[LoopVectorize] Miscompile with iteration-local scoped alias metadata

[nikic]

C reproducer: https://clang.godbolt.org/z/f6rE4jvhM

IR reproducer: https://llvm.godbolt.org/z/7anbPon8K

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define hidden void @test(ptr %arg, i64 %num) {
entry:
  %icmp = icmp ult i64 %num, 2
  br i1 %icmp, label %exit, label %preheader

preheader:
  %arg.1 = getelementptr inbounds i8, ptr %arg, i64 1
  %end = add i64 %num, -2
  br label %loop

loop:
  %prev.ptr = phi ptr [ %cur.ptr, %loop ], [ %arg, %preheader ]
  %iv = phi i64 [ %iv.next, %loop ], [ 0, %preheader ]
  %cur.ptr = getelementptr inbounds i8, ptr %arg.1, i64 %iv
  call void @llvm.experimental.noalias.scope.decl(metadata !0)
  call void @llvm.experimental.noalias.scope.decl(metadata !3)
  %load.prev = load i8, ptr %prev.ptr, align 1, !alias.scope !0, !noalias !3
  %load.cur = load i8, ptr %cur.ptr, align 1, !alias.scope !3
  %add = add i8 %load.cur, %load.prev
  store i8 %add, ptr %cur.ptr, align 1, !alias.scope !3
  %iv.next = add nuw i64 %iv, 1
  %cmp = icmp eq i64 %iv, %end
  br i1 %cmp, label %exit, label %loop

exit:
  ret void
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: readwrite)
declare void @llvm.experimental.noalias.scope.decl(metadata) #0

attributes #0 = { nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: readwrite) }

!0 = !{!1}
!1 = distinct !{!1, !2}
!2 = distinct !{!2}
!3 = !{!4}
!4 = distinct !{!4, !5}
!5 = distinct !{!5}

The current and previous value are only noalias within the iteration (per the location of the @llvm.experimental.noalias.scope.decl), but the vectorization appears to assume that the underlying objects don't alias.

[nikic]

Well, I guess the actual problem here is in LAA rather than LoopVectorize proper. Debug log:

LAA: Found a loop in test: loop
LAA: Processing memory accesses...
  AST: Alias Set Tracker: 2 alias sets for 2 pointer values.
  AliasSet[0x8e36e00, 1] must alias, No access Memory locations: (ptr %cur.ptr, unknown before-or-after)
  AliasSet[0x8e372b0, 1] must alias, No access Memory locations: (ptr %prev.ptr, unknown before-or-after)

LAA:   Accesses(3):
	  %cur.ptr = getelementptr inbounds i8, ptr %arg.1, i64 %iv (write)
	  %prev.ptr = phi ptr [ %cur.ptr, %loop ], [ %arg, %preheader ] (read-only)
	  %cur.ptr = getelementptr inbounds i8, ptr %arg.1, i64 %iv (read)
Underlying objects for pointer   %cur.ptr = getelementptr inbounds i8, ptr %arg.1, i64 %iv
  ptr %arg
Underlying objects for pointer   %cur.ptr = getelementptr inbounds i8, ptr %arg.1, i64 %iv
  ptr %arg
Underlying objects for pointer   %prev.ptr = phi ptr [ %cur.ptr, %loop ], [ %arg, %preheader ]
  ptr %arg
LAA: May be able to perform a memory runtime check if needed.
LAA: No unsafe dependent memory operations in loop.  We don't need runtime memory checks.

When constructing the AST we get the LocationSize to BeforeOrAfter, but we retain the AATags, which is generally not valid.

@fhahn

[nikic]

/cherry-pick 543cf08 cd7ea4e

[github-actions]

/pull-request #79561

[EugeneZelenko]

@mjbedy: But pull request is still open.

[tstellar]

PR has been created, we will track the status there.