) {
3156 if (!BB)
3157 continue;
3158 for (auto &I : *BB)
3159 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3160 PStoreAddresses.insert(SI->getPointerOperand());
3161 }
3162 for (auto *BB : {QTB, QFB}) {
3163 if (!BB)
3164 continue;
3165 for (auto &I : *BB)
3166 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3167 QStoreAddresses.insert(SI->getPointerOperand());
3168 }
3169
3170 set_intersect(PStoreAddresses, QStoreAddresses);
3171 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3172 // clear what it contains.
3173 auto &CommonAddresses = PStoreAddresses;
3174
3175 bool Changed = false;
3176 for (auto *Address : CommonAddresses)
3177 Changed |= mergeConditionalStoreToAddress(
3178 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3179 return Changed;
3180}
3181
3182/// If we have a conditional branch as a predecessor of another block,
3183/// this function tries to simplify it. We know
3184/// that PBI and BI are both conditional branches, and BI is in one of the
3185/// successor blocks of PBI - PBI branches to BI.
3186static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3187 const DataLayout &DL) {
3188 assert(PBI->isConditional() && BI->isConditional())((PBI->isConditional() && BI->isConditional()) ?
static_cast<void> (0) : __assert_fail ("PBI->isConditional() && BI->isConditional()"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3188, __PRETTY_FUNCTION__))
;
3189 BasicBlock *BB = BI->getParent();
3190
3191 // If this block ends with a branch instruction, and if there is a
3192 // predecessor that ends on a branch of the same condition, make
3193 // this conditional branch redundant.
3194 if (PBI->getCondition() == BI->getCondition() &&
3195 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3196 // Okay, the outcome of this conditional branch is statically
3197 // knowable. If this block had a single pred, handle specially.
3198 if (BB->getSinglePredecessor()) {
3199 // Turn this into a branch on constant.
3200 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3201 BI->setCondition(
3202 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3203 return true; // Nuke the branch on constant.
3204 }
3205
3206 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3207 // in the constant and simplify the block result. Subsequent passes of
3208 // simplifycfg will thread the block.
3209 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3210 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3211 PHINode *NewPN = PHINode::Create(
3212 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3213 BI->getCondition()->getName() + ".pr", &BB->front());
3214 // Okay, we're going to insert the PHI node. Since PBI is not the only
3215 // predecessor, compute the PHI'd conditional value for all of the preds.
3216 // Any predecessor where the condition is not computable we keep symbolic.
3217 for (pred_iterator PI = PB; PI != PE; ++PI) {
3218 BasicBlock *P = *PI;
3219 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3220 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3221 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3222 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3223 NewPN->addIncoming(
3224 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3225 P);
3226 } else {
3227 NewPN->addIncoming(BI->getCondition(), P);
3228 }
3229 }
3230
3231 BI->setCondition(NewPN);
3232 return true;
3233 }
3234 }
3235
3236 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3237 if (CE->canTrap())
3238 return false;
3239
3240 // If both branches are conditional and both contain stores to the same
3241 // address, remove the stores from the conditionals and create a conditional
3242 // merged store at the end.
3243 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3244 return true;
3245
3246 // If this is a conditional branch in an empty block, and if any
3247 // predecessors are a conditional branch to one of our destinations,
3248 // fold the conditions into logical ops and one cond br.
3249
3250 // Ignore dbg intrinsics.
3251 if (&*BB->instructionsWithoutDebug().begin() != BI)
3252 return false;
3253
3254 int PBIOp, BIOp;
3255 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3256 PBIOp = 0;
3257 BIOp = 0;
3258 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3259 PBIOp = 0;
3260 BIOp = 1;
3261 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3262 PBIOp = 1;
3263 BIOp = 0;
3264 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3265 PBIOp = 1;
3266 BIOp = 1;
3267 } else {
3268 return false;
3269 }
3270
3271 // Check to make sure that the other destination of this branch
3272 // isn't BB itself. If so, this is an infinite loop that will
3273 // keep getting unwound.
3274 if (PBI->getSuccessor(PBIOp) == BB)
3275 return false;
3276
3277 // Do not perform this transformation if it would require
3278 // insertion of a large number of select instructions. For targets
3279 // without predication/cmovs, this is a big pessimization.
3280
3281 // Also do not perform this transformation if any phi node in the common
3282 // destination block can trap when reached by BB or PBB (PR17073). In that
3283 // case, it would be unsafe to hoist the operation into a select instruction.
3284
3285 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3286 unsigned NumPhis = 0;
3287 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3288 ++II, ++NumPhis) {
3289 if (NumPhis > 2) // Disable this xform.
3290 return false;
3291
3292 PHINode *PN = cast<PHINode>(II);
3293 Value *BIV = PN->getIncomingValueForBlock(BB);
3294 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3295 if (CE->canTrap())
3296 return false;
3297
3298 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3299 Value *PBIV = PN->getIncomingValue(PBBIdx);
3300 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3301 if (CE->canTrap())
3302 return false;
3303 }
3304
3305 // Finally, if everything is ok, fold the branches to logical ops.
3306 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3307
3308 LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
3309 << "AND: " << *BI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
;
3310
3311 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3312 // branch in it, where one edge (OtherDest) goes back to itself but the other
3313 // exits. We don't *know* that the program avoids the infinite loop
3314 // (even though that seems likely). If we do this xform naively, we'll end up
3315 // recursively unpeeling the loop. Since we know that (after the xform is
3316 // done) that the block *is* infinite if reached, we just make it an obviously
3317 // infinite loop with no cond branch.
3318 if (OtherDest == BB) {
3319 // Insert it at the end of the function, because it's either code,
3320 // or it won't matter if it's hot. :)
3321 BasicBlock *InfLoopBlock =
3322 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3323 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3324 OtherDest = InfLoopBlock;
3325 }
3326
3327 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3328
3329 // BI may have other predecessors. Because of this, we leave
3330 // it alone, but modify PBI.
3331
3332 // Make sure we get to CommonDest on True&True directions.
3333 Value *PBICond = PBI->getCondition();
3334 IRBuilder<NoFolder> Builder(PBI);
3335 if (PBIOp)
3336 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3337
3338 Value *BICond = BI->getCondition();
3339 if (BIOp)
3340 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3341
3342 // Merge the conditions.
3343 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3344
3345 // Modify PBI to branch on the new condition to the new dests.
3346 PBI->setCondition(Cond);
3347 PBI->setSuccessor(0, CommonDest);
3348 PBI->setSuccessor(1, OtherDest);
3349
3350 // Update branch weight for PBI.
3351 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3352 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3353 bool HasWeights =
3354 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3355 SuccTrueWeight, SuccFalseWeight);
3356 if (HasWeights) {
3357 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3358 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3359 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3360 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3361 // The weight to CommonDest should be PredCommon * SuccTotal +
3362 // PredOther * SuccCommon.
3363 // The weight to OtherDest should be PredOther * SuccOther.
3364 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3365 PredOther * SuccCommon,
3366 PredOther * SuccOther};
3367 // Halve the weights if any of them cannot fit in an uint32_t
3368 FitWeights(NewWeights);
3369
3370 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3371 }
3372
3373 // OtherDest may have phi nodes. If so, add an entry from PBI's
3374 // block that are identical to the entries for BI's block.
3375 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3376
3377 // We know that the CommonDest already had an edge from PBI to
3378 // it. If it has PHIs though, the PHIs may have different
3379 // entries for BB and PBI's BB. If so, insert a select to make
3380 // them agree.
3381 for (PHINode &PN : CommonDest->phis()) {
3382 Value *BIV = PN.getIncomingValueForBlock(BB);
3383 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3384 Value *PBIV = PN.getIncomingValue(PBBIdx);
3385 if (BIV != PBIV) {
3386 // Insert a select in PBI to pick the right value.
3387 SelectInst *NV = cast<SelectInst>(
3388 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3389 PN.setIncomingValue(PBBIdx, NV);
3390 // Although the select has the same condition as PBI, the original branch
3391 // weights for PBI do not apply to the new select because the select's
3392 // 'logical' edges are incoming edges of the phi that is eliminated, not
3393 // the outgoing edges of PBI.
3394 if (HasWeights) {
3395 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3396 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3397 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3398 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3399 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3400 // The weight to PredOtherDest should be PredOther * SuccCommon.
3401 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3402 PredOther * SuccCommon};
3403
3404 FitWeights(NewWeights);
3405
3406 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3407 }
3408 }
3409 }
3410
3411 LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "INTO: " << *PBI->
getParent(); } } while (false)
;
3412 LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3413
3414 // This basic block is probably dead. We know it has at least
3415 // one fewer predecessor.
3416 return true;
3417}
3418
3419// Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3420// true or to FalseBB if Cond is false.
3421// Takes care of updating the successors and removing the old terminator.
3422// Also makes sure not to introduce new successors by assuming that edges to
3423// non-successor TrueBBs and FalseBBs aren't reachable.
3424static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3425 BasicBlock *TrueBB, BasicBlock *FalseBB,
3426 uint32_t TrueWeight,
3427 uint32_t FalseWeight) {
3428 // Remove any superfluous successor edges from the CFG.
3429 // First, figure out which successors to preserve.
3430 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3431 // successor.
3432 BasicBlock *KeepEdge1 = TrueBB;
3433 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3434
3435 // Then remove the rest.
3436 for (BasicBlock *Succ : successors(OldTerm)) {
3437 // Make sure only to keep exactly one copy of each edge.
3438 if (Succ == KeepEdge1)
3439 KeepEdge1 = nullptr;
3440 else if (Succ == KeepEdge2)
3441 KeepEdge2 = nullptr;
3442 else
3443 Succ->removePredecessor(OldTerm->getParent(),
3444 /*DontDeleteUselessPHIs=*/true);
3445 }
3446
3447 IRBuilder<> Builder(OldTerm);
3448 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3449
3450 // Insert an appropriate new terminator.
3451 if (!KeepEdge1 && !KeepEdge2) {
3452 if (TrueBB == FalseBB)
3453 // We were only looking for one successor, and it was present.
3454 // Create an unconditional branch to it.
3455 Builder.CreateBr(TrueBB);
3456 else {
3457 // We found both of the successors we were looking for.
3458 // Create a conditional branch sharing the condition of the select.
3459 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3460 if (TrueWeight != FalseWeight)
3461 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3462 }
3463 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3464 // Neither of the selected blocks were successors, so this
3465 // terminator must be unreachable.
3466 new UnreachableInst(OldTerm->getContext(), OldTerm);
3467 } else {
3468 // One of the selected values was a successor, but the other wasn't.
3469 // Insert an unconditional branch to the one that was found;
3470 // the edge to the one that wasn't must be unreachable.
3471 if (!KeepEdge1)
3472 // Only TrueBB was found.
3473 Builder.CreateBr(TrueBB);
3474 else
3475 // Only FalseBB was found.
3476 Builder.CreateBr(FalseBB);
3477 }
3478
3479 EraseTerminatorAndDCECond(OldTerm);
3480 return true;
3481}
3482
3483// Replaces
3484// (switch (select cond, X, Y)) on constant X, Y
3485// with a branch - conditional if X and Y lead to distinct BBs,
3486// unconditional otherwise.
3487static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3488 // Check for constant integer values in the select.
3489 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3490 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3491 if (!TrueVal || !FalseVal)
3492 return false;
3493
3494 // Find the relevant condition and destinations.
3495 Value *Condition = Select->getCondition();
3496 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3497 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3498
3499 // Get weight for TrueBB and FalseBB.
3500 uint32_t TrueWeight = 0, FalseWeight = 0;
3501 SmallVector<uint64_t, 8> Weights;
3502 bool HasWeights = HasBranchWeights(SI);
3503 if (HasWeights) {
3504 GetBranchWeights(SI, Weights);
3505 if (Weights.size() == 1 + SI->getNumCases()) {
3506 TrueWeight =
3507 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3508 FalseWeight =
3509 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3510 }
3511 }
3512
3513 // Perform the actual simplification.
3514 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3515 FalseWeight);
3516}
3517
3518// Replaces
3519// (indirectbr (select cond, blockaddress(@fn, BlockA),
3520// blockaddress(@fn, BlockB)))
3521// with
3522// (br cond, BlockA, BlockB).
3523static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3524 // Check that both operands of the select are block addresses.
3525 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3526 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3527 if (!TBA || !FBA)
3528 return false;
3529
3530 // Extract the actual blocks.
3531 BasicBlock *TrueBB = TBA->getBasicBlock();
3532 BasicBlock *FalseBB = FBA->getBasicBlock();
3533
3534 // Perform the actual simplification.
3535 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3536 0);
3537}
3538
3539/// This is called when we find an icmp instruction
3540/// (a seteq/setne with a constant) as the only instruction in a
3541/// block that ends with an uncond branch. We are looking for a very specific
3542/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3543/// this case, we merge the first two "or's of icmp" into a switch, but then the
3544/// default value goes to an uncond block with a seteq in it, we get something
3545/// like:
3546///
3547/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3548/// DEFAULT:
3549/// %tmp = icmp eq i8 %A, 92
3550/// br label %end
3551/// end:
3552/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3553///
3554/// We prefer to split the edge to 'end' so that there is a true/false entry to
3555/// the PHI, merging the third icmp into the switch.
3556bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3557 ICmpInst *ICI, IRBuilder<> &Builder) {
3558 BasicBlock *BB = ICI->getParent();
3559
3560 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3561 // complex.
3562 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3563 return false;
3564
3565 Value *V = ICI->getOperand(0);
3566 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3567
3568 // The pattern we're looking for is where our only predecessor is a switch on
3569 // 'V' and this block is the default case for the switch. In this case we can
3570 // fold the compared value into the switch to simplify things.
3571 BasicBlock *Pred = BB->getSinglePredecessor();
3572 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3573 return false;
3574
3575 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3576 if (SI->getCondition() != V)
3577 return false;
3578
3579 // If BB is reachable on a non-default case, then we simply know the value of
3580 // V in this block. Substitute it and constant fold the icmp instruction
3581 // away.
3582 if (SI->getDefaultDest() != BB) {
3583 ConstantInt *VVal = SI->findCaseDest(BB);
3584 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3584, __PRETTY_FUNCTION__))
;
3585 ICI->setOperand(0, VVal);
3586
3587 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3588 ICI->replaceAllUsesWith(V);
3589 ICI->eraseFromParent();
3590 }
3591 // BB is now empty, so it is likely to simplify away.
3592 return requestResimplify();
3593 }
3594
3595 // Ok, the block is reachable from the default dest. If the constant we're
3596 // comparing exists in one of the other edges, then we can constant fold ICI
3597 // and zap it.
3598 if (SI->findCaseValue(Cst) != SI->case_default()) {
3599 Value *V;
3600 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3601 V = ConstantInt::getFalse(BB->getContext());
3602 else
3603 V = ConstantInt::getTrue(BB->getContext());
3604
3605 ICI->replaceAllUsesWith(V);
3606 ICI->eraseFromParent();
3607 // BB is now empty, so it is likely to simplify away.
3608 return requestResimplify();
3609 }
3610
3611 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3612 // the block.
3613 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3614 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3615 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3616 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3617 return false;
3618
3619 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3620 // true in the PHI.
3621 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3622 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3623
3624 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3625 std::swap(DefaultCst, NewCst);
3626
3627 // Replace ICI (which is used by the PHI for the default value) with true or
3628 // false depending on if it is EQ or NE.
3629 ICI->replaceAllUsesWith(DefaultCst);
3630 ICI->eraseFromParent();
3631
3632 // Okay, the switch goes to this block on a default value. Add an edge from
3633 // the switch to the merge point on the compared value.
3634 BasicBlock *NewBB =
3635 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3636 SmallVector<uint64_t, 8> Weights;
3637 bool HasWeights = HasBranchWeights(SI);
3638 if (HasWeights) {
3639 GetBranchWeights(SI, Weights);
3640 if (Weights.size() == 1 + SI->getNumCases()) {
3641 // Split weight for default case to case for "Cst".
3642 Weights[0] = (Weights[0] + 1) >> 1;
3643 Weights.push_back(Weights[0]);
3644
3645 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3646 setBranchWeights(SI, MDWeights);
3647 }
3648 }
3649 SI->addCase(Cst, NewBB);
3650
3651 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3652 Builder.SetInsertPoint(NewBB);
3653 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3654 Builder.CreateBr(SuccBlock);
3655 PHIUse->addIncoming(NewCst, NewBB);
3656 return true;
3657}
3658
3659/// The specified branch is a conditional branch.
3660/// Check to see if it is branching on an or/and chain of icmp instructions, and
3661/// fold it into a switch instruction if so.
3662static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3663 const DataLayout &DL) {
3664 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3665 if (!Cond)
3666 return false;
3667
3668 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3669 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3670 // 'setne's and'ed together, collect them.
3671
3672 // Try to gather values from a chain of and/or to be turned into a switch
3673 ConstantComparesGatherer ConstantCompare(Cond, DL);
3674 // Unpack the result
3675 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3676 Value *CompVal = ConstantCompare.CompValue;
3677 unsigned UsedICmps = ConstantCompare.UsedICmps;
3678 Value *ExtraCase = ConstantCompare.Extra;
3679
3680 // If we didn't have a multiply compared value, fail.
3681 if (!CompVal)
3682 return false;
3683
3684 // Avoid turning single icmps into a switch.
3685 if (UsedICmps <= 1)
3686 return false;
3687
3688 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3689
3690 // There might be duplicate constants in the list, which the switch
3691 // instruction can't handle, remove them now.
3692 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3693 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3694
3695 // If Extra was used, we require at least two switch values to do the
3696 // transformation. A switch with one value is just a conditional branch.
3697 if (ExtraCase && Values.size() < 2)
3698 return false;
3699
3700 // TODO: Preserve branch weight metadata, similarly to how
3701 // FoldValueComparisonIntoPredecessors preserves it.
3702
3703 // Figure out which block is which destination.
3704 BasicBlock *DefaultBB = BI->getSuccessor(1);
3705 BasicBlock *EdgeBB = BI->getSuccessor(0);
3706 if (!TrueWhenEqual)
3707 std::swap(DefaultBB, EdgeBB);
3708
3709 BasicBlock *BB = BI->getParent();
3710
3711 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)
3712 << " 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)
3713 << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
;
3714
3715 // If there are any extra values that couldn't be folded into the switch
3716 // then we evaluate them with an explicit branch first. Split the block
3717 // right before the condbr to handle it.
3718 if (ExtraCase) {
3719 BasicBlock *NewBB =
3720 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3721 // Remove the uncond branch added to the old block.
3722 Instruction *OldTI = BB->getTerminator();
3723 Builder.SetInsertPoint(OldTI);
3724
3725 if (TrueWhenEqual)
3726 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3727 else
3728 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3729
3730 OldTI->eraseFromParent();
3731
3732 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3733 // for the edge we just added.
3734 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3735
3736 LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCasedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
3737 << "\nEXTRABB = " << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
;
3738 BB = NewBB;
3739 }
3740
3741 Builder.SetInsertPoint(BI);
3742 // Convert pointer to int before we switch.
3743 if (CompVal->getType()->isPointerTy()) {
3744 CompVal = Builder.CreatePtrToInt(
3745 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3746 }
3747
3748 // Create the new switch instruction now.
3749 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3750
3751 // Add all of the 'cases' to the switch instruction.
3752 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3753 New->addCase(Values[i], EdgeBB);
3754
3755 // We added edges from PI to the EdgeBB. As such, if there were any
3756 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3757 // the number of edges added.
3758 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3759 PHINode *PN = cast<PHINode>(BBI);
3760 Value *InVal = PN->getIncomingValueForBlock(BB);
3761 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3762 PN->addIncoming(InVal, BB);
3763 }
3764
3765 // Erase the old branch instruction.
3766 EraseTerminatorAndDCECond(BI);
3767
3768 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)
;
3769 return true;
3770}
3771
3772bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3773 if (isa<PHINode>(RI->getValue()))
3774 return SimplifyCommonResume(RI);
3775 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3776 RI->getValue() == RI->getParent()->getFirstNonPHI())
3777 // The resume must unwind the exception that caused control to branch here.
3778 return SimplifySingleResume(RI);
3779
3780 return false;
3781}
3782
3783// Simplify resume that is shared by several landing pads (phi of landing pad).
3784bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3785 BasicBlock *BB = RI->getParent();
3786
3787 // Check that there are no other instructions except for debug intrinsics
3788 // between the phi of landing pads (RI->getValue()) and resume instruction.
3789 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3790 E = RI->getIterator();
3791 while (++I != E)
3792 if (!isa<DbgInfoIntrinsic>(I))
3793 return false;
3794
3795 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3796 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3797
3798 // Check incoming blocks to see if any of them are trivial.
3799 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3800 Idx++) {
3801 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3802 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3803
3804 // If the block has other successors, we can not delete it because
3805 // it has other dependents.
3806 if (IncomingBB->getUniqueSuccessor() != BB)
3807 continue;
3808
3809 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3810 // Not the landing pad that caused the control to branch here.
3811 if (IncomingValue != LandingPad)
3812 continue;
3813
3814 bool isTrivial = true;
3815
3816 I = IncomingBB->getFirstNonPHI()->getIterator();
3817 E = IncomingBB->getTerminator()->getIterator();
3818 while (++I != E)
3819 if (!isa<DbgInfoIntrinsic>(I)) {
3820 isTrivial = false;
3821 break;
3822 }
3823
3824 if (isTrivial)
3825 TrivialUnwindBlocks.insert(IncomingBB);
3826 }
3827
3828 // If no trivial unwind blocks, don't do any simplifications.
3829 if (TrivialUnwindBlocks.empty())
3830 return false;
3831
3832 // Turn all invokes that unwind here into calls.
3833 for (auto *TrivialBB : TrivialUnwindBlocks) {
3834 // Blocks that will be simplified should be removed from the phi node.
3835 // Note there could be multiple edges to the resume block, and we need
3836 // to remove them all.
3837 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3838 BB->removePredecessor(TrivialBB, true);
3839
3840 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3841 PI != PE;) {
3842 BasicBlock *Pred = *PI++;
3843 removeUnwindEdge(Pred);
3844 }
3845
3846 // In each SimplifyCFG run, only the current processed block can be erased.
3847 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3848 // of erasing TrivialBB, we only remove the branch to the common resume
3849 // block so that we can later erase the resume block since it has no
3850 // predecessors.
3851 TrivialBB->getTerminator()->eraseFromParent();
3852 new UnreachableInst(RI->getContext(), TrivialBB);
3853 }
3854
3855 // Delete the resume block if all its predecessors have been removed.
3856 if (pred_empty(BB))
3857 BB->eraseFromParent();
3858
3859 return !TrivialUnwindBlocks.empty();
3860}
3861
3862// Simplify resume that is only used by a single (non-phi) landing pad.
3863bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3864 BasicBlock *BB = RI->getParent();
3865 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3866 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3867, __PRETTY_FUNCTION__))
3867 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3867, __PRETTY_FUNCTION__))
;
3868
3869 // Check that there are no other instructions except for debug intrinsics.
3870 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3871 while (++I != E)
3872 if (!isa<DbgInfoIntrinsic>(I))
3873 return false;
3874
3875 // Turn all invokes that unwind here into calls and delete the basic block.
3876 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3877 BasicBlock *Pred = *PI++;
3878 removeUnwindEdge(Pred);
3879 }
3880
3881 // The landingpad is now unreachable. Zap it.
3882 if (LoopHeaders)
3883 LoopHeaders->erase(BB);
3884 BB->eraseFromParent();
3885 return true;
3886}
3887
3888static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3889 // If this is a trivial cleanup pad that executes no instructions, it can be
3890 // eliminated. If the cleanup pad continues to the caller, any predecessor
3891 // that is an EH pad will be updated to continue to the caller and any
3892 // predecessor that terminates with an invoke instruction will have its invoke
3893 // instruction converted to a call instruction. If the cleanup pad being
3894 // simplified does not continue to the caller, each predecessor will be
3895 // updated to continue to the unwind destination of the cleanup pad being
3896 // simplified.
3897 BasicBlock *BB = RI->getParent();
3898 CleanupPadInst *CPInst = RI->getCleanupPad();
3899 if (CPInst->getParent() != BB)
3900 // This isn't an empty cleanup.
3901 return false;
3902
3903 // We cannot kill the pad if it has multiple uses. This typically arises
3904 // from unreachable basic blocks.
3905 if (!CPInst->hasOneUse())
3906 return false;
3907
3908 // Check that there are no other instructions except for benign intrinsics.
3909 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3910 while (++I != E) {
3911 auto *II = dyn_cast<IntrinsicInst>(I);
3912 if (!II)
3913 return false;
3914
3915 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3916 switch (IntrinsicID) {
3917 case Intrinsic::dbg_declare:
3918 case Intrinsic::dbg_value:
3919 case Intrinsic::dbg_label:
3920 case Intrinsic::lifetime_end:
3921 break;
3922 default:
3923 return false;
3924 }
3925 }
3926
3927 // If the cleanup return we are simplifying unwinds to the caller, this will
3928 // set UnwindDest to nullptr.
3929 BasicBlock *UnwindDest = RI->getUnwindDest();
3930 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3931
3932 // We're about to remove BB from the control flow. Before we do, sink any
3933 // PHINodes into the unwind destination. Doing this before changing the
3934 // control flow avoids some potentially slow checks, since we can currently
3935 // be certain that UnwindDest and BB have no common predecessors (since they
3936 // are both EH pads).
3937 if (UnwindDest) {
3938 // First, go through the PHI nodes in UnwindDest and update any nodes that
3939 // reference the block we are removing
3940 for (BasicBlock::iterator I = UnwindDest->begin(),
3941 IE = DestEHPad->getIterator();
3942 I != IE; ++I) {
3943 PHINode *DestPN = cast<PHINode>(I);
3944
3945 int Idx = DestPN->getBasicBlockIndex(BB);
3946 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3947 assert(Idx != -1)((Idx != -1) ? static_cast<void> (0) : __assert_fail ("Idx != -1"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3947, __PRETTY_FUNCTION__))
;
3948 // This PHI node has an incoming value that corresponds to a control
3949 // path through the cleanup pad we are removing. If the incoming
3950 // value is in the cleanup pad, it must be a PHINode (because we
3951 // verified above that the block is otherwise empty). Otherwise, the
3952 // value is either a constant or a value that dominates the cleanup
3953 // pad being removed.
3954 //
3955 // Because BB and UnwindDest are both EH pads, all of their
3956 // predecessors must unwind to these blocks, and since no instruction
3957 // can have multiple unwind destinations, there will be no overlap in
3958 // incoming blocks between SrcPN and DestPN.
3959 Value *SrcVal = DestPN->getIncomingValue(Idx);
3960 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3961
3962 // Remove the entry for the block we are deleting.
3963 DestPN->removeIncomingValue(Idx, false);
3964
3965 if (SrcPN && SrcPN->getParent() == BB) {
3966 // If the incoming value was a PHI node in the cleanup pad we are
3967 // removing, we need to merge that PHI node's incoming values into
3968 // DestPN.
3969 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3970 SrcIdx != SrcE; ++SrcIdx) {
3971 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3972 SrcPN->getIncomingBlock(SrcIdx));
3973 }
3974 } else {
3975 // Otherwise, the incoming value came from above BB and
3976 // so we can just reuse it. We must associate all of BB's
3977 // predecessors with this value.
3978 for (auto *pred : predecessors(BB)) {
3979 DestPN->addIncoming(SrcVal, pred);
3980 }
3981 }
3982 }
3983
3984 // Sink any remaining PHI nodes directly into UnwindDest.
3985 Instruction *InsertPt = DestEHPad;
3986 for (BasicBlock::iterator I = BB->begin(),
3987 IE = BB->getFirstNonPHI()->getIterator();
3988 I != IE;) {
3989 // The iterator must be incremented here because the instructions are
3990 // being moved to another block.
3991 PHINode *PN = cast<PHINode>(I++);
3992 if (PN->use_empty())
3993 // If the PHI node has no uses, just leave it. It will be erased
3994 // when we erase BB below.
3995 continue;
3996
3997 // Otherwise, sink this PHI node into UnwindDest.
3998 // Any predecessors to UnwindDest which are not already represented
3999 // must be back edges which inherit the value from the path through
4000 // BB. In this case, the PHI value must reference itself.
4001 for (auto *pred : predecessors(UnwindDest))
4002 if (pred != BB)
4003 PN->addIncoming(PN, pred);
4004 PN->moveBefore(InsertPt);
4005 }
4006 }
4007
4008 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4009 // The iterator must be updated here because we are removing this pred.
4010 BasicBlock *PredBB = *PI++;
4011 if (UnwindDest == nullptr) {
4012 removeUnwindEdge(PredBB);
4013 } else {
4014 Instruction *TI = PredBB->getTerminator();
4015 TI->replaceUsesOfWith(BB, UnwindDest);
4016 }
4017 }
4018
4019 // The cleanup pad is now unreachable. Zap it.
4020 BB->eraseFromParent();
4021 return true;
4022}
4023
4024// Try to merge two cleanuppads together.
4025static bool mergeCleanupPad(CleanupReturnInst *RI) {
4026 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4027 // with.
4028 BasicBlock *UnwindDest = RI->getUnwindDest();
4029 if (!UnwindDest)
4030 return false;
4031
4032 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4033 // be safe to merge without code duplication.
4034 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4035 return false;
4036
4037 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4038 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4039 if (!SuccessorCleanupPad)
4040 return false;
4041
4042 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4043 // Replace any uses of the successor cleanupad with the predecessor pad
4044 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4045 // funclet bundle operands.
4046 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4047 // Remove the old cleanuppad.
4048 SuccessorCleanupPad->eraseFromParent();
4049 // Now, we simply replace the cleanupret with a branch to the unwind
4050 // destination.
4051 BranchInst::Create(UnwindDest, RI->getParent());
4052 RI->eraseFromParent();
4053
4054 return true;
4055}
4056
4057bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4058 // It is possible to transiantly have an undef cleanuppad operand because we
4059 // have deleted some, but not all, dead blocks.
4060 // Eventually, this block will be deleted.
4061 if (isa<UndefValue>(RI->getOperand(0)))
4062 return false;
4063
4064 if (mergeCleanupPad(RI))
4065 return true;
4066
4067 if (removeEmptyCleanup(RI))
4068 return true;
4069
4070 return false;
4071}
4072
4073bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4074 BasicBlock *BB = RI->getParent();
4075 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4076 return false;
4077
4078 // Find predecessors that end with branches.
4079 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4080 SmallVector<BranchInst *, 8> CondBranchPreds;
4081 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4082 BasicBlock *P = *PI;
4083 Instruction *PTI = P->getTerminator();
4084 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4085 if (BI->isUnconditional())
4086 UncondBranchPreds.push_back(P);
4087 else
4088 CondBranchPreds.push_back(BI);
4089 }
4090 }
4091
4092 // If we found some, do the transformation!
4093 if (!UncondBranchPreds.empty() && DupRet) {
4094 while (!UncondBranchPreds.empty()) {
4095 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4096 LLVM_DEBUG(dbgs() << "FOLDING: " << *BBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
4097 << "INTO UNCOND BRANCH PRED: " << *Pred)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
;
4098 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4099 }
4100
4101 // If we eliminated all predecessors of the block, delete the block now.
4102 if (pred_empty(BB)) {
4103 // We know there are no successors, so just nuke the block.
4104 if (LoopHeaders)
4105 LoopHeaders->erase(BB);
4106 BB->eraseFromParent();
4107 }
4108
4109 return true;
4110 }
4111
4112 // Check out all of the conditional branches going to this return
4113 // instruction. If any of them just select between returns, change the
4114 // branch itself into a select/return pair.
4115 while (!CondBranchPreds.empty()) {
4116 BranchInst *BI = CondBranchPreds.pop_back_val();
4117
4118 // Check to see if the non-BB successor is also a return block.
4119 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4120 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4121 SimplifyCondBranchToTwoReturns(BI, Builder))
4122 return true;
4123 }
4124 return false;
4125}
4126
4127bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4128 BasicBlock *BB = UI->getParent();
4129
4130 bool Changed = false;
4131
4132 // If there are any instructions immediately before the unreachable that can
4133 // be removed, do so.
4134 while (UI->getIterator() != BB->begin()) {
4135 BasicBlock::iterator BBI = UI->getIterator();
4136 --BBI;
4137 // Do not delete instructions that can have side effects which might cause
4138 // the unreachable to not be reachable; specifically, calls and volatile
4139 // operations may have this effect.
4140 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4141 break;
4142
4143 if (BBI->mayHaveSideEffects()) {
4144 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4145 if (SI->isVolatile())
4146 break;
4147 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4148 if (LI->isVolatile())
4149 break;
4150 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4151 if (RMWI->isVolatile())
4152 break;
4153 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4154 if (CXI->isVolatile())
4155 break;
4156 } else if (isa<CatchPadInst>(BBI)) {
4157 // A catchpad may invoke exception object constructors and such, which
4158 // in some languages can be arbitrary code, so be conservative by
4159 // default.
4160 // For CoreCLR, it just involves a type test, so can be removed.
4161 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4162 EHPersonality::CoreCLR)
4163 break;
4164 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4165 !isa<LandingPadInst>(BBI)) {
4166 break;
4167 }
4168 // Note that deleting LandingPad's here is in fact okay, although it
4169 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4170 // all the predecessors of this block will be the unwind edges of Invokes,
4171 // and we can therefore guarantee this block will be erased.
4172 }
4173
4174 // Delete this instruction (any uses are guaranteed to be dead)
4175 if (!BBI->use_empty())
4176 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4177 BBI->eraseFromParent();
4178 Changed = true;
4179 }
4180
4181 // If the unreachable instruction is the first in the block, take a gander
4182 // at all of the predecessors of this instruction, and simplify them.
4183 if (&BB->front() != UI)
4184 return Changed;
4185
4186 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4187 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4188 Instruction *TI = Preds[i]->getTerminator();
4189 IRBuilder<> Builder(TI);
4190 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4191 if (BI->isUnconditional()) {
4192 if (BI->getSuccessor(0) == BB) {
4193 new UnreachableInst(TI->getContext(), TI);
4194 TI->eraseFromParent();
4195 Changed = true;
4196 }
4197 } else {
4198 if (BI->getSuccessor(0) == BB) {
4199 Builder.CreateBr(BI->getSuccessor(1));
4200 EraseTerminatorAndDCECond(BI);
4201 } else if (BI->getSuccessor(1) == BB) {
4202 Builder.CreateBr(BI->getSuccessor(0));
4203 EraseTerminatorAndDCECond(BI);
4204 Changed = true;
4205 }
4206 }
4207 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4208 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4209 if (i->getCaseSuccessor() != BB) {
4210 ++i;
4211 continue;
4212 }
4213 BB->removePredecessor(SI->getParent());
4214 i = SI->removeCase(i);
4215 e = SI->case_end();
4216 Changed = true;
4217 }
4218 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4219 if (II->getUnwindDest() == BB) {
4220 removeUnwindEdge(TI->getParent());
4221 Changed = true;
4222 }
4223 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4224 if (CSI->getUnwindDest() == BB) {
4225 removeUnwindEdge(TI->getParent());
4226 Changed = true;
4227 continue;
4228 }
4229
4230 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4231 E = CSI->handler_end();
4232 I != E; ++I) {
4233 if (*I == BB) {
4234 CSI->removeHandler(I);
4235 --I;
4236 --E;
4237 Changed = true;
4238 }
4239 }
4240 if (CSI->getNumHandlers() == 0) {
4241 BasicBlock *CatchSwitchBB = CSI->getParent();
4242 if (CSI->hasUnwindDest()) {
4243 // Redirect preds to the unwind dest
4244 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4245 } else {
4246 // Rewrite all preds to unwind to caller (or from invoke to call).
4247 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4248 for (BasicBlock *EHPred : EHPreds)
4249 removeUnwindEdge(EHPred);
4250 }
4251 // The catchswitch is no longer reachable.
4252 new UnreachableInst(CSI->getContext(), CSI);
4253 CSI->eraseFromParent();
4254 Changed = true;
4255 }
4256 } else if (isa<CleanupReturnInst>(TI)) {
4257 new UnreachableInst(TI->getContext(), TI);
4258 TI->eraseFromParent();
4259 Changed = true;
4260 }
4261 }
4262
4263 // If this block is now dead, remove it.
4264 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4265 // We know there are no successors, so just nuke the block.
4266 if (LoopHeaders)
4267 LoopHeaders->erase(BB);
4268 BB->eraseFromParent();
4269 return true;
4270 }
4271
4272 return Changed;
4273}
4274
4275static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4276 assert(Cases.size() >= 1)((Cases.size() >= 1) ? static_cast<void> (0) : __assert_fail
("Cases.size() >= 1", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4276, __PRETTY_FUNCTION__))
;
4277
4278 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4279 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4280 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4281 return false;
4282 }
4283 return true;
4284}
4285
4286/// Turn a switch with two reachable destinations into an integer range
4287/// comparison and branch.
4288static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4289 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4289, __PRETTY_FUNCTION__))
;
4290
4291 bool HasDefault =
4292 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4293
4294 // Partition the cases into two sets with different destinations.
4295 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4296 BasicBlock *DestB = nullptr;
4297 SmallVector<ConstantInt *, 16> CasesA;
4298 SmallVector<ConstantInt *, 16> CasesB;
4299
4300 for (auto Case : SI->cases()) {
4301 BasicBlock *Dest = Case.getCaseSuccessor();
4302 if (!DestA)
4303 DestA = Dest;
4304 if (Dest == DestA) {
4305 CasesA.push_back(Case.getCaseValue());
4306 continue;
4307 }
4308 if (!DestB)
4309 DestB = Dest;
4310 if (Dest == DestB) {
4311 CasesB.push_back(Case.getCaseValue());
4312 continue;
4313 }
4314 return false; // More than two destinations.
4315 }
4316
4317 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4318, __PRETTY_FUNCTION__))
4318 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4318, __PRETTY_FUNCTION__))
;
4319 assert(DestA != DestB)((DestA != DestB) ? static_cast<void> (0) : __assert_fail
("DestA != DestB", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4319, __PRETTY_FUNCTION__))
;
4320 assert(DestB != SI->getDefaultDest())((DestB != SI->getDefaultDest()) ? static_cast<void>
(0) : __assert_fail ("DestB != SI->getDefaultDest()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4320, __PRETTY_FUNCTION__))
;
4321 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4321, __PRETTY_FUNCTION__))
;
4322 assert(!CasesA.empty() || HasDefault)((!CasesA.empty() || HasDefault) ? static_cast<void> (0
) : __assert_fail ("!CasesA.empty() || HasDefault", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4322, __PRETTY_FUNCTION__))
;
4323
4324 // Figure out if one of the sets of cases form a contiguous range.
4325 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4326 BasicBlock *ContiguousDest = nullptr;
4327 BasicBlock *OtherDest = nullptr;
4328 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4329 ContiguousCases = &CasesA;
4330 ContiguousDest = DestA;
4331 OtherDest = DestB;
4332 } else if (CasesAreContiguous(CasesB)) {
4333 ContiguousCases = &CasesB;
4334 ContiguousDest = DestB;
4335 OtherDest = DestA;
4336 } else
4337 return false;
4338
4339 // Start building the compare and branch.
4340
4341 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4342 Constant *NumCases =
4343 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4344
4345 Value *Sub = SI->getCondition();
4346 if (!Offset->isNullValue())
4347 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4348
4349 Value *Cmp;
4350 // If NumCases overflowed, then all possible values jump to the successor.
4351 if (NumCases->isNullValue() && !ContiguousCases->empty())
4352 Cmp = ConstantInt::getTrue(SI->getContext());
4353 else
4354 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4355 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4356
4357 // Update weight for the newly-created conditional branch.
4358 if (HasBranchWeights(SI)) {
4359 SmallVector<uint64_t, 8> Weights;
4360 GetBranchWeights(SI, Weights);
4361 if (Weights.size() == 1 + SI->getNumCases()) {
4362 uint64_t TrueWeight = 0;
4363 uint64_t FalseWeight = 0;
4364 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4365 if (SI->getSuccessor(I) == ContiguousDest)
4366 TrueWeight += Weights[I];
4367 else
4368 FalseWeight += Weights[I];
4369 }
4370 while (TrueWeight > UINT32_MAX(4294967295U) || FalseWeight > UINT32_MAX(4294967295U)) {
4371 TrueWeight /= 2;
4372 FalseWeight /= 2;
4373 }
4374 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4375 }
4376 }
4377
4378 // Prune obsolete incoming values off the successors' PHI nodes.
4379 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4380 unsigned PreviousEdges = ContiguousCases->size();
4381 if (ContiguousDest == SI->getDefaultDest())
4382 ++PreviousEdges;
4383 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4384 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4385 }
4386 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4387 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4388 if (OtherDest == SI->getDefaultDest())
4389 ++PreviousEdges;
4390 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4391 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4392 }
4393
4394 // Drop the switch.
4395 SI->eraseFromParent();
4396
4397 return true;
4398}
4399
4400/// Compute masked bits for the condition of a switch
4401/// and use it to remove dead cases.
4402static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4403 const DataLayout &DL) {
4404 Value *Cond = SI->getCondition();
4405 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4406 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4407
4408 // We can also eliminate cases by determining that their values are outside of
4409 // the limited range of the condition based on how many significant (non-sign)
4410 // bits are in the condition value.
4411 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4412 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4413
4414 // Gather dead cases.
4415 SmallVector<ConstantInt *, 8> DeadCases;
4416 for (auto &Case : SI->cases()) {
4417 const APInt &CaseVal = Case.getCaseValue()->getValue();
4418 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4419 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4420 DeadCases.push_back(Case.getCaseValue());
4421 LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
4422 << " is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
;
4423 }
4424 }
4425
4426 // If we can prove that the cases must cover all possible values, the
4427 // default destination becomes dead and we can remove it. If we know some
4428 // of the bits in the value, we can use that to more precisely compute the
4429 // number of possible unique case values.
4430 bool HasDefault =
4431 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4432 const unsigned NumUnknownBits =
4433 Bits - (Known.Zero | Known.One).countPopulation();
4434 assert(NumUnknownBits <= Bits)((NumUnknownBits <= Bits) ? static_cast<void> (0) : __assert_fail
("NumUnknownBits <= Bits", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4434, __PRETTY_FUNCTION__))
;
4435 if (HasDefault && DeadCases.empty() &&
4436 NumUnknownBits < 64 /* avoid overflow */ &&
4437 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4438 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)
;
4439 BasicBlock *NewDefault =
4440 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4441 SI->setDefaultDest(&*NewDefault);
4442 SplitBlock(&*NewDefault, &NewDefault->front());
4443 auto *OldTI = NewDefault->getTerminator();
4444 new UnreachableInst(SI->getContext(), OldTI);
4445 EraseTerminatorAndDCECond(OldTI);
4446 return true;
4447 }
4448
4449 SmallVector<uint64_t, 8> Weights;
4450 bool HasWeight = HasBranchWeights(SI);
4451 if (HasWeight) {
4452 GetBranchWeights(SI, Weights);
4453 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4454 }
4455
4456 // Remove dead cases from the switch.
4457 for (ConstantInt *DeadCase : DeadCases) {
4458 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4459 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4460, __PRETTY_FUNCTION__))
4460 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4460, __PRETTY_FUNCTION__))
;
4461 if (HasWeight) {
4462 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4463 Weights.pop_back();
4464 }
4465
4466 // Prune unused values from PHI nodes.
4467 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4468 SI->removeCase(CaseI);
4469 }
4470 if (HasWeight && Weights.size() >= 2) {
4471 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4472 setBranchWeights(SI, MDWeights);
4473 }
4474
4475 return !DeadCases.empty();
4476}
4477
4478/// If BB would be eligible for simplification by
4479/// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4480/// by an unconditional branch), look at the phi node for BB in the successor
4481/// block and see if the incoming value is equal to CaseValue. If so, return
4482/// the phi node, and set PhiIndex to BB's index in the phi node.
4483static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4484 BasicBlock *BB, int *PhiIndex) {
4485 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4486 return nullptr; // BB must be empty to be a candidate for simplification.
4487 if (!BB->getSinglePredecessor())
4488 return nullptr; // BB must be dominated by the switch.
4489
4490 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4491 if (!Branch || !Branch->isUnconditional())
4492 return nullptr; // Terminator must be unconditional branch.
4493
4494 BasicBlock *Succ = Branch->getSuccessor(0);
4495
4496 for (PHINode &PHI : Succ->phis()) {
4497 int Idx = PHI.getBasicBlockIndex(BB);
4498 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4498, __PRETTY_FUNCTION__))
;
4499
4500 Value *InValue = PHI.getIncomingValue(Idx);
4501 if (InValue != CaseValue)
4502 continue;
4503
4504 *PhiIndex = Idx;
4505 return &PHI;
4506 }
4507
4508 return nullptr;
4509}
4510
4511/// Try to forward the condition of a switch instruction to a phi node
4512/// dominated by the switch, if that would mean that some of the destination
4513/// blocks of the switch can be folded away. Return true if a change is made.
4514static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4515 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4516
4517 ForwardingNodesMap ForwardingNodes;
4518 BasicBlock *SwitchBlock = SI->getParent();
4519 bool Changed = false;
4520 for (auto &Case : SI->cases()) {
4521 ConstantInt *CaseValue = Case.getCaseValue();
4522 BasicBlock *CaseDest = Case.getCaseSuccessor();
4523
4524 // Replace phi operands in successor blocks that are using the constant case
4525 // value rather than the switch condition variable:
4526 // switchbb:
4527 // switch i32 %x, label %default [
4528 // i32 17, label %succ
4529 // ...
4530 // succ:
4531 // %r = phi i32 ... [ 17, %switchbb ] ...
4532 // -->
4533 // %r = phi i32 ... [ %x, %switchbb ] ...
4534
4535 for (PHINode &Phi : CaseDest->phis()) {
4536 // This only works if there is exactly 1 incoming edge from the switch to
4537 // a phi. If there is >1, that means multiple cases of the switch map to 1
4538 // value in the phi, and that phi value is not the switch condition. Thus,
4539 // this transform would not make sense (the phi would be invalid because
4540 // a phi can't have different incoming values from the same block).
4541 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4542 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4543 count(Phi.blocks(), SwitchBlock) == 1) {
4544 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4545 Changed = true;
4546 }
4547 }
4548
4549 // Collect phi nodes that are indirectly using this switch's case constants.
4550 int PhiIdx;
4551 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4552 ForwardingNodes[Phi].push_back(PhiIdx);
4553 }
4554
4555 for (auto &ForwardingNode : ForwardingNodes) {
4556 PHINode *Phi = ForwardingNode.first;
4557 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4558 if (Indexes.size() < 2)
4559 continue;
4560
4561 for (int Index : Indexes)
4562 Phi->setIncomingValue(Index, SI->getCondition());
4563 Changed = true;
4564 }
4565
4566 return Changed;
4567}
4568
4569/// Return true if the backend will be able to handle
4570/// initializing an array of constants like C.
4571static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4572 if (C->isThreadDependent())
4573 return false;
4574 if (C->isDLLImportDependent())
4575 return false;
4576
4577 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4578 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4579 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4580 return false;
4581
4582 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4583 if (!CE->isGEPWithNoNotionalOverIndexing())
4584 return false;
4585 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4586 return false;
4587 }
4588
4589 if (!TTI.shouldBuildLookupTablesForConstant(C))
4590 return false;
4591
4592 return true;
4593}
4594
4595/// If V is a Constant, return it. Otherwise, try to look up
4596/// its constant value in ConstantPool, returning 0 if it's not there.
4597static Constant *
4598LookupConstant(Value *V,
4599 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4600 if (Constant *C = dyn_cast<Constant>(V))
4601 return C;
4602 return ConstantPool.lookup(V);
4603}
4604
4605/// Try to fold instruction I into a constant. This works for
4606/// simple instructions such as binary operations where both operands are
4607/// constant or can be replaced by constants from the ConstantPool. Returns the
4608/// resulting constant on success, 0 otherwise.
4609static Constant *
4610ConstantFold(Instruction *I, const DataLayout &DL,
4611 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4612 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4613 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4614 if (!A)
4615 return nullptr;
4616 if (A->isAllOnesValue())
4617 return LookupConstant(Select->getTrueValue(), ConstantPool);
4618 if (A->isNullValue())
4619 return LookupConstant(Select->getFalseValue(), ConstantPool);
4620 return nullptr;
4621 }
4622
4623 SmallVector<Constant *, 4> COps;
4624 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4625 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4626 COps.push_back(A);
4627 else
4628 return nullptr;
4629 }
4630
4631 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4632 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4633 COps[1], DL);
4634 }
4635
4636 return ConstantFoldInstOperands(I, COps, DL);
4637}
4638
4639/// Try to determine the resulting constant values in phi nodes
4640/// at the common destination basic block, *CommonDest, for one of the case
4641/// destionations CaseDest corresponding to value CaseVal (0 for the default
4642/// case), of a switch instruction SI.
4643static bool
4644GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4645 BasicBlock **CommonDest,
4646 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4647 const DataLayout &DL, const TargetTransformInfo &TTI) {
4648 // The block from which we enter the common destination.
4649 BasicBlock *Pred = SI->getParent();
4650
4651 // If CaseDest is empty except for some side-effect free instructions through
4652 // which we can constant-propagate the CaseVal, continue to its successor.
4653 SmallDenseMap<Value *, Constant *> ConstantPool;
4654 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4655 for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4656 if (I.isTerminator()) {
4657 // If the terminator is a simple branch, continue to the next block.
4658 if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4659 return false;
4660 Pred = CaseDest;
4661 CaseDest = I.getSuccessor(0);
4662 } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4663 // Instruction is side-effect free and constant.
4664
4665 // If the instruction has uses outside this block or a phi node slot for
4666 // the block, it is not safe to bypass the instruction since it would then
4667 // no longer dominate all its uses.
4668 for (auto &Use : I.uses()) {
4669 User *User = Use.getUser();
4670 if (Instruction *I = dyn_cast<Instruction>(User))
4671 if (I->getParent() == CaseDest)
4672 continue;
4673 if (PHINode *Phi = dyn_cast<PHINode>(User))
4674 if (Phi->getIncomingBlock(Use) == CaseDest)
4675 continue;
4676 return false;
4677 }
4678
4679 ConstantPool.insert(std::make_pair(&I, C));
4680 } else {
4681 break;
4682 }
4683 }
4684
4685 // If we did not have a CommonDest before, use the current one.
4686 if (!*CommonDest)
4687 *CommonDest = CaseDest;
4688 // If the destination isn't the common one, abort.
4689 if (CaseDest != *CommonDest)
4690 return false;
4691
4692 // Get the values for this case from phi nodes in the destination block.
4693 for (PHINode &PHI : (*CommonDest)->phis()) {
4694 int Idx = PHI.getBasicBlockIndex(Pred);
4695 if (Idx == -1)
4696 continue;
4697
4698 Constant *ConstVal =
4699 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4700 if (!ConstVal)
4701 return false;
4702
4703 // Be conservative about which kinds of constants we support.
4704 if (!ValidLookupTableConstant(ConstVal, TTI))
4705 return false;
4706
4707 Res.push_back(std::make_pair(&PHI, ConstVal));
4708 }
4709
4710 return Res.size() > 0;
4711}
4712
4713// Helper function used to add CaseVal to the list of cases that generate
4714// Result. Returns the updated number of cases that generate this result.
4715static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4716 SwitchCaseResultVectorTy &UniqueResults,
4717 Constant *Result) {
4718 for (auto &I : UniqueResults) {
4719 if (I.first == Result) {
4720 I.second.push_back(CaseVal);
4721 return I.second.size();
4722 }
4723 }
4724 UniqueResults.push_back(
4725 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4726 return 1;
4727}
4728
4729// Helper function that initializes a map containing
4730// results for the PHI node of the common destination block for a switch
4731// instruction. Returns false if multiple PHI nodes have been found or if
4732// there is not a common destination block for the switch.
4733static bool
4734InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4735 SwitchCaseResultVectorTy &UniqueResults,
4736 Constant *&DefaultResult, const DataLayout &DL,
4737 const TargetTransformInfo &TTI,
4738 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4739 for (auto &I : SI->cases()) {
4740 ConstantInt *CaseVal = I.getCaseValue();
4741
4742 // Resulting value at phi nodes for this case value.
4743 SwitchCaseResultsTy Results;
4744 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4745 DL, TTI))
4746 return false;
4747
4748 // Only one value per case is permitted.
4749 if (Results.size() > 1)
4750 return false;
4751
4752 // Add the case->result mapping to UniqueResults.
4753 const uintptr_t NumCasesForResult =
4754 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4755
4756 // Early out if there are too many cases for this result.
4757 if (NumCasesForResult > MaxCasesPerResult)
4758 return false;
4759
4760 // Early out if there are too many unique results.
4761 if (UniqueResults.size() > MaxUniqueResults)
4762 return false;
4763
4764 // Check the PHI consistency.
4765 if (!PHI)
4766 PHI = Results[0].first;
4767 else if (PHI != Results[0].first)
4768 return false;
4769 }
4770 // Find the default result value.
4771 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4772 BasicBlock *DefaultDest = SI->getDefaultDest();
4773 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4774 DL, TTI);
4775 // If the default value is not found abort unless the default destination
4776 // is unreachable.
4777 DefaultResult =
4778 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4779 if ((!DefaultResult &&
4780 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4781 return false;
4782
4783 return true;
4784}
4785
4786// Helper function that checks if it is possible to transform a switch with only
4787// two cases (or two cases + default) that produces a result into a select.
4788// Example:
4789// switch (a) {
4790// case 10: %0 = icmp eq i32 %a, 10
4791// return 10; %1 = select i1 %0, i32 10, i32 4
4792// case 20: ----> %2 = icmp eq i32 %a, 20
4793// return 2; %3 = select i1 %2, i32 2, i32 %1
4794// default:
4795// return 4;
4796// }
4797static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4798 Constant *DefaultResult, Value *Condition,
4799 IRBuilder<> &Builder) {
4800 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4801, __PRETTY_FUNCTION__))
4801 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4801, __PRETTY_FUNCTION__))
;
4802 // If we are selecting between only two cases transform into a simple
4803 // select or a two-way select if default is possible.
4804 if (ResultVector[0].second.size() == 1 &&
4805 ResultVector[1].second.size() == 1) {
4806 ConstantInt *const FirstCase = ResultVector[0].second[0];
4807 ConstantInt *const SecondCase = ResultVector[1].second[0];
4808
4809 bool DefaultCanTrigger = DefaultResult;
4810 Value *SelectValue = ResultVector[1].first;
4811 if (DefaultCanTrigger) {
4812 Value *const ValueCompare =
4813 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4814 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4815 DefaultResult, "switch.select");
4816 }
4817 Value *const ValueCompare =
4818 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4819 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4820 SelectValue, "switch.select");
4821 }
4822
4823 return nullptr;
4824}
4825
4826// Helper function to cleanup a switch instruction that has been converted into
4827// a select, fixing up PHI nodes and basic blocks.
4828static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4829 Value *SelectValue,
4830 IRBuilder<> &Builder) {
4831 BasicBlock *SelectBB = SI->getParent();
4832 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4833 PHI->removeIncomingValue(SelectBB);
4834 PHI->addIncoming(SelectValue, SelectBB);
4835
4836 Builder.CreateBr(PHI->getParent());
4837
4838 // Remove the switch.
4839 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4840 BasicBlock *Succ = SI->getSuccessor(i);
4841
4842 if (Succ == PHI->getParent())
4843 continue;
4844 Succ->removePredecessor(SelectBB);
4845 }
4846 SI->eraseFromParent();
4847}
4848
4849/// If the switch is only used to initialize one or more
4850/// phi nodes in a common successor block with only two different
4851/// constant values, replace the switch with select.
4852static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4853 const DataLayout &DL,
4854 const TargetTransformInfo &TTI) {
4855 Value *const Cond = SI->getCondition();
4856 PHINode *PHI = nullptr;
4857 BasicBlock *CommonDest = nullptr;
4858 Constant *DefaultResult;
4859 SwitchCaseResultVectorTy UniqueResults;
4860 // Collect all the cases that will deliver the same value from the switch.
4861 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4862 DL, TTI, 2, 1))
4863 return false;
4864 // Selects choose between maximum two values.
4865 if (UniqueResults.size() != 2)
4866 return false;
4867 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4867, __PRETTY_FUNCTION__))
;
4868
4869 Builder.SetInsertPoint(SI);
4870 Value *SelectValue =
4871 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4872 if (SelectValue) {
4873 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4874 return true;
4875 }
4876 // The switch couldn't be converted into a select.
4877 return false;
4878}
4879
4880namespace {
4881
4882/// This class represents a lookup table that can be used to replace a switch.
4883class SwitchLookupTable {
4884public:
4885 /// Create a lookup table to use as a switch replacement with the contents
4886 /// of Values, using DefaultValue to fill any holes in the table.
4887 SwitchLookupTable(
4888 Module &M, uint64_t TableSize, ConstantInt *Offset,
4889 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4890 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4891
4892 /// Build instructions with Builder to retrieve the value at
4893 /// the position given by Index in the lookup table.
4894 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4895
4896 /// Return true if a table with TableSize elements of
4897 /// type ElementType would fit in a target-legal register.
4898 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4899 Type *ElementType);
4900
4901private:
4902 // Depending on the contents of the table, it can be represented in
4903 // different ways.
4904 enum {
4905 // For tables where each element contains the same value, we just have to
4906 // store that single value and return it for each lookup.
4907 SingleValueKind,
4908
4909 // For tables where there is a linear relationship between table index
4910 // and values. We calculate the result with a simple multiplication
4911 // and addition instead of a table lookup.
4912 LinearMapKind,
4913
4914 // For small tables with integer elements, we can pack them into a bitmap
4915 // that fits into a target-legal register. Values are retrieved by
4916 // shift and mask operations.
4917 BitMapKind,
4918
4919 // The table is stored as an array of values. Values are retrieved by load
4920 // instructions from the table.
4921 ArrayKind
4922 } Kind;
4923
4924 // For SingleValueKind, this is the single value.
4925 Constant *SingleValue = nullptr;
4926
4927 // For BitMapKind, this is the bitmap.
4928 ConstantInt *BitMap = nullptr;
4929 IntegerType *BitMapElementTy = nullptr;
4930
4931 // For LinearMapKind, these are the constants used to derive the value.
4932 ConstantInt *LinearOffset = nullptr;
4933 ConstantInt *LinearMultiplier = nullptr;
4934
4935 // For ArrayKind, this is the array.
4936 GlobalVariable *Array = nullptr;
4937};
4938
4939} // end anonymous namespace
4940
4941SwitchLookupTable::SwitchLookupTable(
4942 Module &M, uint64_t TableSize, ConstantInt *Offset,
4943 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4944 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4945 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4945, __PRETTY_FUNCTION__))
;
4946 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4946, __PRETTY_FUNCTION__))
;
4947
4948 // If all values in the table are equal, this is that value.
4949 SingleValue = Values.begin()->second;
4950
4951 Type *ValueType = Values.begin()->second->getType();
4952
4953 // Build up the table contents.
4954 SmallVector<Constant *, 64> TableContents(TableSize);
4955 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4956 ConstantInt *CaseVal = Values[I].first;
4957 Constant *CaseRes = Values[I].second;
4958 assert(CaseRes->getType() == ValueType)((CaseRes->getType() == ValueType) ? static_cast<void>
(0) : __assert_fail ("CaseRes->getType() == ValueType", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4958, __PRETTY_FUNCTION__))
;
4959
4960 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4961 TableContents[Idx] = CaseRes;
4962
4963 if (CaseRes != SingleValue)
4964 SingleValue = nullptr;
4965 }
4966
4967 // Fill in any holes in the table with the default result.
4968 if (Values.size() < TableSize) {
4969 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4970, __PRETTY_FUNCTION__))
4970 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4970, __PRETTY_FUNCTION__))
;
4971 assert(DefaultValue->getType() == ValueType)((DefaultValue->getType() == ValueType) ? static_cast<void
> (0) : __assert_fail ("DefaultValue->getType() == ValueType"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4971, __PRETTY_FUNCTION__))
;
4972 for (uint64_t I = 0; I < TableSize; ++I) {
4973 if (!TableContents[I])
4974 TableContents[I] = DefaultValue;
4975 }
4976
4977 if (DefaultValue != SingleValue)
4978 SingleValue = nullptr;
4979 }
4980
4981 // If each element in the table contains the same value, we only need to store
4982 // that single value.
4983 if (SingleValue) {
4984 Kind = SingleValueKind;
4985 return;
4986 }
4987
4988 // Check if we can derive the value with a linear transformation from the
4989 // table index.
4990 if (isa<IntegerType>(ValueType)) {
4991 bool LinearMappingPossible = true;
4992 APInt PrevVal;
4993 APInt DistToPrev;
4994 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4994, __PRETTY_FUNCTION__))
;
4995 // Check if there is the same distance between two consecutive values.
4996 for (uint64_t I = 0; I < TableSize; ++I) {
4997 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4998 if (!ConstVal) {
4999 // This is an undef. We could deal with it, but undefs in lookup tables
5000 // are very seldom. It's probably not worth the additional complexity.
5001 LinearMappingPossible = false;
5002 break;
5003 }
5004 const APInt &Val = ConstVal->getValue();
5005 if (I != 0) {
5006 APInt Dist = Val - PrevVal;
5007 if (I == 1) {
5008 DistToPrev = Dist;
5009 } else if (Dist != DistToPrev) {
5010 LinearMappingPossible = false;
5011 break;
5012 }
5013 }
5014 PrevVal = Val;
5015 }
5016 if (LinearMappingPossible) {
5017 LinearOffset = cast<ConstantInt>(TableContents[0]);
5018 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5019 Kind = LinearMapKind;
5020 ++NumLinearMaps;
5021 return;
5022 }
5023 }
5024
5025 // If the type is integer and the table fits in a register, build a bitmap.
5026 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5027 IntegerType *IT = cast<IntegerType>(ValueType);
5028 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5029 for (uint64_t I = TableSize; I > 0; --I) {
5030 TableInt <<= IT->getBitWidth();
5031 // Insert values into the bitmap. Undef values are set to zero.
5032 if (!isa<UndefValue>(TableContents[I - 1])) {
5033 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5034 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5035 }
5036 }
5037 BitMap = ConstantInt::get(M.getContext(), TableInt);
5038 BitMapElementTy = IT;
5039 Kind = BitMapKind;
5040 ++NumBitMaps;
5041 return;
5042 }
5043
5044 // Store the table in an array.
5045 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5046 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5047
5048 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5049 GlobalVariable::PrivateLinkage, Initializer,
5050 "switch.table." + FuncName);
5051 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5052 // Set the alignment to that of an array items. We will be only loading one
5053 // value out of it.
5054 Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5055 Kind = ArrayKind;
5056}
5057
5058Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5059 switch (Kind) {
5060 case SingleValueKind:
5061 return SingleValue;
5062 case LinearMapKind: {
5063 // Derive the result value from the input value.
5064 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5065 false, "switch.idx.cast");
5066 if (!LinearMultiplier->isOne())
5067 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5068 if (!LinearOffset->isZero())
5069 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5070 return Result;
5071 }
5072 case BitMapKind: {
5073 // Type of the bitmap (e.g. i59).
5074 IntegerType *MapTy = BitMap->getType();
5075
5076 // Cast Index to the same type as the bitmap.
5077 // Note: The Index is <= the number of elements in the table, so
5078 // truncating it to the width of the bitmask is safe.
5079 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5080
5081 // Multiply the shift amount by the element width.
5082 ShiftAmt = Builder.CreateMul(
5083 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5084 "switch.shiftamt");
5085
5086 // Shift down.
5087 Value *DownShifted =
5088 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5089 // Mask off.
5090 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5091 }
5092 case ArrayKind: {
5093 // Make sure the table index will not overflow when treated as signed.
5094 IntegerType *IT = cast<IntegerType>(Index->getType());
5095 uint64_t TableSize =
5096 Array->getInitializer()->getType()->getArrayNumElements();
5097 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5098 Index = Builder.CreateZExt(
5099 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5100 "switch.tableidx.zext");
5101
5102 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5103 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5104 GEPIndices, "switch.gep");
5105 return Builder.CreateLoad(GEP, "switch.load");
5106 }
5107 }
5108 llvm_unreachable("Unknown lookup table kind!")::llvm::llvm_unreachable_internal("Unknown lookup table kind!"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5108)
;
5109}
5110
5111bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5112 uint64_t TableSize,
5113 Type *ElementType) {
5114 auto *IT = dyn_cast<IntegerType>(ElementType);
5115 if (!IT)
5116 return false;
5117 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5118 // are <= 15, we could try to narrow the type.
5119
5120 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5121 if (TableSize >= UINT_MAX(2147483647 *2U +1U) / IT->getBitWidth())
5122 return false;
5123 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5124}
5125
5126/// Determine whether a lookup table should be built for this switch, based on
5127/// the number of cases, size of the table, and the types of the results.
5128static bool
5129ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5130 const TargetTransformInfo &TTI, const DataLayout &DL,
5131 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5132 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX(18446744073709551615UL) / 10)
5133 return false; // TableSize overflowed, or mul below might overflow.
5134
5135 bool AllTablesFitInRegister = true;
5136 bool HasIllegalType = false;
5137 for (const auto &I : ResultTypes) {
5138 Type *Ty = I.second;
5139
5140 // Saturate this flag to true.
5141 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5142
5143 // Saturate this flag to false.
5144 AllTablesFitInRegister =
5145 AllTablesFitInRegister &&
5146 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5147
5148 // If both flags saturate, we're done. NOTE: This *only* works with
5149 // saturating flags, and all flags have to saturate first due to the
5150 // non-deterministic behavior of iterating over a dense map.
5151 if (HasIllegalType && !AllTablesFitInRegister)
5152 break;
5153 }
5154
5155 // If each table would fit in a register, we should build it anyway.
5156 if (AllTablesFitInRegister)
5157 return true;
5158
5159 // Don't build a table that doesn't fit in-register if it has illegal types.
5160 if (HasIllegalType)
5161 return false;
5162
5163 // The table density should be at least 40%. This is the same criterion as for
5164 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5165 // FIXME: Find the best cut-off.
5166 return SI->getNumCases() * 10 >= TableSize * 4;
5167}
5168
5169/// Try to reuse the switch table index compare. Following pattern:
5170/// \code
5171/// if (idx < tablesize)
5172/// r = table[idx]; // table does not contain default_value
5173/// else
5174/// r = default_value;
5175/// if (r != default_value)
5176/// ...
5177/// \endcode
5178/// Is optimized to:
5179/// \code
5180/// cond = idx < tablesize;
5181/// if (cond)
5182/// r = table[idx];
5183/// else
5184/// r = default_value;
5185/// if (cond)
5186/// ...
5187/// \endcode
5188/// Jump threading will then eliminate the second if(cond).
5189static void reuseTableCompare(
5190 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5191 Constant *DefaultValue,
5192 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5193 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5194 if (!CmpInst)
5195 return;
5196
5197 // We require that the compare is in the same block as the phi so that jump
5198 // threading can do its work afterwards.
5199 if (CmpInst->getParent() != PhiBlock)
5200 return;
5201
5202 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5203 if (!CmpOp1)
5204 return;
5205
5206 Value *RangeCmp = RangeCheckBranch->getCondition();
5207 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5208 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5209
5210 // Check if the compare with the default value is constant true or false.
5211 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5212 DefaultValue, CmpOp1, true);
5213 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5214 return;
5215
5216 // Check if the compare with the case values is distinct from the default
5217 // compare result.
5218 for (auto ValuePair : Values) {
5219 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5220 ValuePair.second, CmpOp1, true);
5221 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5222 return;
5223 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5224, __PRETTY_FUNCTION__))
5224 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5224, __PRETTY_FUNCTION__))
;
5225 }
5226
5227 // Check if the branch instruction dominates the phi node. It's a simple
5228 // dominance check, but sufficient for our needs.
5229 // Although this check is invariant in the calling loops, it's better to do it
5230 // at this late stage. Practically we do it at most once for a switch.
5231 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5232 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5233 BasicBlock *Pred = *PI;
5234 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5235 return;
5236 }
5237
5238 if (DefaultConst == FalseConst) {
5239 // The compare yields the same result. We can replace it.
5240 CmpInst->replaceAllUsesWith(RangeCmp);
5241 ++NumTableCmpReuses;
5242 } else {
5243 // The compare yields the same result, just inverted. We can replace it.
5244 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5245 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5246 RangeCheckBranch);
5247 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5248 ++NumTableCmpReuses;
5249 }
5250}
5251
5252/// If the switch is only used to initialize one or more phi nodes in a common
5253/// successor block with different constant values, replace the switch with
5254/// lookup tables.
5255static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5256 const DataLayout &DL,
5257 const TargetTransformInfo &TTI) {
5258 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5258, __PRETTY_FUNCTION__))
;
5259
5260 Function *Fn = SI->getParent()->getParent();
5261 // Only build lookup table when we have a target that supports it or the
5262 // attribute is not set.
5263 if (!TTI.shouldBuildLookupTables() ||
5264 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5265 return false;
5266
5267 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5268 // split off a dense part and build a lookup table for that.
5269
5270 // FIXME: This creates arrays of GEPs to constant strings, which means each
5271 // GEP needs a runtime relocation in PIC code. We should just build one big
5272 // string and lookup indices into that.
5273
5274 // Ignore switches with less than three cases. Lookup tables will not make
5275 // them faster, so we don't analyze them.
5276 if (SI->getNumCases() < 3)
5277 return false;
5278
5279 // Figure out the corresponding result for each case value and phi node in the
5280 // common destination, as well as the min and max case values.
5281 assert(!empty(SI->cases()))((!empty(SI->cases())) ? static_cast<void> (0) : __assert_fail
("!empty(SI->cases())", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5281, __PRETTY_FUNCTION__))
;
5282 SwitchInst::CaseIt CI = SI->case_begin();
5283 ConstantInt *MinCaseVal = CI->getCaseValue();
5284 ConstantInt *MaxCaseVal = CI->getCaseValue();
5285
5286 BasicBlock *CommonDest = nullptr;
5287
5288 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5289 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5290
5291 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5292 SmallDenseMap<PHINode *, Type *> ResultTypes;
5293 SmallVector<PHINode *, 4> PHIs;
5294
5295 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5296 ConstantInt *CaseVal = CI->getCaseValue();
5297 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5298 MinCaseVal = CaseVal;
5299 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5300 MaxCaseVal = CaseVal;
5301
5302 // Resulting value at phi nodes for this case value.
5303 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5304 ResultsTy Results;
5305 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5306 Results, DL, TTI))
5307 return false;
5308
5309 // Append the result from this case to the list for each phi.
5310 for (const auto &I : Results) {
5311 PHINode *PHI = I.first;
5312 Constant *Value = I.second;
5313 if (!ResultLists.count(PHI))
5314 PHIs.push_back(PHI);
5315 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5316 }
5317 }
5318
5319 // Keep track of the result types.
5320 for (PHINode *PHI : PHIs) {
5321 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5322 }
5323
5324 uint64_t NumResults = ResultLists[PHIs[0]].size();
5325 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5326 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5327 bool TableHasHoles = (NumResults < TableSize);
5328
5329 // If the table has holes, we need a constant result for the default case
5330 // or a bitmask that fits in a register.
5331 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5332 bool HasDefaultResults =
5333 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5334 DefaultResultsList, DL, TTI);
5335
5336 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5337 if (NeedMask) {
5338 // As an extra penalty for the validity test we require more cases.
5339 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5340 return false;
5341 if (!DL.fitsInLegalInteger(TableSize))
5342 return false;
5343 }
5344
5345 for (const auto &I : DefaultResultsList) {
5346 PHINode *PHI = I.first;
5347 Constant *Result = I.second;
5348 DefaultResults[PHI] = Result;
5349 }
5350
5351 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5352 return false;
5353
5354 // Create the BB that does the lookups.
5355 Module &Mod = *CommonDest->getParent()->getParent();
5356 BasicBlock *LookupBB = BasicBlock::Create(
5357 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5358
5359 // Compute the table index value.
5360 Builder.SetInsertPoint(SI);
5361 Value *TableIndex;
5362 if (MinCaseVal->isNullValue())
5363 TableIndex = SI->getCondition();
5364 else
5365 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5366 "switch.tableidx");
5367
5368 // Compute the maximum table size representable by the integer type we are
5369 // switching upon.
5370 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5371 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX(18446744073709551615UL) : 1ULL << CaseSize;
5372 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5374, __PRETTY_FUNCTION__))
5373 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5374, __PRETTY_FUNCTION__))
5374 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5374, __PRETTY_FUNCTION__))
;
5375
5376 // If the default destination is unreachable, or if the lookup table covers
5377 // all values of the conditional variable, branch directly to the lookup table
5378 // BB. Otherwise, check that the condition is within the case range.
5379 const bool DefaultIsReachable =
5380 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5381 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5382 BranchInst *RangeCheckBranch = nullptr;
5383
5384 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5385 Builder.CreateBr(LookupBB);
5386 // Note: We call removeProdecessor later since we need to be able to get the
5387 // PHI value for the default case in case we're using a bit mask.
5388 } else {
5389 Value *Cmp = Builder.CreateICmpULT(
5390 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5391 RangeCheckBranch =
5392 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5393 }
5394
5395 // Populate the BB that does the lookups.
5396 Builder.SetInsertPoint(LookupBB);
5397
5398 if (NeedMask) {
5399 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5400 // re-purposed to do the hole check, and we create a new LookupBB.
5401 BasicBlock *MaskBB = LookupBB;
5402 MaskBB->setName("switch.hole_check");
5403 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5404 CommonDest->getParent(), CommonDest);
5405
5406 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5407 // unnecessary illegal types.
5408 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5409 APInt MaskInt(TableSizePowOf2, 0);
5410 APInt One(TableSizePowOf2, 1);
5411 // Build bitmask; fill in a 1 bit for every case.
5412 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5413 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5414 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5415 .getLimitedValue();
5416 MaskInt |= One << Idx;
5417 }
5418 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5419
5420 // Get the TableIndex'th bit of the bitmask.
5421 // If this bit is 0 (meaning hole) jump to the default destination,
5422 // else continue with table lookup.
5423 IntegerType *MapTy = TableMask->getType();
5424 Value *MaskIndex =
5425 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5426 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5427 Value *LoBit = Builder.CreateTrunc(
5428 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5429 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5430
5431 Builder.SetInsertPoint(LookupBB);
5432 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5433 }
5434
5435 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5436 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5437 // do not delete PHINodes here.
5438 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5439 /*DontDeleteUselessPHIs=*/true);
5440 }
5441
5442 bool ReturnedEarly = false;
5443 for (PHINode *PHI : PHIs) {
5444 const ResultListTy &ResultList = ResultLists[PHI];
5445
5446 // If using a bitmask, use any value to fill the lookup table holes.
5447 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5448 StringRef FuncName = Fn->getName();
5449 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5450 FuncName);
5451
5452 Value *Result = Table.BuildLookup(TableIndex, Builder);
5453
5454 // If the result is used to return immediately from the function, we want to
5455 // do that right here.
5456 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5457 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5458 Builder.CreateRet(Result);
5459 ReturnedEarly = true;
5460 break;
5461 }
5462
5463 // Do a small peephole optimization: re-use the switch table compare if
5464 // possible.
5465 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5466 BasicBlock *PhiBlock = PHI->getParent();
5467 // Search for compare instructions which use the phi.
5468 for (auto *User : PHI->users()) {
5469 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5470 }
5471 }
5472
5473 PHI->addIncoming(Result, LookupBB);
5474 }
5475
5476 if (!ReturnedEarly)
5477 Builder.CreateBr(CommonDest);
5478
5479 // Remove the switch.
5480 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5481 BasicBlock *Succ = SI->getSuccessor(i);
5482
5483 if (Succ == SI->getDefaultDest())
5484 continue;
5485 Succ->removePredecessor(SI->getParent());
5486 }
5487 SI->eraseFromParent();
5488
5489 ++NumLookupTables;
5490 if (NeedMask)
5491 ++NumLookupTablesHoles;
5492 return true;
5493}
5494
5495static bool isSwitchDense(ArrayRef<int64_t> Values) {
5496 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5497 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5498 uint64_t Range = Diff + 1;
5499 uint64_t NumCases = Values.size();
5500 // 40% is the default density for building a jump table in optsize/minsize mode.
5501 uint64_t MinDensity = 40;
5502
5503 return NumCases * 100 >= Range * MinDensity;
5504}
5505
5506/// Try to transform a switch that has "holes" in it to a contiguous sequence
5507/// of cases.
5508///
5509/// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5510/// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5511///
5512/// This converts a sparse switch into a dense switch which allows better
5513/// lowering and could also allow transforming into a lookup table.
5514static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5515 const DataLayout &DL,
5516 const TargetTransformInfo &TTI) {
5517 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5518 if (CondTy->getIntegerBitWidth() > 64 ||
5519 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5520 return false;
5521 // Only bother with this optimization if there are more than 3 switch cases;
5522 // SDAG will only bother creating jump tables for 4 or more cases.
5523 if (SI->getNumCases() < 4)
5524 return false;
5525
5526 // This transform is agnostic to the signedness of the input or case values. We
5527 // can treat the case values as signed or unsigned. We can optimize more common
5528 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5529 // as signed.
5530 SmallVector<int64_t,4> Values;
5531 for (auto &C : SI->cases())
5532 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5533 llvm::sort(Values);
5534
5535 // If the switch is already dense, there's nothing useful to do here.
5536 if (isSwitchDense(Values))
5537 return false;
5538
5539 // First, transform the values such that they start at zero and ascend.
5540 int64_t Base = Values[0];
5541 for (auto &V : Values)
5542 V -= (uint64_t)(Base);
5543
5544 // Now we have signed numbers that have been shifted so that, given enough
5545 // precision, there are no negative values. Since the rest of the transform
5546 // is bitwise only, we switch now to an unsigned representation.
5547 uint64_t GCD = 0;
5548 for (auto &V : Values)
5549 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5550
5551 // This transform can be done speculatively because it is so cheap - it results
5552 // in a single rotate operation being inserted. This can only happen if the
5553 // factor extracted is a power of 2.
5554 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5555 // inverse of GCD and then perform this transform.
5556 // FIXME: It's possible that optimizing a switch on powers of two might also
5557 // be beneficial - flag values are often powers of two and we could use a CLZ
5558 // as the key function.
5559 if (GCD <= 1 || !isPowerOf2_64(GCD))
5560 // No common divisor found or too expensive to compute key function.
5561 return false;
5562
5563 unsigned Shift = Log2_64(GCD);
5564 for (auto &V : Values)
5565 V = (int64_t)((uint64_t)V >> Shift);
5566
5567 if (!isSwitchDense(Values))
5568 // Transform didn't create a dense switch.
5569 return false;
5570
5571 // The obvious transform is to shift the switch condition right and emit a
5572 // check that the condition actually cleanly divided by GCD, i.e.
5573 // C & (1 << Shift - 1) == 0
5574 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5575 //
5576 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5577 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5578 // are nonzero then the switch condition will be very large and will hit the
5579 // default case.
5580
5581 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5582 Builder.SetInsertPoint(SI);
5583 auto *ShiftC = ConstantInt::get(Ty, Shift);
5584 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5585 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5586 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5587 auto *Rot = Builder.CreateOr(LShr, Shl);
5588 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5589
5590 for (auto Case : SI->cases()) {
5591 auto *Orig = Case.getCaseValue();
5592 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5593 Case.setValue(
5594 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5595 }
5596 return true;
5597}
5598
5599bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5600 BasicBlock *BB = SI->getParent();
5601
5602 if (isValueEqualityComparison(SI)) {
5603 // If we only have one predecessor, and if it is a branch on this value,
5604 // see if that predecessor totally determines the outcome of this switch.
5605 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5606 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5607 return requestResimplify();
5608
5609 Value *Cond = SI->getCondition();
5610 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5611 if (SimplifySwitchOnSelect(SI, Select))
5612 return requestResimplify();
5613
5614 // If the block only contains the switch, see if we can fold the block
5615 // away into any preds.
5616 if (SI == &*BB->instructionsWithoutDebug().begin())
5617 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5618 return requestResimplify();
5619 }
5620
5621 // Try to transform the switch into an icmp and a branch.
5622 if (TurnSwitchRangeIntoICmp(SI, Builder))
5623 return requestResimplify();
5624
5625 // Remove unreachable cases.
5626 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5627 return requestResimplify();
5628
5629 if (switchToSelect(SI, Builder, DL, TTI))
5630 return requestResimplify();
5631
5632 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5633 return requestResimplify();
5634
5635 // The conversion from switch to lookup tables results in difficult-to-analyze
5636 // code and makes pruning branches much harder. This is a problem if the
5637 // switch expression itself can still be restricted as a result of inlining or
5638 // CVP. Therefore, only apply this transformation during late stages of the
5639 // optimisation pipeline.
5640 if (Options.ConvertSwitchToLookupTable &&
5641 SwitchToLookupTable(SI, Builder, DL, TTI))
5642 return requestResimplify();
5643
5644 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5645 return requestResimplify();
5646
5647 return false;
5648}
5649
5650bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5651 BasicBlock *BB = IBI->getParent();
5652 bool Changed = false;
5653
5654 // Eliminate redundant destinations.
5655 SmallPtrSet<Value *, 8> Succs;
5656 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5657 BasicBlock *Dest = IBI->getDestination(i);
5658 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5659 Dest->removePredecessor(BB);
5660 IBI->removeDestination(i);
5661 --i;
5662 --e;
5663 Changed = true;
5664 }
5665 }
5666
5667 if (IBI->getNumDestinations() == 0) {
5668 // If the indirectbr has no successors, change it to unreachable.
5669 new UnreachableInst(IBI->getContext(), IBI);
5670 EraseTerminatorAndDCECond(IBI);
5671 return true;
5672 }
5673
5674 if (IBI->getNumDestinations() == 1) {
5675 // If the indirectbr has one successor, change it to a direct branch.
5676 BranchInst::Create(IBI->getDestination(0), IBI);
5677 EraseTerminatorAndDCECond(IBI);
5678 return true;
5679 }
5680
5681 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5682 if (SimplifyIndirectBrOnSelect(IBI, SI))
5683 return requestResimplify();
5684 }
5685 return Changed;
5686}
5687
5688/// Given an block with only a single landing pad and a unconditional branch
5689/// try to find another basic block which this one can be merged with. This
5690/// handles cases where we have multiple invokes with unique landing pads, but
5691/// a shared handler.
5692///
5693/// We specifically choose to not worry about merging non-empty blocks
5694/// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5695/// practice, the optimizer produces empty landing pad blocks quite frequently
5696/// when dealing with exception dense code. (see: instcombine, gvn, if-else
5697/// sinking in this file)
5698///
5699/// This is primarily a code size optimization. We need to avoid performing
5700/// any transform which might inhibit optimization (such as our ability to
5701/// specialize a particular handler via tail commoning). We do this by not
5702/// merging any blocks which require us to introduce a phi. Since the same
5703/// values are flowing through both blocks, we don't lose any ability to
5704/// specialize. If anything, we make such specialization more likely.
5705///
5706/// TODO - This transformation could remove entries from a phi in the target
5707/// block when the inputs in the phi are the same for the two blocks being
5708/// merged. In some cases, this could result in removal of the PHI entirely.
5709static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5710 BasicBlock *BB) {
5711 auto Succ = BB->getUniqueSuccessor();
5712 assert(Succ)((Succ) ? static_cast<void> (0) : __assert_fail ("Succ"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5712, __PRETTY_FUNCTION__))
;
5713 // If there's a phi in the successor block, we'd likely have to introduce
5714 // a phi into the merged landing pad block.
5715 if (isa<PHINode>(*Succ->begin()))
5716 return false;
5717
5718 for (BasicBlock *OtherPred : predecessors(Succ)) {
5719 if (BB == OtherPred)
5720 continue;
5721 BasicBlock::iterator I = OtherPred->begin();
5722 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5723 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5724 continue;
5725 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5726 ;
5727 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5728 if (!BI2 || !BI2->isIdenticalTo(BI))
5729 continue;
5730
5731 // We've found an identical block. Update our predecessors to take that
5732 // path instead and make ourselves dead.
5733 SmallPtrSet<BasicBlock *, 16> Preds;
5734 Preds.insert(pred_begin(BB), pred_end(BB));
5735 for (BasicBlock *Pred : Preds) {
5736 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5737 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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5738, __PRETTY_FUNCTION__))
5738 "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-8~svn350071/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5738, __PRETTY_FUNCTION__))
;
5739 II->setUnwindDest(OtherPred);
5740 }
5741
5742 // The debug info in OtherPred doesn't cover the merged control flow that
5743 // used to go through BB. We need to delete it or update it.
5744 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5745 Instruction &Inst = *I;
5746 I++;
5747 if (isa<DbgInfoIntrinsic>(Inst))
5748 Inst.eraseFromParent();
5749 }
5750
5751 SmallPtrSet<BasicBlock *, 16> Succs;
5752 Succs.insert(succ_begin(BB), succ_end(BB));
5753 for (BasicBlock *Succ : Succs) {
5754 Succ->removePredecessor(BB);
5755 }
5756
5757 IRBuilder<> Builder(BI);
5758 Builder.CreateUnreachable();
5759 BI->eraseFromParent();
5760 return true;
5761 }
5762 return false;
5763}
5764
5765bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5766 IRBuilder<> &Builder) {
5767 BasicBlock *BB = BI->getParent();
5768 BasicBlock *Succ = BI->getSuccessor(0);
5769
5770 // If the Terminator is the only non-phi instruction, simplify the block.
5771 // If LoopHeader is provided, check if the block or its successor is a loop
5772 // header. (This is for early invocations before loop simplify and
5773 // vectorization to keep canonical loop forms for nested loops. These blocks
5774 // can be eliminated when the pass is invoked later in the back-end.)
5775 // Note that if BB has only one predecessor then we do not introduce new
5776 // backedge, so we can eliminate BB.
5777 bool NeedCanonicalLoop =
5778 Options.NeedCanonicalLoop &&
5779 (LoopHeaders && BB->hasNPredecessorsOrMore(2) &&
5780 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5781 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5782 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5783 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5784 return true;
5785
5786 // If the only instruction in the block is a seteq/setne comparison against a
5787 // constant, try to simplify the block.
5788 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5789 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5790 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5791 ;
5792 if (I->isTerminator() &&
5793 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder))
5794 return true;
5795 }
5796
5797 // See if we can merge an empty landing pad block with another which is
5798 // equivalent.
5799 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5800 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5801 ;
5802 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5803 return true;
5804 }
5805
5806 // If this basic block is ONLY a compare and a branch, and if a predecessor
5807 // branches to us and our successor, fold the comparison into the
5808 // predecessor and use logical operations to update the incoming value
5809 // for PHI nodes in common successor.
5810 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5811 return requestResimplify();
5812 return false;
5813}
5814
5815static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5816 BasicBlock *PredPred = nullptr;
5817 for (auto *P : predecessors(BB)) {
5818 BasicBlock *PPred = P->getSinglePredecessor();
5819 if (!PPred || (PredPred && PredPred != PPred))
5820 return nullptr;
5821 PredPred = PPred;
5822 }
5823 return PredPred;
5824}
5825
5826bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5827 BasicBlock *BB = BI->getParent();
5828 const Function *Fn = BB->getParent();
5829 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
5830 return false;
5831
5832 // Conditional branch
5833 if (isValueEqualityComparison(BI)) {
5834 // If we only have one predecessor, and if it is a branch on this value,
5835 // see if that predecessor totally determines the outcome of this
5836 // switch.
5837 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5838 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5839 return requestResimplify();
5840
5841 // This block must be empty, except for the setcond inst, if it exists.
5842 // Ignore dbg intrinsics.
5843 auto I = BB->instructionsWithoutDebug().begin();
5844 if (&*I == BI) {
5845 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5846 return requestResimplify();
5847 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5848 ++I;
5849 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5850 return requestResimplify();
5851 }
5852 }
5853
5854 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5855 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5856 return true;
5857
5858 // If this basic block has dominating predecessor blocks and the dominating
5859 // blocks' conditions imply BI's condition, we know the direction of BI.
5860 Optional<bool> Imp = isImpliedByDomCondition(BI->getCondition(), BI, DL);
5861 if (Imp) {
5862 // Turn this into a branch on constant.
5863 auto *OldCond = BI->getCondition();
5864 ConstantInt *TorF = *Imp ? ConstantInt::getTrue(BB->getContext())
5865 : ConstantInt::getFalse(BB->getContext());
5866 BI->setCondition(TorF);
5867 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5868 return requestResimplify();
5869 }
5870
5871 // If this basic block is ONLY a compare and a branch, and if a predecessor
5872 // branches to us and one of our successors, fold the comparison into the
5873 // predecessor and use logical operations to pick the right destination.
5874 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5875 return requestResimplify();
5876
5877 // We have a conditional branch to two blocks that are only reachable
5878 // from BI. We know that the condbr dominates the two blocks, so see if
5879 // there is any identical code in the "then" and "else" blocks. If so, we
5880 // can hoist it up to the branching block.
5881 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5882 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5883 if (HoistThenElseCodeToIf(BI, TTI))
5884 return requestResimplify();
5885 } else {
5886 // If Successor #1 has multiple preds, we may be able to conditionally
5887 // execute Successor #0 if it branches to Successor #1.
5888 Instruction *Succ0TI = BI->getSuccessor(0)->getTerminator();
5889 if (Succ0TI->getNumSuccessors() == 1 &&
5890 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5891 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5892 return requestResimplify();
5893 }
5894 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5895 // If Successor #0 has multiple preds, we may be able to conditionally
5896 // execute Successor #1 if it branches to Successor #0.
5897 Instruction *Succ1TI = BI->getSuccessor(1)->getTerminator();
5898 if (Succ1TI->getNumSuccessors() == 1 &&
5899 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5900 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5901 return requestResimplify();