1 //===- FlatternCFG.cpp - Code to perform CFG flattening -------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // Reduce conditional branches in CFG.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/ADT/SmallPtrSet.h"
14 #include "llvm/Analysis/AliasAnalysis.h"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/Analysis/ValueTracking.h"
17 #include "llvm/IR/BasicBlock.h"
18 #include "llvm/IR/IRBuilder.h"
19 #include "llvm/IR/InstrTypes.h"
20 #include "llvm/IR/Instruction.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/Value.h"
23 #include "llvm/Support/Casting.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #define DEBUG_TYPE "flatten-cfg"
38 /// Use parallel-and or parallel-or to generate conditions for
39 /// conditional branches.
40 bool FlattenParallelAndOr(BasicBlock
*BB
, IRBuilder
<> &Builder
);
42 /// If \param BB is the merge block of an if-region, attempt to merge
43 /// the if-region with an adjacent if-region upstream if two if-regions
44 /// contain identical instructions.
45 bool MergeIfRegion(BasicBlock
*BB
, IRBuilder
<> &Builder
);
47 /// Compare a pair of blocks: \p Block1 and \p Block2, which
48 /// are from two if-regions, where \p Head2 is the entry block of the 2nd
49 /// if-region. \returns true if \p Block1 and \p Block2 contain identical
50 /// instructions, and have no memory reference alias with \p Head2.
51 /// This is used as a legality check for merging if-regions.
52 bool CompareIfRegionBlock(BasicBlock
*Block1
, BasicBlock
*Block2
,
56 FlattenCFGOpt(AliasAnalysis
*AA
) : AA(AA
) {}
58 bool run(BasicBlock
*BB
);
61 } // end anonymous namespace
63 /// If \param [in] BB has more than one predecessor that is a conditional
64 /// branch, attempt to use parallel and/or for the branch condition. \returns
69 /// %cmp10 = fcmp une float %tmp1, %tmp2
70 /// br i1 %cmp10, label %if.then, label %lor.rhs
74 /// %cmp11 = fcmp une float %tmp3, %tmp4
75 /// br i1 %cmp11, label %if.then, label %ifend
77 /// if.end: // the merge block
80 /// if.then: // has two predecessors, both of them contains conditional branch.
86 /// %cmp10 = fcmp une float %tmp1, %tmp2
88 /// %cmp11 = fcmp une float %tmp3, %tmp4
89 /// %cmp12 = or i1 %cmp10, %cmp11 // parallel-or mode.
90 /// br i1 %cmp12, label %if.then, label %ifend
99 /// Current implementation handles two cases.
100 /// Case 1: BB is on the else-path.
106 /// BB3 \ | where, BB1, BB2 contain conditional branches.
107 /// \ | / BB3 contains unconditional branch.
108 /// \ | / BB4 corresponds to BB which is also the merge.
112 /// Corresponding source code:
114 /// if (a == b && c == d)
115 /// statement; // BB3
117 /// Case 2: BB is on the then-path.
122 /// \ / | where BB1, BB2 contain conditional branches.
123 /// BB => BB3 | BB3 contains unconditiona branch and corresponds
124 /// \ / to BB. BB4 is the merge.
127 /// Corresponding source code:
129 /// if (a == b || c == d)
130 /// statement; // BB3
132 /// In both cases, BB is the common successor of conditional branches.
133 /// In Case 1, BB (BB4) has an unconditional branch (BB3) as
134 /// its predecessor. In Case 2, BB (BB3) only has conditional branches
135 /// as its predecessors.
136 bool FlattenCFGOpt::FlattenParallelAndOr(BasicBlock
*BB
, IRBuilder
<> &Builder
) {
137 PHINode
*PHI
= dyn_cast
<PHINode
>(BB
->begin());
139 return false; // For simplicity, avoid cases containing PHI nodes.
141 BasicBlock
*LastCondBlock
= nullptr;
142 BasicBlock
*FirstCondBlock
= nullptr;
143 BasicBlock
*UnCondBlock
= nullptr;
146 // Check predecessors of \param BB.
147 SmallPtrSet
<BasicBlock
*, 16> Preds(pred_begin(BB
), pred_end(BB
));
148 for (BasicBlock
*Pred
: Preds
) {
149 BranchInst
*PBI
= dyn_cast
<BranchInst
>(Pred
->getTerminator());
151 // All predecessors should terminate with a branch.
155 BasicBlock
*PP
= Pred
->getSinglePredecessor();
157 if (PBI
->isUnconditional()) {
158 // Case 1: Pred (BB3) is an unconditional block, it should
159 // have a single predecessor (BB2) that is also a predecessor
160 // of \param BB (BB4) and should not have address-taken.
161 // There should exist only one such unconditional
162 // branch among the predecessors.
163 if (UnCondBlock
|| !PP
|| !Preds
.contains(PP
) ||
164 Pred
->hasAddressTaken())
171 // Only conditional branches are allowed beyond this point.
172 assert(PBI
->isConditional());
174 // Condition's unique use should be the branch instruction.
175 Value
*PC
= PBI
->getCondition();
176 if (!PC
|| !PC
->hasOneUse())
179 if (PP
&& Preds
.count(PP
)) {
180 // These are internal condition blocks to be merged from, e.g.,
181 // BB2 in both cases.
182 // Should not be address-taken.
183 if (Pred
->hasAddressTaken())
186 // Instructions in the internal condition blocks should be safe
188 for (BasicBlock::iterator BI
= Pred
->begin(), BE
= PBI
->getIterator();
190 Instruction
*CI
= &*BI
++;
191 if (isa
<PHINode
>(CI
) || !isSafeToSpeculativelyExecute(CI
))
195 // This is the condition block to be merged into, e.g. BB1 in
199 FirstCondBlock
= Pred
;
202 // Find whether BB is uniformly on the true (or false) path
203 // for all of its predecessors.
204 BasicBlock
*PS1
= PBI
->getSuccessor(0);
205 BasicBlock
*PS2
= PBI
->getSuccessor(1);
206 BasicBlock
*PS
= (PS1
== BB
) ? PS2
: PS1
;
207 int CIdx
= (PS1
== BB
) ? 0 : 1;
211 else if (CIdx
!= Idx
)
214 // PS is the successor which is not BB. Check successors to identify
215 // the last conditional branch.
216 if (!Preds
.contains(PS
)) {
218 LastCondBlock
= Pred
;
221 BranchInst
*BPS
= dyn_cast
<BranchInst
>(PS
->getTerminator());
222 if (BPS
&& BPS
->isUnconditional()) {
223 // Case 1: PS(BB3) should be an unconditional branch.
224 LastCondBlock
= Pred
;
229 if (!FirstCondBlock
|| !LastCondBlock
|| (FirstCondBlock
== LastCondBlock
))
232 Instruction
*TBB
= LastCondBlock
->getTerminator();
233 BasicBlock
*PS1
= TBB
->getSuccessor(0);
234 BasicBlock
*PS2
= TBB
->getSuccessor(1);
235 BranchInst
*PBI1
= dyn_cast
<BranchInst
>(PS1
->getTerminator());
236 BranchInst
*PBI2
= dyn_cast
<BranchInst
>(PS2
->getTerminator());
238 // If PS1 does not jump into PS2, but PS2 jumps into PS1,
239 // attempt branch inversion.
240 if (!PBI1
|| !PBI1
->isUnconditional() ||
241 (PS1
->getTerminator()->getSuccessor(0) != PS2
)) {
242 // Check whether PS2 jumps into PS1.
243 if (!PBI2
|| !PBI2
->isUnconditional() ||
244 (PS2
->getTerminator()->getSuccessor(0) != PS1
))
247 // Do branch inversion.
248 BasicBlock
*CurrBlock
= LastCondBlock
;
249 bool EverChanged
= false;
250 for (; CurrBlock
!= FirstCondBlock
;
251 CurrBlock
= CurrBlock
->getSinglePredecessor()) {
252 auto *BI
= cast
<BranchInst
>(CurrBlock
->getTerminator());
253 auto *CI
= dyn_cast
<CmpInst
>(BI
->getCondition());
257 CmpInst::Predicate Predicate
= CI
->getPredicate();
258 // Canonicalize icmp_ne -> icmp_eq, fcmp_one -> fcmp_oeq
259 if ((Predicate
== CmpInst::ICMP_NE
) || (Predicate
== CmpInst::FCMP_ONE
)) {
260 CI
->setPredicate(ICmpInst::getInversePredicate(Predicate
));
261 BI
->swapSuccessors();
268 // PS1 must have a conditional branch.
269 if (!PBI1
|| !PBI1
->isUnconditional())
272 // PS2 should not contain PHI node.
273 PHI
= dyn_cast
<PHINode
>(PS2
->begin());
277 // Do the transformation.
279 BranchInst
*PBI
= cast
<BranchInst
>(FirstCondBlock
->getTerminator());
280 bool Iteration
= true;
281 IRBuilder
<>::InsertPointGuard
Guard(Builder
);
282 Value
*PC
= PBI
->getCondition();
285 CB
= PBI
->getSuccessor(1 - Idx
);
286 // Delete the conditional branch.
287 FirstCondBlock
->back().eraseFromParent();
288 FirstCondBlock
->splice(FirstCondBlock
->end(), CB
);
289 PBI
= cast
<BranchInst
>(FirstCondBlock
->getTerminator());
290 Value
*CC
= PBI
->getCondition();
292 Builder
.SetInsertPoint(PBI
);
295 // Case 2, use parallel or.
296 NC
= Builder
.CreateOr(PC
, CC
);
298 // Case 1, use parallel and.
299 NC
= Builder
.CreateAnd(PC
, CC
);
301 PBI
->replaceUsesOfWith(CC
, NC
);
303 if (CB
== LastCondBlock
)
305 // Remove internal conditional branches.
306 CB
->dropAllReferences();
307 // make CB unreachable and let downstream to delete the block.
308 new UnreachableInst(CB
->getContext(), CB
);
311 LLVM_DEBUG(dbgs() << "Use parallel and/or in:\n" << *FirstCondBlock
);
315 /// Compare blocks from two if-regions, where \param Head2 is the entry of the
316 /// 2nd if-region. \param Block1 is a block in the 1st if-region to compare.
317 /// \param Block2 is a block in the 2nd if-region to compare. \returns true if
318 /// Block1 and Block2 have identical instructions and do not have
319 /// memory reference alias with Head2.
320 bool FlattenCFGOpt::CompareIfRegionBlock(BasicBlock
*Block1
, BasicBlock
*Block2
,
322 Instruction
*PTI2
= Head2
->getTerminator();
323 Instruction
*PBI2
= &Head2
->front();
325 // Check whether instructions in Block1 and Block2 are identical
326 // and do not alias with instructions in Head2.
327 BasicBlock::iterator iter1
= Block1
->begin();
328 BasicBlock::iterator end1
= Block1
->getTerminator()->getIterator();
329 BasicBlock::iterator iter2
= Block2
->begin();
330 BasicBlock::iterator end2
= Block2
->getTerminator()->getIterator();
339 if (!iter1
->isIdenticalTo(&*iter2
))
342 // Illegal to remove instructions with side effects except
343 // non-volatile stores.
344 if (iter1
->mayHaveSideEffects()) {
345 Instruction
*CurI
= &*iter1
;
346 StoreInst
*SI
= dyn_cast
<StoreInst
>(CurI
);
347 if (!SI
|| SI
->isVolatile())
351 // For simplicity and speed, data dependency check can be
352 // avoided if read from memory doesn't exist.
353 if (iter1
->mayReadFromMemory())
356 if (iter1
->mayWriteToMemory()) {
357 for (BasicBlock::iterator
BI(PBI2
), BE(PTI2
); BI
!= BE
; ++BI
) {
358 if (BI
->mayReadFromMemory() || BI
->mayWriteToMemory()) {
359 // Check alias with Head2.
360 if (!AA
|| !AA
->isNoAlias(&*iter1
, &*BI
))
372 /// Check whether \param BB is the merge block of a if-region. If yes, check
373 /// whether there exists an adjacent if-region upstream, the two if-regions
374 /// contain identical instructions and can be legally merged. \returns true if
375 /// the two if-regions are merged.
404 /// We always take the form of the first if-region. This means that if the
405 /// statement in the first if-region, is in the "then-path", while in the second
406 /// if-region it is in the "else-path", then we convert the second to the first
407 /// form, by inverting the condition and the branch successors. The same
408 /// approach goes for the opposite case.
409 bool FlattenCFGOpt::MergeIfRegion(BasicBlock
*BB
, IRBuilder
<> &Builder
) {
410 // We cannot merge the if-region if the merge point has phi nodes.
411 if (isa
<PHINode
>(BB
->front()))
414 BasicBlock
*IfTrue2
, *IfFalse2
;
415 BranchInst
*DomBI2
= GetIfCondition(BB
, IfTrue2
, IfFalse2
);
418 Instruction
*CInst2
= dyn_cast
<Instruction
>(DomBI2
->getCondition());
422 BasicBlock
*SecondEntryBlock
= CInst2
->getParent();
423 if (SecondEntryBlock
->hasAddressTaken())
426 BasicBlock
*IfTrue1
, *IfFalse1
;
427 BranchInst
*DomBI1
= GetIfCondition(SecondEntryBlock
, IfTrue1
, IfFalse1
);
430 Instruction
*CInst1
= dyn_cast
<Instruction
>(DomBI1
->getCondition());
434 BasicBlock
*FirstEntryBlock
= CInst1
->getParent();
435 // Don't die trying to process degenerate/unreachable code.
436 if (FirstEntryBlock
== SecondEntryBlock
)
439 // Either then-path or else-path should be empty.
440 bool InvertCond2
= false;
441 BinaryOperator::BinaryOps CombineOp
;
442 if (IfFalse1
== FirstEntryBlock
) {
443 // The else-path is empty, so we must use "or" operation to combine the
445 CombineOp
= BinaryOperator::Or
;
446 if (IfFalse2
!= SecondEntryBlock
) {
447 if (IfTrue2
!= SecondEntryBlock
)
451 std::swap(IfTrue2
, IfFalse2
);
454 if (!CompareIfRegionBlock(IfTrue1
, IfTrue2
, SecondEntryBlock
))
456 } else if (IfTrue1
== FirstEntryBlock
) {
457 // The then-path is empty, so we must use "and" operation to combine the
459 CombineOp
= BinaryOperator::And
;
460 if (IfTrue2
!= SecondEntryBlock
) {
461 if (IfFalse2
!= SecondEntryBlock
)
465 std::swap(IfTrue2
, IfFalse2
);
468 if (!CompareIfRegionBlock(IfFalse1
, IfFalse2
, SecondEntryBlock
))
473 Instruction
*PTI2
= SecondEntryBlock
->getTerminator();
474 Instruction
*PBI2
= &SecondEntryBlock
->front();
476 // Check whether \param SecondEntryBlock has side-effect and is safe to
478 for (BasicBlock::iterator
BI(PBI2
), BE(PTI2
); BI
!= BE
; ++BI
) {
479 Instruction
*CI
= &*BI
;
480 if (isa
<PHINode
>(CI
) || CI
->mayHaveSideEffects() ||
481 !isSafeToSpeculativelyExecute(CI
))
485 // Merge \param SecondEntryBlock into \param FirstEntryBlock.
486 FirstEntryBlock
->back().eraseFromParent();
487 FirstEntryBlock
->splice(FirstEntryBlock
->end(), SecondEntryBlock
);
488 BranchInst
*PBI
= cast
<BranchInst
>(FirstEntryBlock
->getTerminator());
489 assert(PBI
->getCondition() == CInst2
);
490 BasicBlock
*SaveInsertBB
= Builder
.GetInsertBlock();
491 BasicBlock::iterator SaveInsertPt
= Builder
.GetInsertPoint();
492 Builder
.SetInsertPoint(PBI
);
494 InvertBranch(PBI
, Builder
);
496 Value
*NC
= Builder
.CreateBinOp(CombineOp
, CInst1
, PBI
->getCondition());
497 PBI
->replaceUsesOfWith(PBI
->getCondition(), NC
);
498 Builder
.SetInsertPoint(SaveInsertBB
, SaveInsertPt
);
501 if (IfTrue1
!= FirstEntryBlock
) {
502 IfTrue1
->dropAllReferences();
503 IfTrue1
->eraseFromParent();
507 if (IfFalse1
!= FirstEntryBlock
) {
508 IfFalse1
->dropAllReferences();
509 IfFalse1
->eraseFromParent();
512 // Remove \param SecondEntryBlock
513 SecondEntryBlock
->dropAllReferences();
514 SecondEntryBlock
->eraseFromParent();
515 LLVM_DEBUG(dbgs() << "If conditions merged into:\n" << *FirstEntryBlock
);
519 bool FlattenCFGOpt::run(BasicBlock
*BB
) {
520 assert(BB
&& BB
->getParent() && "Block not embedded in function!");
521 assert(BB
->getTerminator() && "Degenerate basic block encountered!");
523 IRBuilder
<> Builder(BB
);
525 if (FlattenParallelAndOr(BB
, Builder
) || MergeIfRegion(BB
, Builder
))
530 /// FlattenCFG - This function is used to flatten a CFG. For
531 /// example, it uses parallel-and and parallel-or mode to collapse
532 /// if-conditions and merge if-regions with identical statements.
533 bool llvm::FlattenCFG(BasicBlock
*BB
, AAResults
*AA
) {
534 return FlattenCFGOpt(AA
).run(BB
);