From: Serge Pavlov Date: Tue, 24 Jan 2017 06:58:39 +0000 (+0000) Subject: Update domtree incrementally in loop peeling. X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=aeb6f08dc22af684eec3c868e042e6674644a859;p=llvm Update domtree incrementally in loop peeling. With this change dominator tree remains in sync after each step of loop peeling. Differential Revision: https://reviews.llvm.org/D29029 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@292895 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Transforms/Utils/LoopUnrollPeel.cpp b/lib/Transforms/Utils/LoopUnrollPeel.cpp index f021617f2ff..3cb9e6e2e02 100644 --- a/lib/Transforms/Utils/LoopUnrollPeel.cpp +++ b/lib/Transforms/Utils/LoopUnrollPeel.cpp @@ -165,7 +165,8 @@ static void cloneLoopBlocks(Loop *L, unsigned IterNumber, BasicBlock *InsertTop, BasicBlock *InsertBot, BasicBlock *Exit, SmallVectorImpl &NewBlocks, LoopBlocksDFS &LoopBlocks, ValueToValueMapTy &VMap, - ValueToValueMapTy &LVMap, LoopInfo *LI) { + ValueToValueMapTy &LVMap, DominatorTree *DT, + LoopInfo *LI) { BasicBlock *Header = L->getHeader(); BasicBlock *Latch = L->getLoopLatch(); @@ -186,6 +187,17 @@ static void cloneLoopBlocks(Loop *L, unsigned IterNumber, BasicBlock *InsertTop, ParentLoop->addBasicBlockToLoop(NewBB, *LI); VMap[*BB] = NewBB; + + // If dominator tree is available, insert nodes to represent cloned blocks. + if (DT) { + if (Header == *BB) + DT->addNewBlock(NewBB, InsertTop); + else { + DomTreeNode *IDom = DT->getNode(*BB)->getIDom(); + // VMap must contain entry for IDom, as the iteration order is RPO. + DT->addNewBlock(NewBB, cast(VMap[IDom->getBlock()])); + } + } } // Hook-up the control flow for the newly inserted blocks. @@ -199,11 +211,13 @@ static void cloneLoopBlocks(Loop *L, unsigned IterNumber, BasicBlock *InsertTop, // The backedge now goes to the "bottom", which is either the loop's real // header (for the last peeled iteration) or the copied header of the next // iteration (for every other iteration) - BranchInst *LatchBR = - cast(cast(VMap[Latch])->getTerminator()); + BasicBlock *NewLatch = cast(VMap[Latch]); + BranchInst *LatchBR = cast(NewLatch->getTerminator()); unsigned HeaderIdx = (LatchBR->getSuccessor(0) == Header ? 0 : 1); LatchBR->setSuccessor(HeaderIdx, InsertBot); LatchBR->setSuccessor(1 - HeaderIdx, Exit); + if (DT) + DT->changeImmediateDominator(InsertBot, NewLatch); // The new copy of the loop body starts with a bunch of PHI nodes // that pick an incoming value from either the preheader, or the previous @@ -359,7 +373,19 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, LoopInfo *LI, CurHeaderWeight = 1; cloneLoopBlocks(L, Iter, InsertTop, InsertBot, Exit, - NewBlocks, LoopBlocks, VMap, LVMap, LI); + NewBlocks, LoopBlocks, VMap, LVMap, DT, LI); + if (DT) { + // Latches of the cloned loops dominate over the loop exit, so idom of the + // latter is the first cloned loop body, as original PreHeader dominates + // the original loop body. + if (Iter == 0) + DT->changeImmediateDominator(Exit, cast(LVMap[Latch])); +#ifndef NDEBUG + if (VerifyDomInfo) + DT->verifyDomTree(); +#endif + } + updateBranchWeights(InsertBot, cast(VMap[LatchBR]), Iter, PeelCount, ExitWeight); @@ -405,9 +431,6 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, LoopInfo *LI, LatchBR->setMetadata(LLVMContext::MD_prof, WeightNode); } - // FIXME: Incrementally update domtree. - DT->recalculate(*L->getHeader()->getParent()); - // If the loop is nested, we changed the parent loop, update SE. if (Loop *ParentLoop = L->getParentLoop()) { SE->forgetLoop(ParentLoop); diff --git a/test/Transforms/LoopUnroll/peel-loop.ll b/test/Transforms/LoopUnroll/peel-loop.ll index 24912202238..bf0801fc760 100644 --- a/test/Transforms/LoopUnroll/peel-loop.ll +++ b/test/Transforms/LoopUnroll/peel-loop.ll @@ -1,4 +1,4 @@ -; RUN: opt < %s -S -loop-unroll -unroll-force-peel-count=3 -simplifycfg -instcombine | FileCheck %s +; RUN: opt < %s -S -loop-unroll -unroll-force-peel-count=3 -verify-dom-info -simplifycfg -instcombine | FileCheck %s ; Basic loop peeling - check that we can peel-off the first 3 loop iterations ; when explicitly requested.