llvm dev - Oct 2010 - [LLVMdev] Unexpected Failure with 'make check'

Hi all,

I hit an 'Unexpected Failure' in the dejagnu testsuite with the 
Transforms/GVN/null-aliases-nothing.ll test. That test has the following
checks:
; CHECK: load
; CHECK-NOT: load
; CHECK: ret void

This check failed as I was working with unaligned loads, and had my SVN
tree in the folder /home/kraiskil/llvm/unaligned_load/. opt puts that
path into the ModuleID as a comment at the start of the output. This
means that the CHECK:load matches the path in the comment, and the
CHECK-NOT:load matches the load instruction the author intended.


So which of the following is to blame?
1) FileCheck checks comments. 
2) opt inserting the comment with the path. 
3) the test itself should be more verbose (e.g. have a 
'CHECK: define void @test1' to align to the function to check). 
4) No error, this is just The Way Things Are, and we should be aware of
it.


I would have thought 3) is the one to fix, even tough it means all
(most) tests potentially may have 'false negatives' at some point in
time. But there was a different opinion on the irc (and I was in a hurry
and couldn't finish that discussion).


kalle

-- 
"Modern computing machines are very complex objects"
                      -von Neumann introducing the NORC, 1954

Hi Kalle,
> So which of the following is to blame?
> 1) FileCheck checks comments.
> 2) opt inserting the comment with the path.
probably points 1) and/or 2) I reckon.

Ciao,

Duncan.
> 3) the test itself should be more verbose (e.g. have a
> 'CHECK: define void @test1' to align to the function to check).
> 4) No error, this is just The Way Things Are, and we should be aware of
> it.
>
>
> I would have thought 3) is the one to fix, even tough it means all
> (most) tests potentially may have 'false negatives' at some point
in
> time. But there was a different opinion on the irc (and I was in a hurry
> and couldn't finish that discussion).
>
>
> kalle
>