This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH, GCC 4.7] Backport fix for PR tree-optimization/53708
- From: Peter Bergner <bergner at vnet dot ibm dot com>
- To: "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>
- Cc: Richard Biener <richard dot guenther at gmail dot com>, "libc-alpha at sourceware dot org" <libc-alpha at sourceware dot org>
- Date: Tue, 30 Oct 2012 13:43:33 -0500
- Subject: Re: [PATCH, GCC 4.7] Backport fix for PR tree-optimization/53708
- References: <1351616337.4778.7.camel@otta.rchland.ibm.com>
On Tue, 2012-10-30 at 11:58 -0500, Peter Bergner wrote:
> I'm hitting the same bug as in PR53708 when compiling GLIBC's dlfcn.c when
> vectorization is enabled on powerpc64-linux. A reduced test case is:
>
> bergner@bns:~/gcc/BUGS> cat foo.i
> static void (*const init_array []) (void)
> __attribute__ ((section (".init_array"), aligned (sizeof (void *)), used))
> = { 0 };
>
> bergner@bns:~/gcc/BUGS> /home/bergner/gcc/build/gcc-fsf-4_7-base/gcc/xgcc
> -B/home/bergner/gcc/build/gcc-fsf-4_7-base/gcc -S -m64 -O3 -maltivec foo.i -o
> bad.s
>
> bergner@bns:~/gcc/BUGS> /home/bergner/gcc/build/gcc-fsf-4_7-pr53708/gcc/xgcc
> -B/home/bergner/gcc/build/gcc-fsf-4_7-pr53708/gcc -S -m64 -O3 -maltivec foo.i
> -o good.s
>
> bergner@bns:~/gcc/BUGS> diff -u bad.s good.s
> --- bad.s 2012-10-30 10:41:15.000000000 -0500
> +++ good.s 2012-10-30 10:41:23.000000000 -0500
> @@ -2,7 +2,7 @@
> .section ".toc","aw"
> .section ".text"
> .section .init_array,"a"
> - .align 4
> + .align 3
> .type init_array, @object
> .size init_array, 8
> init_array:
>
> The above is bad, because the extra alignment causes the linker to add some
> null padding to the init_array and the loader isn't expecting that and ends
> up segv'ing. I'd like to backport Richard's patch below to the 4.7 branch.
> The patch bootstrapped and regtested on powerpc64-linux with no regressions.
Commenting on Richard's question from the bugzilla:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53708#c10
I suppose if attribute((__aligned__)) truly does just set a minimum alignment
value (and the documentation seems to say that) and the compiler is free to
arbitrarily increase it, then the GLIBC code to scan the init_array needs to
be tolerant of null values in init_array.
Does everyone agree with that assessment?
Peter