Bug Summary

File:lib/Transforms/Utils/SimplifyCFG.cpp
Warning:line 2375, 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-eagerly-assume -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 -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-7/lib/clang/7.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/include -I /build/llvm-toolchain-snapshot-7~svn329677/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/backward -internal-isystem /usr/include/clang/7.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-7/lib/clang/7.0.0/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-7~svn329677/build-llvm/lib/Transforms/Utils -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-checker optin.performance.Padding -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-04-11-031539-24776-1 -x c++ /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp
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/SmallSet.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/Statistic.h"
25#include "llvm/ADT/StringRef.h"
26#include "llvm/Analysis/AssumptionCache.h"
27#include "llvm/Analysis/ConstantFolding.h"
28#include "llvm/Analysis/EHPersonalities.h"
29#include "llvm/Analysis/InstructionSimplify.h"
30#include "llvm/Analysis/TargetTransformInfo.h"
31#include "llvm/Analysis/Utils/Local.h"
32#include "llvm/Analysis/ValueTracking.h"
33#include "llvm/IR/Attributes.h"
34#include "llvm/IR/BasicBlock.h"
35#include "llvm/IR/CFG.h"
36#include "llvm/IR/CallSite.h"
37#include "llvm/IR/Constant.h"
38#include "llvm/IR/ConstantRange.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/DerivedTypes.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/GlobalValue.h"
44#include "llvm/IR/GlobalVariable.h"
45#include "llvm/IR/IRBuilder.h"
46#include "llvm/IR/InstrTypes.h"
47#include "llvm/IR/Instruction.h"
48#include "llvm/IR/Instructions.h"
49#include "llvm/IR/IntrinsicInst.h"
50#include "llvm/IR/Intrinsics.h"
51#include "llvm/IR/LLVMContext.h"
52#include "llvm/IR/MDBuilder.h"
53#include "llvm/IR/Metadata.h"
54#include "llvm/IR/Module.h"
55#include "llvm/IR/NoFolder.h"
56#include "llvm/IR/Operator.h"
57#include "llvm/IR/PatternMatch.h"
58#include "llvm/IR/Type.h"
59#include "llvm/IR/Use.h"
60#include "llvm/IR/User.h"
61#include "llvm/IR/Value.h"
62#include "llvm/Support/Casting.h"
63#include "llvm/Support/CommandLine.h"
64#include "llvm/Support/Debug.h"
65#include "llvm/Support/ErrorHandling.h"
66#include "llvm/Support/KnownBits.h"
67#include "llvm/Support/MathExtras.h"
68#include "llvm/Support/raw_ostream.h"
69#include "llvm/Transforms/Utils/BasicBlockUtils.h"
70#include "llvm/Transforms/Utils/ValueMapper.h"
71#include <algorithm>
72#include <cassert>
73#include <climits>
74#include <cstddef>
75#include <cstdint>
76#include <iterator>
77#include <map>
78#include <set>
79#include <tuple>
80#include <utility>
81#include <vector>
82
83using namespace llvm;
84using namespace PatternMatch;
85
86#define DEBUG_TYPE"simplifycfg" "simplifycfg"
87
88// Chosen as 2 so as to be cheap, but still to have enough power to fold
89// a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90// To catch this, we need to fold a compare and a select, hence '2' being the
91// minimum reasonable default.
92static cl::opt<unsigned> PHINodeFoldingThreshold(
93 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94 cl::desc(
95 "Control the amount of phi node folding to perform (default = 2)"));
96
97static cl::opt<bool> DupRet(
98 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99 cl::desc("Duplicate return instructions into unconditional branches"));
100
101static cl::opt<bool>
102 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103 cl::desc("Sink common instructions down to the end block"));
104
105static cl::opt<bool> HoistCondStores(
106 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107 cl::desc("Hoist conditional stores if an unconditional store precedes"));
108
109static cl::opt<bool> MergeCondStores(
110 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111 cl::desc("Hoist conditional stores even if an unconditional store does not "
112 "precede - hoist multiple conditional stores into a single "
113 "predicated store"));
114
115static cl::opt<bool> MergeCondStoresAggressively(
116 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117 cl::desc("When merging conditional stores, do so even if the resultant "
118 "basic blocks are unlikely to be if-converted as a result"));
119
120static cl::opt<bool> SpeculateOneExpensiveInst(
121 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122 cl::desc("Allow exactly one expensive instruction to be speculatively "
123 "executed"));
124
125static cl::opt<unsigned> MaxSpeculationDepth(
126 "max-speculation-depth", cl::Hidden, cl::init(10),
127 cl::desc("Limit maximum recursion depth when calculating costs of "
128 "speculatively executed instructions"));
129
130STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps")static llvm::Statistic NumBitMaps = {"simplifycfg", "NumBitMaps"
, "Number of switch instructions turned into bitmaps", {0}, {
false}}
;
131STATISTIC(NumLinearMaps,static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
132 "Number of switch instructions turned into linear mapping")static llvm::Statistic NumLinearMaps = {"simplifycfg", "NumLinearMaps"
, "Number of switch instructions turned into linear mapping",
{0}, {false}}
;
133STATISTIC(NumLookupTables,static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
134 "Number of switch instructions turned into lookup tables")static llvm::Statistic NumLookupTables = {"simplifycfg", "NumLookupTables"
, "Number of switch instructions turned into lookup tables", {
0}, {false}}
;
135STATISTIC(static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
136 NumLookupTablesHoles,static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
137 "Number of switch instructions turned into lookup tables (holes checked)")static llvm::Statistic NumLookupTablesHoles = {"simplifycfg",
"NumLookupTablesHoles", "Number of switch instructions turned into lookup tables (holes checked)"
, {0}, {false}}
;
138STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares")static llvm::Statistic NumTableCmpReuses = {"simplifycfg", "NumTableCmpReuses"
, "Number of reused switch table lookup compares", {0}, {false
}}
;
139STATISTIC(NumSinkCommons,static llvm::Statistic NumSinkCommons = {"simplifycfg", "NumSinkCommons"
, "Number of common instructions sunk down to the end block",
{0}, {false}}
140 "Number of common instructions sunk down to the end block")static llvm::Statistic NumSinkCommons = {"simplifycfg", "NumSinkCommons"
, "Number of common instructions sunk down to the end block",
{0}, {false}}
;
141STATISTIC(NumSpeculations, "Number of speculative executed instructions")static llvm::Statistic NumSpeculations = {"simplifycfg", "NumSpeculations"
, "Number of speculative executed instructions", {0}, {false}
}
;
142
143namespace {
144
145// The first field contains the value that the switch produces when a certain
146// case group is selected, and the second field is a vector containing the
147// cases composing the case group.
148using SwitchCaseResultVectorTy =
149 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
150
151// The first field contains the phi node that generates a result of the switch
152// and the second field contains the value generated for a certain case in the
153// switch for that PHI.
154using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155
156/// ValueEqualityComparisonCase - Represents a case of a switch.
157struct ValueEqualityComparisonCase {
158 ConstantInt *Value;
159 BasicBlock *Dest;
160
161 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162 : Value(Value), Dest(Dest) {}
163
164 bool operator<(ValueEqualityComparisonCase RHS) const {
165 // Comparing pointers is ok as we only rely on the order for uniquing.
166 return Value < RHS.Value;
167 }
168
169 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170};
171
172class SimplifyCFGOpt {
173 const TargetTransformInfo &TTI;
174 const DataLayout &DL;
175 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176 const SimplifyCFGOptions &Options;
177
178 Value *isValueEqualityComparison(TerminatorInst *TI);
179 BasicBlock *GetValueEqualityComparisonCases(
180 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
182 BasicBlock *Pred,
183 IRBuilder<> &Builder);
184 bool FoldValueComparisonIntoPredecessors(TerminatorInst *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
198public:
199 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
200 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
201 const SimplifyCFGOptions &Opts)
202 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
203
204 bool run(BasicBlock *BB);
205};
206
207} // end anonymous namespace
208
209/// Return true if it is safe to merge these two
210/// terminator instructions together.
211static bool
212SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
213 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
214 if (SI1 == SI2)
215 return false; // Can't merge with self!
216
217 // It is not safe to merge these two switch instructions if they have a common
218 // successor, and if that successor has a PHI node, and if *that* PHI node has
219 // conflicting incoming values from the two switch blocks.
220 BasicBlock *SI1BB = SI1->getParent();
221 BasicBlock *SI2BB = SI2->getParent();
222
223 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
224 bool Fail = false;
225 for (BasicBlock *Succ : successors(SI2BB))
226 if (SI1Succs.count(Succ))
227 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228 PHINode *PN = cast<PHINode>(BBI);
229 if (PN->getIncomingValueForBlock(SI1BB) !=
230 PN->getIncomingValueForBlock(SI2BB)) {
231 if (FailBlocks)
232 FailBlocks->insert(Succ);
233 Fail = true;
234 }
235 }
236
237 return !Fail;
238}
239
240/// Return true if it is safe and profitable to merge these two terminator
241/// instructions together, where SI1 is an unconditional branch. PhiNodes will
242/// store all PHI nodes in common successors.
243static bool
244isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
245 Instruction *Cond,
246 SmallVectorImpl<PHINode *> &PhiNodes) {
247 if (SI1 == SI2)
248 return false; // Can't merge with self!
249 assert(SI1->isUnconditional() && SI2->isConditional())(static_cast <bool> (SI1->isUnconditional() &&
SI2->isConditional()) ? void (0) : __assert_fail ("SI1->isUnconditional() && SI2->isConditional()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 249, __extension__ __PRETTY_FUNCTION__))
;
250
251 // We fold the unconditional branch if we can easily update all PHI nodes in
252 // common successors:
253 // 1> We have a constant incoming value for the conditional branch;
254 // 2> We have "Cond" as the incoming value for the unconditional branch;
255 // 3> SI2->getCondition() and Cond have same operands.
256 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
257 if (!Ci2)
258 return false;
259 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260 Cond->getOperand(1) == Ci2->getOperand(1)) &&
261 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262 Cond->getOperand(1) == Ci2->getOperand(0)))
263 return false;
264
265 BasicBlock *SI1BB = SI1->getParent();
266 BasicBlock *SI2BB = SI2->getParent();
267 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268 for (BasicBlock *Succ : successors(SI2BB))
269 if (SI1Succs.count(Succ))
270 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271 PHINode *PN = cast<PHINode>(BBI);
272 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
274 return false;
275 PhiNodes.push_back(PN);
276 }
277 return true;
278}
279
280/// Update PHI nodes in Succ to indicate that there will now be entries in it
281/// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282/// will be the same as those coming in from ExistPred, an existing predecessor
283/// of Succ.
284static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285 BasicBlock *ExistPred) {
286 for (PHINode &PN : Succ->phis())
287 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
288}
289
290/// Compute an abstract "cost" of speculating the given instruction,
291/// which is assumed to be safe to speculate. TCC_Free means cheap,
292/// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
293/// expensive.
294static unsigned ComputeSpeculationCost(const User *I,
295 const TargetTransformInfo &TTI) {
296 assert(isSafeToSpeculativelyExecute(I) &&(static_cast <bool> (isSafeToSpeculativelyExecute(I) &&
"Instruction is not safe to speculatively execute!") ? void (
0) : __assert_fail ("isSafeToSpeculativelyExecute(I) && \"Instruction is not safe to speculatively execute!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 297, __extension__ __PRETTY_FUNCTION__))
297 "Instruction is not safe to speculatively execute!")(static_cast <bool> (isSafeToSpeculativelyExecute(I) &&
"Instruction is not safe to speculatively execute!") ? void (
0) : __assert_fail ("isSafeToSpeculativelyExecute(I) && \"Instruction is not safe to speculatively execute!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 297, __extension__ __PRETTY_FUNCTION__))
;
298 return TTI.getUserCost(I);
299}
300
301/// If we have a merge point of an "if condition" as accepted above,
302/// return true if the specified value dominates the block. We
303/// don't handle the true generality of domination here, just a special case
304/// which works well enough for us.
305///
306/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
307/// see if V (which must be an instruction) and its recursive operands
308/// that do not dominate BB have a combined cost lower than CostRemaining and
309/// are non-trapping. If both are true, the instruction is inserted into the
310/// set and true is returned.
311///
312/// The cost for most non-trapping instructions is defined as 1 except for
313/// Select whose cost is 2.
314///
315/// After this function returns, CostRemaining is decreased by the cost of
316/// V plus its non-dominating operands. If that cost is greater than
317/// CostRemaining, false is returned and CostRemaining is undefined.
318static bool DominatesMergePoint(Value *V, BasicBlock *BB,
319 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
320 unsigned &CostRemaining,
321 const TargetTransformInfo &TTI,
322 unsigned Depth = 0) {
323 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
324 // so limit the recursion depth.
325 // TODO: While this recursion limit does prevent pathological behavior, it
326 // would be better to track visited instructions to avoid cycles.
327 if (Depth == MaxSpeculationDepth)
328 return false;
329
330 Instruction *I = dyn_cast<Instruction>(V);
331 if (!I) {
332 // Non-instructions all dominate instructions, but not all constantexprs
333 // can be executed unconditionally.
334 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
335 if (C->canTrap())
336 return false;
337 return true;
338 }
339 BasicBlock *PBB = I->getParent();
340
341 // We don't want to allow weird loops that might have the "if condition" in
342 // the bottom of this block.
343 if (PBB == BB)
344 return false;
345
346 // If this instruction is defined in a block that contains an unconditional
347 // branch to BB, then it must be in the 'conditional' part of the "if
348 // statement". If not, it definitely dominates the region.
349 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
350 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
351 return true;
352
353 // If we aren't allowing aggressive promotion anymore, then don't consider
354 // instructions in the 'if region'.
355 if (!AggressiveInsts)
356 return false;
357
358 // If we have seen this instruction before, don't count it again.
359 if (AggressiveInsts->count(I))
360 return true;
361
362 // Okay, it looks like the instruction IS in the "condition". Check to
363 // see if it's a cheap instruction to unconditionally compute, and if it
364 // only uses stuff defined outside of the condition. If so, hoist it out.
365 if (!isSafeToSpeculativelyExecute(I))
366 return false;
367
368 unsigned Cost = ComputeSpeculationCost(I, TTI);
369
370 // Allow exactly one instruction to be speculated regardless of its cost
371 // (as long as it is safe to do so).
372 // This is intended to flatten the CFG even if the instruction is a division
373 // or other expensive operation. The speculation of an expensive instruction
374 // is expected to be undone in CodeGenPrepare if the speculation has not
375 // enabled further IR optimizations.
376 if (Cost > CostRemaining &&
377 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
378 return false;
379
380 // Avoid unsigned wrap.
381 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
382
383 // Okay, we can only really hoist these out if their operands do
384 // not take us over the cost threshold.
385 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
386 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
387 Depth + 1))
388 return false;
389 // Okay, it's safe to do this! Remember this instruction.
390 AggressiveInsts->insert(I);
391 return true;
392}
393
394/// Extract ConstantInt from value, looking through IntToPtr
395/// and PointerNullValue. Return NULL if value is not a constant int.
396static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
397 // Normal constant int.
398 ConstantInt *CI = dyn_cast<ConstantInt>(V);
399 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
400 return CI;
401
402 // This is some kind of pointer constant. Turn it into a pointer-sized
403 // ConstantInt if possible.
404 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
405
406 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
407 if (isa<ConstantPointerNull>(V))
408 return ConstantInt::get(PtrTy, 0);
409
410 // IntToPtr const int.
411 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
412 if (CE->getOpcode() == Instruction::IntToPtr)
413 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
414 // The constant is very likely to have the right type already.
415 if (CI->getType() == PtrTy)
416 return CI;
417 else
418 return cast<ConstantInt>(
419 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
420 }
421 return nullptr;
422}
423
424namespace {
425
426/// Given a chain of or (||) or and (&&) comparison of a value against a
427/// constant, this will try to recover the information required for a switch
428/// structure.
429/// It will depth-first traverse the chain of comparison, seeking for patterns
430/// like %a == 12 or %a < 4 and combine them to produce a set of integer
431/// representing the different cases for the switch.
432/// Note that if the chain is composed of '||' it will build the set of elements
433/// that matches the comparisons (i.e. any of this value validate the chain)
434/// while for a chain of '&&' it will build the set elements that make the test
435/// fail.
436struct ConstantComparesGatherer {
437 const DataLayout &DL;
438
439 /// Value found for the switch comparison
440 Value *CompValue = nullptr;
441
442 /// Extra clause to be checked before the switch
443 Value *Extra = nullptr;
444
445 /// Set of integers to match in switch
446 SmallVector<ConstantInt *, 8> Vals;
447
448 /// Number of comparisons matched in the and/or chain
449 unsigned UsedICmps = 0;
450
451 /// Construct and compute the result for the comparison instruction Cond
452 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
453 gather(Cond);
454 }
455
456 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
457 ConstantComparesGatherer &
458 operator=(const ConstantComparesGatherer &) = delete;
459
460private:
461 /// Try to set the current value used for the comparison, it succeeds only if
462 /// it wasn't set before or if the new value is the same as the old one
463 bool setValueOnce(Value *NewVal) {
464 if (CompValue && CompValue != NewVal)
465 return false;
466 CompValue = NewVal;
467 return (CompValue != nullptr);
468 }
469
470 /// Try to match Instruction "I" as a comparison against a constant and
471 /// populates the array Vals with the set of values that match (or do not
472 /// match depending on isEQ).
473 /// Return false on failure. On success, the Value the comparison matched
474 /// against is placed in CompValue.
475 /// If CompValue is already set, the function is expected to fail if a match
476 /// is found but the value compared to is different.
477 bool matchInstruction(Instruction *I, bool isEQ) {
478 // If this is an icmp against a constant, handle this as one of the cases.
479 ICmpInst *ICI;
480 ConstantInt *C;
481 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
482 (C = GetConstantInt(I->getOperand(1), DL)))) {
483 return false;
484 }
485
486 Value *RHSVal;
487 const APInt *RHSC;
488
489 // Pattern match a special case
490 // (x & ~2^z) == y --> x == y || x == y|2^z
491 // This undoes a transformation done by instcombine to fuse 2 compares.
492 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
493 // It's a little bit hard to see why the following transformations are
494 // correct. Here is a CVC3 program to verify them for 64-bit values:
495
496 /*
497 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
498 x : BITVECTOR(64);
499 y : BITVECTOR(64);
500 z : BITVECTOR(64);
501 mask : BITVECTOR(64) = BVSHL(ONE, z);
502 QUERY( (y & ~mask = y) =>
503 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
504 );
505 QUERY( (y | mask = y) =>
506 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
507 );
508 */
509
510 // Please note that each pattern must be a dual implication (<--> or
511 // iff). One directional implication can create spurious matches. If the
512 // implication is only one-way, an unsatisfiable condition on the left
513 // side can imply a satisfiable condition on the right side. Dual
514 // implication ensures that satisfiable conditions are transformed to
515 // other satisfiable conditions and unsatisfiable conditions are
516 // transformed to other unsatisfiable conditions.
517
518 // Here is a concrete example of a unsatisfiable condition on the left
519 // implying a satisfiable condition on the right:
520 //
521 // mask = (1 << z)
522 // (x & ~mask) == y --> (x == y || x == (y | mask))
523 //
524 // Substituting y = 3, z = 0 yields:
525 // (x & -2) == 3 --> (x == 3 || x == 2)
526
527 // Pattern match a special case:
528 /*
529 QUERY( (y & ~mask = y) =>
530 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
531 );
532 */
533 if (match(ICI->getOperand(0),
534 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
535 APInt Mask = ~*RHSC;
536 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
537 // If we already have a value for the switch, it has to match!
538 if (!setValueOnce(RHSVal))
539 return false;
540
541 Vals.push_back(C);
542 Vals.push_back(
543 ConstantInt::get(C->getContext(),
544 C->getValue() | Mask));
545 UsedICmps++;
546 return true;
547 }
548 }
549
550 // Pattern match a special case:
551 /*
552 QUERY( (y | mask = y) =>
553 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
554 );
555 */
556 if (match(ICI->getOperand(0),
557 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
558 APInt Mask = *RHSC;
559 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
560 // If we already have a value for the switch, it has to match!
561 if (!setValueOnce(RHSVal))
562 return false;
563
564 Vals.push_back(C);
565 Vals.push_back(ConstantInt::get(C->getContext(),
566 C->getValue() & ~Mask));
567 UsedICmps++;
568 return true;
569 }
570 }
571
572 // If we already have a value for the switch, it has to match!
573 if (!setValueOnce(ICI->getOperand(0)))
574 return false;
575
576 UsedICmps++;
577 Vals.push_back(C);
578 return ICI->getOperand(0);
579 }
580
581 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
582 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
583 ICI->getPredicate(), C->getValue());
584
585 // Shift the range if the compare is fed by an add. This is the range
586 // compare idiom as emitted by instcombine.
587 Value *CandidateVal = I->getOperand(0);
588 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
589 Span = Span.subtract(*RHSC);
590 CandidateVal = RHSVal;
591 }
592
593 // If this is an and/!= check, then we are looking to build the set of
594 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
595 // x != 0 && x != 1.
596 if (!isEQ)
597 Span = Span.inverse();
598
599 // If there are a ton of values, we don't want to make a ginormous switch.
600 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
601 return false;
602 }
603
604 // If we already have a value for the switch, it has to match!
605 if (!setValueOnce(CandidateVal))
606 return false;
607
608 // Add all values from the range to the set
609 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
610 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
611
612 UsedICmps++;
613 return true;
614 }
615
616 /// Given a potentially 'or'd or 'and'd together collection of icmp
617 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
618 /// the value being compared, and stick the list constants into the Vals
619 /// vector.
620 /// One "Extra" case is allowed to differ from the other.
621 void gather(Value *V) {
622 Instruction *I = dyn_cast<Instruction>(V);
623 bool isEQ = (I->getOpcode() == Instruction::Or);
624
625 // Keep a stack (SmallVector for efficiency) for depth-first traversal
626 SmallVector<Value *, 8> DFT;
627 SmallPtrSet<Value *, 8> Visited;
628
629 // Initialize
630 Visited.insert(V);
631 DFT.push_back(V);
632
633 while (!DFT.empty()) {
634 V = DFT.pop_back_val();
635
636 if (Instruction *I = dyn_cast<Instruction>(V)) {
637 // If it is a || (or && depending on isEQ), process the operands.
638 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
639 if (Visited.insert(I->getOperand(1)).second)
640 DFT.push_back(I->getOperand(1));
641 if (Visited.insert(I->getOperand(0)).second)
642 DFT.push_back(I->getOperand(0));
643 continue;
644 }
645
646 // Try to match the current instruction
647 if (matchInstruction(I, isEQ))
648 // Match succeed, continue the loop
649 continue;
650 }
651
652 // One element of the sequence of || (or &&) could not be match as a
653 // comparison against the same value as the others.
654 // We allow only one "Extra" case to be checked before the switch
655 if (!Extra) {
656 Extra = V;
657 continue;
658 }
659 // Failed to parse a proper sequence, abort now
660 CompValue = nullptr;
661 break;
662 }
663 }
664};
665
666} // end anonymous namespace
667
668static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
669 Instruction *Cond = nullptr;
670 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
671 Cond = dyn_cast<Instruction>(SI->getCondition());
672 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
673 if (BI->isConditional())
674 Cond = dyn_cast<Instruction>(BI->getCondition());
675 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
676 Cond = dyn_cast<Instruction>(IBI->getAddress());
677 }
678
679 TI->eraseFromParent();
680 if (Cond)
681 RecursivelyDeleteTriviallyDeadInstructions(Cond);
682}
683
684/// Return true if the specified terminator checks
685/// to see if a value is equal to constant integer value.
686Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
687 Value *CV = nullptr;
688 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
689 // Do not permit merging of large switch instructions into their
690 // predecessors unless there is only one predecessor.
691 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
692 pred_end(SI->getParent())) <=
693 128)
694 CV = SI->getCondition();
695 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
696 if (BI->isConditional() && BI->getCondition()->hasOneUse())
697 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
698 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
699 CV = ICI->getOperand(0);
700 }
701
702 // Unwrap any lossless ptrtoint cast.
703 if (CV) {
704 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
705 Value *Ptr = PTII->getPointerOperand();
706 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
707 CV = Ptr;
708 }
709 }
710 return CV;
711}
712
713/// Given a value comparison instruction,
714/// decode all of the 'cases' that it represents and return the 'default' block.
715BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
716 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
717 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
718 Cases.reserve(SI->getNumCases());
719 for (auto Case : SI->cases())
720 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
721 Case.getCaseSuccessor()));
722 return SI->getDefaultDest();
723 }
724
725 BranchInst *BI = cast<BranchInst>(TI);
726 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
727 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
728 Cases.push_back(ValueEqualityComparisonCase(
729 GetConstantInt(ICI->getOperand(1), DL), Succ));
730 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
731}
732
733/// Given a vector of bb/value pairs, remove any entries
734/// in the list that match the specified block.
735static void
736EliminateBlockCases(BasicBlock *BB,
737 std::vector<ValueEqualityComparisonCase> &Cases) {
738 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
739}
740
741/// Return true if there are any keys in C1 that exist in C2 as well.
742static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
743 std::vector<ValueEqualityComparisonCase> &C2) {
744 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
745
746 // Make V1 be smaller than V2.
747 if (V1->size() > V2->size())
748 std::swap(V1, V2);
749
750 if (V1->empty())
751 return false;
752 if (V1->size() == 1) {
753 // Just scan V2.
754 ConstantInt *TheVal = (*V1)[0].Value;
755 for (unsigned i = 0, e = V2->size(); i != e; ++i)
756 if (TheVal == (*V2)[i].Value)
757 return true;
758 }
759
760 // Otherwise, just sort both lists and compare element by element.
761 array_pod_sort(V1->begin(), V1->end());
762 array_pod_sort(V2->begin(), V2->end());
763 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
764 while (i1 != e1 && i2 != e2) {
765 if ((*V1)[i1].Value == (*V2)[i2].Value)
766 return true;
767 if ((*V1)[i1].Value < (*V2)[i2].Value)
768 ++i1;
769 else
770 ++i2;
771 }
772 return false;
773}
774
775// Set branch weights on SwitchInst. This sets the metadata if there is at
776// least one non-zero weight.
777static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
778 // Check that there is at least one non-zero weight. Otherwise, pass
779 // nullptr to setMetadata which will erase the existing metadata.
780 MDNode *N = nullptr;
781 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
782 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
783 SI->setMetadata(LLVMContext::MD_prof, N);
784}
785
786// Similar to the above, but for branch and select instructions that take
787// exactly 2 weights.
788static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
789 uint32_t FalseWeight) {
790 assert(isa<BranchInst>(I) || isa<SelectInst>(I))(static_cast <bool> (isa<BranchInst>(I) || isa<
SelectInst>(I)) ? void (0) : __assert_fail ("isa<BranchInst>(I) || isa<SelectInst>(I)"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 790, __extension__ __PRETTY_FUNCTION__))
;
791 // Check that there is at least one non-zero weight. Otherwise, pass
792 // nullptr to setMetadata which will erase the existing metadata.
793 MDNode *N = nullptr;
794 if (TrueWeight || FalseWeight)
795 N = MDBuilder(I->getParent()->getContext())
796 .createBranchWeights(TrueWeight, FalseWeight);
797 I->setMetadata(LLVMContext::MD_prof, N);
798}
799
800/// If TI is known to be a terminator instruction and its block is known to
801/// only have a single predecessor block, check to see if that predecessor is
802/// also a value comparison with the same value, and if that comparison
803/// determines the outcome of this comparison. If so, simplify TI. This does a
804/// very limited form of jump threading.
805bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
806 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
807 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
808 if (!PredVal)
809 return false; // Not a value comparison in predecessor.
810
811 Value *ThisVal = isValueEqualityComparison(TI);
812 assert(ThisVal && "This isn't a value comparison!!")(static_cast <bool> (ThisVal && "This isn't a value comparison!!"
) ? void (0) : __assert_fail ("ThisVal && \"This isn't a value comparison!!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 812, __extension__ __PRETTY_FUNCTION__))
;
813 if (ThisVal != PredVal)
814 return false; // Different predicates.
815
816 // TODO: Preserve branch weight metadata, similarly to how
817 // FoldValueComparisonIntoPredecessors preserves it.
818
819 // Find out information about when control will move from Pred to TI's block.
820 std::vector<ValueEqualityComparisonCase> PredCases;
821 BasicBlock *PredDef =
822 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
823 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
824
825 // Find information about how control leaves this block.
826 std::vector<ValueEqualityComparisonCase> ThisCases;
827 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
828 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
829
830 // If TI's block is the default block from Pred's comparison, potentially
831 // simplify TI based on this knowledge.
832 if (PredDef == TI->getParent()) {
833 // If we are here, we know that the value is none of those cases listed in
834 // PredCases. If there are any cases in ThisCases that are in PredCases, we
835 // can simplify TI.
836 if (!ValuesOverlap(PredCases, ThisCases))
837 return false;
838
839 if (isa<BranchInst>(TI)) {
840 // Okay, one of the successors of this condbr is dead. Convert it to a
841 // uncond br.
842 assert(ThisCases.size() == 1 && "Branch can only have one case!")(static_cast <bool> (ThisCases.size() == 1 && "Branch can only have one case!"
) ? void (0) : __assert_fail ("ThisCases.size() == 1 && \"Branch can only have one case!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 842, __extension__ __PRETTY_FUNCTION__))
;
843 // Insert the new branch.
844 Instruction *NI = Builder.CreateBr(ThisDef);
845 (void)NI;
846
847 // Remove PHI node entries for the dead edge.
848 ThisCases[0].Dest->removePredecessor(TI->getParent());
849
850 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)
851 << "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)
852 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
853
854 EraseTerminatorInstAndDCECond(TI);
855 return true;
856 }
857
858 SwitchInst *SI = cast<SwitchInst>(TI);
859 // Okay, TI has cases that are statically dead, prune them away.
860 SmallPtrSet<Constant *, 16> DeadCases;
861 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
862 DeadCases.insert(PredCases[i].Value);
863
864 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)
865 << "Through successor TI: " << *TI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI; } } while (false)
;
866
867 // Collect branch weights into a vector.
868 SmallVector<uint32_t, 8> Weights;
869 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
870 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
871 if (HasWeight)
872 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
873 ++MD_i) {
874 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
875 Weights.push_back(CI->getValue().getZExtValue());
876 }
877 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
878 --i;
879 if (DeadCases.count(i->getCaseValue())) {
880 if (HasWeight) {
881 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
882 Weights.pop_back();
883 }
884 i->getCaseSuccessor()->removePredecessor(TI->getParent());
885 SI->removeCase(i);
886 }
887 }
888 if (HasWeight && Weights.size() >= 2)
889 setBranchWeights(SI, Weights);
890
891 DEBUG(dbgs() << "Leaving: " << *TI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Leaving: " << *TI <<
"\n"; } } while (false)
;
892 return true;
893 }
894
895 // Otherwise, TI's block must correspond to some matched value. Find out
896 // which value (or set of values) this is.
897 ConstantInt *TIV = nullptr;
898 BasicBlock *TIBB = TI->getParent();
899 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
900 if (PredCases[i].Dest == TIBB) {
901 if (TIV)
902 return false; // Cannot handle multiple values coming to this block.
903 TIV = PredCases[i].Value;
904 }
905 assert(TIV && "No edge from pred to succ?")(static_cast <bool> (TIV && "No edge from pred to succ?"
) ? void (0) : __assert_fail ("TIV && \"No edge from pred to succ?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 905, __extension__ __PRETTY_FUNCTION__))
;
906
907 // Okay, we found the one constant that our value can be if we get into TI's
908 // BB. Find out which successor will unconditionally be branched to.
909 BasicBlock *TheRealDest = nullptr;
910 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
911 if (ThisCases[i].Value == TIV) {
912 TheRealDest = ThisCases[i].Dest;
913 break;
914 }
915
916 // If not handled by any explicit cases, it is handled by the default case.
917 if (!TheRealDest)
918 TheRealDest = ThisDef;
919
920 // Remove PHI node entries for dead edges.
921 BasicBlock *CheckEdge = TheRealDest;
922 for (BasicBlock *Succ : successors(TIBB))
923 if (Succ != CheckEdge)
924 Succ->removePredecessor(TIBB);
925 else
926 CheckEdge = nullptr;
927
928 // Insert the new branch.
929 Instruction *NI = Builder.CreateBr(TheRealDest);
930 (void)NI;
931
932 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)
933 << "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)
934 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Threading pred instr: " <<
*Pred->getTerminator() << "Through successor TI: " <<
*TI << "Leaving: " << *NI << "\n"; } } while
(false)
;
935
936 EraseTerminatorInstAndDCECond(TI);
937 return true;
938}
939
940namespace {
941
942/// This class implements a stable ordering of constant
943/// integers that does not depend on their address. This is important for
944/// applications that sort ConstantInt's to ensure uniqueness.
945struct ConstantIntOrdering {
946 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
947 return LHS->getValue().ult(RHS->getValue());
948 }
949};
950
951} // end anonymous namespace
952
953static int ConstantIntSortPredicate(ConstantInt *const *P1,
954 ConstantInt *const *P2) {
955 const ConstantInt *LHS = *P1;
956 const ConstantInt *RHS = *P2;
957 if (LHS == RHS)
958 return 0;
959 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
960}
961
962static inline bool HasBranchWeights(const Instruction *I) {
963 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
964 if (ProfMD && ProfMD->getOperand(0))
965 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
966 return MDS->getString().equals("branch_weights");
967
968 return false;
969}
970
971/// Get Weights of a given TerminatorInst, the default weight is at the front
972/// of the vector. If TI is a conditional eq, we need to swap the branch-weight
973/// metadata.
974static void GetBranchWeights(TerminatorInst *TI,
975 SmallVectorImpl<uint64_t> &Weights) {
976 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
977 assert(MD)(static_cast <bool> (MD) ? void (0) : __assert_fail ("MD"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 977, __extension__ __PRETTY_FUNCTION__))
;
978 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
979 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
980 Weights.push_back(CI->getValue().getZExtValue());
981 }
982
983 // If TI is a conditional eq, the default case is the false case,
984 // and the corresponding branch-weight data is at index 2. We swap the
985 // default weight to be the first entry.
986 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
987 assert(Weights.size() == 2)(static_cast <bool> (Weights.size() == 2) ? void (0) : __assert_fail
("Weights.size() == 2", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 987, __extension__ __PRETTY_FUNCTION__))
;
988 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
989 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
990 std::swap(Weights.front(), Weights.back());
991 }
992}
993
994/// Keep halving the weights until all can fit in uint32_t.
995static void FitWeights(MutableArrayRef<uint64_t> Weights) {
996 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
997 if (Max > UINT_MAX(2147483647 *2U +1U)) {
998 unsigned Offset = 32 - countLeadingZeros(Max);
999 for (uint64_t &I : Weights)
1000 I >>= Offset;
1001 }
1002}
1003
1004/// The specified terminator is a value equality comparison instruction
1005/// (either a switch or a branch on "X == c").
1006/// See if any of the predecessors of the terminator block are value comparisons
1007/// on the same value. If so, and if safe to do so, fold them together.
1008bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
1009 IRBuilder<> &Builder) {
1010 BasicBlock *BB = TI->getParent();
1011 Value *CV = isValueEqualityComparison(TI); // CondVal
1012 assert(CV && "Not a comparison?")(static_cast <bool> (CV && "Not a comparison?")
? void (0) : __assert_fail ("CV && \"Not a comparison?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1012, __extension__ __PRETTY_FUNCTION__))
;
1013 bool Changed = false;
1014
1015 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1016 while (!Preds.empty()) {
1017 BasicBlock *Pred = Preds.pop_back_val();
1018
1019 // See if the predecessor is a comparison with the same value.
1020 TerminatorInst *PTI = Pred->getTerminator();
1021 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1022
1023 if (PCV == CV && TI != PTI) {
1024 SmallSetVector<BasicBlock*, 4> FailBlocks;
1025 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1026 for (auto *Succ : FailBlocks) {
1027 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1028 return false;
1029 }
1030 }
1031
1032 // Figure out which 'cases' to copy from SI to PSI.
1033 std::vector<ValueEqualityComparisonCase> BBCases;
1034 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1035
1036 std::vector<ValueEqualityComparisonCase> PredCases;
1037 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1038
1039 // Based on whether the default edge from PTI goes to BB or not, fill in
1040 // PredCases and PredDefault with the new switch cases we would like to
1041 // build.
1042 SmallVector<BasicBlock *, 8> NewSuccessors;
1043
1044 // Update the branch weight metadata along the way
1045 SmallVector<uint64_t, 8> Weights;
1046 bool PredHasWeights = HasBranchWeights(PTI);
1047 bool SuccHasWeights = HasBranchWeights(TI);
1048
1049 if (PredHasWeights) {
1050 GetBranchWeights(PTI, Weights);
1051 // branch-weight metadata is inconsistent here.
1052 if (Weights.size() != 1 + PredCases.size())
1053 PredHasWeights = SuccHasWeights = false;
1054 } else if (SuccHasWeights)
1055 // If there are no predecessor weights but there are successor weights,
1056 // populate Weights with 1, which will later be scaled to the sum of
1057 // successor's weights
1058 Weights.assign(1 + PredCases.size(), 1);
1059
1060 SmallVector<uint64_t, 8> SuccWeights;
1061 if (SuccHasWeights) {
1062 GetBranchWeights(TI, SuccWeights);
1063 // branch-weight metadata is inconsistent here.
1064 if (SuccWeights.size() != 1 + BBCases.size())
1065 PredHasWeights = SuccHasWeights = false;
1066 } else if (PredHasWeights)
1067 SuccWeights.assign(1 + BBCases.size(), 1);
1068
1069 if (PredDefault == BB) {
1070 // If this is the default destination from PTI, only the edges in TI
1071 // that don't occur in PTI, or that branch to BB will be activated.
1072 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1073 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1074 if (PredCases[i].Dest != BB)
1075 PTIHandled.insert(PredCases[i].Value);
1076 else {
1077 // The default destination is BB, we don't need explicit targets.
1078 std::swap(PredCases[i], PredCases.back());
1079
1080 if (PredHasWeights || SuccHasWeights) {
1081 // Increase weight for the default case.
1082 Weights[0] += Weights[i + 1];
1083 std::swap(Weights[i + 1], Weights.back());
1084 Weights.pop_back();
1085 }
1086
1087 PredCases.pop_back();
1088 --i;
1089 --e;
1090 }
1091
1092 // Reconstruct the new switch statement we will be building.
1093 if (PredDefault != BBDefault) {
1094 PredDefault->removePredecessor(Pred);
1095 PredDefault = BBDefault;
1096 NewSuccessors.push_back(BBDefault);
1097 }
1098
1099 unsigned CasesFromPred = Weights.size();
1100 uint64_t ValidTotalSuccWeight = 0;
1101 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1102 if (!PTIHandled.count(BBCases[i].Value) &&
1103 BBCases[i].Dest != BBDefault) {
1104 PredCases.push_back(BBCases[i]);
1105 NewSuccessors.push_back(BBCases[i].Dest);
1106 if (SuccHasWeights || PredHasWeights) {
1107 // The default weight is at index 0, so weight for the ith case
1108 // should be at index i+1. Scale the cases from successor by
1109 // PredDefaultWeight (Weights[0]).
1110 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1111 ValidTotalSuccWeight += SuccWeights[i + 1];
1112 }
1113 }
1114
1115 if (SuccHasWeights || PredHasWeights) {
1116 ValidTotalSuccWeight += SuccWeights[0];
1117 // Scale the cases from predecessor by ValidTotalSuccWeight.
1118 for (unsigned i = 1; i < CasesFromPred; ++i)
1119 Weights[i] *= ValidTotalSuccWeight;
1120 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1121 Weights[0] *= SuccWeights[0];
1122 }
1123 } else {
1124 // If this is not the default destination from PSI, only the edges
1125 // in SI that occur in PSI with a destination of BB will be
1126 // activated.
1127 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1128 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1129 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1130 if (PredCases[i].Dest == BB) {
1131 PTIHandled.insert(PredCases[i].Value);
1132
1133 if (PredHasWeights || SuccHasWeights) {
1134 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1135 std::swap(Weights[i + 1], Weights.back());
1136 Weights.pop_back();
1137 }
1138
1139 std::swap(PredCases[i], PredCases.back());
1140 PredCases.pop_back();
1141 --i;
1142 --e;
1143 }
1144
1145 // Okay, now we know which constants were sent to BB from the
1146 // predecessor. Figure out where they will all go now.
1147 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1148 if (PTIHandled.count(BBCases[i].Value)) {
1149 // If this is one we are capable of getting...
1150 if (PredHasWeights || SuccHasWeights)
1151 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1152 PredCases.push_back(BBCases[i]);
1153 NewSuccessors.push_back(BBCases[i].Dest);
1154 PTIHandled.erase(
1155 BBCases[i].Value); // This constant is taken care of
1156 }
1157
1158 // If there are any constants vectored to BB that TI doesn't handle,
1159 // they must go to the default destination of TI.
1160 for (ConstantInt *I : PTIHandled) {
1161 if (PredHasWeights || SuccHasWeights)
1162 Weights.push_back(WeightsForHandled[I]);
1163 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1164 NewSuccessors.push_back(BBDefault);
1165 }
1166 }
1167
1168 // Okay, at this point, we know which new successor Pred will get. Make
1169 // sure we update the number of entries in the PHI nodes for these
1170 // successors.
1171 for (BasicBlock *NewSuccessor : NewSuccessors)
1172 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1173
1174 Builder.SetInsertPoint(PTI);
1175 // Convert pointer to int before we switch.
1176 if (CV->getType()->isPointerTy()) {
1177 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1178 "magicptr");
1179 }
1180
1181 // Now that the successors are updated, create the new Switch instruction.
1182 SwitchInst *NewSI =
1183 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1184 NewSI->setDebugLoc(PTI->getDebugLoc());
1185 for (ValueEqualityComparisonCase &V : PredCases)
1186 NewSI->addCase(V.Value, V.Dest);
1187
1188 if (PredHasWeights || SuccHasWeights) {
1189 // Halve the weights if any of them cannot fit in an uint32_t
1190 FitWeights(Weights);
1191
1192 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1193
1194 setBranchWeights(NewSI, MDWeights);
1195 }
1196
1197 EraseTerminatorInstAndDCECond(PTI);
1198
1199 // Okay, last check. If BB is still a successor of PSI, then we must
1200 // have an infinite loop case. If so, add an infinitely looping block
1201 // to handle the case to preserve the behavior of the code.
1202 BasicBlock *InfLoopBlock = nullptr;
1203 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1204 if (NewSI->getSuccessor(i) == BB) {
1205 if (!InfLoopBlock) {
1206 // Insert it at the end of the function, because it's either code,
1207 // or it won't matter if it's hot. :)
1208 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1209 BB->getParent());
1210 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1211 }
1212 NewSI->setSuccessor(i, InfLoopBlock);
1213 }
1214
1215 Changed = true;
1216 }
1217 }
1218 return Changed;
1219}
1220
1221// If we would need to insert a select that uses the value of this invoke
1222// (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1223// can't hoist the invoke, as there is nowhere to put the select in this case.
1224static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1225 Instruction *I1, Instruction *I2) {
1226 for (BasicBlock *Succ : successors(BB1)) {
1227 for (const PHINode &PN : Succ->phis()) {
1228 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1229 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1230 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1231 return false;
1232 }
1233 }
1234 }
1235 return true;
1236}
1237
1238static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1239
1240/// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1241/// in the two blocks up into the branch block. The caller of this function
1242/// guarantees that BI's block dominates BB1 and BB2.
1243static bool HoistThenElseCodeToIf(BranchInst *BI,
1244 const TargetTransformInfo &TTI) {
1245 // This does very trivial matching, with limited scanning, to find identical
1246 // instructions in the two blocks. In particular, we don't want to get into
1247 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1248 // such, we currently just scan for obviously identical instructions in an
1249 // identical order.
1250 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1251 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1252
1253 BasicBlock::iterator BB1_Itr = BB1->begin();
1254 BasicBlock::iterator BB2_Itr = BB2->begin();
1255
1256 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1257 // Skip debug info if it is not identical.
1258 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1259 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1260 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1261 while (isa<DbgInfoIntrinsic>(I1))
1262 I1 = &*BB1_Itr++;
1263 while (isa<DbgInfoIntrinsic>(I2))
1264 I2 = &*BB2_Itr++;
1265 }
1266 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1267 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1268 return false;
1269
1270 BasicBlock *BIParent = BI->getParent();
1271
1272 bool Changed = false;
1273 do {
1274 // If we are hoisting the terminator instruction, don't move one (making a
1275 // broken BB), instead clone it, and remove BI.
1276 if (isa<TerminatorInst>(I1))
1277 goto HoistTerminator;
1278
1279 // If we're going to hoist a call, make sure that the two instructions we're
1280 // commoning/hoisting are both marked with musttail, or neither of them is
1281 // marked as such. Otherwise, we might end up in a situation where we hoist
1282 // from a block where the terminator is a `ret` to a block where the terminator
1283 // is a `br`, and `musttail` calls expect to be followed by a return.
1284 auto *C1 = dyn_cast<CallInst>(I1);
1285 auto *C2 = dyn_cast<CallInst>(I2);
1286 if (C1 && C2)
1287 if (C1->isMustTailCall() != C2->isMustTailCall())
1288 return Changed;
1289
1290 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1291 return Changed;
1292
1293 if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1294 assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2))(static_cast <bool> (isa<DbgInfoIntrinsic>(I1) &&
isa<DbgInfoIntrinsic>(I2)) ? void (0) : __assert_fail (
"isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2)"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1294, __extension__ __PRETTY_FUNCTION__))
;
1295 // The debug location is an integral part of a debug info intrinsic
1296 // and can't be separated from it or replaced. Instead of attempting
1297 // to merge locations, simply hoist both copies of the intrinsic.
1298 BIParent->getInstList().splice(BI->getIterator(),
1299 BB1->getInstList(), I1);
1300 BIParent->getInstList().splice(BI->getIterator(),
1301 BB2->getInstList(), I2);
1302 Changed = true;
1303 } else {
1304 // For a normal instruction, we just move one to right before the branch,
1305 // then replace all uses of the other with the first. Finally, we remove
1306 // the now redundant second instruction.
1307 BIParent->getInstList().splice(BI->getIterator(),
1308 BB1->getInstList(), I1);
1309 if (!I2->use_empty())
1310 I2->replaceAllUsesWith(I1);
1311 I1->andIRFlags(I2);
1312 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1313 LLVMContext::MD_range,
1314 LLVMContext::MD_fpmath,
1315 LLVMContext::MD_invariant_load,
1316 LLVMContext::MD_nonnull,
1317 LLVMContext::MD_invariant_group,
1318 LLVMContext::MD_align,
1319 LLVMContext::MD_dereferenceable,
1320 LLVMContext::MD_dereferenceable_or_null,
1321 LLVMContext::MD_mem_parallel_loop_access};
1322 combineMetadata(I1, I2, KnownIDs);
1323
1324 // I1 and I2 are being combined into a single instruction. Its debug
1325 // location is the merged locations of the original instructions.
1326 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1327
1328 I2->eraseFromParent();
1329 Changed = true;
1330 }
1331
1332 I1 = &*BB1_Itr++;
1333 I2 = &*BB2_Itr++;
1334 // Skip debug info if it is not identical.
1335 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1336 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1337 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1338 while (isa<DbgInfoIntrinsic>(I1))
1339 I1 = &*BB1_Itr++;
1340 while (isa<DbgInfoIntrinsic>(I2))
1341 I2 = &*BB2_Itr++;
1342 }
1343 } while (I1->isIdenticalToWhenDefined(I2));
1344
1345 return true;
1346
1347HoistTerminator:
1348 // It may not be possible to hoist an invoke.
1349 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1350 return Changed;
1351
1352 for (BasicBlock *Succ : successors(BB1)) {
1353 for (PHINode &PN : Succ->phis()) {
1354 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1355 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1356 if (BB1V == BB2V)
1357 continue;
1358
1359 // Check for passingValueIsAlwaysUndefined here because we would rather
1360 // eliminate undefined control flow then converting it to a select.
1361 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1362 passingValueIsAlwaysUndefined(BB2V, &PN))
1363 return Changed;
1364
1365 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1366 return Changed;
1367 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1368 return Changed;
1369 }
1370 }
1371
1372 // Okay, it is safe to hoist the terminator.
1373 Instruction *NT = I1->clone();
1374 BIParent->getInstList().insert(BI->getIterator(), NT);
1375 if (!NT->getType()->isVoidTy()) {
1376 I1->replaceAllUsesWith(NT);
1377 I2->replaceAllUsesWith(NT);
1378 NT->takeName(I1);
1379 }
1380
1381 IRBuilder<NoFolder> Builder(NT);
1382 // Hoisting one of the terminators from our successor is a great thing.
1383 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1384 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1385 // nodes, so we insert select instruction to compute the final result.
1386 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1387 for (BasicBlock *Succ : successors(BB1)) {
1388 for (PHINode &PN : Succ->phis()) {
1389 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1390 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1391 if (BB1V == BB2V)
1392 continue;
1393
1394 // These values do not agree. Insert a select instruction before NT
1395 // that determines the right value.
1396 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1397 if (!SI)
1398 SI = cast<SelectInst>(
1399 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1400 BB1V->getName() + "." + BB2V->getName(), BI));
1401
1402 // Make the PHI node use the select for all incoming values for BB1/BB2
1403 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1404 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1405 PN.setIncomingValue(i, SI);
1406 }
1407 }
1408
1409 // Update any PHI nodes in our new successors.
1410 for (BasicBlock *Succ : successors(BB1))
1411 AddPredecessorToBlock(Succ, BIParent, BB1);
1412
1413 EraseTerminatorInstAndDCECond(BI);
1414 return true;
1415}
1416
1417// All instructions in Insts belong to different blocks that all unconditionally
1418// branch to a common successor. Analyze each instruction and return true if it
1419// would be possible to sink them into their successor, creating one common
1420// instruction instead. For every value that would be required to be provided by
1421// PHI node (because an operand varies in each input block), add to PHIOperands.
1422static bool canSinkInstructions(
1423 ArrayRef<Instruction *> Insts,
1424 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1425 // Prune out obviously bad instructions to move. Any non-store instruction
1426 // must have exactly one use, and we check later that use is by a single,
1427 // common PHI instruction in the successor.
1428 for (auto *I : Insts) {
1429 // These instructions may change or break semantics if moved.
1430 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1431 I->getType()->isTokenTy())
1432 return false;
1433
1434 // Conservatively return false if I is an inline-asm instruction. Sinking
1435 // and merging inline-asm instructions can potentially create arguments
1436 // that cannot satisfy the inline-asm constraints.
1437 if (const auto *C = dyn_cast<CallInst>(I))
1438 if (C->isInlineAsm())
1439 return false;
1440
1441 // Everything must have only one use too, apart from stores which
1442 // have no uses.
1443 if (!isa<StoreInst>(I) && !I->hasOneUse())
1444 return false;
1445 }
1446
1447 const Instruction *I0 = Insts.front();
1448 for (auto *I : Insts)
1449 if (!I->isSameOperationAs(I0))
1450 return false;
1451
1452 // All instructions in Insts are known to be the same opcode. If they aren't
1453 // stores, check the only user of each is a PHI or in the same block as the
1454 // instruction, because if a user is in the same block as an instruction
1455 // we're contemplating sinking, it must already be determined to be sinkable.
1456 if (!isa<StoreInst>(I0)) {
1457 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1458 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1459 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1460 auto *U = cast<Instruction>(*I->user_begin());
1461 return (PNUse &&
1462 PNUse->getParent() == Succ &&
1463 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1464 U->getParent() == I->getParent();
1465 }))
1466 return false;
1467 }
1468
1469 // Because SROA can't handle speculating stores of selects, try not
1470 // to sink loads or stores of allocas when we'd have to create a PHI for
1471 // the address operand. Also, because it is likely that loads or stores
1472 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1473 // This can cause code churn which can have unintended consequences down
1474 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1475 // FIXME: This is a workaround for a deficiency in SROA - see
1476 // https://llvm.org/bugs/show_bug.cgi?id=30188
1477 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1478 return isa<AllocaInst>(I->getOperand(1));
1479 }))
1480 return false;
1481 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1482 return isa<AllocaInst>(I->getOperand(0));
1483 }))
1484 return false;
1485
1486 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1487 if (I0->getOperand(OI)->getType()->isTokenTy())
1488 // Don't touch any operand of token type.
1489 return false;
1490
1491 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1492 assert(I->getNumOperands() == I0->getNumOperands())(static_cast <bool> (I->getNumOperands() == I0->getNumOperands
()) ? void (0) : __assert_fail ("I->getNumOperands() == I0->getNumOperands()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1492, __extension__ __PRETTY_FUNCTION__))
;
1493 return I->getOperand(OI) == I0->getOperand(OI);
1494 };
1495 if (!all_of(Insts, SameAsI0)) {
1496 if (!canReplaceOperandWithVariable(I0, OI))
1497 // We can't create a PHI from this GEP.
1498 return false;
1499 // Don't create indirect calls! The called value is the final operand.
1500 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1501 // FIXME: if the call was *already* indirect, we should do this.
1502 return false;
1503 }
1504 for (auto *I : Insts)
1505 PHIOperands[I].push_back(I->getOperand(OI));
1506 }
1507 }
1508 return true;
1509}
1510
1511// Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1512// instruction of every block in Blocks to their common successor, commoning
1513// into one instruction.
1514static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1515 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1516
1517 // canSinkLastInstruction returning true guarantees that every block has at
1518 // least one non-terminator instruction.
1519 SmallVector<Instruction*,4> Insts;
1520 for (auto *BB : Blocks) {
1521 Instruction *I = BB->getTerminator();
1522 do {
1523 I = I->getPrevNode();
1524 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1525 if (!isa<DbgInfoIntrinsic>(I))
1526 Insts.push_back(I);
1527 }
1528
1529 // The only checking we need to do now is that all users of all instructions
1530 // are the same PHI node. canSinkLastInstruction should have checked this but
1531 // it is slightly over-aggressive - it gets confused by commutative instructions
1532 // so double-check it here.
1533 Instruction *I0 = Insts.front();
1534 if (!isa<StoreInst>(I0)) {
1535 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1536 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1537 auto *U = cast<Instruction>(*I->user_begin());
1538 return U == PNUse;
1539 }))
1540 return false;
1541 }
1542
1543 // We don't need to do any more checking here; canSinkLastInstruction should
1544 // have done it all for us.
1545 SmallVector<Value*, 4> NewOperands;
1546 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1547 // This check is different to that in canSinkLastInstruction. There, we
1548 // cared about the global view once simplifycfg (and instcombine) have
1549 // completed - it takes into account PHIs that become trivially
1550 // simplifiable. However here we need a more local view; if an operand
1551 // differs we create a PHI and rely on instcombine to clean up the very
1552 // small mess we may make.
1553 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1554 return I->getOperand(O) != I0->getOperand(O);
1555 });
1556 if (!NeedPHI) {
1557 NewOperands.push_back(I0->getOperand(O));
1558 continue;
1559 }
1560
1561 // Create a new PHI in the successor block and populate it.
1562 auto *Op = I0->getOperand(O);
1563 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!")(static_cast <bool> (!Op->getType()->isTokenTy() &&
"Can't PHI tokens!") ? void (0) : __assert_fail ("!Op->getType()->isTokenTy() && \"Can't PHI tokens!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1563, __extension__ __PRETTY_FUNCTION__))
;
1564 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1565 Op->getName() + ".sink", &BBEnd->front());
1566 for (auto *I : Insts)
1567 PN->addIncoming(I->getOperand(O), I->getParent());
1568 NewOperands.push_back(PN);
1569 }
1570
1571 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1572 // and move it to the start of the successor block.
1573 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1574 I0->getOperandUse(O).set(NewOperands[O]);
1575 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1576
1577 // Update metadata and IR flags, and merge debug locations.
1578 for (auto *I : Insts)
1579 if (I != I0) {
1580 // The debug location for the "common" instruction is the merged locations
1581 // of all the commoned instructions. We start with the original location
1582 // of the "common" instruction and iteratively merge each location in the
1583 // loop below.
1584 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1585 // However, as N-way merge for CallInst is rare, so we use simplified API
1586 // instead of using complex API for N-way merge.
1587 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1588 combineMetadataForCSE(I0, I);
1589 I0->andIRFlags(I);
1590 }
1591
1592 if (!isa<StoreInst>(I0)) {
1593 // canSinkLastInstruction checked that all instructions were used by
1594 // one and only one PHI node. Find that now, RAUW it to our common
1595 // instruction and nuke it.
1596 assert(I0->hasOneUse())(static_cast <bool> (I0->hasOneUse()) ? void (0) : __assert_fail
("I0->hasOneUse()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1596, __extension__ __PRETTY_FUNCTION__))
;
1597 auto *PN = cast<PHINode>(*I0->user_begin());
1598 PN->replaceAllUsesWith(I0);
1599 PN->eraseFromParent();
1600 }
1601
1602 // Finally nuke all instructions apart from the common instruction.
1603 for (auto *I : Insts)
1604 if (I != I0)
1605 I->eraseFromParent();
1606
1607 return true;
1608}
1609
1610namespace {
1611
1612 // LockstepReverseIterator - Iterates through instructions
1613 // in a set of blocks in reverse order from the first non-terminator.
1614 // For example (assume all blocks have size n):
1615 // LockstepReverseIterator I([B1, B2, B3]);
1616 // *I-- = [B1[n], B2[n], B3[n]];
1617 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1618 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1619 // ...
1620 class LockstepReverseIterator {
1621 ArrayRef<BasicBlock*> Blocks;
1622 SmallVector<Instruction*,4> Insts;
1623 bool Fail;
1624
1625 public:
1626 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1627 reset();
1628 }
1629
1630 void reset() {
1631 Fail = false;
1632 Insts.clear();
1633 for (auto *BB : Blocks) {
1634 Instruction *Inst = BB->getTerminator();
1635 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1636 Inst = Inst->getPrevNode();
1637 if (!Inst) {
1638 // Block wasn't big enough.
1639 Fail = true;
1640 return;
1641 }
1642 Insts.push_back(Inst);
1643 }
1644 }
1645
1646 bool isValid() const {
1647 return !Fail;
1648 }
1649
1650 void operator--() {
1651 if (Fail)
1652 return;
1653 for (auto *&Inst : Insts) {
1654 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1655 Inst = Inst->getPrevNode();
1656 // Already at beginning of block.
1657 if (!Inst) {
1658 Fail = true;
1659 return;
1660 }
1661 }
1662 }
1663
1664 ArrayRef<Instruction*> operator * () const {
1665 return Insts;
1666 }
1667 };
1668
1669} // end anonymous namespace
1670
1671/// Check whether BB's predecessors end with unconditional branches. If it is
1672/// true, sink any common code from the predecessors to BB.
1673/// We also allow one predecessor to end with conditional branch (but no more
1674/// than one).
1675static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1676 // We support two situations:
1677 // (1) all incoming arcs are unconditional
1678 // (2) one incoming arc is conditional
1679 //
1680 // (2) is very common in switch defaults and
1681 // else-if patterns;
1682 //
1683 // if (a) f(1);
1684 // else if (b) f(2);
1685 //
1686 // produces:
1687 //
1688 // [if]
1689 // / \
1690 // [f(1)] [if]
1691 // | | \
1692 // | | |
1693 // | [f(2)]|
1694 // \ | /
1695 // [ end ]
1696 //
1697 // [end] has two unconditional predecessor arcs and one conditional. The
1698 // conditional refers to the implicit empty 'else' arc. This conditional
1699 // arc can also be caused by an empty default block in a switch.
1700 //
1701 // In this case, we attempt to sink code from all *unconditional* arcs.
1702 // If we can sink instructions from these arcs (determined during the scan
1703 // phase below) we insert a common successor for all unconditional arcs and
1704 // connect that to [end], to enable sinking:
1705 //
1706 // [if]
1707 // / \
1708 // [x(1)] [if]
1709 // | | \
1710 // | | \
1711 // | [x(2)] |
1712 // \ / |
1713 // [sink.split] |
1714 // \ /
1715 // [ end ]
1716 //
1717 SmallVector<BasicBlock*,4> UnconditionalPreds;
1718 Instruction *Cond = nullptr;
1719 for (auto *B : predecessors(BB)) {
1720 auto *T = B->getTerminator();
1721 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1722 UnconditionalPreds.push_back(B);
1723 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1724 Cond = T;
1725 else
1726 return false;
1727 }
1728 if (UnconditionalPreds.size() < 2)
1729 return false;
1730
1731 bool Changed = false;
1732 // We take a two-step approach to tail sinking. First we scan from the end of
1733 // each block upwards in lockstep. If the n'th instruction from the end of each
1734 // block can be sunk, those instructions are added to ValuesToSink and we
1735 // carry on. If we can sink an instruction but need to PHI-merge some operands
1736 // (because they're not identical in each instruction) we add these to
1737 // PHIOperands.
1738 unsigned ScanIdx = 0;
1739 SmallPtrSet<Value*,4> InstructionsToSink;
1740 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1741 LockstepReverseIterator LRI(UnconditionalPreds);
1742 while (LRI.isValid() &&
1743 canSinkInstructions(*LRI, PHIOperands)) {
1744 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: instruction can be sunk: "
<< *(*LRI)[0] << "\n"; } } while (false)
;
1745 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1746 ++ScanIdx;
1747 --LRI;
1748 }
1749
1750 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1751 unsigned NumPHIdValues = 0;
1752 for (auto *I : *LRI)
1753 for (auto *V : PHIOperands[I])
1754 if (InstructionsToSink.count(V) == 0)
1755 ++NumPHIdValues;
1756 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: #phid values: " <<
NumPHIdValues << "\n"; } } while (false)
;
1757 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1758 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1759 NumPHIInsts++;
1760
1761 return NumPHIInsts <= 1;
1762 };
1763
1764 if (ScanIdx > 0 && Cond) {
1765 // Check if we would actually sink anything first! This mutates the CFG and
1766 // adds an extra block. The goal in doing this is to allow instructions that
1767 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1768 // (such as trunc, add) can be sunk and predicated already. So we check that
1769 // we're going to sink at least one non-speculatable instruction.
1770 LRI.reset();
1771 unsigned Idx = 0;
1772 bool Profitable = false;
1773 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1774 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1775 Profitable = true;
1776 break;
1777 }
1778 --LRI;
1779 ++Idx;
1780 }
1781 if (!Profitable)
1782 return false;
1783
1784 DEBUG(dbgs() << "SINK: Splitting edge\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Splitting edge\n"; }
} while (false)
;
1785 // We have a conditional edge and we're going to sink some instructions.
1786 // Insert a new block postdominating all blocks we're going to sink from.
1787 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1788 // Edges couldn't be split.
1789 return false;
1790 Changed = true;
1791 }
1792
1793 // Now that we've analyzed all potential sinking candidates, perform the
1794 // actual sink. We iteratively sink the last non-terminator of the source
1795 // blocks into their common successor unless doing so would require too
1796 // many PHI instructions to be generated (currently only one PHI is allowed
1797 // per sunk instruction).
1798 //
1799 // We can use InstructionsToSink to discount values needing PHI-merging that will
1800 // actually be sunk in a later iteration. This allows us to be more
1801 // aggressive in what we sink. This does allow a false positive where we
1802 // sink presuming a later value will also be sunk, but stop half way through
1803 // and never actually sink it which means we produce more PHIs than intended.
1804 // This is unlikely in practice though.
1805 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1806 DEBUG(dbgs() << "SINK: Sink: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1807 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
1808 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SINK: Sink: " << *UnconditionalPreds
[0]->getTerminator()->getPrevNode() << "\n"; } } while
(false)
;
1809
1810 // Because we've sunk every instruction in turn, the current instruction to
1811 // sink is always at index 0.
1812 LRI.reset();
1813 if (!ProfitableToSinkInstruction(LRI)) {
1814 // Too many PHIs would be created.
1815 DEBUG(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)
;
1816 break;
1817 }
1818
1819 if (!sinkLastInstruction(UnconditionalPreds))
1820 return Changed;
1821 NumSinkCommons++;
1822 Changed = true;
1823 }
1824 return Changed;
1825}
1826
1827/// \brief Determine if we can hoist sink a sole store instruction out of a
1828/// conditional block.
1829///
1830/// We are looking for code like the following:
1831/// BrBB:
1832/// store i32 %add, i32* %arrayidx2
1833/// ... // No other stores or function calls (we could be calling a memory
1834/// ... // function).
1835/// %cmp = icmp ult %x, %y
1836/// br i1 %cmp, label %EndBB, label %ThenBB
1837/// ThenBB:
1838/// store i32 %add5, i32* %arrayidx2
1839/// br label EndBB
1840/// EndBB:
1841/// ...
1842/// We are going to transform this into:
1843/// BrBB:
1844/// store i32 %add, i32* %arrayidx2
1845/// ... //
1846/// %cmp = icmp ult %x, %y
1847/// %add.add5 = select i1 %cmp, i32 %add, %add5
1848/// store i32 %add.add5, i32* %arrayidx2
1849/// ...
1850///
1851/// \return The pointer to the value of the previous store if the store can be
1852/// hoisted into the predecessor block. 0 otherwise.
1853static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1854 BasicBlock *StoreBB, BasicBlock *EndBB) {
1855 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1856 if (!StoreToHoist)
1857 return nullptr;
1858
1859 // Volatile or atomic.
1860 if (!StoreToHoist->isSimple())
1861 return nullptr;
1862
1863 Value *StorePtr = StoreToHoist->getPointerOperand();
1864
1865 // Look for a store to the same pointer in BrBB.
1866 unsigned MaxNumInstToLookAt = 9;
1867 for (Instruction &CurI : reverse(*BrBB)) {
1868 if (!MaxNumInstToLookAt)
1869 break;
1870 // Skip debug info.
1871 if (isa<DbgInfoIntrinsic>(CurI))
1872 continue;
1873 --MaxNumInstToLookAt;
1874
1875 // Could be calling an instruction that affects memory like free().
1876 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1877 return nullptr;
1878
1879 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1880 // Found the previous store make sure it stores to the same location.
1881 if (SI->getPointerOperand() == StorePtr)
1882 // Found the previous store, return its value operand.
1883 return SI->getValueOperand();
1884 return nullptr; // Unknown store.
1885 }
1886 }
1887
1888 return nullptr;
1889}
1890
1891/// \brief Speculate a conditional basic block flattening the CFG.
1892///
1893/// Note that this is a very risky transform currently. Speculating
1894/// instructions like this is most often not desirable. Instead, there is an MI
1895/// pass which can do it with full awareness of the resource constraints.
1896/// However, some cases are "obvious" and we should do directly. An example of
1897/// this is speculating a single, reasonably cheap instruction.
1898///
1899/// There is only one distinct advantage to flattening the CFG at the IR level:
1900/// it makes very common but simplistic optimizations such as are common in
1901/// instcombine and the DAG combiner more powerful by removing CFG edges and
1902/// modeling their effects with easier to reason about SSA value graphs.
1903///
1904///
1905/// An illustration of this transform is turning this IR:
1906/// \code
1907/// BB:
1908/// %cmp = icmp ult %x, %y
1909/// br i1 %cmp, label %EndBB, label %ThenBB
1910/// ThenBB:
1911/// %sub = sub %x, %y
1912/// br label BB2
1913/// EndBB:
1914/// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1915/// ...
1916/// \endcode
1917///
1918/// Into this IR:
1919/// \code
1920/// BB:
1921/// %cmp = icmp ult %x, %y
1922/// %sub = sub %x, %y
1923/// %cond = select i1 %cmp, 0, %sub
1924/// ...
1925/// \endcode
1926///
1927/// \returns true if the conditional block is removed.
1928static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1929 const TargetTransformInfo &TTI) {
1930 // Be conservative for now. FP select instruction can often be expensive.
1931 Value *BrCond = BI->getCondition();
1932 if (isa<FCmpInst>(BrCond))
1933 return false;
1934
1935 BasicBlock *BB = BI->getParent();
1936 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1937
1938 // If ThenBB is actually on the false edge of the conditional branch, remember
1939 // to swap the select operands later.
1940 bool Invert = false;
1941 if (ThenBB != BI->getSuccessor(0)) {
1942 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?")(static_cast <bool> (ThenBB == BI->getSuccessor(1) &&
"No edge from 'if' block?") ? void (0) : __assert_fail ("ThenBB == BI->getSuccessor(1) && \"No edge from 'if' block?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1942, __extension__ __PRETTY_FUNCTION__))
;
1943 Invert = true;
1944 }
1945 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block")(static_cast <bool> (EndBB == BI->getSuccessor(!Invert
) && "No edge from to end block") ? void (0) : __assert_fail
("EndBB == BI->getSuccessor(!Invert) && \"No edge from to end block\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 1945, __extension__ __PRETTY_FUNCTION__))
;
1946
1947 // Keep a count of how many times instructions are used within CondBB when
1948 // they are candidates for sinking into CondBB. Specifically:
1949 // - They are defined in BB, and
1950 // - They have no side effects, and
1951 // - All of their uses are in CondBB.
1952 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1953
1954 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1955
1956 unsigned SpeculationCost = 0;
1957 Value *SpeculatedStoreValue = nullptr;
1958 StoreInst *SpeculatedStore = nullptr;
1959 for (BasicBlock::iterator BBI = ThenBB->begin(),
1960 BBE = std::prev(ThenBB->end());
1961 BBI != BBE; ++BBI) {
1962 Instruction *I = &*BBI;
1963 // Skip debug info.
1964 if (isa<DbgInfoIntrinsic>(I)) {
1965 SpeculatedDbgIntrinsics.push_back(I);
1966 continue;
1967 }
1968
1969 // Only speculatively execute a single instruction (not counting the
1970 // terminator) for now.
1971 ++SpeculationCost;
1972 if (SpeculationCost > 1)
1973 return false;
1974
1975 // Don't hoist the instruction if it's unsafe or expensive.
1976 if (!isSafeToSpeculativelyExecute(I) &&
1977 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1978 I, BB, ThenBB, EndBB))))
1979 return false;
1980 if (!SpeculatedStoreValue &&
1981 ComputeSpeculationCost(I, TTI) >
1982 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1983 return false;
1984
1985 // Store the store speculation candidate.
1986 if (SpeculatedStoreValue)
1987 SpeculatedStore = cast<StoreInst>(I);
1988
1989 // Do not hoist the instruction if any of its operands are defined but not
1990 // used in BB. The transformation will prevent the operand from
1991 // being sunk into the use block.
1992 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1993 Instruction *OpI = dyn_cast<Instruction>(*i);
1994 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1995 continue; // Not a candidate for sinking.
1996
1997 ++SinkCandidateUseCounts[OpI];
1998 }
1999 }
2000
2001 // Consider any sink candidates which are only used in CondBB as costs for
2002 // speculation. Note, while we iterate over a DenseMap here, we are summing
2003 // and so iteration order isn't significant.
2004 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2005 I = SinkCandidateUseCounts.begin(),
2006 E = SinkCandidateUseCounts.end();
2007 I != E; ++I)
2008 if (I->first->getNumUses() == I->second) {
2009 ++SpeculationCost;
2010 if (SpeculationCost > 1)
2011 return false;
2012 }
2013
2014 // Check that the PHI nodes can be converted to selects.
2015 bool HaveRewritablePHIs = false;
2016 for (PHINode &PN : EndBB->phis()) {
2017 Value *OrigV = PN.getIncomingValueForBlock(BB);
2018 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2019
2020 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2021 // Skip PHIs which are trivial.
2022 if (ThenV == OrigV)
2023 continue;
2024
2025 // Don't convert to selects if we could remove undefined behavior instead.
2026 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2027 passingValueIsAlwaysUndefined(ThenV, &PN))
2028 return false;
2029
2030 HaveRewritablePHIs = true;
2031 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2032 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2033 if (!OrigCE && !ThenCE)
2034 continue; // Known safe and cheap.
2035
2036 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2037 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2038 return false;
2039 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2040 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2041 unsigned MaxCost =
2042 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2043 if (OrigCost + ThenCost > MaxCost)
2044 return false;
2045
2046 // Account for the cost of an unfolded ConstantExpr which could end up
2047 // getting expanded into Instructions.
2048 // FIXME: This doesn't account for how many operations are combined in the
2049 // constant expression.
2050 ++SpeculationCost;
2051 if (SpeculationCost > 1)
2052 return false;
2053 }
2054
2055 // If there are no PHIs to process, bail early. This helps ensure idempotence
2056 // as well.
2057 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2058 return false;
2059
2060 // If we get here, we can hoist the instruction and if-convert.
2061 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SPECULATIVELY EXECUTING BB"
<< *ThenBB << "\n";; } } while (false)
;
2062
2063 // Insert a select of the value of the speculated store.
2064 if (SpeculatedStoreValue) {
2065 IRBuilder<NoFolder> Builder(BI);
2066 Value *TrueV = SpeculatedStore->getValueOperand();
2067 Value *FalseV = SpeculatedStoreValue;
2068 if (Invert)
2069 std::swap(TrueV, FalseV);
2070 Value *S = Builder.CreateSelect(
2071 BrCond, TrueV, FalseV, "spec.store.select", BI);
2072 SpeculatedStore->setOperand(0, S);
2073 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2074 SpeculatedStore->getDebugLoc());
2075 }
2076
2077 // Metadata can be dependent on the condition we are hoisting above.
2078 // Conservatively strip all metadata on the instruction.
2079 for (auto &I : *ThenBB)
2080 I.dropUnknownNonDebugMetadata();
2081
2082 // Hoist the instructions.
2083 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2084 ThenBB->begin(), std::prev(ThenBB->end()));
2085
2086 // Insert selects and rewrite the PHI operands.
2087 IRBuilder<NoFolder> Builder(BI);
2088 for (PHINode &PN : EndBB->phis()) {
2089 unsigned OrigI = PN.getBasicBlockIndex(BB);
2090 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2091 Value *OrigV = PN.getIncomingValue(OrigI);
2092 Value *ThenV = PN.getIncomingValue(ThenI);
2093
2094 // Skip PHIs which are trivial.
2095 if (OrigV == ThenV)
2096 continue;
2097
2098 // Create a select whose true value is the speculatively executed value and
2099 // false value is the preexisting value. Swap them if the branch
2100 // destinations were inverted.
2101 Value *TrueV = ThenV, *FalseV = OrigV;
2102 if (Invert)
2103 std::swap(TrueV, FalseV);
2104 Value *V = Builder.CreateSelect(
2105 BrCond, TrueV, FalseV, "spec.select", BI);
2106 PN.setIncomingValue(OrigI, V);
2107 PN.setIncomingValue(ThenI, V);
2108 }
2109
2110 // Remove speculated dbg intrinsics.
2111 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2112 // dbg value for the different flows and inserting it after the select.
2113 for (Instruction *I : SpeculatedDbgIntrinsics)
2114 I->eraseFromParent();
2115
2116 ++NumSpeculations;
2117 return true;
2118}
2119
2120/// Return true if we can thread a branch across this block.
2121static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2122 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2123 unsigned Size = 0;
2124
2125 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2126 if (isa<DbgInfoIntrinsic>(BBI))
2127 continue;
2128 if (Size > 10)
2129 return false; // Don't clone large BB's.
2130 ++Size;
2131
2132 // We can only support instructions that do not define values that are
2133 // live outside of the current basic block.
2134 for (User *U : BBI->users()) {
2135 Instruction *UI = cast<Instruction>(U);
2136 if (UI->getParent() != BB || isa<PHINode>(UI))
2137 return false;
2138 }
2139
2140 // Looks ok, continue checking.
2141 }
2142
2143 return true;
2144}
2145
2146/// If we have a conditional branch on a PHI node value that is defined in the
2147/// same block as the branch and if any PHI entries are constants, thread edges
2148/// corresponding to that entry to be branches to their ultimate destination.
2149static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2150 AssumptionCache *AC) {
2151 BasicBlock *BB = BI->getParent();
2152 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2153 // NOTE: we currently cannot transform this case if the PHI node is used
2154 // outside of the block.
2155 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2156 return false;
2157
2158 // Degenerate case of a single entry PHI.
2159 if (PN->getNumIncomingValues() == 1) {
2160 FoldSingleEntryPHINodes(PN->getParent());
2161 return true;
2162 }
2163
2164 // Now we know that this block has multiple preds and two succs.
2165 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2166 return false;
2167
2168 // Can't fold blocks that contain noduplicate or convergent calls.
2169 if (any_of(*BB, [](const Instruction &I) {
2170 const CallInst *CI = dyn_cast<CallInst>(&I);
2171 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2172 }))
2173 return false;
2174
2175 // Okay, this is a simple enough basic block. See if any phi values are
2176 // constants.
2177 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2178 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2179 if (!CB || !CB->getType()->isIntegerTy(1))
2180 continue;
2181
2182 // Okay, we now know that all edges from PredBB should be revectored to
2183 // branch to RealDest.
2184 BasicBlock *PredBB = PN->getIncomingBlock(i);
2185 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2186
2187 if (RealDest == BB)
2188 continue; // Skip self loops.
2189 // Skip if the predecessor's terminator is an indirect branch.
2190 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2191 continue;
2192
2193 // The dest block might have PHI nodes, other predecessors and other
2194 // difficult cases. Instead of being smart about this, just insert a new
2195 // block that jumps to the destination block, effectively splitting
2196 // the edge we are about to create.
2197 BasicBlock *EdgeBB =
2198 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2199 RealDest->getParent(), RealDest);
2200 BranchInst::Create(RealDest, EdgeBB);
2201
2202 // Update PHI nodes.
2203 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2204
2205 // BB may have instructions that are being threaded over. Clone these
2206 // instructions into EdgeBB. We know that there will be no uses of the
2207 // cloned instructions outside of EdgeBB.
2208 BasicBlock::iterator InsertPt = EdgeBB->begin();
2209 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2210 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2211 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2212 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2213 continue;
2214 }
2215 // Clone the instruction.
2216 Instruction *N = BBI->clone();
2217 if (BBI->hasName())
2218 N->setName(BBI->getName() + ".c");
2219
2220 // Update operands due to translation.
2221 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2222 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2223 if (PI != TranslateMap.end())
2224 *i = PI->second;
2225 }
2226
2227 // Check for trivial simplification.
2228 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2229 if (!BBI->use_empty())
2230 TranslateMap[&*BBI] = V;
2231 if (!N->mayHaveSideEffects()) {
2232 N->deleteValue(); // Instruction folded away, don't need actual inst
2233 N = nullptr;
2234 }
2235 } else {
2236 if (!BBI->use_empty())
2237 TranslateMap[&*BBI] = N;
2238 }
2239 // Insert the new instruction into its new home.
2240 if (N)
2241 EdgeBB->getInstList().insert(InsertPt, N);
2242
2243 // Register the new instruction with the assumption cache if necessary.
2244 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2245 if (II->getIntrinsicID() == Intrinsic::assume)
2246 AC->registerAssumption(II);
2247 }
2248
2249 // Loop over all of the edges from PredBB to BB, changing them to branch
2250 // to EdgeBB instead.
2251 TerminatorInst *PredBBTI = PredBB->getTerminator();
2252 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2253 if (PredBBTI->getSuccessor(i) == BB) {
2254 BB->removePredecessor(PredBB);
2255 PredBBTI->setSuccessor(i, EdgeBB);
2256 }
2257
2258 // Recurse, simplifying any other constants.
2259 return FoldCondBranchOnPHI(BI, DL, AC) | true;
2260 }
2261
2262 return false;
2263}
2264
2265/// Given a BB that starts with the specified two-entry PHI node,
2266/// see if we can eliminate it.
2267static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2268 const DataLayout &DL) {
2269 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2270 // statement", which has a very simple dominance structure. Basically, we
2271 // are trying to find the condition that is being branched on, which
2272 // subsequently causes this merge to happen. We really want control
2273 // dependence information for this check, but simplifycfg can't keep it up
2274 // to date, and this catches most of the cases we care about anyway.
2275 BasicBlock *BB = PN->getParent();
2276 const Function *Fn = BB->getParent();
2277 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
26
Assuming 'Fn' is null
2278 return false;
2279
2280 BasicBlock *IfTrue, *IfFalse;
2281 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2282 if (!IfCond ||
27
Assuming 'IfCond' is non-null
28
Taking false branch
2283 // Don't bother if the branch will be constant folded trivially.
2284 isa<ConstantInt>(IfCond))
2285 return false;
2286
2287 // Okay, we found that we can merge this two-entry phi node into a select.
2288 // Doing so would require us to fold *all* two entry phi nodes in this block.
2289 // At some point this becomes non-profitable (particularly if the target
2290 // doesn't support cmov's). Only do this transformation if there are two or
2291 // fewer PHI nodes in this block.
2292 unsigned NumPhis = 0;
2293 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
29
Loop condition is false. Execution continues on line 2300
2294 if (NumPhis > 2)
2295 return false;
2296
2297 // Loop over the PHI's seeing if we can promote them all to select
2298 // instructions. While we are at it, keep track of the instructions
2299 // that need to be moved to the dominating block.
2300 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2301 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2302 MaxCostVal1 = PHINodeFoldingThreshold;
2303 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2304 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2305
2306 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
30
Loop condition is false. Execution continues on line 2323
2307 PHINode *PN = cast<PHINode>(II++);
2308 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2309 PN->replaceAllUsesWith(V);
2310 PN->eraseFromParent();
2311 continue;
2312 }
2313
2314 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2315 MaxCostVal0, TTI) ||
2316 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2317 MaxCostVal1, TTI))
2318 return false;
2319 }
2320
2321 // If we folded the first phi, PN dangles at this point. Refresh it. If
2322 // we ran out of PHIs then we simplified them all.
2323 PN = dyn_cast<PHINode>(BB->begin());
2324 if (!PN)
31
Assuming 'PN' is non-null
32
Taking false branch
2325 return true;
2326
2327 // Don't fold i1 branches on PHIs which contain binary operators. These can
2328 // often be turned into switches and other things.
2329 if (PN->getType()->isIntegerTy(1) &&
33
Assuming the condition is false
34
Taking false branch
2330 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2331 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2332 isa<BinaryOperator>(IfCond)))
2333 return false;
2334
2335 // If all PHI nodes are promotable, check to make sure that all instructions
2336 // in the predecessor blocks can be promoted as well. If not, we won't be able
2337 // to get rid of the control flow, so it's not worth promoting to select
2338 // instructions.
2339 BasicBlock *DomBlock = nullptr;
35
'DomBlock' initialized to a null pointer value
2340 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2341 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2342 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
36
Taking true branch
2343 IfBlock1 = nullptr;
2344 } else {
2345 DomBlock = *pred_begin(IfBlock1);
2346 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2347 ++I)
2348 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2349 // This is not an aggressive instruction that we can promote.
2350 // Because of this, we won't be able to get rid of the control flow, so
2351 // the xform is not worth it.
2352 return false;
2353 }
2354 }
2355
2356 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
37
Taking true branch
2357 IfBlock2 = nullptr;
2358 } else {
2359 DomBlock = *pred_begin(IfBlock2);
2360 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2361 ++I)
2362 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2363 // This is not an aggressive instruction that we can promote.
2364 // Because of this, we won't be able to get rid of the control flow, so
2365 // the xform is not worth it.
2366 return false;
2367 }
2368 }
2369
2370 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOUND IF CONDITION! " <<
*IfCond << " T: " << IfTrue->getName() <<
" F: " << IfFalse->getName() << "\n"; } } while
(false)
2371 << IfTrue->getName() << " 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)
;
2372
2373 // If we can still promote the PHI nodes after this gauntlet of tests,
2374 // do all of the PHI's now.
2375 Instruction *InsertPt = DomBlock->getTerminator();
38
Called C++ object pointer is null
2376 IRBuilder<NoFolder> Builder(InsertPt);
2377
2378 // Move all 'aggressive' instructions, which are defined in the
2379 // conditional parts of the if's up to the dominating block.
2380 if (IfBlock1) {
2381 for (auto &I : *IfBlock1)
2382 I.dropUnknownNonDebugMetadata();
2383 DomBlock->getInstList().splice(InsertPt->getIterator(),
2384 IfBlock1->getInstList(), IfBlock1->begin(),
2385 IfBlock1->getTerminator()->getIterator());
2386 }
2387 if (IfBlock2) {
2388 for (auto &I : *IfBlock2)
2389 I.dropUnknownNonDebugMetadata();
2390 DomBlock->getInstList().splice(InsertPt->getIterator(),
2391 IfBlock2->getInstList(), IfBlock2->begin(),
2392 IfBlock2->getTerminator()->getIterator());
2393 }
2394
2395 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2396 // Change the PHI node into a select instruction.
2397 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2398 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2399
2400 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2401 PN->replaceAllUsesWith(Sel);
2402 Sel->takeName(PN);
2403 PN->eraseFromParent();
2404 }
2405
2406 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2407 // has been flattened. Change DomBlock to jump directly to our new block to
2408 // avoid other simplifycfg's kicking in on the diamond.
2409 TerminatorInst *OldTI = DomBlock->getTerminator();
2410 Builder.SetInsertPoint(OldTI);
2411 Builder.CreateBr(BB);
2412 OldTI->eraseFromParent();
2413 return true;
2414}
2415
2416/// If we found a conditional branch that goes to two returning blocks,
2417/// try to merge them together into one return,
2418/// introducing a select if the return values disagree.
2419static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2420 IRBuilder<> &Builder) {
2421 assert(BI->isConditional() && "Must be a conditional branch")(static_cast <bool> (BI->isConditional() && "Must be a conditional branch"
) ? void (0) : __assert_fail ("BI->isConditional() && \"Must be a conditional branch\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2421, __extension__ __PRETTY_FUNCTION__))
;
2422 BasicBlock *TrueSucc = BI->getSuccessor(0);
2423 BasicBlock *FalseSucc = BI->getSuccessor(1);
2424 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2425 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2426
2427 // Check to ensure both blocks are empty (just a return) or optionally empty
2428 // with PHI nodes. If there are other instructions, merging would cause extra
2429 // computation on one path or the other.
2430 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2431 return false;
2432 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2433 return false;
2434
2435 Builder.SetInsertPoint(BI);
2436 // Okay, we found a branch that is going to two return nodes. If
2437 // there is no return value for this function, just change the
2438 // branch into a return.
2439 if (FalseRet->getNumOperands() == 0) {
2440 TrueSucc->removePredecessor(BI->getParent());
2441 FalseSucc->removePredecessor(BI->getParent());
2442 Builder.CreateRetVoid();
2443 EraseTerminatorInstAndDCECond(BI);
2444 return true;
2445 }
2446
2447 // Otherwise, figure out what the true and false return values are
2448 // so we can insert a new select instruction.
2449 Value *TrueValue = TrueRet->getReturnValue();
2450 Value *FalseValue = FalseRet->getReturnValue();
2451
2452 // Unwrap any PHI nodes in the return blocks.
2453 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2454 if (TVPN->getParent() == TrueSucc)
2455 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2456 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2457 if (FVPN->getParent() == FalseSucc)
2458 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2459
2460 // In order for this transformation to be safe, we must be able to
2461 // unconditionally execute both operands to the return. This is
2462 // normally the case, but we could have a potentially-trapping
2463 // constant expression that prevents this transformation from being
2464 // safe.
2465 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2466 if (TCV->canTrap())
2467 return false;
2468 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2469 if (FCV->canTrap())
2470 return false;
2471
2472 // Okay, we collected all the mapped values and checked them for sanity, and
2473 // defined to really do this transformation. First, update the CFG.
2474 TrueSucc->removePredecessor(BI->getParent());
2475 FalseSucc->removePredecessor(BI->getParent());
2476
2477 // Insert select instructions where needed.
2478 Value *BrCond = BI->getCondition();
2479 if (TrueValue) {
2480 // Insert a select if the results differ.
2481 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2482 } else if (isa<UndefValue>(TrueValue)) {
2483 TrueValue = FalseValue;
2484 } else {
2485 TrueValue =
2486 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2487 }
2488 }
2489
2490 Value *RI =
2491 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2492
2493 (void)RI;
2494
2495 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)
2496 << "\n " << *BI << "NewRet = " << *RIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "NewRet = " << *
RI << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "
<< *FalseSucc; } } while (false)
2497 << "TRUEBLOCK: " << *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)
;
2498
2499 EraseTerminatorInstAndDCECond(BI);
2500
2501 return true;
2502}
2503
2504/// Return true if the given instruction is available
2505/// in its predecessor block. If yes, the instruction will be removed.
2506static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2507 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2508 return false;
2509 for (Instruction &I : *PB) {
2510 Instruction *PBI = &I;
2511 // Check whether Inst and PBI generate the same value.
2512 if (Inst->isIdenticalTo(PBI)) {
2513 Inst->replaceAllUsesWith(PBI);
2514 Inst->eraseFromParent();
2515 return true;
2516 }
2517 }
2518 return false;
2519}
2520
2521/// Return true if either PBI or BI has branch weight available, and store
2522/// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2523/// not have branch weight, use 1:1 as its weight.
2524static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2525 uint64_t &PredTrueWeight,
2526 uint64_t &PredFalseWeight,
2527 uint64_t &SuccTrueWeight,
2528 uint64_t &SuccFalseWeight) {
2529 bool PredHasWeights =
2530 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2531 bool SuccHasWeights =
2532 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2533 if (PredHasWeights || SuccHasWeights) {
2534 if (!PredHasWeights)
2535 PredTrueWeight = PredFalseWeight = 1;
2536 if (!SuccHasWeights)
2537 SuccTrueWeight = SuccFalseWeight = 1;
2538 return true;
2539 } else {
2540 return false;
2541 }
2542}
2543
2544/// If this basic block is simple enough, and if a predecessor branches to us
2545/// and one of our successors, fold the block into the predecessor and use
2546/// logical operations to pick the right destination.
2547bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2548 BasicBlock *BB = BI->getParent();
2549
2550 Instruction *Cond = nullptr;
2551 if (BI->isConditional())
2552 Cond = dyn_cast<Instruction>(BI->getCondition());
2553 else {
2554 // For unconditional branch, check for a simple CFG pattern, where
2555 // BB has a single predecessor and BB's successor is also its predecessor's
2556 // successor. If such pattern exists, check for CSE between BB and its
2557 // predecessor.
2558 if (BasicBlock *PB = BB->getSinglePredecessor())
2559 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2560 if (PBI->isConditional() &&
2561 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2562 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2563 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2564 Instruction *Curr = &*I++;
2565 if (isa<CmpInst>(Curr)) {
2566 Cond = Curr;
2567 break;
2568 }
2569 // Quit if we can't remove this instruction.
2570 if (!checkCSEInPredecessor(Curr, PB))
2571 return false;
2572 }
2573 }
2574
2575 if (!Cond)
2576 return false;
2577 }
2578
2579 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2580 Cond->getParent() != BB || !Cond->hasOneUse())
2581 return false;
2582
2583 // Make sure the instruction after the condition is the cond branch.
2584 BasicBlock::iterator CondIt = ++Cond->getIterator();
2585
2586 // Ignore dbg intrinsics.
2587 while (isa<DbgInfoIntrinsic>(CondIt))
2588 ++CondIt;
2589
2590 if (&*CondIt != BI)
2591 return false;
2592
2593 // Only allow this transformation if computing the condition doesn't involve
2594 // too many instructions and these involved instructions can be executed
2595 // unconditionally. We denote all involved instructions except the condition
2596 // as "bonus instructions", and only allow this transformation when the
2597 // number of the bonus instructions does not exceed a certain threshold.
2598 unsigned NumBonusInsts = 0;
2599 for (auto I = BB->begin(); Cond != &*I; ++I) {
2600 // Ignore dbg intrinsics.
2601 if (isa<DbgInfoIntrinsic>(I))
2602 continue;
2603 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2604 return false;
2605 // I has only one use and can be executed unconditionally.
2606 Instruction *User = dyn_cast<Instruction>(I->user_back());
2607 if (User == nullptr || User->getParent() != BB)
2608 return false;
2609 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2610 // to use any other instruction, User must be an instruction between next(I)
2611 // and Cond.
2612 ++NumBonusInsts;
2613 // Early exits once we reach the limit.
2614 if (NumBonusInsts > BonusInstThreshold)
2615 return false;
2616 }
2617
2618 // Cond is known to be a compare or binary operator. Check to make sure that
2619 // neither operand is a potentially-trapping constant expression.
2620 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2621 if (CE->canTrap())
2622 return false;
2623 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2624 if (CE->canTrap())
2625 return false;
2626
2627 // Finally, don't infinitely unroll conditional loops.
2628 BasicBlock *TrueDest = BI->getSuccessor(0);
2629 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2630 if (TrueDest == BB || FalseDest == BB)
2631 return false;
2632
2633 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2634 BasicBlock *PredBlock = *PI;
2635 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2636
2637 // Check that we have two conditional branches. If there is a PHI node in
2638 // the common successor, verify that the same value flows in from both
2639 // blocks.
2640 SmallVector<PHINode *, 4> PHIs;
2641 if (!PBI || PBI->isUnconditional() ||
2642 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2643 (!BI->isConditional() &&
2644 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2645 continue;
2646
2647 // Determine if the two branches share a common destination.
2648 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2649 bool InvertPredCond = false;
2650
2651 if (BI->isConditional()) {
2652 if (PBI->getSuccessor(0) == TrueDest) {
2653 Opc = Instruction::Or;
2654 } else if (PBI->getSuccessor(1) == FalseDest) {
2655 Opc = Instruction::And;
2656 } else if (PBI->getSuccessor(0) == FalseDest) {
2657 Opc = Instruction::And;
2658 InvertPredCond = true;
2659 } else if (PBI->getSuccessor(1) == TrueDest) {
2660 Opc = Instruction::Or;
2661 InvertPredCond = true;
2662 } else {
2663 continue;
2664 }
2665 } else {
2666 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2667 continue;
2668 }
2669
2670 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)
;
2671 IRBuilder<> Builder(PBI);
2672
2673 // If we need to invert the condition in the pred block to match, do so now.
2674 if (InvertPredCond) {
2675 Value *NewCond = PBI->getCondition();
2676
2677 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2678 CmpInst *CI = cast<CmpInst>(NewCond);
2679 CI->setPredicate(CI->getInversePredicate());
2680 } else {
2681 NewCond =
2682 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2683 }
2684
2685 PBI->setCondition(NewCond);
2686 PBI->swapSuccessors();
2687 }
2688
2689 // If we have bonus instructions, clone them into the predecessor block.
2690 // Note that there may be multiple predecessor blocks, so we cannot move
2691 // bonus instructions to a predecessor block.
2692 ValueToValueMapTy VMap; // maps original values to cloned values
2693 // We already make sure Cond is the last instruction before BI. Therefore,
2694 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2695 // instructions.
2696 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2697 if (isa<DbgInfoIntrinsic>(BonusInst))
2698 continue;
2699 Instruction *NewBonusInst = BonusInst->clone();
2700 RemapInstruction(NewBonusInst, VMap,
2701 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2702 VMap[&*BonusInst] = NewBonusInst;
2703
2704 // If we moved a load, we cannot any longer claim any knowledge about
2705 // its potential value. The previous information might have been valid
2706 // only given the branch precondition.
2707 // For an analogous reason, we must also drop all the metadata whose
2708 // semantics we don't understand.
2709 NewBonusInst->dropUnknownNonDebugMetadata();
2710
2711 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2712 NewBonusInst->takeName(&*BonusInst);
2713 BonusInst->setName(BonusInst->getName() + ".old");
2714 }
2715
2716 // Clone Cond into the predecessor basic block, and or/and the
2717 // two conditions together.
2718 Instruction *New = Cond->clone();
2719 RemapInstruction(New, VMap,
2720 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2721 PredBlock->getInstList().insert(PBI->getIterator(), New);
2722 New->takeName(Cond);
2723 Cond->setName(New->getName() + ".old");
2724
2725 if (BI->isConditional()) {
2726 Instruction *NewCond = cast<Instruction>(
2727 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2728 PBI->setCondition(NewCond);
2729
2730 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2731 bool HasWeights =
2732 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2733 SuccTrueWeight, SuccFalseWeight);
2734 SmallVector<uint64_t, 8> NewWeights;
2735
2736 if (PBI->getSuccessor(0) == BB) {
2737 if (HasWeights) {
2738 // PBI: br i1 %x, BB, FalseDest
2739 // BI: br i1 %y, TrueDest, FalseDest
2740 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2741 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2742 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2743 // TrueWeight for PBI * FalseWeight for BI.
2744 // We assume that total weights of a BranchInst can fit into 32 bits.
2745 // Therefore, we will not have overflow using 64-bit arithmetic.
2746 NewWeights.push_back(PredFalseWeight *
2747 (SuccFalseWeight + SuccTrueWeight) +
2748 PredTrueWeight * SuccFalseWeight);
2749 }
2750 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2751 PBI->setSuccessor(0, TrueDest);
2752 }
2753 if (PBI->getSuccessor(1) == BB) {
2754 if (HasWeights) {
2755 // PBI: br i1 %x, TrueDest, BB
2756 // BI: br i1 %y, TrueDest, FalseDest
2757 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2758 // FalseWeight for PBI * TrueWeight for BI.
2759 NewWeights.push_back(PredTrueWeight *
2760 (SuccFalseWeight + SuccTrueWeight) +
2761 PredFalseWeight * SuccTrueWeight);
2762 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2763 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2764 }
2765 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2766 PBI->setSuccessor(1, FalseDest);
2767 }
2768 if (NewWeights.size() == 2) {
2769 // Halve the weights if any of them cannot fit in an uint32_t
2770 FitWeights(NewWeights);
2771
2772 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2773 NewWeights.end());
2774 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2775 } else
2776 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2777 } else {
2778 // Update PHI nodes in the common successors.
2779 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2780 ConstantInt *PBI_C = cast<ConstantInt>(
2781 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2782 assert(PBI_C->getType()->isIntegerTy(1))(static_cast <bool> (PBI_C->getType()->isIntegerTy
(1)) ? void (0) : __assert_fail ("PBI_C->getType()->isIntegerTy(1)"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2782, __extension__ __PRETTY_FUNCTION__))
;
2783 Instruction *MergedCond = nullptr;
2784 if (PBI->getSuccessor(0) == TrueDest) {
2785 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2786 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2787 // is false: !PBI_Cond and BI_Value
2788 Instruction *NotCond = cast<Instruction>(
2789 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2790 MergedCond = cast<Instruction>(
2791 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2792 if (PBI_C->isOne())
2793 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2794 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2795 } else {
2796 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2797 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2798 // is false: PBI_Cond and BI_Value
2799 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2800 Instruction::And, PBI->getCondition(), New, "and.cond"));
2801 if (PBI_C->isOne()) {
2802 Instruction *NotCond = cast<Instruction>(
2803 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2804 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2805 Instruction::Or, NotCond, MergedCond, "or.cond"));
2806 }
2807 }
2808 // Update PHI Node.
2809 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2810 MergedCond);
2811 }
2812 // Change PBI from Conditional to Unconditional.
2813 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2814 EraseTerminatorInstAndDCECond(PBI);
2815 PBI = New_PBI;
2816 }
2817
2818 // If BI was a loop latch, it may have had associated loop metadata.
2819 // We need to copy it to the new latch, that is, PBI.
2820 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2821 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2822
2823 // TODO: If BB is reachable from all paths through PredBlock, then we
2824 // could replace PBI's branch probabilities with BI's.
2825
2826 // Copy any debug value intrinsics into the end of PredBlock.
2827 for (Instruction &I : *BB)
2828 if (isa<DbgInfoIntrinsic>(I))
2829 I.clone()->insertBefore(PBI);
2830
2831 return true;
2832 }
2833 return false;
2834}
2835
2836// If there is only one store in BB1 and BB2, return it, otherwise return
2837// nullptr.
2838static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2839 StoreInst *S = nullptr;
2840 for (auto *BB : {BB1, BB2}) {
2841 if (!BB)
2842 continue;
2843 for (auto &I : *BB)
2844 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2845 if (S)
2846 // Multiple stores seen.
2847 return nullptr;
2848 else
2849 S = SI;
2850 }
2851 }
2852 return S;
2853}
2854
2855static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2856 Value *AlternativeV = nullptr) {
2857 // PHI is going to be a PHI node that allows the value V that is defined in
2858 // BB to be referenced in BB's only successor.
2859 //
2860 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2861 // doesn't matter to us what the other operand is (it'll never get used). We
2862 // could just create a new PHI with an undef incoming value, but that could
2863 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2864 // other PHI. So here we directly look for some PHI in BB's successor with V
2865 // as an incoming operand. If we find one, we use it, else we create a new
2866 // one.
2867 //
2868 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2869 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2870 // where OtherBB is the single other predecessor of BB's only successor.
2871 PHINode *PHI = nullptr;
2872 BasicBlock *Succ = BB->getSingleSuccessor();
2873
2874 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2875 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2876 PHI = cast<PHINode>(I);
2877 if (!AlternativeV)
2878 break;
2879
2880 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2)(static_cast <bool> (std::distance(pred_begin(Succ), pred_end
(Succ)) == 2) ? void (0) : __assert_fail ("std::distance(pred_begin(Succ), pred_end(Succ)) == 2"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 2880, __extension__ __PRETTY_FUNCTION__))
;
2881 auto PredI = pred_begin(Succ);
2882 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2883 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2884 break;
2885 PHI = nullptr;
2886 }
2887 if (PHI)
2888 return PHI;
2889
2890 // If V is not an instruction defined in BB, just return it.
2891 if (!AlternativeV &&
2892 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2893 return V;
2894
2895 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2896 PHI->addIncoming(V, BB);
2897 for (BasicBlock *PredBB : predecessors(Succ))
2898 if (PredBB != BB)
2899 PHI->addIncoming(
2900 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2901 return PHI;
2902}
2903
2904static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2905 BasicBlock *QTB, BasicBlock *QFB,
2906 BasicBlock *PostBB, Value *Address,
2907 bool InvertPCond, bool InvertQCond,
2908 const DataLayout &DL) {
2909 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2910 return Operator::getOpcode(&I) == Instruction::BitCast &&
2911 I.getType()->isPointerTy();
2912 };
2913
2914 // If we're not in aggressive mode, we only optimize if we have some
2915 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2916 auto IsWorthwhile = [&](BasicBlock *BB) {
2917 if (!BB)
2918 return true;
2919 // Heuristic: if the block can be if-converted/phi-folded and the
2920 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2921 // thread this store.
2922 unsigned N = 0;
2923 for (auto &I : *BB) {
2924 // Cheap instructions viable for folding.
2925 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2926 isa<StoreInst>(I))
2927 ++N;
2928 // Free instructions.
2929 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2930 IsaBitcastOfPointerType(I))
2931 continue;
2932 else
2933 return false;
2934 }
2935 // The store we want to merge is counted in N, so add 1 to make sure
2936 // we're counting the instructions that would be left.
2937 return N <= (PHINodeFoldingThreshold + 1);
2938 };
2939
2940 if (!MergeCondStoresAggressively &&
2941 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2942 !IsWorthwhile(QFB)))
2943 return false;
2944
2945 // For every pointer, there must be exactly two stores, one coming from
2946 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2947 // store (to any address) in PTB,PFB or QTB,QFB.
2948 // FIXME: We could relax this restriction with a bit more work and performance
2949 // testing.
2950 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2951 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2952 if (!PStore || !QStore)
2953 return false;
2954
2955 // Now check the stores are compatible.
2956 if (!QStore->isUnordered() || !PStore->isUnordered())
2957 return false;
2958
2959 // Check that sinking the store won't cause program behavior changes. Sinking
2960 // the store out of the Q blocks won't change any behavior as we're sinking
2961 // from a block to its unconditional successor. But we're moving a store from
2962 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2963 // So we need to check that there are no aliasing loads or stores in
2964 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2965 // operations between PStore and the end of its parent block.
2966 //
2967 // The ideal way to do this is to query AliasAnalysis, but we don't
2968 // preserve AA currently so that is dangerous. Be super safe and just
2969 // check there are no other memory operations at all.
2970 for (auto &I : *QFB->getSinglePredecessor())
2971 if (I.mayReadOrWriteMemory())
2972 return false;
2973 for (auto &I : *QFB)
2974 if (&I != QStore && I.mayReadOrWriteMemory())
2975 return false;
2976 if (QTB)
2977 for (auto &I : *QTB)
2978 if (&I != QStore && I.mayReadOrWriteMemory())
2979 return false;
2980 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2981 I != E; ++I)
2982 if (&*I != PStore && I->mayReadOrWriteMemory())
2983 return false;
2984
2985 // If PostBB has more than two predecessors, we need to split it so we can
2986 // sink the store.
2987 if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2988 // We know that QFB's only successor is PostBB. And QFB has a single
2989 // predecessor. If QTB exists, then its only successor is also PostBB.
2990 // If QTB does not exist, then QFB's only predecessor has a conditional
2991 // branch to QFB and PostBB.
2992 BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
2993 BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
2994 "condstore.split");
2995 if (!NewBB)
2996 return false;
2997 PostBB = NewBB;
2998 }
2999
3000 // OK, we're going to sink the stores to PostBB. The store has to be
3001 // conditional though, so first create the predicate.
3002 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3003 ->getCondition();
3004 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3005 ->getCondition();
3006
3007 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
3008 PStore->getParent());
3009 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
3010 QStore->getParent(), PPHI);
3011
3012 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3013
3014 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3015 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3016
3017 if (InvertPCond)
3018 PPred = QB.CreateNot(PPred);
3019 if (InvertQCond)
3020 QPred = QB.CreateNot(QPred);
3021 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3022
3023 auto *T =
3024 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3025 QB.SetInsertPoint(T);
3026 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3027 AAMDNodes AAMD;
3028 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3029 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3030 SI->setAAMetadata(AAMD);
3031 unsigned PAlignment = PStore->getAlignment();
3032 unsigned QAlignment = QStore->getAlignment();
3033 unsigned TypeAlignment =
3034 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3035 unsigned MinAlignment;
3036 unsigned MaxAlignment;
3037 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3038 // Choose the minimum alignment. If we could prove both stores execute, we
3039 // could use biggest one. In this case, though, we only know that one of the
3040 // stores executes. And we don't know it's safe to take the alignment from a
3041 // store that doesn't execute.
3042 if (MinAlignment != 0) {
3043 // Choose the minimum of all non-zero alignments.
3044 SI->setAlignment(MinAlignment);
3045 } else if (MaxAlignment != 0) {
3046 // Choose the minimal alignment between the non-zero alignment and the ABI
3047 // default alignment for the type of the stored value.
3048 SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3049 } else {
3050 // If both alignments are zero, use ABI default alignment for the type of
3051 // the stored value.
3052 SI->setAlignment(TypeAlignment);
3053 }
3054
3055 QStore->eraseFromParent();
3056 PStore->eraseFromParent();
3057
3058 return true;
3059}
3060
3061static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3062 const DataLayout &DL) {
3063 // The intention here is to find diamonds or triangles (see below) where each
3064 // conditional block contains a store to the same address. Both of these
3065 // stores are conditional, so they can't be unconditionally sunk. But it may
3066 // be profitable to speculatively sink the stores into one merged store at the
3067 // end, and predicate the merged store on the union of the two conditions of
3068 // PBI and QBI.
3069 //
3070 // This can reduce the number of stores executed if both of the conditions are
3071 // true, and can allow the blocks to become small enough to be if-converted.
3072 // This optimization will also chain, so that ladders of test-and-set
3073 // sequences can be if-converted away.
3074 //
3075 // We only deal with simple diamonds or triangles:
3076 //
3077 // PBI or PBI or a combination of the two
3078 // / \ | \
3079 // PTB PFB | PFB
3080 // \ / | /
3081 // QBI QBI
3082 // / \ | \
3083 // QTB QFB | QFB
3084 // \ / | /
3085 // PostBB PostBB
3086 //
3087 // We model triangles as a type of diamond with a nullptr "true" block.
3088 // Triangles are canonicalized so that the fallthrough edge is represented by
3089 // a true condition, as in the diagram above.
3090 BasicBlock *PTB = PBI->getSuccessor(0);
3091 BasicBlock *PFB = PBI->getSuccessor(1);
3092 BasicBlock *QTB = QBI->getSuccessor(0);
3093 BasicBlock *QFB = QBI->getSuccessor(1);
3094 BasicBlock *PostBB = QFB->getSingleSuccessor();
3095
3096 // Make sure we have a good guess for PostBB. If QTB's only successor is
3097 // QFB, then QFB is a better PostBB.
3098 if (QTB->getSingleSuccessor() == QFB)
3099 PostBB = QFB;
3100
3101 // If we couldn't find a good PostBB, stop.
3102 if (!PostBB)
3103 return false;
3104
3105 bool InvertPCond = false, InvertQCond = false;
3106 // Canonicalize fallthroughs to the true branches.
3107 if (PFB == QBI->getParent()) {
3108 std::swap(PFB, PTB);
3109 InvertPCond = true;
3110 }
3111 if (QFB == PostBB) {
3112 std::swap(QFB, QTB);
3113 InvertQCond = true;
3114 }
3115
3116 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3117 // and QFB may not. Model fallthroughs as a nullptr block.
3118 if (PTB == QBI->getParent())
3119 PTB = nullptr;
3120 if (QTB == PostBB)
3121 QTB = nullptr;
3122
3123 // Legality bailouts. We must have at least the non-fallthrough blocks and
3124 // the post-dominating block, and the non-fallthroughs must only have one
3125 // predecessor.
3126 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3127 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3128 };
3129 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3130 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3131 return false;
3132 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3133 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3134 return false;
3135 if (!QBI->getParent()->hasNUses(2))
3136 return false;
3137
3138 // OK, this is a sequence of two diamonds or triangles.
3139 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3140 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3141 for (auto *BB : {PTB, PFB}) {
3142 if (!BB)
3143 continue;
3144 for (auto &I : *BB)
3145 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3146 PStoreAddresses.insert(SI->getPointerOperand());
3147 }
3148 for (auto *BB : {QTB, QFB}) {
3149 if (!BB)
3150 continue;
3151 for (auto &I : *BB)
3152 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3153 QStoreAddresses.insert(SI->getPointerOperand());
3154 }
3155
3156 set_intersect(PStoreAddresses, QStoreAddresses);
3157 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3158 // clear what it contains.
3159 auto &CommonAddresses = PStoreAddresses;
3160
3161 bool Changed = false;
3162 for (auto *Address : CommonAddresses)
3163 Changed |= mergeConditionalStoreToAddress(
3164 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3165 return Changed;
3166}
3167
3168/// If we have a conditional branch as a predecessor of another block,
3169/// this function tries to simplify it. We know
3170/// that PBI and BI are both conditional branches, and BI is in one of the
3171/// successor blocks of PBI - PBI branches to BI.
3172static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3173 const DataLayout &DL) {
3174 assert(PBI->isConditional() && BI->isConditional())(static_cast <bool> (PBI->isConditional() &&
BI->isConditional()) ? void (0) : __assert_fail ("PBI->isConditional() && BI->isConditional()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3174, __extension__ __PRETTY_FUNCTION__))
;
3175 BasicBlock *BB = BI->getParent();
3176
3177 // If this block ends with a branch instruction, and if there is a
3178 // predecessor that ends on a branch of the same condition, make
3179 // this conditional branch redundant.
3180 if (PBI->getCondition() == BI->getCondition() &&
3181 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3182 // Okay, the outcome of this conditional branch is statically
3183 // knowable. If this block had a single pred, handle specially.
3184 if (BB->getSinglePredecessor()) {
3185 // Turn this into a branch on constant.
3186 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3187 BI->setCondition(
3188 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3189 return true; // Nuke the branch on constant.
3190 }
3191
3192 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3193 // in the constant and simplify the block result. Subsequent passes of
3194 // simplifycfg will thread the block.
3195 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3196 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3197 PHINode *NewPN = PHINode::Create(
3198 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3199 BI->getCondition()->getName() + ".pr", &BB->front());
3200 // Okay, we're going to insert the PHI node. Since PBI is not the only
3201 // predecessor, compute the PHI'd conditional value for all of the preds.
3202 // Any predecessor where the condition is not computable we keep symbolic.
3203 for (pred_iterator PI = PB; PI != PE; ++PI) {
3204 BasicBlock *P = *PI;
3205 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3206 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3207 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3208 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3209 NewPN->addIncoming(
3210 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3211 P);
3212 } else {
3213 NewPN->addIncoming(BI->getCondition(), P);
3214 }
3215 }
3216
3217 BI->setCondition(NewPN);
3218 return true;
3219 }
3220 }
3221
3222 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3223 if (CE->canTrap())
3224 return false;
3225
3226 // If both branches are conditional and both contain stores to the same
3227 // address, remove the stores from the conditionals and create a conditional
3228 // merged store at the end.
3229 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3230 return true;
3231
3232 // If this is a conditional branch in an empty block, and if any
3233 // predecessors are a conditional branch to one of our destinations,
3234 // fold the conditions into logical ops and one cond br.
3235 BasicBlock::iterator BBI = BB->begin();
3236 // Ignore dbg intrinsics.
3237 while (isa<DbgInfoIntrinsic>(BBI))
3238 ++BBI;
3239 if (&*BBI != BI)
3240 return false;
3241
3242 int PBIOp, BIOp;
3243 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3244 PBIOp = 0;
3245 BIOp = 0;
3246 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3247 PBIOp = 0;
3248 BIOp = 1;
3249 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3250 PBIOp = 1;
3251 BIOp = 0;
3252 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3253 PBIOp = 1;
3254 BIOp = 1;
3255 } else {
3256 return false;
3257 }
3258
3259 // Check to make sure that the other destination of this branch
3260 // isn't BB itself. If so, this is an infinite loop that will
3261 // keep getting unwound.
3262 if (PBI->getSuccessor(PBIOp) == BB)
3263 return false;
3264
3265 // Do not perform this transformation if it would require
3266 // insertion of a large number of select instructions. For targets
3267 // without predication/cmovs, this is a big pessimization.
3268
3269 // Also do not perform this transformation if any phi node in the common
3270 // destination block can trap when reached by BB or PBB (PR17073). In that
3271 // case, it would be unsafe to hoist the operation into a select instruction.
3272
3273 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3274 unsigned NumPhis = 0;
3275 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3276 ++II, ++NumPhis) {
3277 if (NumPhis > 2) // Disable this xform.
3278 return false;
3279
3280 PHINode *PN = cast<PHINode>(II);
3281 Value *BIV = PN->getIncomingValueForBlock(BB);
3282 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3283 if (CE->canTrap())
3284 return false;
3285
3286 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3287 Value *PBIV = PN->getIncomingValue(PBBIdx);
3288 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3289 if (CE->canTrap())
3290 return false;
3291 }
3292
3293 // Finally, if everything is ok, fold the branches to logical ops.
3294 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3295
3296 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
3297 << "AND: " << *BI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING BRs:" << *PBI
->getParent() << "AND: " << *BI->getParent(
); } } while (false)
;
3298
3299 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3300 // branch in it, where one edge (OtherDest) goes back to itself but the other
3301 // exits. We don't *know* that the program avoids the infinite loop
3302 // (even though that seems likely). If we do this xform naively, we'll end up
3303 // recursively unpeeling the loop. Since we know that (after the xform is
3304 // done) that the block *is* infinite if reached, we just make it an obviously
3305 // infinite loop with no cond branch.
3306 if (OtherDest == BB) {
3307 // Insert it at the end of the function, because it's either code,
3308 // or it won't matter if it's hot. :)
3309 BasicBlock *InfLoopBlock =
3310 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3311 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3312 OtherDest = InfLoopBlock;
3313 }
3314
3315 DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3316
3317 // BI may have other predecessors. Because of this, we leave
3318 // it alone, but modify PBI.
3319
3320 // Make sure we get to CommonDest on True&True directions.
3321 Value *PBICond = PBI->getCondition();
3322 IRBuilder<NoFolder> Builder(PBI);
3323 if (PBIOp)
3324 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3325
3326 Value *BICond = BI->getCondition();
3327 if (BIOp)
3328 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3329
3330 // Merge the conditions.
3331 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3332
3333 // Modify PBI to branch on the new condition to the new dests.
3334 PBI->setCondition(Cond);
3335 PBI->setSuccessor(0, CommonDest);
3336 PBI->setSuccessor(1, OtherDest);
3337
3338 // Update branch weight for PBI.
3339 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3340 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3341 bool HasWeights =
3342 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3343 SuccTrueWeight, SuccFalseWeight);
3344 if (HasWeights) {
3345 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3346 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3347 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3348 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3349 // The weight to CommonDest should be PredCommon * SuccTotal +
3350 // PredOther * SuccCommon.
3351 // The weight to OtherDest should be PredOther * SuccOther.
3352 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3353 PredOther * SuccCommon,
3354 PredOther * SuccOther};
3355 // Halve the weights if any of them cannot fit in an uint32_t
3356 FitWeights(NewWeights);
3357
3358 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3359 }
3360
3361 // OtherDest may have phi nodes. If so, add an entry from PBI's
3362 // block that are identical to the entries for BI's block.
3363 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3364
3365 // We know that the CommonDest already had an edge from PBI to
3366 // it. If it has PHIs though, the PHIs may have different
3367 // entries for BB and PBI's BB. If so, insert a select to make
3368 // them agree.
3369 for (PHINode &PN : CommonDest->phis()) {
3370 Value *BIV = PN.getIncomingValueForBlock(BB);
3371 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3372 Value *PBIV = PN.getIncomingValue(PBBIdx);
3373 if (BIV != PBIV) {
3374 // Insert a select in PBI to pick the right value.
3375 SelectInst *NV = cast<SelectInst>(
3376 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3377 PN.setIncomingValue(PBBIdx, NV);
3378 // Although the select has the same condition as PBI, the original branch
3379 // weights for PBI do not apply to the new select because the select's
3380 // 'logical' edges are incoming edges of the phi that is eliminated, not
3381 // the outgoing edges of PBI.
3382 if (HasWeights) {
3383 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3384 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3385 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3386 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3387 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3388 // The weight to PredOtherDest should be PredOther * SuccCommon.
3389 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3390 PredOther * SuccCommon};
3391
3392 FitWeights(NewWeights);
3393
3394 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3395 }
3396 }
3397 }
3398
3399 DEBUG(dbgs() << "INTO: " << *PBI->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "INTO: " << *PBI->
getParent(); } } while (false)
;
3400 DEBUG(dbgs() << *PBI->getParent()->getParent())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << *PBI->getParent()->getParent
(); } } while (false)
;
3401
3402 // This basic block is probably dead. We know it has at least
3403 // one fewer predecessor.
3404 return true;
3405}
3406
3407// Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3408// true or to FalseBB if Cond is false.
3409// Takes care of updating the successors and removing the old terminator.
3410// Also makes sure not to introduce new successors by assuming that edges to
3411// non-successor TrueBBs and FalseBBs aren't reachable.
3412static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3413 BasicBlock *TrueBB, BasicBlock *FalseBB,
3414 uint32_t TrueWeight,
3415 uint32_t FalseWeight) {
3416 // Remove any superfluous successor edges from the CFG.
3417 // First, figure out which successors to preserve.
3418 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3419 // successor.
3420 BasicBlock *KeepEdge1 = TrueBB;
3421 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3422
3423 // Then remove the rest.
3424 for (BasicBlock *Succ : OldTerm->successors()) {
3425 // Make sure only to keep exactly one copy of each edge.
3426 if (Succ == KeepEdge1)
3427 KeepEdge1 = nullptr;
3428 else if (Succ == KeepEdge2)
3429 KeepEdge2 = nullptr;
3430 else
3431 Succ->removePredecessor(OldTerm->getParent(),
3432 /*DontDeleteUselessPHIs=*/true);
3433 }
3434
3435 IRBuilder<> Builder(OldTerm);
3436 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3437
3438 // Insert an appropriate new terminator.
3439 if (!KeepEdge1 && !KeepEdge2) {
3440 if (TrueBB == FalseBB)
3441 // We were only looking for one successor, and it was present.
3442 // Create an unconditional branch to it.
3443 Builder.CreateBr(TrueBB);
3444 else {
3445 // We found both of the successors we were looking for.
3446 // Create a conditional branch sharing the condition of the select.
3447 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3448 if (TrueWeight != FalseWeight)
3449 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3450 }
3451 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3452 // Neither of the selected blocks were successors, so this
3453 // terminator must be unreachable.
3454 new UnreachableInst(OldTerm->getContext(), OldTerm);
3455 } else {
3456 // One of the selected values was a successor, but the other wasn't.
3457 // Insert an unconditional branch to the one that was found;
3458 // the edge to the one that wasn't must be unreachable.
3459 if (!KeepEdge1)
3460 // Only TrueBB was found.
3461 Builder.CreateBr(TrueBB);
3462 else
3463 // Only FalseBB was found.
3464 Builder.CreateBr(FalseBB);
3465 }
3466
3467 EraseTerminatorInstAndDCECond(OldTerm);
3468 return true;
3469}
3470
3471// Replaces
3472// (switch (select cond, X, Y)) on constant X, Y
3473// with a branch - conditional if X and Y lead to distinct BBs,
3474// unconditional otherwise.
3475static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3476 // Check for constant integer values in the select.
3477 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3478 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3479 if (!TrueVal || !FalseVal)
3480 return false;
3481
3482 // Find the relevant condition and destinations.
3483 Value *Condition = Select->getCondition();
3484 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3485 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3486
3487 // Get weight for TrueBB and FalseBB.
3488 uint32_t TrueWeight = 0, FalseWeight = 0;
3489 SmallVector<uint64_t, 8> Weights;
3490 bool HasWeights = HasBranchWeights(SI);
3491 if (HasWeights) {
3492 GetBranchWeights(SI, Weights);
3493 if (Weights.size() == 1 + SI->getNumCases()) {
3494 TrueWeight =
3495 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3496 FalseWeight =
3497 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3498 }
3499 }
3500
3501 // Perform the actual simplification.
3502 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3503 FalseWeight);
3504}
3505
3506// Replaces
3507// (indirectbr (select cond, blockaddress(@fn, BlockA),
3508// blockaddress(@fn, BlockB)))
3509// with
3510// (br cond, BlockA, BlockB).
3511static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3512 // Check that both operands of the select are block addresses.
3513 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3514 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3515 if (!TBA || !FBA)
3516 return false;
3517
3518 // Extract the actual blocks.
3519 BasicBlock *TrueBB = TBA->getBasicBlock();
3520 BasicBlock *FalseBB = FBA->getBasicBlock();
3521
3522 // Perform the actual simplification.
3523 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3524 0);
3525}
3526
3527/// This is called when we find an icmp instruction
3528/// (a seteq/setne with a constant) as the only instruction in a
3529/// block that ends with an uncond branch. We are looking for a very specific
3530/// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3531/// this case, we merge the first two "or's of icmp" into a switch, but then the
3532/// default value goes to an uncond block with a seteq in it, we get something
3533/// like:
3534///
3535/// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3536/// DEFAULT:
3537/// %tmp = icmp eq i8 %A, 92
3538/// br label %end
3539/// end:
3540/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3541///
3542/// We prefer to split the edge to 'end' so that there is a true/false entry to
3543/// the PHI, merging the third icmp into the switch.
3544static bool tryToSimplifyUncondBranchWithICmpInIt(
3545 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3546 const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3547 BasicBlock *BB = ICI->getParent();
3548
3549 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3550 // complex.
3551 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
1
Assuming the condition is false
2
Taking false branch
3552 return false;
3553
3554 Value *V = ICI->getOperand(0);
3555 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3556
3557 // The pattern we're looking for is where our only predecessor is a switch on
3558 // 'V' and this block is the default case for the switch. In this case we can
3559 // fold the compared value into the switch to simplify things.
3560 BasicBlock *Pred = BB->getSinglePredecessor();
3561 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3
Assuming 'Pred' is non-null
4
Taking false branch
3562 return false;
3563
3564 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3565 if (SI->getCondition() != V)
5
Assuming the condition is false
6
Taking false branch
3566 return false;
3567
3568 // If BB is reachable on a non-default case, then we simply know the value of
3569 // V in this block. Substitute it and constant fold the icmp instruction
3570 // away.
3571 if (SI->getDefaultDest() != BB) {
7
Assuming the condition is false
8
Taking false branch
3572 ConstantInt *VVal = SI->findCaseDest(BB);
3573 assert(VVal && "Should have a unique destination value")(static_cast <bool> (VVal && "Should have a unique destination value"
) ? void (0) : __assert_fail ("VVal && \"Should have a unique destination value\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3573, __extension__ __PRETTY_FUNCTION__))
;
3574 ICI->setOperand(0, VVal);
3575
3576 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3577 ICI->replaceAllUsesWith(V);
3578 ICI->eraseFromParent();
3579 }
3580 // BB is now empty, so it is likely to simplify away.
3581 return simplifyCFG(BB, TTI, Options) | true;
3582 }
3583
3584 // Ok, the block is reachable from the default dest. If the constant we're
3585 // comparing exists in one of the other edges, then we can constant fold ICI
3586 // and zap it.
3587 if (SI->findCaseValue(Cst) != SI->case_default()) {
9
Assuming the condition is true
10
Taking true branch
3588 Value *V;
3589 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
11
Assuming the condition is false
12
Taking false branch
3590 V = ConstantInt::getFalse(BB->getContext());
3591 else
3592 V = ConstantInt::getTrue(BB->getContext());
3593
3594 ICI->replaceAllUsesWith(V);
3595 ICI->eraseFromParent();
3596 // BB is now empty, so it is likely to simplify away.
3597 return simplifyCFG(BB, TTI, Options) | true;
13
Calling 'simplifyCFG'
3598 }
3599
3600 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3601 // the block.
3602 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3603 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3604 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3605 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3606 return false;
3607
3608 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3609 // true in the PHI.
3610 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3611 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3612
3613 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3614 std::swap(DefaultCst, NewCst);
3615
3616 // Replace ICI (which is used by the PHI for the default value) with true or
3617 // false depending on if it is EQ or NE.
3618 ICI->replaceAllUsesWith(DefaultCst);
3619 ICI->eraseFromParent();
3620
3621 // Okay, the switch goes to this block on a default value. Add an edge from
3622 // the switch to the merge point on the compared value.
3623 BasicBlock *NewBB =
3624 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3625 SmallVector<uint64_t, 8> Weights;
3626 bool HasWeights = HasBranchWeights(SI);
3627 if (HasWeights) {
3628 GetBranchWeights(SI, Weights);
3629 if (Weights.size() == 1 + SI->getNumCases()) {
3630 // Split weight for default case to case for "Cst".
3631 Weights[0] = (Weights[0] + 1) >> 1;
3632 Weights.push_back(Weights[0]);
3633
3634 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3635 setBranchWeights(SI, MDWeights);
3636 }
3637 }
3638 SI->addCase(Cst, NewBB);
3639
3640 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3641 Builder.SetInsertPoint(NewBB);
3642 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3643 Builder.CreateBr(SuccBlock);
3644 PHIUse->addIncoming(NewCst, NewBB);
3645 return true;
3646}
3647
3648/// The specified branch is a conditional branch.
3649/// Check to see if it is branching on an or/and chain of icmp instructions, and
3650/// fold it into a switch instruction if so.
3651static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3652 const DataLayout &DL) {
3653 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3654 if (!Cond)
3655 return false;
3656
3657 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3658 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3659 // 'setne's and'ed together, collect them.
3660
3661 // Try to gather values from a chain of and/or to be turned into a switch
3662 ConstantComparesGatherer ConstantCompare(Cond, DL);
3663 // Unpack the result
3664 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3665 Value *CompVal = ConstantCompare.CompValue;
3666 unsigned UsedICmps = ConstantCompare.UsedICmps;
3667 Value *ExtraCase = ConstantCompare.Extra;
3668
3669 // If we didn't have a multiply compared value, fail.
3670 if (!CompVal)
3671 return false;
3672
3673 // Avoid turning single icmps into a switch.
3674 if (UsedICmps <= 1)
3675 return false;
3676
3677 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3678
3679 // There might be duplicate constants in the list, which the switch
3680 // instruction can't handle, remove them now.
3681 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3682 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3683
3684 // If Extra was used, we require at least two switch values to do the
3685 // transformation. A switch with one value is just a conditional branch.
3686 if (ExtraCase && Values.size() < 2)
3687 return false;
3688
3689 // TODO: Preserve branch weight metadata, similarly to how
3690 // FoldValueComparisonIntoPredecessors preserves it.
3691
3692 // Figure out which block is which destination.
3693 BasicBlock *DefaultBB = BI->getSuccessor(1);
3694 BasicBlock *EdgeBB = BI->getSuccessor(0);
3695 if (!TrueWhenEqual)
3696 std::swap(DefaultBB, EdgeBB);
3697
3698 BasicBlock *BB = BI->getParent();
3699
3700 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)
3701 << " 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)
3702 << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Converting 'icmp' chain with "
<< Values.size() << " cases into SWITCH. BB is:\n"
<< *BB; } } while (false)
;
3703
3704 // If there are any extra values that couldn't be folded into the switch
3705 // then we evaluate them with an explicit branch first. Split the block
3706 // right before the condbr to handle it.
3707 if (ExtraCase) {
3708 BasicBlock *NewBB =
3709 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3710 // Remove the uncond branch added to the old block.
3711 TerminatorInst *OldTI = BB->getTerminator();
3712 Builder.SetInsertPoint(OldTI);
3713
3714 if (TrueWhenEqual)
3715 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3716 else
3717 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3718
3719 OldTI->eraseFromParent();
3720
3721 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3722 // for the edge we just added.
3723 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3724
3725 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCasedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
3726 << "\nEXTRABB = " << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << " ** 'icmp' chain unhandled condition: "
<< *ExtraCase << "\nEXTRABB = " << *BB; } }
while (false)
;
3727 BB = NewBB;
3728 }
3729
3730 Builder.SetInsertPoint(BI);
3731 // Convert pointer to int before we switch.
3732 if (CompVal->getType()->isPointerTy()) {
3733 CompVal = Builder.CreatePtrToInt(
3734 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3735 }
3736
3737 // Create the new switch instruction now.
3738 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3739
3740 // Add all of the 'cases' to the switch instruction.
3741 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3742 New->addCase(Values[i], EdgeBB);
3743
3744 // We added edges from PI to the EdgeBB. As such, if there were any
3745 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3746 // the number of edges added.
3747 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3748 PHINode *PN = cast<PHINode>(BBI);
3749 Value *InVal = PN->getIncomingValueForBlock(BB);
3750 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3751 PN->addIncoming(InVal, BB);
3752 }
3753
3754 // Erase the old branch instruction.
3755 EraseTerminatorInstAndDCECond(BI);
3756
3757 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)
;
3758 return true;
3759}
3760
3761bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3762 if (isa<PHINode>(RI->getValue()))
3763 return SimplifyCommonResume(RI);
3764 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3765 RI->getValue() == RI->getParent()->getFirstNonPHI())
3766 // The resume must unwind the exception that caused control to branch here.
3767 return SimplifySingleResume(RI);
3768
3769 return false;
3770}
3771
3772// Simplify resume that is shared by several landing pads (phi of landing pad).
3773bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3774 BasicBlock *BB = RI->getParent();
3775
3776 // Check that there are no other instructions except for debug intrinsics
3777 // between the phi of landing pads (RI->getValue()) and resume instruction.
3778 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3779 E = RI->getIterator();
3780 while (++I != E)
3781 if (!isa<DbgInfoIntrinsic>(I))
3782 return false;
3783
3784 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3785 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3786
3787 // Check incoming blocks to see if any of them are trivial.
3788 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3789 Idx++) {
3790 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3791 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3792
3793 // If the block has other successors, we can not delete it because
3794 // it has other dependents.
3795 if (IncomingBB->getUniqueSuccessor() != BB)
3796 continue;
3797
3798 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3799 // Not the landing pad that caused the control to branch here.
3800 if (IncomingValue != LandingPad)
3801 continue;
3802
3803 bool isTrivial = true;
3804
3805 I = IncomingBB->getFirstNonPHI()->getIterator();
3806 E = IncomingBB->getTerminator()->getIterator();
3807 while (++I != E)
3808 if (!isa<DbgInfoIntrinsic>(I)) {
3809 isTrivial = false;
3810 break;
3811 }
3812
3813 if (isTrivial)
3814 TrivialUnwindBlocks.insert(IncomingBB);
3815 }
3816
3817 // If no trivial unwind blocks, don't do any simplifications.
3818 if (TrivialUnwindBlocks.empty())
3819 return false;
3820
3821 // Turn all invokes that unwind here into calls.
3822 for (auto *TrivialBB : TrivialUnwindBlocks) {
3823 // Blocks that will be simplified should be removed from the phi node.
3824 // Note there could be multiple edges to the resume block, and we need
3825 // to remove them all.
3826 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3827 BB->removePredecessor(TrivialBB, true);
3828
3829 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3830 PI != PE;) {
3831 BasicBlock *Pred = *PI++;
3832 removeUnwindEdge(Pred);
3833 }
3834
3835 // In each SimplifyCFG run, only the current processed block can be erased.
3836 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3837 // of erasing TrivialBB, we only remove the branch to the common resume
3838 // block so that we can later erase the resume block since it has no
3839 // predecessors.
3840 TrivialBB->getTerminator()->eraseFromParent();
3841 new UnreachableInst(RI->getContext(), TrivialBB);
3842 }
3843
3844 // Delete the resume block if all its predecessors have been removed.
3845 if (pred_empty(BB))
3846 BB->eraseFromParent();
3847
3848 return !TrivialUnwindBlocks.empty();
3849}
3850
3851// Simplify resume that is only used by a single (non-phi) landing pad.
3852bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3853 BasicBlock *BB = RI->getParent();
3854 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3855 assert(RI->getValue() == LPInst &&(static_cast <bool> (RI->getValue() == LPInst &&
"Resume must unwind the exception that caused control to here"
) ? void (0) : __assert_fail ("RI->getValue() == LPInst && \"Resume must unwind the exception that caused control to here\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3856, __extension__ __PRETTY_FUNCTION__))
3856 "Resume must unwind the exception that caused control to here")(static_cast <bool> (RI->getValue() == LPInst &&
"Resume must unwind the exception that caused control to here"
) ? void (0) : __assert_fail ("RI->getValue() == LPInst && \"Resume must unwind the exception that caused control to here\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3856, __extension__ __PRETTY_FUNCTION__))
;
3857
3858 // Check that there are no other instructions except for debug intrinsics.
3859 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3860 while (++I != E)
3861 if (!isa<DbgInfoIntrinsic>(I))
3862 return false;
3863
3864 // Turn all invokes that unwind here into calls and delete the basic block.
3865 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3866 BasicBlock *Pred = *PI++;
3867 removeUnwindEdge(Pred);
3868 }
3869
3870 // The landingpad is now unreachable. Zap it.
3871 BB->eraseFromParent();
3872 if (LoopHeaders)
3873 LoopHeaders->erase(BB);
3874 return true;
3875}
3876
3877static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3878 // If this is a trivial cleanup pad that executes no instructions, it can be
3879 // eliminated. If the cleanup pad continues to the caller, any predecessor
3880 // that is an EH pad will be updated to continue to the caller and any
3881 // predecessor that terminates with an invoke instruction will have its invoke
3882 // instruction converted to a call instruction. If the cleanup pad being
3883 // simplified does not continue to the caller, each predecessor will be
3884 // updated to continue to the unwind destination of the cleanup pad being
3885 // simplified.
3886 BasicBlock *BB = RI->getParent();
3887 CleanupPadInst *CPInst = RI->getCleanupPad();
3888 if (CPInst->getParent() != BB)
3889 // This isn't an empty cleanup.
3890 return false;
3891
3892 // We cannot kill the pad if it has multiple uses. This typically arises
3893 // from unreachable basic blocks.
3894 if (!CPInst->hasOneUse())
3895 return false;
3896
3897 // Check that there are no other instructions except for benign intrinsics.
3898 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3899 while (++I != E) {
3900 auto *II = dyn_cast<IntrinsicInst>(I);
3901 if (!II)
3902 return false;
3903
3904 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3905 switch (IntrinsicID) {
3906 case Intrinsic::dbg_declare:
3907 case Intrinsic::dbg_value:
3908 case Intrinsic::lifetime_end:
3909 break;
3910 default:
3911 return false;
3912 }
3913 }
3914
3915 // If the cleanup return we are simplifying unwinds to the caller, this will
3916 // set UnwindDest to nullptr.
3917 BasicBlock *UnwindDest = RI->getUnwindDest();
3918 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3919
3920 // We're about to remove BB from the control flow. Before we do, sink any
3921 // PHINodes into the unwind destination. Doing this before changing the
3922 // control flow avoids some potentially slow checks, since we can currently
3923 // be certain that UnwindDest and BB have no common predecessors (since they
3924 // are both EH pads).
3925 if (UnwindDest) {
3926 // First, go through the PHI nodes in UnwindDest and update any nodes that
3927 // reference the block we are removing
3928 for (BasicBlock::iterator I = UnwindDest->begin(),
3929 IE = DestEHPad->getIterator();
3930 I != IE; ++I) {
3931 PHINode *DestPN = cast<PHINode>(I);
3932
3933 int Idx = DestPN->getBasicBlockIndex(BB);
3934 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3935 assert(Idx != -1)(static_cast <bool> (Idx != -1) ? void (0) : __assert_fail
("Idx != -1", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 3935, __extension__ __PRETTY_FUNCTION__))
;
3936 // This PHI node has an incoming value that corresponds to a control
3937 // path through the cleanup pad we are removing. If the incoming
3938 // value is in the cleanup pad, it must be a PHINode (because we
3939 // verified above that the block is otherwise empty). Otherwise, the
3940 // value is either a constant or a value that dominates the cleanup
3941 // pad being removed.
3942 //
3943 // Because BB and UnwindDest are both EH pads, all of their
3944 // predecessors must unwind to these blocks, and since no instruction
3945 // can have multiple unwind destinations, there will be no overlap in
3946 // incoming blocks between SrcPN and DestPN.
3947 Value *SrcVal = DestPN->getIncomingValue(Idx);
3948 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3949
3950 // Remove the entry for the block we are deleting.
3951 DestPN->removeIncomingValue(Idx, false);
3952
3953 if (SrcPN && SrcPN->getParent() == BB) {
3954 // If the incoming value was a PHI node in the cleanup pad we are
3955 // removing, we need to merge that PHI node's incoming values into
3956 // DestPN.
3957 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3958 SrcIdx != SrcE; ++SrcIdx) {
3959 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3960 SrcPN->getIncomingBlock(SrcIdx));
3961 }
3962 } else {
3963 // Otherwise, the incoming value came from above BB and
3964 // so we can just reuse it. We must associate all of BB's
3965 // predecessors with this value.
3966 for (auto *pred : predecessors(BB)) {
3967 DestPN->addIncoming(SrcVal, pred);
3968 }
3969 }
3970 }
3971
3972 // Sink any remaining PHI nodes directly into UnwindDest.
3973 Instruction *InsertPt = DestEHPad;
3974 for (BasicBlock::iterator I = BB->begin(),
3975 IE = BB->getFirstNonPHI()->getIterator();
3976 I != IE;) {
3977 // The iterator must be incremented here because the instructions are
3978 // being moved to another block.
3979 PHINode *PN = cast<PHINode>(I++);
3980 if (PN->use_empty())
3981 // If the PHI node has no uses, just leave it. It will be erased
3982 // when we erase BB below.
3983 continue;
3984
3985 // Otherwise, sink this PHI node into UnwindDest.
3986 // Any predecessors to UnwindDest which are not already represented
3987 // must be back edges which inherit the value from the path through
3988 // BB. In this case, the PHI value must reference itself.
3989 for (auto *pred : predecessors(UnwindDest))
3990 if (pred != BB)
3991 PN->addIncoming(PN, pred);
3992 PN->moveBefore(InsertPt);
3993 }
3994 }
3995
3996 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3997 // The iterator must be updated here because we are removing this pred.
3998 BasicBlock *PredBB = *PI++;
3999 if (UnwindDest == nullptr) {
4000 removeUnwindEdge(PredBB);
4001 } else {
4002 TerminatorInst *TI = PredBB->getTerminator();
4003 TI->replaceUsesOfWith(BB, UnwindDest);
4004 }
4005 }
4006
4007 // The cleanup pad is now unreachable. Zap it.
4008 BB->eraseFromParent();
4009 return true;
4010}
4011
4012// Try to merge two cleanuppads together.
4013static bool mergeCleanupPad(CleanupReturnInst *RI) {
4014 // Skip any cleanuprets which unwind to caller, there is nothing to merge
4015 // with.
4016 BasicBlock *UnwindDest = RI->getUnwindDest();
4017 if (!UnwindDest)
4018 return false;
4019
4020 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4021 // be safe to merge without code duplication.
4022 if (UnwindDest->getSinglePredecessor() != RI->getParent())
4023 return false;
4024
4025 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4026 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4027 if (!SuccessorCleanupPad)
4028 return false;
4029
4030 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4031 // Replace any uses of the successor cleanupad with the predecessor pad
4032 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4033 // funclet bundle operands.
4034 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4035 // Remove the old cleanuppad.
4036 SuccessorCleanupPad->eraseFromParent();
4037 // Now, we simply replace the cleanupret with a branch to the unwind
4038 // destination.
4039 BranchInst::Create(UnwindDest, RI->getParent());
4040 RI->eraseFromParent();
4041
4042 return true;
4043}
4044
4045bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4046 // It is possible to transiantly have an undef cleanuppad operand because we
4047 // have deleted some, but not all, dead blocks.
4048 // Eventually, this block will be deleted.
4049 if (isa<UndefValue>(RI->getOperand(0)))
4050 return false;
4051
4052 if (mergeCleanupPad(RI))
4053 return true;
4054
4055 if (removeEmptyCleanup(RI))
4056 return true;
4057
4058 return false;
4059}
4060
4061bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4062 BasicBlock *BB = RI->getParent();
4063 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4064 return false;
4065
4066 // Find predecessors that end with branches.
4067 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4068 SmallVector<BranchInst *, 8> CondBranchPreds;
4069 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4070 BasicBlock *P = *PI;
4071 TerminatorInst *PTI = P->getTerminator();
4072 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4073 if (BI->isUnconditional())
4074 UncondBranchPreds.push_back(P);
4075 else
4076 CondBranchPreds.push_back(BI);
4077 }
4078 }
4079
4080 // If we found some, do the transformation!
4081 if (!UncondBranchPreds.empty() && DupRet) {
4082 while (!UncondBranchPreds.empty()) {
4083 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4084 DEBUG(dbgs() << "FOLDING: " << *BBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
4085 << "INTO UNCOND BRANCH PRED: " << *Pred)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "FOLDING: " << *BB <<
"INTO UNCOND BRANCH PRED: " << *Pred; } } while (false
)
;
4086 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4087 }
4088
4089 // If we eliminated all predecessors of the block, delete the block now.
4090 if (pred_empty(BB)) {
4091 // We know there are no successors, so just nuke the block.
4092 BB->eraseFromParent();
4093 if (LoopHeaders)
4094 LoopHeaders->erase(BB);
4095 }
4096
4097 return true;
4098 }
4099
4100 // Check out all of the conditional branches going to this return
4101 // instruction. If any of them just select between returns, change the
4102 // branch itself into a select/return pair.
4103 while (!CondBranchPreds.empty()) {
4104 BranchInst *BI = CondBranchPreds.pop_back_val();
4105
4106 // Check to see if the non-BB successor is also a return block.
4107 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4108 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4109 SimplifyCondBranchToTwoReturns(BI, Builder))
4110 return true;
4111 }
4112 return false;
4113}
4114
4115bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4116 BasicBlock *BB = UI->getParent();
4117
4118 bool Changed = false;
4119
4120 // If there are any instructions immediately before the unreachable that can
4121 // be removed, do so.
4122 while (UI->getIterator() != BB->begin()) {
4123 BasicBlock::iterator BBI = UI->getIterator();
4124 --BBI;
4125 // Do not delete instructions that can have side effects which might cause
4126 // the unreachable to not be reachable; specifically, calls and volatile
4127 // operations may have this effect.
4128 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4129 break;
4130
4131 if (BBI->mayHaveSideEffects()) {
4132 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4133 if (SI->isVolatile())
4134 break;
4135 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4136 if (LI->isVolatile())
4137 break;
4138 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4139 if (RMWI->isVolatile())
4140 break;
4141 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4142 if (CXI->isVolatile())
4143 break;
4144 } else if (isa<CatchPadInst>(BBI)) {
4145 // A catchpad may invoke exception object constructors and such, which
4146 // in some languages can be arbitrary code, so be conservative by
4147 // default.
4148 // For CoreCLR, it just involves a type test, so can be removed.
4149 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4150 EHPersonality::CoreCLR)
4151 break;
4152 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4153 !isa<LandingPadInst>(BBI)) {
4154 break;
4155 }
4156 // Note that deleting LandingPad's here is in fact okay, although it
4157 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4158 // all the predecessors of this block will be the unwind edges of Invokes,
4159 // and we can therefore guarantee this block will be erased.
4160 }
4161
4162 // Delete this instruction (any uses are guaranteed to be dead)
4163 if (!BBI->use_empty())
4164 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4165 BBI->eraseFromParent();
4166 Changed = true;
4167 }
4168
4169 // If the unreachable instruction is the first in the block, take a gander
4170 // at all of the predecessors of this instruction, and simplify them.
4171 if (&BB->front() != UI)
4172 return Changed;
4173
4174 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4175 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4176 TerminatorInst *TI = Preds[i]->getTerminator();
4177 IRBuilder<> Builder(TI);
4178 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4179 if (BI->isUnconditional()) {
4180 if (BI->getSuccessor(0) == BB) {
4181 new UnreachableInst(TI->getContext(), TI);
4182 TI->eraseFromParent();
4183 Changed = true;
4184 }
4185 } else {
4186 if (BI->getSuccessor(0) == BB) {
4187 Builder.CreateBr(BI->getSuccessor(1));
4188 EraseTerminatorInstAndDCECond(BI);
4189 } else if (BI->getSuccessor(1) == BB) {
4190 Builder.CreateBr(BI->getSuccessor(0));
4191 EraseTerminatorInstAndDCECond(BI);
4192 Changed = true;
4193 }
4194 }
4195 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4196 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4197 if (i->getCaseSuccessor() != BB) {
4198 ++i;
4199 continue;
4200 }
4201 BB->removePredecessor(SI->getParent());
4202 i = SI->removeCase(i);
4203 e = SI->case_end();
4204 Changed = true;
4205 }
4206 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4207 if (II->getUnwindDest() == BB) {
4208 removeUnwindEdge(TI->getParent());
4209 Changed = true;
4210 }
4211 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4212 if (CSI->getUnwindDest() == BB) {
4213 removeUnwindEdge(TI->getParent());
4214 Changed = true;
4215 continue;
4216 }
4217
4218 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4219 E = CSI->handler_end();
4220 I != E; ++I) {
4221 if (*I == BB) {
4222 CSI->removeHandler(I);
4223 --I;
4224 --E;
4225 Changed = true;
4226 }
4227 }
4228 if (CSI->getNumHandlers() == 0) {
4229 BasicBlock *CatchSwitchBB = CSI->getParent();
4230 if (CSI->hasUnwindDest()) {
4231 // Redirect preds to the unwind dest
4232 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4233 } else {
4234 // Rewrite all preds to unwind to caller (or from invoke to call).
4235 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4236 for (BasicBlock *EHPred : EHPreds)
4237 removeUnwindEdge(EHPred);
4238 }
4239 // The catchswitch is no longer reachable.
4240 new UnreachableInst(CSI->getContext(), CSI);
4241 CSI->eraseFromParent();
4242 Changed = true;
4243 }
4244 } else if (isa<CleanupReturnInst>(TI)) {
4245 new UnreachableInst(TI->getContext(), TI);
4246 TI->eraseFromParent();
4247 Changed = true;
4248 }
4249 }
4250
4251 // If this block is now dead, remove it.
4252 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4253 // We know there are no successors, so just nuke the block.
4254 BB->eraseFromParent();
4255 if (LoopHeaders)
4256 LoopHeaders->erase(BB);
4257 return true;
4258 }
4259
4260 return Changed;
4261}
4262
4263static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4264 assert(Cases.size() >= 1)(static_cast <bool> (Cases.size() >= 1) ? void (0) :
__assert_fail ("Cases.size() >= 1", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4264, __extension__ __PRETTY_FUNCTION__))
;
4265
4266 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4267 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4268 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4269 return false;
4270 }
4271 return true;
4272}
4273
4274/// Turn a switch with two reachable destinations into an integer range
4275/// comparison and branch.
4276static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4277 assert(SI->getNumCases() > 1 && "Degenerate switch?")(static_cast <bool> (SI->getNumCases() > 1 &&
"Degenerate switch?") ? void (0) : __assert_fail ("SI->getNumCases() > 1 && \"Degenerate switch?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4277, __extension__ __PRETTY_FUNCTION__))
;
4278
4279 bool HasDefault =
4280 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4281
4282 // Partition the cases into two sets with different destinations.
4283 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4284 BasicBlock *DestB = nullptr;
4285 SmallVector<ConstantInt *, 16> CasesA;
4286 SmallVector<ConstantInt *, 16> CasesB;
4287
4288 for (auto Case : SI->cases()) {
4289 BasicBlock *Dest = Case.getCaseSuccessor();
4290 if (!DestA)
4291 DestA = Dest;
4292 if (Dest == DestA) {
4293 CasesA.push_back(Case.getCaseValue());
4294 continue;
4295 }
4296 if (!DestB)
4297 DestB = Dest;
4298 if (Dest == DestB) {
4299 CasesB.push_back(Case.getCaseValue());
4300 continue;
4301 }
4302 return false; // More than two destinations.
4303 }
4304
4305 assert(DestA && DestB &&(static_cast <bool> (DestA && DestB && "Single-destination switch should have been folded."
) ? void (0) : __assert_fail ("DestA && DestB && \"Single-destination switch should have been folded.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4306, __extension__ __PRETTY_FUNCTION__))
4306 "Single-destination switch should have been folded.")(static_cast <bool> (DestA && DestB && "Single-destination switch should have been folded."
) ? void (0) : __assert_fail ("DestA && DestB && \"Single-destination switch should have been folded.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4306, __extension__ __PRETTY_FUNCTION__))
;
4307 assert(DestA != DestB)(static_cast <bool> (DestA != DestB) ? void (0) : __assert_fail
("DestA != DestB", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4307, __extension__ __PRETTY_FUNCTION__))
;
4308 assert(DestB != SI->getDefaultDest())(static_cast <bool> (DestB != SI->getDefaultDest()) ?
void (0) : __assert_fail ("DestB != SI->getDefaultDest()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4308, __extension__ __PRETTY_FUNCTION__))
;
4309 assert(!CasesB.empty() && "There must be non-default cases.")(static_cast <bool> (!CasesB.empty() && "There must be non-default cases."
) ? void (0) : __assert_fail ("!CasesB.empty() && \"There must be non-default cases.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4309, __extension__ __PRETTY_FUNCTION__))
;
4310 assert(!CasesA.empty() || HasDefault)(static_cast <bool> (!CasesA.empty() || HasDefault) ? void
(0) : __assert_fail ("!CasesA.empty() || HasDefault", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4310, __extension__ __PRETTY_FUNCTION__))
;
4311
4312 // Figure out if one of the sets of cases form a contiguous range.
4313 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4314 BasicBlock *ContiguousDest = nullptr;
4315 BasicBlock *OtherDest = nullptr;
4316 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4317 ContiguousCases = &CasesA;
4318 ContiguousDest = DestA;
4319 OtherDest = DestB;
4320 } else if (CasesAreContiguous(CasesB)) {
4321 ContiguousCases = &CasesB;
4322 ContiguousDest = DestB;
4323 OtherDest = DestA;
4324 } else
4325 return false;
4326
4327 // Start building the compare and branch.
4328
4329 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4330 Constant *NumCases =
4331 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4332
4333 Value *Sub = SI->getCondition();
4334 if (!Offset->isNullValue())
4335 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4336
4337 Value *Cmp;
4338 // If NumCases overflowed, then all possible values jump to the successor.
4339 if (NumCases->isNullValue() && !ContiguousCases->empty())
4340 Cmp = ConstantInt::getTrue(SI->getContext());
4341 else
4342 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4343 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4344
4345 // Update weight for the newly-created conditional branch.
4346 if (HasBranchWeights(SI)) {
4347 SmallVector<uint64_t, 8> Weights;
4348 GetBranchWeights(SI, Weights);
4349 if (Weights.size() == 1 + SI->getNumCases()) {
4350 uint64_t TrueWeight = 0;
4351 uint64_t FalseWeight = 0;
4352 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4353 if (SI->getSuccessor(I) == ContiguousDest)
4354 TrueWeight += Weights[I];
4355 else
4356 FalseWeight += Weights[I];
4357 }
4358 while (TrueWeight > UINT32_MAX(4294967295U) || FalseWeight > UINT32_MAX(4294967295U)) {
4359 TrueWeight /= 2;
4360 FalseWeight /= 2;
4361 }
4362 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4363 }
4364 }
4365
4366 // Prune obsolete incoming values off the successors' PHI nodes.
4367 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4368 unsigned PreviousEdges = ContiguousCases->size();
4369 if (ContiguousDest == SI->getDefaultDest())
4370 ++PreviousEdges;
4371 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4372 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4373 }
4374 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4375 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4376 if (OtherDest == SI->getDefaultDest())
4377 ++PreviousEdges;
4378 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4379 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4380 }
4381
4382 // Drop the switch.
4383 SI->eraseFromParent();
4384
4385 return true;
4386}
4387
4388/// Compute masked bits for the condition of a switch
4389/// and use it to remove dead cases.
4390static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4391 const DataLayout &DL) {
4392 Value *Cond = SI->getCondition();
4393 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4394 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4395
4396 // We can also eliminate cases by determining that their values are outside of
4397 // the limited range of the condition based on how many significant (non-sign)
4398 // bits are in the condition value.
4399 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4400 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4401
4402 // Gather dead cases.
4403 SmallVector<ConstantInt *, 8> DeadCases;
4404 for (auto &Case : SI->cases()) {
4405 const APInt &CaseVal = Case.getCaseValue()->getValue();
4406 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4407 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4408 DeadCases.push_back(Case.getCaseValue());
4409 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch case "
<< CaseVal << " is dead.\n"; } } while (false)
;
4410 }
4411 }
4412
4413 // If we can prove that the cases must cover all possible values, the
4414 // default destination becomes dead and we can remove it. If we know some
4415 // of the bits in the value, we can use that to more precisely compute the
4416 // number of possible unique case values.
4417 bool HasDefault =
4418 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4419 const unsigned NumUnknownBits =
4420 Bits - (Known.Zero | Known.One).countPopulation();
4421 assert(NumUnknownBits <= Bits)(static_cast <bool> (NumUnknownBits <= Bits) ? void (
0) : __assert_fail ("NumUnknownBits <= Bits", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4421, __extension__ __PRETTY_FUNCTION__))
;
4422 if (HasDefault && DeadCases.empty() &&
4423 NumUnknownBits < 64 /* avoid overflow */ &&
4424 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4425 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "SimplifyCFG: switch default is dead.\n"
; } } while (false)
;
4426 BasicBlock *NewDefault =
4427 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4428 SI->setDefaultDest(&*NewDefault);
4429 SplitBlock(&*NewDefault, &NewDefault->front());
4430 auto *OldTI = NewDefault->getTerminator();
4431 new UnreachableInst(SI->getContext(), OldTI);
4432 EraseTerminatorInstAndDCECond(OldTI);
4433 return true;
4434 }
4435
4436 SmallVector<uint64_t, 8> Weights;
4437 bool HasWeight = HasBranchWeights(SI);
4438 if (HasWeight) {
4439 GetBranchWeights(SI, Weights);
4440 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4441 }
4442
4443 // Remove dead cases from the switch.
4444 for (ConstantInt *DeadCase : DeadCases) {
4445 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4446 assert(CaseI != SI->case_default() &&(static_cast <bool> (CaseI != SI->case_default() &&
"Case was not found. Probably mistake in DeadCases forming."
) ? void (0) : __assert_fail ("CaseI != SI->case_default() && \"Case was not found. Probably mistake in DeadCases forming.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4447, __extension__ __PRETTY_FUNCTION__))
4447 "Case was not found. Probably mistake in DeadCases forming.")(static_cast <bool> (CaseI != SI->case_default() &&
"Case was not found. Probably mistake in DeadCases forming."
) ? void (0) : __assert_fail ("CaseI != SI->case_default() && \"Case was not found. Probably mistake in DeadCases forming.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4447, __extension__ __PRETTY_FUNCTION__))
;
4448 if (HasWeight) {
4449 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4450 Weights.pop_back();
4451 }
4452
4453 // Prune unused values from PHI nodes.
4454 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4455 SI->removeCase(CaseI);
4456 }
4457 if (HasWeight && Weights.size() >= 2) {
4458 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4459 setBranchWeights(SI, MDWeights);
4460 }
4461
4462 return !DeadCases.empty();
4463}
4464
4465/// If BB would be eligible for simplification by
4466/// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4467/// by an unconditional branch), look at the phi node for BB in the successor
4468/// block and see if the incoming value is equal to CaseValue. If so, return
4469/// the phi node, and set PhiIndex to BB's index in the phi node.
4470static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4471 BasicBlock *BB, int *PhiIndex) {
4472 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4473 return nullptr; // BB must be empty to be a candidate for simplification.
4474 if (!BB->getSinglePredecessor())
4475 return nullptr; // BB must be dominated by the switch.
4476
4477 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4478 if (!Branch || !Branch->isUnconditional())
4479 return nullptr; // Terminator must be unconditional branch.
4480
4481 BasicBlock *Succ = Branch->getSuccessor(0);
4482
4483 for (PHINode &PHI : Succ->phis()) {
4484 int Idx = PHI.getBasicBlockIndex(BB);
4485 assert(Idx >= 0 && "PHI has no entry for predecessor?")(static_cast <bool> (Idx >= 0 && "PHI has no entry for predecessor?"
) ? void (0) : __assert_fail ("Idx >= 0 && \"PHI has no entry for predecessor?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4485, __extension__ __PRETTY_FUNCTION__))
;
4486
4487 Value *InValue = PHI.getIncomingValue(Idx);
4488 if (InValue != CaseValue)
4489 continue;
4490
4491 *PhiIndex = Idx;
4492 return &PHI;
4493 }
4494
4495 return nullptr;
4496}
4497
4498/// Try to forward the condition of a switch instruction to a phi node
4499/// dominated by the switch, if that would mean that some of the destination
4500/// blocks of the switch can be folded away. Return true if a change is made.
4501static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4502 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4503
4504 ForwardingNodesMap ForwardingNodes;
4505 BasicBlock *SwitchBlock = SI->getParent();
4506 bool Changed = false;
4507 for (auto &Case : SI->cases()) {
4508 ConstantInt *CaseValue = Case.getCaseValue();
4509 BasicBlock *CaseDest = Case.getCaseSuccessor();
4510
4511 // Replace phi operands in successor blocks that are using the constant case
4512 // value rather than the switch condition variable:
4513 // switchbb:
4514 // switch i32 %x, label %default [
4515 // i32 17, label %succ
4516 // ...
4517 // succ:
4518 // %r = phi i32 ... [ 17, %switchbb ] ...
4519 // -->
4520 // %r = phi i32 ... [ %x, %switchbb ] ...
4521
4522 for (PHINode &Phi : CaseDest->phis()) {
4523 // This only works if there is exactly 1 incoming edge from the switch to
4524 // a phi. If there is >1, that means multiple cases of the switch map to 1
4525 // value in the phi, and that phi value is not the switch condition. Thus,
4526 // this transform would not make sense (the phi would be invalid because
4527 // a phi can't have different incoming values from the same block).
4528 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4529 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4530 count(Phi.blocks(), SwitchBlock) == 1) {
4531 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4532 Changed = true;
4533 }
4534 }
4535
4536 // Collect phi nodes that are indirectly using this switch's case constants.
4537 int PhiIdx;
4538 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4539 ForwardingNodes[Phi].push_back(PhiIdx);
4540 }
4541
4542 for (auto &ForwardingNode : ForwardingNodes) {
4543 PHINode *Phi = ForwardingNode.first;
4544 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4545 if (Indexes.size() < 2)
4546 continue;
4547
4548 for (int Index : Indexes)
4549 Phi->setIncomingValue(Index, SI->getCondition());
4550 Changed = true;
4551 }
4552
4553 return Changed;
4554}
4555
4556/// Return true if the backend will be able to handle
4557/// initializing an array of constants like C.
4558static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4559 if (C->isThreadDependent())
4560 return false;
4561 if (C->isDLLImportDependent())
4562 return false;
4563
4564 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4565 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4566 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4567 return false;
4568
4569 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4570 if (!CE->isGEPWithNoNotionalOverIndexing())
4571 return false;
4572 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4573 return false;
4574 }
4575
4576 if (!TTI.shouldBuildLookupTablesForConstant(C))
4577 return false;
4578
4579 return true;
4580}
4581
4582/// If V is a Constant, return it. Otherwise, try to look up
4583/// its constant value in ConstantPool, returning 0 if it's not there.
4584static Constant *
4585LookupConstant(Value *V,
4586 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4587 if (Constant *C = dyn_cast<Constant>(V))
4588 return C;
4589 return ConstantPool.lookup(V);
4590}
4591
4592/// Try to fold instruction I into a constant. This works for
4593/// simple instructions such as binary operations where both operands are
4594/// constant or can be replaced by constants from the ConstantPool. Returns the
4595/// resulting constant on success, 0 otherwise.
4596static Constant *
4597ConstantFold(Instruction *I, const DataLayout &DL,
4598 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4599 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4600 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4601 if (!A)
4602 return nullptr;
4603 if (A->isAllOnesValue())
4604 return LookupConstant(Select->getTrueValue(), ConstantPool);
4605 if (A->isNullValue())
4606 return LookupConstant(Select->getFalseValue(), ConstantPool);
4607 return nullptr;
4608 }
4609
4610 SmallVector<Constant *, 4> COps;
4611 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4612 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4613 COps.push_back(A);
4614 else
4615 return nullptr;
4616 }
4617
4618 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4619 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4620 COps[1], DL);
4621 }
4622
4623 return ConstantFoldInstOperands(I, COps, DL);
4624}
4625
4626/// Try to determine the resulting constant values in phi nodes
4627/// at the common destination basic block, *CommonDest, for one of the case
4628/// destionations CaseDest corresponding to value CaseVal (0 for the default
4629/// case), of a switch instruction SI.
4630static bool
4631GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4632 BasicBlock **CommonDest,
4633 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4634 const DataLayout &DL, const TargetTransformInfo &TTI) {
4635 // The block from which we enter the common destination.
4636 BasicBlock *Pred = SI->getParent();
4637
4638 // If CaseDest is empty except for some side-effect free instructions through
4639 // which we can constant-propagate the CaseVal, continue to its successor.
4640 SmallDenseMap<Value *, Constant *> ConstantPool;
4641 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4642 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4643 ++I) {
4644 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4645 // If the terminator is a simple branch, continue to the next block.
4646 if (T->getNumSuccessors() != 1 || T->isExceptional())
4647 return false;
4648 Pred = CaseDest;
4649 CaseDest = T->getSuccessor(0);
4650 } else if (isa<DbgInfoIntrinsic>(I)) {
4651 // Skip debug intrinsic.
4652 continue;
4653 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4654 // Instruction is side-effect free and constant.
4655
4656 // If the instruction has uses outside this block or a phi node slot for
4657 // the block, it is not safe to bypass the instruction since it would then
4658 // no longer dominate all its uses.
4659 for (auto &Use : I->uses()) {
4660 User *User = Use.getUser();
4661 if (Instruction *I = dyn_cast<Instruction>(User))
4662 if (I->getParent() == CaseDest)
4663 continue;
4664 if (PHINode *Phi = dyn_cast<PHINode>(User))
4665 if (Phi->getIncomingBlock(Use) == CaseDest)
4666 continue;
4667 return false;
4668 }
4669
4670 ConstantPool.insert(std::make_pair(&*I, C));
4671 } else {
4672 break;
4673 }
4674 }
4675
4676 // If we did not have a CommonDest before, use the current one.
4677 if (!*CommonDest)
4678 *CommonDest = CaseDest;
4679 // If the destination isn't the common one, abort.
4680 if (CaseDest != *CommonDest)
4681 return false;
4682
4683 // Get the values for this case from phi nodes in the destination block.
4684 for (PHINode &PHI : (*CommonDest)->phis()) {
4685 int Idx = PHI.getBasicBlockIndex(Pred);
4686 if (Idx == -1)
4687 continue;
4688
4689 Constant *ConstVal =
4690 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4691 if (!ConstVal)
4692 return false;
4693
4694 // Be conservative about which kinds of constants we support.
4695 if (!ValidLookupTableConstant(ConstVal, TTI))
4696 return false;
4697
4698 Res.push_back(std::make_pair(&PHI, ConstVal));
4699 }
4700
4701 return Res.size() > 0;
4702}
4703
4704// Helper function used to add CaseVal to the list of cases that generate
4705// Result. Returns the updated number of cases that generate this result.
4706static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4707 SwitchCaseResultVectorTy &UniqueResults,
4708 Constant *Result) {
4709 for (auto &I : UniqueResults) {
4710 if (I.first == Result) {
4711 I.second.push_back(CaseVal);
4712 return I.second.size();
4713 }
4714 }
4715 UniqueResults.push_back(
4716 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4717 return 1;
4718}
4719
4720// Helper function that initializes a map containing
4721// results for the PHI node of the common destination block for a switch
4722// instruction. Returns false if multiple PHI nodes have been found or if
4723// there is not a common destination block for the switch.
4724static bool
4725InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI, BasicBlock *&CommonDest,
4726 SwitchCaseResultVectorTy &UniqueResults,
4727 Constant *&DefaultResult, const DataLayout &DL,
4728 const TargetTransformInfo &TTI,
4729 uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4730 for (auto &I : SI->cases()) {
4731 ConstantInt *CaseVal = I.getCaseValue();
4732
4733 // Resulting value at phi nodes for this case value.
4734 SwitchCaseResultsTy Results;
4735 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4736 DL, TTI))
4737 return false;
4738
4739 // Only one value per case is permitted.
4740 if (Results.size() > 1)
4741 return false;
4742
4743 // Add the case->result mapping to UniqueResults.
4744 const uintptr_t NumCasesForResult =
4745 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4746
4747 // Early out if there are too many cases for this result.
4748 if (NumCasesForResult > MaxCasesPerResult)
4749 return false;
4750
4751 // Early out if there are too many unique results.
4752 if (UniqueResults.size() > MaxUniqueResults)
4753 return false;
4754
4755 // Check the PHI consistency.
4756 if (!PHI)
4757 PHI = Results[0].first;
4758 else if (PHI != Results[0].first)
4759 return false;
4760 }
4761 // Find the default result value.
4762 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4763 BasicBlock *DefaultDest = SI->getDefaultDest();
4764 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4765 DL, TTI);
4766 // If the default value is not found abort unless the default destination
4767 // is unreachable.
4768 DefaultResult =
4769 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4770 if ((!DefaultResult &&
4771 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4772 return false;
4773
4774 return true;
4775}
4776
4777// Helper function that checks if it is possible to transform a switch with only
4778// two cases (or two cases + default) that produces a result into a select.
4779// Example:
4780// switch (a) {
4781// case 10: %0 = icmp eq i32 %a, 10
4782// return 10; %1 = select i1 %0, i32 10, i32 4
4783// case 20: ----> %2 = icmp eq i32 %a, 20
4784// return 2; %3 = select i1 %2, i32 2, i32 %1
4785// default:
4786// return 4;
4787// }
4788static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4789 Constant *DefaultResult, Value *Condition,
4790 IRBuilder<> &Builder) {
4791 assert(ResultVector.size() == 2 &&(static_cast <bool> (ResultVector.size() == 2 &&
"We should have exactly two unique results at this point") ?
void (0) : __assert_fail ("ResultVector.size() == 2 && \"We should have exactly two unique results at this point\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4792, __extension__ __PRETTY_FUNCTION__))
4792 "We should have exactly two unique results at this point")(static_cast <bool> (ResultVector.size() == 2 &&
"We should have exactly two unique results at this point") ?
void (0) : __assert_fail ("ResultVector.size() == 2 && \"We should have exactly two unique results at this point\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4792, __extension__ __PRETTY_FUNCTION__))
;
4793 // If we are selecting between only two cases transform into a simple
4794 // select or a two-way select if default is possible.
4795 if (ResultVector[0].second.size() == 1 &&
4796 ResultVector[1].second.size() == 1) {
4797 ConstantInt *const FirstCase = ResultVector[0].second[0];
4798 ConstantInt *const SecondCase = ResultVector[1].second[0];
4799
4800 bool DefaultCanTrigger = DefaultResult;
4801 Value *SelectValue = ResultVector[1].first;
4802 if (DefaultCanTrigger) {
4803 Value *const ValueCompare =
4804 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4805 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4806 DefaultResult, "switch.select");
4807 }
4808 Value *const ValueCompare =
4809 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4810 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4811 SelectValue, "switch.select");
4812 }
4813
4814 return nullptr;
4815}
4816
4817// Helper function to cleanup a switch instruction that has been converted into
4818// a select, fixing up PHI nodes and basic blocks.
4819static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4820 Value *SelectValue,
4821 IRBuilder<> &Builder) {
4822 BasicBlock *SelectBB = SI->getParent();
4823 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4824 PHI->removeIncomingValue(SelectBB);
4825 PHI->addIncoming(SelectValue, SelectBB);
4826
4827 Builder.CreateBr(PHI->getParent());
4828
4829 // Remove the switch.
4830 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4831 BasicBlock *Succ = SI->getSuccessor(i);
4832
4833 if (Succ == PHI->getParent())
4834 continue;
4835 Succ->removePredecessor(SelectBB);
4836 }
4837 SI->eraseFromParent();
4838}
4839
4840/// If the switch is only used to initialize one or more
4841/// phi nodes in a common successor block with only two different
4842/// constant values, replace the switch with select.
4843static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4844 const DataLayout &DL,
4845 const TargetTransformInfo &TTI) {
4846 Value *const Cond = SI->getCondition();
4847 PHINode *PHI = nullptr;
4848 BasicBlock *CommonDest = nullptr;
4849 Constant *DefaultResult;
4850 SwitchCaseResultVectorTy UniqueResults;
4851 // Collect all the cases that will deliver the same value from the switch.
4852 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4853 DL, TTI, 2, 1))
4854 return false;
4855 // Selects choose between maximum two values.
4856 if (UniqueResults.size() != 2)
4857 return false;
4858 assert(PHI != nullptr && "PHI for value select not found")(static_cast <bool> (PHI != nullptr && "PHI for value select not found"
) ? void (0) : __assert_fail ("PHI != nullptr && \"PHI for value select not found\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4858, __extension__ __PRETTY_FUNCTION__))
;
4859
4860 Builder.SetInsertPoint(SI);
4861 Value *SelectValue =
4862 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4863 if (SelectValue) {
4864 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4865 return true;
4866 }
4867 // The switch couldn't be converted into a select.
4868 return false;
4869}
4870
4871namespace {
4872
4873/// This class represents a lookup table that can be used to replace a switch.
4874class SwitchLookupTable {
4875public:
4876 /// Create a lookup table to use as a switch replacement with the contents
4877 /// of Values, using DefaultValue to fill any holes in the table.
4878 SwitchLookupTable(
4879 Module &M, uint64_t TableSize, ConstantInt *Offset,
4880 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4881 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4882
4883 /// Build instructions with Builder to retrieve the value at
4884 /// the position given by Index in the lookup table.
4885 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4886
4887 /// Return true if a table with TableSize elements of
4888 /// type ElementType would fit in a target-legal register.
4889 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4890 Type *ElementType);
4891
4892private:
4893 // Depending on the contents of the table, it can be represented in
4894 // different ways.
4895 enum {
4896 // For tables where each element contains the same value, we just have to
4897 // store that single value and return it for each lookup.
4898 SingleValueKind,
4899
4900 // For tables where there is a linear relationship between table index
4901 // and values. We calculate the result with a simple multiplication
4902 // and addition instead of a table lookup.
4903 LinearMapKind,
4904
4905 // For small tables with integer elements, we can pack them into a bitmap
4906 // that fits into a target-legal register. Values are retrieved by
4907 // shift and mask operations.
4908 BitMapKind,
4909
4910 // The table is stored as an array of values. Values are retrieved by load
4911 // instructions from the table.
4912 ArrayKind
4913 } Kind;
4914
4915 // For SingleValueKind, this is the single value.
4916 Constant *SingleValue = nullptr;
4917
4918 // For BitMapKind, this is the bitmap.
4919 ConstantInt *BitMap = nullptr;
4920 IntegerType *BitMapElementTy = nullptr;
4921
4922 // For LinearMapKind, these are the constants used to derive the value.
4923 ConstantInt *LinearOffset = nullptr;
4924 ConstantInt *LinearMultiplier = nullptr;
4925
4926 // For ArrayKind, this is the array.
4927 GlobalVariable *Array = nullptr;
4928};
4929
4930} // end anonymous namespace
4931
4932SwitchLookupTable::SwitchLookupTable(
4933 Module &M, uint64_t TableSize, ConstantInt *Offset,
4934 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4935 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4936 assert(Values.size() && "Can't build lookup table without values!")(static_cast <bool> (Values.size() && "Can't build lookup table without values!"
) ? void (0) : __assert_fail ("Values.size() && \"Can't build lookup table without values!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4936, __extension__ __PRETTY_FUNCTION__))
;
4937 assert(TableSize >= Values.size() && "Can't fit values in table!")(static_cast <bool> (TableSize >= Values.size() &&
"Can't fit values in table!") ? void (0) : __assert_fail ("TableSize >= Values.size() && \"Can't fit values in table!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4937, __extension__ __PRETTY_FUNCTION__))
;
4938
4939 // If all values in the table are equal, this is that value.
4940 SingleValue = Values.begin()->second;
4941
4942 Type *ValueType = Values.begin()->second->getType();
4943
4944 // Build up the table contents.
4945 SmallVector<Constant *, 64> TableContents(TableSize);
4946 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4947 ConstantInt *CaseVal = Values[I].first;
4948 Constant *CaseRes = Values[I].second;
4949 assert(CaseRes->getType() == ValueType)(static_cast <bool> (CaseRes->getType() == ValueType
) ? void (0) : __assert_fail ("CaseRes->getType() == ValueType"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4949, __extension__ __PRETTY_FUNCTION__))
;
4950
4951 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4952 TableContents[Idx] = CaseRes;
4953
4954 if (CaseRes != SingleValue)
4955 SingleValue = nullptr;
4956 }
4957
4958 // Fill in any holes in the table with the default result.
4959 if (Values.size() < TableSize) {
4960 assert(DefaultValue &&(static_cast <bool> (DefaultValue && "Need a default value to fill the lookup table holes."
) ? void (0) : __assert_fail ("DefaultValue && \"Need a default value to fill the lookup table holes.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4961, __extension__ __PRETTY_FUNCTION__))
4961 "Need a default value to fill the lookup table holes.")(static_cast <bool> (DefaultValue && "Need a default value to fill the lookup table holes."
) ? void (0) : __assert_fail ("DefaultValue && \"Need a default value to fill the lookup table holes.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4961, __extension__ __PRETTY_FUNCTION__))
;
4962 assert(DefaultValue->getType() == ValueType)(static_cast <bool> (DefaultValue->getType() == ValueType
) ? void (0) : __assert_fail ("DefaultValue->getType() == ValueType"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4962, __extension__ __PRETTY_FUNCTION__))
;
4963 for (uint64_t I = 0; I < TableSize; ++I) {
4964 if (!TableContents[I])
4965 TableContents[I] = DefaultValue;
4966 }
4967
4968 if (DefaultValue != SingleValue)
4969 SingleValue = nullptr;
4970 }
4971
4972 // If each element in the table contains the same value, we only need to store
4973 // that single value.
4974 if (SingleValue) {
4975 Kind = SingleValueKind;
4976 return;
4977 }
4978
4979 // Check if we can derive the value with a linear transformation from the
4980 // table index.
4981 if (isa<IntegerType>(ValueType)) {
4982 bool LinearMappingPossible = true;
4983 APInt PrevVal;
4984 APInt DistToPrev;
4985 assert(TableSize >= 2 && "Should be a SingleValue table.")(static_cast <bool> (TableSize >= 2 && "Should be a SingleValue table."
) ? void (0) : __assert_fail ("TableSize >= 2 && \"Should be a SingleValue table.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 4985, __extension__ __PRETTY_FUNCTION__))
;
4986 // Check if there is the same distance between two consecutive values.
4987 for (uint64_t I = 0; I < TableSize; ++I) {
4988 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4989 if (!ConstVal) {
4990 // This is an undef. We could deal with it, but undefs in lookup tables
4991 // are very seldom. It's probably not worth the additional complexity.
4992 LinearMappingPossible = false;
4993 break;
4994 }
4995 const APInt &Val = ConstVal->getValue();
4996 if (I != 0) {
4997 APInt Dist = Val - PrevVal;
4998 if (I == 1) {
4999 DistToPrev = Dist;
5000 } else if (Dist != DistToPrev) {
5001 LinearMappingPossible = false;
5002 break;
5003 }
5004 }
5005 PrevVal = Val;
5006 }
5007 if (LinearMappingPossible) {
5008 LinearOffset = cast<ConstantInt>(TableContents[0]);
5009 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5010 Kind = LinearMapKind;
5011 ++NumLinearMaps;
5012 return;
5013 }
5014 }
5015
5016 // If the type is integer and the table fits in a register, build a bitmap.
5017 if (WouldFitInRegister(DL, TableSize, ValueType)) {
5018 IntegerType *IT = cast<IntegerType>(ValueType);
5019 APInt TableInt(TableSize * IT->getBitWidth(), 0);
5020 for (uint64_t I = TableSize; I > 0; --I) {
5021 TableInt <<= IT->getBitWidth();
5022 // Insert values into the bitmap. Undef values are set to zero.
5023 if (!isa<UndefValue>(TableContents[I - 1])) {
5024 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5025 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5026 }
5027 }
5028 BitMap = ConstantInt::get(M.getContext(), TableInt);
5029 BitMapElementTy = IT;
5030 Kind = BitMapKind;
5031 ++NumBitMaps;
5032 return;
5033 }
5034
5035 // Store the table in an array.
5036 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5037 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5038
5039 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5040 GlobalVariable::PrivateLinkage, Initializer,
5041 "switch.table." + FuncName);
5042 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5043 Kind = ArrayKind;
5044}
5045
5046Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5047 switch (Kind) {
5048 case SingleValueKind:
5049 return SingleValue;
5050 case LinearMapKind: {
5051 // Derive the result value from the input value.
5052 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5053 false, "switch.idx.cast");
5054 if (!LinearMultiplier->isOne())
5055 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5056 if (!LinearOffset->isZero())
5057 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5058 return Result;
5059 }
5060 case BitMapKind: {
5061 // Type of the bitmap (e.g. i59).
5062 IntegerType *MapTy = BitMap->getType();
5063
5064 // Cast Index to the same type as the bitmap.
5065 // Note: The Index is <= the number of elements in the table, so
5066 // truncating it to the width of the bitmask is safe.
5067 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5068
5069 // Multiply the shift amount by the element width.
5070 ShiftAmt = Builder.CreateMul(
5071 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5072 "switch.shiftamt");
5073
5074 // Shift down.
5075 Value *DownShifted =
5076 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5077 // Mask off.
5078 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5079 }
5080 case ArrayKind: {
5081 // Make sure the table index will not overflow when treated as signed.
5082 IntegerType *IT = cast<IntegerType>(Index->getType());
5083 uint64_t TableSize =
5084 Array->getInitializer()->getType()->getArrayNumElements();
5085 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5086 Index = Builder.CreateZExt(
5087 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5088 "switch.tableidx.zext");
5089
5090 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5091 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5092 GEPIndices, "switch.gep");
5093 return Builder.CreateLoad(GEP, "switch.load");
5094 }
5095 }
5096 llvm_unreachable("Unknown lookup table kind!")::llvm::llvm_unreachable_internal("Unknown lookup table kind!"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5096)
;
5097}
5098
5099bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5100 uint64_t TableSize,
5101 Type *ElementType) {
5102 auto *IT = dyn_cast<IntegerType>(ElementType);
5103 if (!IT)
5104 return false;
5105 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5106 // are <= 15, we could try to narrow the type.
5107
5108 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5109 if (TableSize >= UINT_MAX(2147483647 *2U +1U) / IT->getBitWidth())
5110 return false;
5111 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5112}
5113
5114/// Determine whether a lookup table should be built for this switch, based on
5115/// the number of cases, size of the table, and the types of the results.
5116static bool
5117ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5118 const TargetTransformInfo &TTI, const DataLayout &DL,
5119 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5120 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX(18446744073709551615UL) / 10)
5121 return false; // TableSize overflowed, or mul below might overflow.
5122
5123 bool AllTablesFitInRegister = true;
5124 bool HasIllegalType = false;
5125 for (const auto &I : ResultTypes) {
5126 Type *Ty = I.second;
5127
5128 // Saturate this flag to true.
5129 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5130
5131 // Saturate this flag to false.
5132 AllTablesFitInRegister =
5133 AllTablesFitInRegister &&
5134 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5135
5136 // If both flags saturate, we're done. NOTE: This *only* works with
5137 // saturating flags, and all flags have to saturate first due to the
5138 // non-deterministic behavior of iterating over a dense map.
5139 if (HasIllegalType && !AllTablesFitInRegister)
5140 break;
5141 }
5142
5143 // If each table would fit in a register, we should build it anyway.
5144 if (AllTablesFitInRegister)
5145 return true;
5146
5147 // Don't build a table that doesn't fit in-register if it has illegal types.
5148 if (HasIllegalType)
5149 return false;
5150
5151 // The table density should be at least 40%. This is the same criterion as for
5152 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5153 // FIXME: Find the best cut-off.
5154 return SI->getNumCases() * 10 >= TableSize * 4;
5155}
5156
5157/// Try to reuse the switch table index compare. Following pattern:
5158/// \code
5159/// if (idx < tablesize)
5160/// r = table[idx]; // table does not contain default_value
5161/// else
5162/// r = default_value;
5163/// if (r != default_value)
5164/// ...
5165/// \endcode
5166/// Is optimized to:
5167/// \code
5168/// cond = idx < tablesize;
5169/// if (cond)
5170/// r = table[idx];
5171/// else
5172/// r = default_value;
5173/// if (cond)
5174/// ...
5175/// \endcode
5176/// Jump threading will then eliminate the second if(cond).
5177static void reuseTableCompare(
5178 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5179 Constant *DefaultValue,
5180 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5181 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5182 if (!CmpInst)
5183 return;
5184
5185 // We require that the compare is in the same block as the phi so that jump
5186 // threading can do its work afterwards.
5187 if (CmpInst->getParent() != PhiBlock)
5188 return;
5189
5190 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5191 if (!CmpOp1)
5192 return;
5193
5194 Value *RangeCmp = RangeCheckBranch->getCondition();
5195 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5196 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5197
5198 // Check if the compare with the default value is constant true or false.
5199 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5200 DefaultValue, CmpOp1, true);
5201 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5202 return;
5203
5204 // Check if the compare with the case values is distinct from the default
5205 // compare result.
5206 for (auto ValuePair : Values) {
5207 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5208 ValuePair.second, CmpOp1, true);
5209 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5210 return;
5211 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&(static_cast <bool> ((CaseConst == TrueConst || CaseConst
== FalseConst) && "Expect true or false as compare result."
) ? void (0) : __assert_fail ("(CaseConst == TrueConst || CaseConst == FalseConst) && \"Expect true or false as compare result.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5212, __extension__ __PRETTY_FUNCTION__))
5212 "Expect true or false as compare result.")(static_cast <bool> ((CaseConst == TrueConst || CaseConst
== FalseConst) && "Expect true or false as compare result."
) ? void (0) : __assert_fail ("(CaseConst == TrueConst || CaseConst == FalseConst) && \"Expect true or false as compare result.\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5212, __extension__ __PRETTY_FUNCTION__))
;
5213 }
5214
5215 // Check if the branch instruction dominates the phi node. It's a simple
5216 // dominance check, but sufficient for our needs.
5217 // Although this check is invariant in the calling loops, it's better to do it
5218 // at this late stage. Practically we do it at most once for a switch.
5219 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5220 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5221 BasicBlock *Pred = *PI;
5222 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5223 return;
5224 }
5225
5226 if (DefaultConst == FalseConst) {
5227 // The compare yields the same result. We can replace it.
5228 CmpInst->replaceAllUsesWith(RangeCmp);
5229 ++NumTableCmpReuses;
5230 } else {
5231 // The compare yields the same result, just inverted. We can replace it.
5232 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5233 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5234 RangeCheckBranch);
5235 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5236 ++NumTableCmpReuses;
5237 }
5238}
5239
5240/// If the switch is only used to initialize one or more phi nodes in a common
5241/// successor block with different constant values, replace the switch with
5242/// lookup tables.
5243static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5244 const DataLayout &DL,
5245 const TargetTransformInfo &TTI) {
5246 assert(SI->getNumCases() > 1 && "Degenerate switch?")(static_cast <bool> (SI->getNumCases() > 1 &&
"Degenerate switch?") ? void (0) : __assert_fail ("SI->getNumCases() > 1 && \"Degenerate switch?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5246, __extension__ __PRETTY_FUNCTION__))
;
5247
5248 Function *Fn = SI->getParent()->getParent();
5249 // Only build lookup table when we have a target that supports it or the
5250 // attribute is not set.
5251 if (!TTI.shouldBuildLookupTables() ||
5252 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5253 return false;
5254
5255 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5256 // split off a dense part and build a lookup table for that.
5257
5258 // FIXME: This creates arrays of GEPs to constant strings, which means each
5259 // GEP needs a runtime relocation in PIC code. We should just build one big
5260 // string and lookup indices into that.
5261
5262 // Ignore switches with less than three cases. Lookup tables will not make
5263 // them faster, so we don't analyze them.
5264 if (SI->getNumCases() < 3)
5265 return false;
5266
5267 // Figure out the corresponding result for each case value and phi node in the
5268 // common destination, as well as the min and max case values.
5269 assert(SI->case_begin() != SI->case_end())(static_cast <bool> (SI->case_begin() != SI->case_end
()) ? void (0) : __assert_fail ("SI->case_begin() != SI->case_end()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5269, __extension__ __PRETTY_FUNCTION__))
;
5270 SwitchInst::CaseIt CI = SI->case_begin();
5271 ConstantInt *MinCaseVal = CI->getCaseValue();
5272 ConstantInt *MaxCaseVal = CI->getCaseValue();
5273
5274 BasicBlock *CommonDest = nullptr;
5275
5276 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5277 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5278
5279 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5280 SmallDenseMap<PHINode *, Type *> ResultTypes;
5281 SmallVector<PHINode *, 4> PHIs;
5282
5283 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5284 ConstantInt *CaseVal = CI->getCaseValue();
5285 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5286 MinCaseVal = CaseVal;
5287 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5288 MaxCaseVal = CaseVal;
5289
5290 // Resulting value at phi nodes for this case value.
5291 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5292 ResultsTy Results;
5293 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5294 Results, DL, TTI))
5295 return false;
5296
5297 // Append the result from this case to the list for each phi.
5298 for (const auto &I : Results) {
5299 PHINode *PHI = I.first;
5300 Constant *Value = I.second;
5301 if (!ResultLists.count(PHI))
5302 PHIs.push_back(PHI);
5303 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5304 }
5305 }
5306
5307 // Keep track of the result types.
5308 for (PHINode *PHI : PHIs) {
5309 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5310 }
5311
5312 uint64_t NumResults = ResultLists[PHIs[0]].size();
5313 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5314 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5315 bool TableHasHoles = (NumResults < TableSize);
5316
5317 // If the table has holes, we need a constant result for the default case
5318 // or a bitmask that fits in a register.
5319 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5320 bool HasDefaultResults =
5321 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5322 DefaultResultsList, DL, TTI);
5323
5324 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5325 if (NeedMask) {
5326 // As an extra penalty for the validity test we require more cases.
5327 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5328 return false;
5329 if (!DL.fitsInLegalInteger(TableSize))
5330 return false;
5331 }
5332
5333 for (const auto &I : DefaultResultsList) {
5334 PHINode *PHI = I.first;
5335 Constant *Result = I.second;
5336 DefaultResults[PHI] = Result;
5337 }
5338
5339 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5340 return false;
5341
5342 // Create the BB that does the lookups.
5343 Module &Mod = *CommonDest->getParent()->getParent();
5344 BasicBlock *LookupBB = BasicBlock::Create(
5345 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5346
5347 // Compute the table index value.
5348 Builder.SetInsertPoint(SI);
5349 Value *TableIndex;
5350 if (MinCaseVal->isNullValue())
5351 TableIndex = SI->getCondition();
5352 else
5353 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5354 "switch.tableidx");
5355
5356 // Compute the maximum table size representable by the integer type we are
5357 // switching upon.
5358 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5359 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX(18446744073709551615UL) : 1ULL << CaseSize;
5360 assert(MaxTableSize >= TableSize &&(static_cast <bool> (MaxTableSize >= TableSize &&
"It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? 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-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5362, __extension__ __PRETTY_FUNCTION__))
5361 "It is impossible for a switch to have more entries than the max "(static_cast <bool> (MaxTableSize >= TableSize &&
"It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? 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-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5362, __extension__ __PRETTY_FUNCTION__))
5362 "representable value of its input integer type's size.")(static_cast <bool> (MaxTableSize >= TableSize &&
"It is impossible for a switch to have more entries than the max "
"representable value of its input integer type's size.") ? 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-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5362, __extension__ __PRETTY_FUNCTION__))
;
5363
5364 // If the default destination is unreachable, or if the lookup table covers
5365 // all values of the conditional variable, branch directly to the lookup table
5366 // BB. Otherwise, check that the condition is within the case range.
5367 const bool DefaultIsReachable =
5368 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5369 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5370 BranchInst *RangeCheckBranch = nullptr;
5371
5372 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5373 Builder.CreateBr(LookupBB);
5374 // Note: We call removeProdecessor later since we need to be able to get the
5375 // PHI value for the default case in case we're using a bit mask.
5376 } else {
5377 Value *Cmp = Builder.CreateICmpULT(
5378 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5379 RangeCheckBranch =
5380 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5381 }
5382
5383 // Populate the BB that does the lookups.
5384 Builder.SetInsertPoint(LookupBB);
5385
5386 if (NeedMask) {
5387 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5388 // re-purposed to do the hole check, and we create a new LookupBB.
5389 BasicBlock *MaskBB = LookupBB;
5390 MaskBB->setName("switch.hole_check");
5391 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5392 CommonDest->getParent(), CommonDest);
5393
5394 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5395 // unnecessary illegal types.
5396 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5397 APInt MaskInt(TableSizePowOf2, 0);
5398 APInt One(TableSizePowOf2, 1);
5399 // Build bitmask; fill in a 1 bit for every case.
5400 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5401 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5402 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5403 .getLimitedValue();
5404 MaskInt |= One << Idx;
5405 }
5406 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5407
5408 // Get the TableIndex'th bit of the bitmask.
5409 // If this bit is 0 (meaning hole) jump to the default destination,
5410 // else continue with table lookup.
5411 IntegerType *MapTy = TableMask->getType();
5412 Value *MaskIndex =
5413 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5414 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5415 Value *LoBit = Builder.CreateTrunc(
5416 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5417 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5418
5419 Builder.SetInsertPoint(LookupBB);
5420 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5421 }
5422
5423 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5424 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5425 // do not delete PHINodes here.
5426 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5427 /*DontDeleteUselessPHIs=*/true);
5428 }
5429
5430 bool ReturnedEarly = false;
5431 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5432 PHINode *PHI = PHIs[I];
5433 const ResultListTy &ResultList = ResultLists[PHI];
5434
5435 // If using a bitmask, use any value to fill the lookup table holes.
5436 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5437 StringRef FuncName = Fn->getName();
5438 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5439 FuncName);
5440
5441 Value *Result = Table.BuildLookup(TableIndex, Builder);
5442
5443 // If the result is used to return immediately from the function, we want to
5444 // do that right here.
5445 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5446 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5447 Builder.CreateRet(Result);
5448 ReturnedEarly = true;
5449 break;
5450 }
5451
5452 // Do a small peephole optimization: re-use the switch table compare if
5453 // possible.
5454 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5455 BasicBlock *PhiBlock = PHI->getParent();
5456 // Search for compare instructions which use the phi.
5457 for (auto *User : PHI->users()) {
5458 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5459 }
5460 }
5461
5462 PHI->addIncoming(Result, LookupBB);
5463 }
5464
5465 if (!ReturnedEarly)
5466 Builder.CreateBr(CommonDest);
5467
5468 // Remove the switch.
5469 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5470 BasicBlock *Succ = SI->getSuccessor(i);
5471
5472 if (Succ == SI->getDefaultDest())
5473 continue;
5474 Succ->removePredecessor(SI->getParent());
5475 }
5476 SI->eraseFromParent();
5477
5478 ++NumLookupTables;
5479 if (NeedMask)
5480 ++NumLookupTablesHoles;
5481 return true;
5482}
5483
5484static bool isSwitchDense(ArrayRef<int64_t> Values) {
5485 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5486 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5487 uint64_t Range = Diff + 1;
5488 uint64_t NumCases = Values.size();
5489 // 40% is the default density for building a jump table in optsize/minsize mode.
5490 uint64_t MinDensity = 40;
5491
5492 return NumCases * 100 >= Range * MinDensity;
5493}
5494
5495/// Try to transform a switch that has "holes" in it to a contiguous sequence
5496/// of cases.
5497///
5498/// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5499/// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5500///
5501/// This converts a sparse switch into a dense switch which allows better
5502/// lowering and could also allow transforming into a lookup table.
5503static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5504 const DataLayout &DL,
5505 const TargetTransformInfo &TTI) {
5506 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5507 if (CondTy->getIntegerBitWidth() > 64 ||
5508 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5509 return false;
5510 // Only bother with this optimization if there are more than 3 switch cases;
5511 // SDAG will only bother creating jump tables for 4 or more cases.
5512 if (SI->getNumCases() < 4)
5513 return false;
5514
5515 // This transform is agnostic to the signedness of the input or case values. We
5516 // can treat the case values as signed or unsigned. We can optimize more common
5517 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5518 // as signed.
5519 SmallVector<int64_t,4> Values;
5520 for (auto &C : SI->cases())
5521 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5522 std::sort(Values.begin(), Values.end());
5523
5524 // If the switch is already dense, there's nothing useful to do here.
5525 if (isSwitchDense(Values))
5526 return false;
5527
5528 // First, transform the values such that they start at zero and ascend.
5529 int64_t Base = Values[0];
5530 for (auto &V : Values)
5531 V -= (uint64_t)(Base);
5532
5533 // Now we have signed numbers that have been shifted so that, given enough
5534 // precision, there are no negative values. Since the rest of the transform
5535 // is bitwise only, we switch now to an unsigned representation.
5536 uint64_t GCD = 0;
5537 for (auto &V : Values)
5538 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5539
5540 // This transform can be done speculatively because it is so cheap - it results
5541 // in a single rotate operation being inserted. This can only happen if the
5542 // factor extracted is a power of 2.
5543 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5544 // inverse of GCD and then perform this transform.
5545 // FIXME: It's possible that optimizing a switch on powers of two might also
5546 // be beneficial - flag values are often powers of two and we could use a CLZ
5547 // as the key function.
5548 if (GCD <= 1 || !isPowerOf2_64(GCD))
5549 // No common divisor found or too expensive to compute key function.
5550 return false;
5551
5552 unsigned Shift = Log2_64(GCD);
5553 for (auto &V : Values)
5554 V = (int64_t)((uint64_t)V >> Shift);
5555
5556 if (!isSwitchDense(Values))
5557 // Transform didn't create a dense switch.
5558 return false;
5559
5560 // The obvious transform is to shift the switch condition right and emit a
5561 // check that the condition actually cleanly divided by GCD, i.e.
5562 // C & (1 << Shift - 1) == 0
5563 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5564 //
5565 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5566 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5567 // are nonzero then the switch condition will be very large and will hit the
5568 // default case.
5569
5570 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5571 Builder.SetInsertPoint(SI);
5572 auto *ShiftC = ConstantInt::get(Ty, Shift);
5573 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5574 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5575 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5576 auto *Rot = Builder.CreateOr(LShr, Shl);
5577 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5578
5579 for (auto Case : SI->cases()) {
5580 auto *Orig = Case.getCaseValue();
5581 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5582 Case.setValue(
5583 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5584 }
5585 return true;
5586}
5587
5588bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5589 BasicBlock *BB = SI->getParent();
5590
5591 if (isValueEqualityComparison(SI)) {
5592 // If we only have one predecessor, and if it is a branch on this value,
5593 // see if that predecessor totally determines the outcome of this switch.
5594 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5595 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5596 return simplifyCFG(BB, TTI, Options) | true;
5597
5598 Value *Cond = SI->getCondition();
5599 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5600 if (SimplifySwitchOnSelect(SI, Select))
5601 return simplifyCFG(BB, TTI, Options) | true;
5602
5603 // If the block only contains the switch, see if we can fold the block
5604 // away into any preds.
5605 BasicBlock::iterator BBI = BB->begin();
5606 // Ignore dbg intrinsics.
5607 while (isa<DbgInfoIntrinsic>(BBI))
5608 ++BBI;
5609 if (SI == &*BBI)
5610 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5611 return simplifyCFG(BB, TTI, Options) | true;
5612 }
5613
5614 // Try to transform the switch into an icmp and a branch.
5615 if (TurnSwitchRangeIntoICmp(SI, Builder))
5616 return simplifyCFG(BB, TTI, Options) | true;
5617
5618 // Remove unreachable cases.
5619 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5620 return simplifyCFG(BB, TTI, Options) | true;
5621
5622 if (switchToSelect(SI, Builder, DL, TTI))
5623 return simplifyCFG(BB, TTI, Options) | true;
5624
5625 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5626 return simplifyCFG(BB, TTI, Options) | true;
5627
5628 // The conversion from switch to lookup tables results in difficult-to-analyze
5629 // code and makes pruning branches much harder. This is a problem if the
5630 // switch expression itself can still be restricted as a result of inlining or
5631 // CVP. Therefore, only apply this transformation during late stages of the
5632 // optimisation pipeline.
5633 if (Options.ConvertSwitchToLookupTable &&
5634 SwitchToLookupTable(SI, Builder, DL, TTI))
5635 return simplifyCFG(BB, TTI, Options) | true;
5636
5637 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5638 return simplifyCFG(BB, TTI, Options) | true;
5639
5640 return false;
5641}
5642
5643bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5644 BasicBlock *BB = IBI->getParent();
5645 bool Changed = false;
5646
5647 // Eliminate redundant destinations.
5648 SmallPtrSet<Value *, 8> Succs;
5649 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5650 BasicBlock *Dest = IBI->getDestination(i);
5651 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5652 Dest->removePredecessor(BB);
5653 IBI->removeDestination(i);
5654 --i;
5655 --e;
5656 Changed = true;
5657 }
5658 }
5659
5660 if (IBI->getNumDestinations() == 0) {
5661 // If the indirectbr has no successors, change it to unreachable.
5662 new UnreachableInst(IBI->getContext(), IBI);
5663 EraseTerminatorInstAndDCECond(IBI);
5664 return true;
5665 }
5666
5667 if (IBI->getNumDestinations() == 1) {
5668 // If the indirectbr has one successor, change it to a direct branch.
5669 BranchInst::Create(IBI->getDestination(0), IBI);
5670 EraseTerminatorInstAndDCECond(IBI);
5671 return true;
5672 }
5673
5674 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5675 if (SimplifyIndirectBrOnSelect(IBI, SI))
5676 return simplifyCFG(BB, TTI, Options) | true;
5677 }
5678 return Changed;
5679}
5680
5681/// Given an block with only a single landing pad and a unconditional branch
5682/// try to find another basic block which this one can be merged with. This
5683/// handles cases where we have multiple invokes with unique landing pads, but
5684/// a shared handler.
5685///
5686/// We specifically choose to not worry about merging non-empty blocks
5687/// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5688/// practice, the optimizer produces empty landing pad blocks quite frequently
5689/// when dealing with exception dense code. (see: instcombine, gvn, if-else
5690/// sinking in this file)
5691///
5692/// This is primarily a code size optimization. We need to avoid performing
5693/// any transform which might inhibit optimization (such as our ability to
5694/// specialize a particular handler via tail commoning). We do this by not
5695/// merging any blocks which require us to introduce a phi. Since the same
5696/// values are flowing through both blocks, we don't loose any ability to
5697/// specialize. If anything, we make such specialization more likely.
5698///
5699/// TODO - This transformation could remove entries from a phi in the target
5700/// block when the inputs in the phi are the same for the two blocks being
5701/// merged. In some cases, this could result in removal of the PHI entirely.
5702static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5703 BasicBlock *BB) {
5704 auto Succ = BB->getUniqueSuccessor();
5705 assert(Succ)(static_cast <bool> (Succ) ? void (0) : __assert_fail (
"Succ", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5705, __extension__ __PRETTY_FUNCTION__))
;
5706 // If there's a phi in the successor block, we'd likely have to introduce
5707 // a phi into the merged landing pad block.
5708 if (isa<PHINode>(*Succ->begin()))
5709 return false;
5710
5711 for (BasicBlock *OtherPred : predecessors(Succ)) {
5712 if (BB == OtherPred)
5713 continue;
5714 BasicBlock::iterator I = OtherPred->begin();
5715 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5716 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5717 continue;
5718 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5719 ;
5720 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5721 if (!BI2 || !BI2->isIdenticalTo(BI))
5722 continue;
5723
5724 // We've found an identical block. Update our predecessors to take that
5725 // path instead and make ourselves dead.
5726 SmallSet<BasicBlock *, 16> Preds;
5727 Preds.insert(pred_begin(BB), pred_end(BB));
5728 for (BasicBlock *Pred : Preds) {
5729 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5730 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&(static_cast <bool> (II->getNormalDest() != BB &&
II->getUnwindDest() == BB && "unexpected successor"
) ? void (0) : __assert_fail ("II->getNormalDest() != BB && II->getUnwindDest() == BB && \"unexpected successor\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5731, __extension__ __PRETTY_FUNCTION__))
5731 "unexpected successor")(static_cast <bool> (II->getNormalDest() != BB &&
II->getUnwindDest() == BB && "unexpected successor"
) ? void (0) : __assert_fail ("II->getNormalDest() != BB && II->getUnwindDest() == BB && \"unexpected successor\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5731, __extension__ __PRETTY_FUNCTION__))
;
5732 II->setUnwindDest(OtherPred);
5733 }
5734
5735 // The debug info in OtherPred doesn't cover the merged control flow that
5736 // used to go through BB. We need to delete it or update it.
5737 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5738 Instruction &Inst = *I;
5739 I++;
5740 if (isa<DbgInfoIntrinsic>(Inst))
5741 Inst.eraseFromParent();
5742 }
5743
5744 SmallSet<BasicBlock *, 16> Succs;
5745 Succs.insert(succ_begin(BB), succ_end(BB));
5746 for (BasicBlock *Succ : Succs) {
5747 Succ->removePredecessor(BB);
5748 }
5749
5750 IRBuilder<> Builder(BI);
5751 Builder.CreateUnreachable();
5752 BI->eraseFromParent();
5753 return true;
5754 }
5755 return false;
5756}
5757
5758bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5759 IRBuilder<> &Builder) {
5760 BasicBlock *BB = BI->getParent();
5761 BasicBlock *Succ = BI->getSuccessor(0);
5762
5763 // If the Terminator is the only non-phi instruction, simplify the block.
5764 // If LoopHeader is provided, check if the block or its successor is a loop
5765 // header. (This is for early invocations before loop simplify and
5766 // vectorization to keep canonical loop forms for nested loops. These blocks
5767 // can be eliminated when the pass is invoked later in the back-end.)
5768 // Note that if BB has only one predecessor then we do not introduce new
5769 // backedge, so we can eliminate BB.
5770 bool NeedCanonicalLoop =
5771 Options.NeedCanonicalLoop &&
5772 (LoopHeaders && std::distance(pred_begin(BB), pred_end(BB)) > 1 &&
5773 (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5774 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5775 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5776 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5777 return true;
5778
5779 // If the only instruction in the block is a seteq/setne comparison against a
5780 // constant, try to simplify the block.
5781 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5782 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5783 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5784 ;
5785 if (I->isTerminator() &&
5786 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI, Options))
5787 return true;
5788 }
5789
5790 // See if we can merge an empty landing pad block with another which is
5791 // equivalent.
5792 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5793 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5794 ;
5795 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5796 return true;
5797 }
5798
5799 // If this basic block is ONLY a compare and a branch, and if a predecessor
5800 // branches to us and our successor, fold the comparison into the
5801 // predecessor and use logical operations to update the incoming value
5802 // for PHI nodes in common successor.
5803 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5804 return simplifyCFG(BB, TTI, Options) | true;
5805 return false;
5806}
5807
5808static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5809 BasicBlock *PredPred = nullptr;
5810 for (auto *P : predecessors(BB)) {
5811 BasicBlock *PPred = P->getSinglePredecessor();
5812 if (!PPred || (PredPred && PredPred != PPred))
5813 return nullptr;
5814 PredPred = PPred;
5815 }
5816 return PredPred;
5817}
5818
5819bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5820 BasicBlock *BB = BI->getParent();
5821 const Function *Fn = BB->getParent();
5822 if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
5823 return false;
5824
5825 // Conditional branch
5826 if (isValueEqualityComparison(BI)) {
5827 // If we only have one predecessor, and if it is a branch on this value,
5828 // see if that predecessor totally determines the outcome of this
5829 // switch.
5830 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5831 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5832 return simplifyCFG(BB, TTI, Options) | true;
5833
5834 // This block must be empty, except for the setcond inst, if it exists.
5835 // Ignore dbg intrinsics.
5836 BasicBlock::iterator I = BB->begin();
5837 // Ignore dbg intrinsics.
5838 while (isa<DbgInfoIntrinsic>(I))
5839 ++I;
5840 if (&*I == BI) {
5841 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5842 return simplifyCFG(BB, TTI, Options) | true;
5843 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5844 ++I;
5845 // Ignore dbg intrinsics.
5846 while (isa<DbgInfoIntrinsic>(I))
5847 ++I;
5848 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5849 return simplifyCFG(BB, TTI, Options) | true;
5850 }
5851 }
5852
5853 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5854 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5855 return true;
5856
5857 // If this basic block has a single dominating predecessor block and the
5858 // dominating block's condition implies BI's condition, we know the direction
5859 // of the BI branch.
5860 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5861 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5862 if (PBI && PBI->isConditional() &&
5863 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
5864 assert(PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)(static_cast <bool> (PBI->getSuccessor(0) == BB || PBI
->getSuccessor(1) == BB) ? void (0) : __assert_fail ("PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 5864, __extension__ __PRETTY_FUNCTION__))
;
5865 bool CondIsTrue = PBI->getSuccessor(0) == BB;
5866 Optional<bool> Implication = isImpliedCondition(
5867 PBI->getCondition(), BI->getCondition(), DL, CondIsTrue);
5868 if (Implication) {
5869 // Turn this into a branch on constant.
5870 auto *OldCond = BI->getCondition();
5871 ConstantInt *CI = *Implication
5872 ? ConstantInt::getTrue(BB->getContext())
5873 : ConstantInt::getFalse(BB->getContext());
5874 BI->setCondition(CI);
5875 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5876 return simplifyCFG(BB, TTI, Options) | true;
5877 }
5878 }
5879 }
5880
5881 // If this basic block is ONLY a compare and a branch, and if a predecessor
5882 // branches to us and one of our successors, fold the comparison into the
5883 // predecessor and use logical operations to pick the right destination.
5884 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5885 return simplifyCFG(BB, TTI, Options) | true;
5886
5887 // We have a conditional branch to two blocks that are only reachable
5888 // from BI. We know that the condbr dominates the two blocks, so see if
5889 // there is any identical code in the "then" and "else" blocks. If so, we
5890 // can hoist it up to the branching block.
5891 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5892 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5893 if (HoistThenElseCodeToIf(BI, TTI))
5894 return simplifyCFG(BB, TTI, Options) | true;
5895 } else {
5896 // If Successor #1 has multiple preds, we may be able to conditionally
5897 // execute Successor #0 if it branches to Successor #1.
5898 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5899 if (Succ0TI->getNumSuccessors() == 1 &&
5900 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5901 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5902 return simplifyCFG(BB, TTI, Options) | true;
5903 }
5904 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5905 // If Successor #0 has multiple preds, we may be able to conditionally
5906 // execute Successor #1 if it branches to Successor #0.
5907 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5908 if (Succ1TI->getNumSuccessors() == 1 &&
5909 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5910 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5911 return simplifyCFG(BB, TTI, Options) | true;
5912 }
5913
5914 // If this is a branch on a phi node in the current block, thread control
5915 // through this block if any PHI node entries are constants.
5916 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5917 if (PN->getParent() == BI->getParent())
5918 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
5919 return simplifyCFG(BB, TTI, Options) | true;
5920
5921 // Scan predecessor blocks for conditional branches.
5922 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5923 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5924 if (PBI != BI && PBI->isConditional())
5925 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5926 return simplifyCFG(BB, TTI, Options) | true;
5927
5928 // Look for diamond patterns.
5929 if (MergeCondStores)
5930 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5931 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5932 if (PBI != BI && PBI->isConditional())
5933 if (mergeConditionalStores(PBI, BI, DL))
5934 return simplifyCFG(BB, TTI, Options) | true;
5935
5936 return false;
5937}
5938
5939/// Check if passing a value to an instruction will cause undefined behavior.
5940static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5941 Constant *C = dyn_cast<Constant>(V);
5942 if (!C)
5943 return false;
5944
5945 if (I->use_empty())
5946 return false;
5947
5948 if (C->isNullValue() || isa<UndefValue>(C)) {
5949 // Only look at the first use, avoid hurting compile time with long uselists
5950 User *Use = *I->user_begin();
5951
5952 // Now make sure that there are no instructions in between that can alter
5953 // control flow (eg. calls)
5954 for (BasicBlock::iterator
5955 i = ++BasicBlock::iterator(I),
5956 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5957 i != UI; ++i)
5958 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5959 return false;
5960
5961 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5962 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5963 if (GEP->getPointerOperand() == I)
5964 return passingValueIsAlwaysUndefined(V, GEP);
5965
5966 // Look through bitcasts.
5967 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5968 return passingValueIsAlwaysUndefined(V, BC);
5969
5970 // Load from null is undefined.
5971 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5972 if (!LI->isVolatile())
5973 return LI->getPointerAddressSpace() == 0;
5974
5975 // Store to null is undefined.
5976 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5977 if (!SI->isVolatile())
5978 return SI->getPointerAddressSpace() == 0 &&
5979 SI->getPointerOperand() == I;
5980
5981 // A call to null is undefined.
5982 if (auto CS = CallSite(Use))
5983 return CS.getCalledValue() == I;
5984 }
5985 return false;
5986}
5987
5988/// If BB has an incoming value that will always trigger undefined behavior
5989/// (eg. null pointer dereference), remove the branch leading here.
5990static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5991 for (PHINode &PHI : BB->phis())
5992 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
5993 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
5994 TerminatorInst *T = PHI.getIncomingBlock(i)->getTerminator();
5995 IRBuilder<> Builder(T);
5996 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5997 BB->removePredecessor(PHI.getIncomingBlock(i));
5998 // Turn uncoditional branches into unreachables and remove the dead
5999 // destination from conditional branches.
6000 if (BI->isUnconditional())
6001 Builder.CreateUnreachable();
6002 else
6003 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
6004 : BI->getSuccessor(0));
6005 BI->eraseFromParent();
6006 return true;
6007 }
6008 // TODO: SwitchInst.
6009 }
6010
6011 return false;
6012}
6013
6014bool SimplifyCFGOpt::run(BasicBlock *BB) {
6015 bool Changed = false;
6016
6017 assert(BB && BB->getParent() && "Block not embedded in function!")(static_cast <bool> (BB && BB->getParent() &&
"Block not embedded in function!") ? void (0) : __assert_fail
("BB && BB->getParent() && \"Block not embedded in function!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 6017, __extension__ __PRETTY_FUNCTION__))
;
6018 assert(BB->getTerminator() && "Degenerate basic block encountered!")(static_cast <bool> (BB->getTerminator() && "Degenerate basic block encountered!"
) ? void (0) : __assert_fail ("BB->getTerminator() && \"Degenerate basic block encountered!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Utils/SimplifyCFG.cpp"
, 6018, __extension__ __PRETTY_FUNCTION__))
;
6019
6020 // Remove basic blocks that have no predecessors (except the entry block)...
6021 // or that just have themself as a predecessor. These are unreachable.
6022 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
15
Assuming the condition is false
17
Taking false branch
6023 BB->getSinglePredecessor() == BB) {
16
Assuming the condition is false
6024 DEBUG(dbgs() << "Removing BB: \n" << *BB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("simplifycfg")) { dbgs() << "Removing BB: \n" <<
*BB; } } while (false)
;
6025 DeleteDeadBlock(BB);
6026 return true;
6027 }
6028
6029 // Check to see if we can constant propagate this terminator instruction
6030 // away...
6031 Changed |= ConstantFoldTerminator(BB, true);
6032
6033 // Check for and eliminate duplicate PHI nodes in this block.
6034 Changed |= EliminateDuplicatePHINodes(BB);
6035
6036 // Check for and remove branches that will always cause undefined behavior.
6037 Changed |= removeUndefIntroducingPredecessor(BB);
6038
6039 // Merge basic blocks into their predecessor if there is only one distinct
6040 // pred, and if there is only one distinct successor of the predecessor, and
6041 // if there are no PHI nodes.
6042 if (MergeBlockIntoPredecessor(BB))
18
Assuming the condition is false
19
Taking false branch
6043 return true;
6044
6045 if (SinkCommon && Options.SinkCommonInsts)
20
Assuming the condition is false
6046 Changed |= SinkCommonCodeFromPredecessors(BB);
6047
6048 IRBuilder<> Builder(BB);
6049
6050 // If there is a trivial two-entry PHI node in this basic block, and we can
6051 // eliminate it, do so now.
6052 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
21
Assuming 'PN' is non-null
22
Taking true branch
6053 if (PN->getNumIncomingValues() == 2)
23
Assuming the condition is true
24
Taking true branch
6054 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
25
Calling 'FoldTwoEntryPHINode'
6055
6056 Builder.SetInsertPoint(BB->getTerminator());
6057 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
6058 if (BI->isUnconditional()) {
6059 if (SimplifyUncondBranch(BI, Builder))
6060 return true;
6061 } else {
6062 if (SimplifyCondBranch(BI, Builder))
6063 return true;
6064 }
6065 } else if (auto *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6066 if (SimplifyReturn(RI, Builder))
6067 return true;
6068 } else if (auto *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6069 if (SimplifyResume(RI, Builder))
6070 return true;
6071 } else if (auto *RI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6072 if (SimplifyCleanupReturn(RI))
6073 return true;
6074 } else if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6075 if (SimplifySwitch(SI, Builder))
6076 return true;
6077 } else if (auto *UI = dyn_cast<UnreachableInst>(BB->getTerminator())) {
6078 if (SimplifyUnreachable(UI))
6079 return true;
6080 } else if (auto *IBI = dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6081 if (SimplifyIndirectBr(IBI))
6082 return true;
6083 }
6084
6085 return Changed;
6086}
6087
6088bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6089 const SimplifyCFGOptions &Options,
6090 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6091 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,
14
Calling 'SimplifyCFGOpt::run'
6092 Options)
6093 .run(BB);
6094}