diff options
author | Dan Gohman <gohman@apple.com> | 2010-02-22 03:59:54 +0000 |
---|---|---|
committer | Dan Gohman <gohman@apple.com> | 2010-02-22 03:59:54 +0000 |
commit | 278f958c61e973975bf63b7a82f3d8322ea16c00 (patch) | |
tree | f7d6053e91ae33afe8c3b62de9df2dc0dbb24f24 | |
parent | 2444080ca4f1f63d647272650aae874360c604cd (diff) | |
download | external_llvm-278f958c61e973975bf63b7a82f3d8322ea16c00.zip external_llvm-278f958c61e973975bf63b7a82f3d8322ea16c00.tar.gz external_llvm-278f958c61e973975bf63b7a82f3d8322ea16c00.tar.bz2 |
When emitting an instruction which depends on both a post-incremented
induction variable value and a loop-variant value, don't force the
insert position to be at the post-increment position, because it may
not be dominated by the loop-variant value. This fixes a
use-before-def problem noticed on PPC.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96774 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Transforms/Scalar/LoopStrengthReduce.cpp | 6 | ||||
-rw-r--r-- | test/CodeGen/PowerPC/lsr-postinc-pos.ll | 32 |
2 files changed, 36 insertions, 2 deletions
diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp index cf3d702..a4bc017 100644 --- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp +++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp @@ -2866,8 +2866,10 @@ Value *LSRInstance::Expand(const LSRFixup &LF, if (AR->getLoop() == LF.PostIncLoop) { Reg = SE.getAddExpr(Reg, AR->getStepRecurrence(SE)); // If the user is inside the loop, insert the code after the increment - // so that it is dominated by its operand. - if (L->contains(LF.UserInst)) + // so that it is dominated by its operand. If the original insert point + // was already dominated by the increment, keep it, because there may + // be loop-variant operands that need to be respected also. + if (L->contains(LF.UserInst) && !DT.dominates(IVIncInsertPos, IP)) IP = IVIncInsertPos; break; } diff --git a/test/CodeGen/PowerPC/lsr-postinc-pos.ll b/test/CodeGen/PowerPC/lsr-postinc-pos.ll new file mode 100644 index 0000000..f441e42 --- /dev/null +++ b/test/CodeGen/PowerPC/lsr-postinc-pos.ll @@ -0,0 +1,32 @@ +; RUN: llc < %s -print-lsr-output |& FileCheck %s + +; The icmp is a post-inc use, and the increment is in %bb11, but the +; scevgep needs to be inserted in %bb so that it is dominated by %t. + +; CHECK: %t = load i8** undef +; CHECK: %scevgep = getelementptr i8* %t, i32 %lsr.iv.next +; CHECK: %c1 = icmp ult i8* %scevgep, undef + +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-f128:64:128-n32" +target triple = "powerpc-apple-darwin9" + +define void @foo() nounwind { +entry: + br label %bb11 + +bb11: + %i = phi i32 [ 0, %entry ], [ %i.next, %bb ] ; <i32> [#uses=3] + %ii = shl i32 %i, 2 ; <i32> [#uses=1] + %c0 = icmp eq i32 %i, undef ; <i1> [#uses=1] + br i1 %c0, label %bb13, label %bb + +bb: + %t = load i8** undef, align 16 ; <i8*> [#uses=1] + %p = getelementptr i8* %t, i32 %ii ; <i8*> [#uses=1] + %c1 = icmp ult i8* %p, undef ; <i1> [#uses=1] + %i.next = add i32 %i, 1 ; <i32> [#uses=1] + br i1 %c1, label %bb11, label %bb13 + +bb13: + unreachable +} |