From 069d6f3396856655d5d4ba155ee16eb0209d38b0 Mon Sep 17 00:00:00 2001 From: Dan Gohman Date: Tue, 2 Mar 2010 01:59:21 +0000 Subject: [PATCH] Non-affine post-inc SCEV expansions have more code which must be emitted after the increment. Make sure the insert position reflects this. This fixes PR6453. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@97537 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Analysis/ScalarEvolutionExpander.cpp | 2 +- lib/Transforms/Scalar/LoopStrengthReduce.cpp | 8 +++- .../LoopStrengthReduce/nonlinear-postinc.ll | 44 +++++++++++++++++++ 3 files changed, 51 insertions(+), 3 deletions(-) create mode 100644 test/Transforms/LoopStrengthReduce/nonlinear-postinc.ll diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp index 9aa8bfbdb17..b6c7ce64481 100644 --- a/lib/Analysis/ScalarEvolutionExpander.cpp +++ b/lib/Analysis/ScalarEvolutionExpander.cpp @@ -1087,7 +1087,7 @@ Value *SCEVExpander::expand(const SCEV *S) { // If the SCEV is computable at this level, insert it into the header // after the PHIs (and after any other instructions that we've inserted // there) so that it is guaranteed to dominate any user inside the loop. - if (L && S->hasComputableLoopEvolution(L)) + if (L && S->hasComputableLoopEvolution(L) && L != PostIncLoop) InsertPt = L->getHeader()->getFirstNonPHI(); while (isInsertedInstruction(InsertPt)) InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp index 8bbf19345ab..14274296de7 100644 --- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp +++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp @@ -2793,8 +2793,12 @@ Value *LSRInstance::Expand(const LSRFixup &LF, if (Instruction *I = dyn_cast(cast(LF.UserInst)->getOperand(1))) Inputs.push_back(I); - if (LF.PostIncLoop && !L->contains(LF.UserInst)) - Inputs.push_back(L->getLoopLatch()->getTerminator()); + if (LF.PostIncLoop) { + if (!L->contains(LF.UserInst)) + Inputs.push_back(L->getLoopLatch()->getTerminator()); + else + Inputs.push_back(IVIncInsertPos); + } // Then, climb up the immediate dominator tree as far as we can go while // still being dominated by the input positions. diff --git a/test/Transforms/LoopStrengthReduce/nonlinear-postinc.ll b/test/Transforms/LoopStrengthReduce/nonlinear-postinc.ll new file mode 100644 index 00000000000..1e63770553d --- /dev/null +++ b/test/Transforms/LoopStrengthReduce/nonlinear-postinc.ll @@ -0,0 +1,44 @@ +; RUN: opt < %s -loop-reduce +; PR6453 + +target datalayout = "e-p:64:64:64" + +define void @_ZNK15PolynomialSpaceILi3EE13compute_indexEjRA3_j() nounwind { +entry: + br label %bb6 + +bb6: + %t4 = phi i32 [ 0, %entry ], [ %t3, %bb5 ] + %t16 = sub i32 undef, %t4 + %t25 = sub i32 undef, %t4 + %t26 = add i32 undef, %t25 + br label %bb4 + +bb4: + %t2 = phi i32 [ %t1, %bb3 ], [ 0, %bb6 ] + %t17 = mul i32 %t2, %t16 + %t18 = zext i32 %t2 to i33 + %t19 = add i32 %t2, -1 + %t20 = zext i32 %t19 to i33 + %t21 = mul i33 %t18, %t20 + %t22 = lshr i33 %t21, 1 + %t23 = trunc i33 %t22 to i32 + %t24 = sub i32 %t17, %t23 + %t27 = add i32 %t24, %t26 + br i1 false, label %bb1, label %bb5 + +bb1: + %t = icmp ugt i32 %t27, undef + br i1 %t, label %bb2, label %bb3 + +bb3: + %t1 = add i32 %t2, 1 + br label %bb4 + +bb5: + %t3 = add i32 %t4, 1 + br label %bb6 + +bb2: + ret void +} -- 2.34.1