[LLVMdev] GEP instruction change

[Dan Gohman]

On Aug 21, 2009, at 2:02 AM, Artur Pietrek wrote:


> Hi All,
> Since few days I observe weird change.
> Consider the following C code
>
> char array[] = "0123456789";
> extern int test(char arr[], int size);
>
> int main(void) {
>   return test(array-1, sizeof(array)-1);
> }

As Eli mentioned, this invokes undefined behavior.

>
> using clang frontend i get this:
>
> %call = call i32 @test(i8* getelementptr inbounds ([11 x i8]*  
> @array, i32 0, i32 -1), i32 10) ; <i32> [#uses=1]

I get the same code for clang and llvm-gcc. Did you rebuild clang?

>
> and using LLVM-GCC this:
> %1 = call i32 @test(i8* getelementptr ([11 x i8]* @array, i32  
> 1561806289, i32 4), i32 10) nounwind ; <i32> [#uses=1]
>
> Should the second getelementptr look like this? Is it correct or a  
> bug?

It's a bug, though a subtle one. The first index is truncated
incorrectly. It should be 390451572, and the second index 3, since
it looks like you're using a 32-bit target and 39045157*11+3 is -1
when interpreted as a 32-bit signed integer. This is now fixed.

> Since that change there's a problem in MSIL backend - it simply  
> triggers pointer overflow assertion.

This sounds like a bug in the MSIL backend regardless.

Dan