456ec1c2f4
We are not handling 'S' scalar dependencies correctly and have at least the following miscompiles related to that: [LoopInterchange] incorrect handling of scalar dependencies and dependence vectors starting with ">" #54176 [LoopInterchange] Interchange breaks program correctness #46867 [LoopInterchange] Loops should not interchanged due to dependencies #47259 [LoopInterchange] Loops should not interchanged due to control flow #47401 This patch does no longer insert the "S" dependency/direction into the dependency matrix, so a dependency is never "S". We seem to have forgotten what the exact meaning is of this dependency type, and don't see why it should be treated differently. We prefer correctness over incorrect and more aggressive results. I.e., this prevents the miscompiles at the expense of handling less cases, i.e. making interchange more pessimistic. However, some of the cases that are now rejected for dependence analysis reasons, were rejected before too but for other reasons (e.g. profitability). So at least for the llvm regression tests, the number of regression are very reasonable. This should be a stopgap. We would like to get interchange enabled by default and thus prefer correctness over unsafe transforms, and later see if we can get solve the regressions.
46 lines
1.8 KiB
LLVM
46 lines
1.8 KiB
LLVM
; REQUIRES: asserts
|
|
; RUN: opt < %s -passes=loop-interchange -S -debug 2>&1 | FileCheck %s
|
|
|
|
; CHECK: Dependency matrix before interchange:
|
|
; CHECK-NEXT: I I
|
|
; CHECK-NEXT: = *
|
|
; CHECK-NEXT: < *
|
|
; CHECK-NEXT: Processing InnerLoopId
|
|
|
|
; This example is taken from github issue #54176
|
|
;
|
|
define void @foo(i32 noundef %n, i32 noundef %m, ptr nocapture noundef %aa, ptr nocapture noundef readonly %bb, ptr nocapture noundef writeonly %cc) {
|
|
entry:
|
|
%arrayidx7 = getelementptr inbounds i8, ptr %aa, i64 512
|
|
br label %for.cond1.preheader
|
|
|
|
for.cond1.preheader:
|
|
%indvars.iv32 = phi i64 [ 1, %entry ], [ %indvars.iv.next33, %for.cond.cleanup3 ]
|
|
%0 = add nsw i64 %indvars.iv32, -1
|
|
%arrayidx9 = getelementptr inbounds [128 x float], ptr %arrayidx7, i64 0, i64 %0
|
|
%arrayidx12 = getelementptr inbounds [128 x float], ptr %arrayidx7, i64 0, i64 %indvars.iv32
|
|
br label %for.body4
|
|
|
|
for.cond.cleanup:
|
|
ret void
|
|
|
|
for.cond.cleanup3:
|
|
%indvars.iv.next33 = add nuw nsw i64 %indvars.iv32, 1
|
|
%exitcond36 = icmp ne i64 %indvars.iv.next33, 128
|
|
br i1 %exitcond36, label %for.cond1.preheader, label %for.cond.cleanup
|
|
|
|
for.body4:
|
|
%indvars.iv = phi i64 [ 1, %for.cond1.preheader ], [ %indvars.iv.next, %for.body4 ]
|
|
%arrayidx6 = getelementptr inbounds [128 x float], ptr %bb, i64 %indvars.iv, i64 %indvars.iv32
|
|
%1 = load float, ptr %arrayidx6, align 4
|
|
%2 = load float, ptr %arrayidx9, align 4
|
|
%add = fadd fast float %2, %1
|
|
store float %add, ptr %arrayidx9, align 4
|
|
%3 = load float, ptr %arrayidx12, align 4
|
|
%arrayidx16 = getelementptr inbounds [128 x float], ptr %cc, i64 %indvars.iv, i64 %indvars.iv32
|
|
store float %3, ptr %arrayidx16, align 4
|
|
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
|
|
%exitcond = icmp ne i64 %indvars.iv.next, 1024
|
|
br i1 %exitcond, label %for.body4, label %for.cond.cleanup3
|
|
}
|