[llvm-branch-commits] [llvm] 4e692c9 - [LoopIdiom] Don't transform loop into memmove when load from body has more than one use
Dawid Jurczak via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Aug 25 08:30:28 PDT 2021
Author: Dawid Jurczak
Date: 2021-08-25T16:19:10+02:00
New Revision: 4e692c97a8c17a6eaf60f81327edff357086aa74
URL: https://github.com/llvm/llvm-project/commit/4e692c97a8c17a6eaf60f81327edff357086aa74
DIFF: https://github.com/llvm/llvm-project/commit/4e692c97a8c17a6eaf60f81327edff357086aa74.diff
LOG: [LoopIdiom] Don't transform loop into memmove when load from body has more than one use
This change fixes issue found by Markus: https://reviews.llvm.org/rG11338e998df1
Before this patch following code was transformed to memmove:
for (int i = 15; i >= 1; i--) {
p[i] = p[i-1];
sum += p[i-1];
}
However load from p[i-1] is used not only by store to p[i] but also by sum computation.
Therefore we cannot emit memmove in loop header.
Differential Revision: https://reviews.llvm.org/D107964
(cherry picked from commit bdcf04246c401aec9bdddf32fabc99fa4834a477)
Added:
Modified:
llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
llvm/test/Transforms/LoopIdiom/basic.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
index 3d60e205b0023..a153f393448c4 100644
--- a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
@@ -1247,6 +1247,11 @@ bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
StoreSize, *AA, Stores);
if (UseMemMove) {
+ // For memmove case it's not enough to guarantee that loop doesn't access
+ // TheStore and TheLoad. Additionally we need to make sure that TheStore is
+ // the only user of TheLoad.
+ if (!TheLoad->hasOneUse())
+ return Changed;
Stores.insert(TheLoad);
if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
BECount, StoreSize, *AA, Stores)) {
diff --git a/llvm/test/Transforms/LoopIdiom/basic.ll b/llvm/test/Transforms/LoopIdiom/basic.ll
index ea6d607c8ef12..c8c6d187a0ccf 100644
--- a/llvm/test/Transforms/LoopIdiom/basic.ll
+++ b/llvm/test/Transforms/LoopIdiom/basic.ll
@@ -1300,14 +1300,10 @@ for.end: ; preds = %for.body, %entry
ret void
}
-;; FIXME: Do not form memmove when load has more than one use.
+;; Do not form memmove when load has more than one use.
define i32 @do_not_form_memmove5(i32* %p) {
; CHECK-LABEL: @do_not_form_memmove5(
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[P2:%.*]] = bitcast i32* [[P:%.*]] to i8*
-; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i32, i32* [[P:%.*]], i64 1
-; CHECK-NEXT: [[SCEVGEP1:%.*]] = bitcast i32* [[SCEVGEP]] to i8*
-; CHECK-NEXT: call void @llvm.memmove.p0i8.p0i8.i64(i8* align 4 [[SCEVGEP1]], i8* align 4 [[P2]], i64 60, i1 false)
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.cond.cleanup:
; CHECK-NEXT: [[ADD_LCSSA:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ]
@@ -1321,6 +1317,7 @@ define i32 @do_not_form_memmove5(i32* %p) {
; CHECK-NEXT: [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[INDEX]] to i64
; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[P:%.*]], i64 [[IDXPROM]]
+; CHECK-NEXT: store i32 [[TMP1]], i32* [[ARRAYIDX2]], align 4
; CHECK-NEXT: [[ADD]] = add nsw i32 [[TMP1]], [[SUM:%.*]]
; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i32 [[INDEX]], 1
; CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY]], label [[FOR_COND_CLEANUP]]
More information about the llvm-branch-commits
mailing list