jeroen.dobbelaere added a comment.
Just to give an example:
int foo(int* restrict *pA, int* restrict *pB) {
int tmp=**pB;
**pA=42;
return tmp - **pB; // **pA and **pB can refer to the same objects
}
This gives following llvm-ir code with the 'full noalias' clang version (after
optimizations):
; Function Attrs: nofree nounwind uwtable
define dso_local i32 @foo(i32** nocapture %pA, i32** nocapture readonly %pB)
local_unnamed_addr #0 !noalias !2 {
entry:
%0 = load i32*, i32** %pB, noalias_sidechannel i32** undef, align 8, !tbaa
!5, !noalias !2
%1 = tail call i32* @llvm.side.noalias.p0i32.p0i8.p0p0i32.p0p0i32.i64(i32*
%0, i8* null, i32** %pB, i32** undef, i64 0, metadata !2), !tbaa !5, !noalias !2
%2 = load i32, i32* %0, noalias_sidechannel i32* %1, align 4, !tbaa !9,
!noalias !2
%3 = load i32*, i32** %pA, noalias_sidechannel i32** undef, align 8, !tbaa
!5, !noalias !2
%4 = tail call i32* @llvm.side.noalias.p0i32.p0i8.p0p0i32.p0p0i32.i64(i32*
%3, i8* null, i32** %pA, i32** undef, i64 0, metadata !2), !tbaa !5, !noalias !2
store i32 42, i32* %3, noalias_sidechannel i32* %4, align 4, !tbaa !9,
!noalias !2
%5 = load i32, i32* %0, noalias_sidechannel i32* %1, align 4, !tbaa !9,
!noalias !2
%sub = sub nsw i32 %2, %5
ret i32 %sub
}
When adding 'noalias' to the arguments (my understanding is that this is what
the attributer would deduce, as pA and pB are only read):
; Function Attrs: nofree nounwind uwtable
define dso_local i32 @foo(i32** noalias nocapture %pA, i32** noalias
nocapture readonly %pB) local_unnamed_addr #0 !noalias !2 {
entry:
%0 = load i32*, i32** %pB, noalias_sidechannel i32** undef, align 8, !tbaa
!5, !noalias !2
%1 = tail call i32* @llvm.side.noalias.p0i32.p0i8.p0p0i32.p0p0i32.i64(i32*
%0, i8* null, i32** %pB, i32** undef, i64 0, metadata !2), !tbaa !5, !noalias !2
%2 = load i32, i32* %0, noalias_sidechannel i32* %1, align 4, !tbaa !9,
!noalias !2
%3 = load i32*, i32** %pA, noalias_sidechannel i32** undef, align 8, !tbaa
!5, !noalias !2
%4 = tail call i32* @llvm.side.noalias.p0i32.p0i8.p0p0i32.p0p0i32.i64(i32*
%3, i8* null, i32** %pA, i32** undef, i64 0, metadata !2), !tbaa !5, !noalias !2
store i32 42, i32* %3, noalias_sidechannel i32* %4, align 4, !tbaa !9,
!noalias !2
%5 = load i32, i32* %0, noalias_sidechannel i32* %1, align 4, !tbaa !9,
!noalias !2
%sub = sub nsw i32 %2, %5
ret i32 %sub
}
and reoptimizing, we now get:
; Function Attrs: nofree nounwind uwtable
define dso_local i32 @foo(i32** noalias nocapture %pA, i32** noalias
nocapture readonly %pB) local_unnamed_addr #0 !noalias !2 {
entry:
%0 = load i32*, i32** %pA, noalias_sidechannel i32** undef, align 8, !tbaa
!5, !noalias !2
%1 = tail call i32* @llvm.side.noalias.p0i32.p0i8.p0p0i32.p0p0i32.i64(i32*
%0, i8* null, i32** %pA, i32** undef, i64 0, metadata !2), !tbaa !5, !noalias !2
store i32 42, i32* %0, noalias_sidechannel i32* %1, align 4, !tbaa !9,
!noalias !2
ret i32 0
}
'%sub' is optimized away because the now the alias analyzer deduces that '%pA'
and '%pB' cannot refer to the same objects. This implies (because of restrict)
that '*%pA' and '*%pB' cannot refer to the same objects and because of that,
the 'store i32' could not have interfered with the 'load i32'.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D74935/new/
https://reviews.llvm.org/D74935
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
