Created attachment 474 [details]
Test source

This is definitely bug in InstCombine.

opt -instcombine produces bad output

Hmm, no cast instructions.

What does instcombine produce? Can you attach it please?

Reid.

Before:

        %ovm = and int %tmp1, 32                ; <int> [#uses=1]
        %ov3 = add int %ovm, 145                ; <int> [#uses=2]
        %ov110 = xor int %ov3, 153
        %hvar174 = add int %ov110, 1

After:

        %ovm = and int %tmp1, 32                ; <int> [#uses=1]
        %ov1102 = or int %ovm, 153              ; <int> [#uses=1]
        %hvar174 = add int %ov1102, 1           ; <int> [#uses=1]

Fixed.  Testcase here: InstCombine/2006-11-27-XorBug.ll  Patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20061127/040292.html

-Chris

Attached testcase is still broken.

In fact, I haven't found any output difference before and after patch.

I get:
$ llvm-as < t.ll | opt -instcombine | llvm-dis
...
int %main() {
entry:
        %v0 = alloca int, align 4               ; <int*> [#uses=2]
        %tmp = call int (sbyte*, ...)* %scanf( sbyte* getelementptr ([3 x sbyte]* %str, int 0, uint 0), int* %v0 
)              ; <int> [#uses=0]
        %tmp1 = load int* %v0           ; <int> [#uses=1]
        %ovm = and int %tmp1, 32                ; <int> [#uses=1]
        %hvar1742 = or int %ovm, 9              ; <int> [#uses=1]
        %tmp3 = call int (sbyte*, ...)* %printf( sbyte* getelementptr ([4 x sbyte]* %str, int 0, uint 0), int %
hvar1742 )               ; <int> [#uses=0]
        br label %return

return:         ; preds = %entry
        ret int 0
}
..

You don't get this?

This same bug also exists in the dag combiner.  Fixed like this:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20061127/040294.html

-Chris