From: Dan Gohman Date: Thu, 23 Apr 2009 15:16:49 +0000 (+0000) Subject: Change SCEVExpander's expandCodeFor to provide more flexibility X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=752ec7da506f5d41c08bd37e195750b57550ce68;p=oota-llvm.git Change SCEVExpander's expandCodeFor to provide more flexibility with the persistent insertion point, and change IndVars to make use of it. This fixes a bug where IndVars was holding on to a stale insertion point and forcing the SCEVExpander to continue to use it. This fixes PR4038. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@69892 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/include/llvm/Analysis/ScalarEvolutionExpander.h b/include/llvm/Analysis/ScalarEvolutionExpander.h index 6e2db29f013..b6b45788ea7 100644 --- a/include/llvm/Analysis/ScalarEvolutionExpander.h +++ b/include/llvm/Analysis/ScalarEvolutionExpander.h @@ -71,13 +71,23 @@ namespace llvm { InsertedInstructions.insert(I); } + void setInsertionPoint(BasicBlock::iterator NewIP) { InsertPt = NewIP; } + BasicBlock::iterator getInsertionPoint() const { return InsertPt; } + /// expandCodeFor - Insert code to directly compute the specified SCEV + /// expression into the program. The inserted code is inserted into the + /// SCEVExpander's current insertion point. + Value *expandCodeFor(SCEVHandle SH, const Type *Ty); + /// expandCodeFor - Insert code to directly compute the specified SCEV /// expression into the program. The inserted code is inserted into the /// specified block. Value *expandCodeFor(SCEVHandle SH, const Type *Ty, - BasicBlock::iterator IP); + BasicBlock::iterator IP) { + setInsertionPoint(IP); + return expandCodeFor(SH, Ty); + } /// InsertCastOfTo - Insert a cast of V to the specified type, doing what /// we can to share the casts. diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp index 80b47d89240..9676ea20af5 100644 --- a/lib/Analysis/ScalarEvolutionExpander.cpp +++ b/lib/Analysis/ScalarEvolutionExpander.cpp @@ -332,12 +332,10 @@ Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { return LHS; } -Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty, - BasicBlock::iterator IP) { +Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty) { // Expand the code for this SCEV. assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && "non-trivial casts should be done with the SCEVs directly!"); - InsertPt = IP; Value *V = expand(SH); return InsertNoopCastOfTo(V, Ty); } diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp index 63ef021eb31..cc3919da846 100644 --- a/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -543,8 +543,7 @@ static Value *getSignExtendedTruncVar(const SCEVAddRecExpr *AR, ScalarEvolution *SE, const Type *LargestType, Loop *L, const Type *myType, - SCEVExpander &Rewriter, - BasicBlock::iterator InsertPt) { + SCEVExpander &Rewriter) { SCEVHandle ExtendedStart = SE->getSignExtendExpr(AR->getStart(), LargestType); SCEVHandle ExtendedStep = @@ -553,15 +552,14 @@ static Value *getSignExtendedTruncVar(const SCEVAddRecExpr *AR, SE->getAddRecExpr(ExtendedStart, ExtendedStep, L); if (LargestType != myType) ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType); - return Rewriter.expandCodeFor(ExtendedAddRec, myType, InsertPt); + return Rewriter.expandCodeFor(ExtendedAddRec, myType); } static Value *getZeroExtendedTruncVar(const SCEVAddRecExpr *AR, ScalarEvolution *SE, const Type *LargestType, Loop *L, const Type *myType, - SCEVExpander &Rewriter, - BasicBlock::iterator InsertPt) { + SCEVExpander &Rewriter) { SCEVHandle ExtendedStart = SE->getZeroExtendExpr(AR->getStart(), LargestType); SCEVHandle ExtendedStep = @@ -570,7 +568,7 @@ static Value *getZeroExtendedTruncVar(const SCEVAddRecExpr *AR, SE->getAddRecExpr(ExtendedStart, ExtendedStep, L); if (LargestType != myType) ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType); - return Rewriter.expandCodeFor(ExtendedAddRec, myType, InsertPt); + return Rewriter.expandCodeFor(ExtendedAddRec, myType); } /// allUsesAreSameTyped - See whether all Uses of I are instructions @@ -699,6 +697,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // recurrences in terms of the induction variable. Start with the auxillary // induction variables, and recursively rewrite any of their uses. BasicBlock::iterator InsertPt = Header->getFirstNonPHI(); + Rewriter.setInsertionPoint(InsertPt); // If there were induction variables of other sizes, cast the primary // induction variable to the right size for them, avoiding the need for the @@ -718,7 +717,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { while (!IndVars.empty()) { PHINode *PN = IndVars.back().first; const SCEVAddRecExpr *AR = cast(IndVars.back().second); - Value *NewVal = Rewriter.expandCodeFor(AR, PN->getType(), InsertPt); + Value *NewVal = Rewriter.expandCodeFor(AR, PN->getType()); DOUT << "INDVARS: Rewrote IV '" << *AR << "' " << *PN << " into = " << *NewVal << "\n"; NewVal->takeName(PN); @@ -732,7 +731,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { Instruction *UInst = dyn_cast(*UI); if (UInst && isa(UInst) && NoSignedWrap) { Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, L, - UInst->getType(), Rewriter, InsertPt); + UInst->getType(), Rewriter); UInst->replaceAllUsesWith(TruncIndVar); DeadInsts.insert(UInst); } @@ -753,8 +752,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { SExtInst* oldSext = dyn_cast(UInst->use_begin()); uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits(); Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, - L, oldSext->getType(), Rewriter, - InsertPt); + L, oldSext->getType(), Rewriter); APInt APnewAddRHS = APInt(AddRHS->getValue()).sext(newBitSize); if (newBitSize > truncSize) APnewAddRHS = APnewAddRHS.trunc(truncSize); @@ -784,8 +782,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { SExtInst* oldSext = dyn_cast(UInst->use_begin()); uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits(); Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, - L, oldSext->getType(), Rewriter, - InsertPt); + L, oldSext->getType(), Rewriter); APInt APnewOrRHS = APInt(RHS->getValue()).sext(newBitSize); if (newBitSize > truncSize) APnewOrRHS = APnewOrRHS.trunc(truncSize); @@ -805,7 +802,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // A zext of a signed variable known not to overflow is still safe. if (UInst && isa(UInst) && (NoUnsignedWrap || NoSignedWrap)) { Value *TruncIndVar = getZeroExtendedTruncVar(AR, SE, LargestType, L, - UInst->getType(), Rewriter, InsertPt); + UInst->getType(), Rewriter); UInst->replaceAllUsesWith(TruncIndVar); DeadInsts.insert(UInst); } @@ -822,7 +819,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { ZExtInst* oldZext = dyn_cast(UInst->use_begin()); uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits(); Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, - L, oldZext->getType(), Rewriter, InsertPt); + L, oldZext->getType(), Rewriter); APInt APnewAndRHS = APInt(AndRHS->getValue()).zext(newBitSize); if (newBitSize > truncSize) APnewAndRHS = APnewAndRHS.trunc(truncSize); @@ -858,7 +855,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { ZExtInst* oldZext = dyn_cast(UInst2->use_begin()); uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits(); Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, - L, oldZext->getType(), Rewriter, InsertPt); + L, oldZext->getType(), Rewriter); ConstantInt* AndRHS = dyn_cast(UInst2->getOperand(1)); APInt APnewAddRHS = APInt(AddRHS->getValue()).zext(newBitSize); if (newBitSize > truncSize) diff --git a/test/Transforms/IndVarSimplify/casted-argument.ll b/test/Transforms/IndVarSimplify/casted-argument.ll index ae41e3a8b7c..0a14c3e787d 100644 --- a/test/Transforms/IndVarSimplify/casted-argument.ll +++ b/test/Transforms/IndVarSimplify/casted-argument.ll @@ -1,5 +1,6 @@ ; RUN: llvm-as < %s | opt -indvars -disable-output ; PR4009 +; PR4038 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32" target triple = "i386-pc-linux-gnu" @@ -21,4 +22,27 @@ loop: ; preds = %loop, %if.else br label %loop } +define void @safe_bcopy_4038(i8* %from, i8* %to, i32 %size) nounwind { +entry: + br i1 false, label %if.else, label %if.then12 + +if.then12: ; preds = %entry + ret void + +if.else: ; preds = %entry + %sub.ptr.rhs.cast40 = ptrtoint i8* %from to i32 ; [#uses=1] + br label %if.end54 + +if.end54: ; preds = %if.end54, %if.else + %sub.ptr4912.pn = phi i8* [ %sub.ptr4912, %if.end54 ], [ null, %if.else ] ; [#uses=1] + %sub.ptr7 = phi i8* [ %sub.ptr, %if.end54 ], [ null, %if.else ] ; [#uses=2] + %sub.ptr.rhs.cast46.pn = ptrtoint i8* %from to i32 ; [#uses=1] + %sub.ptr.lhs.cast45.pn = ptrtoint i8* %to to i32 ; [#uses=1] + %sub.ptr.sub47.pn = sub i32 %sub.ptr.rhs.cast46.pn, %sub.ptr.lhs.cast45.pn ; [#uses=1] + %sub.ptr4912 = getelementptr i8* %sub.ptr4912.pn, i32 %sub.ptr.sub47.pn ; [#uses=2] + tail call void @bcopy(i8* %sub.ptr4912, i8* %sub.ptr7, i32 0) nounwind + %sub.ptr = getelementptr i8* %sub.ptr7, i32 %sub.ptr.rhs.cast40 ; [#uses=1] + br label %if.end54 +} + declare void @bcopy(i8* nocapture) nounwind