[Mesa-dev] [PATCH] mesa: use build flag to ensure stack is realigned on x86

[Matt Turner]

On Sat, Dec 6, 2014 at 8:52 PM, Timothy Arceri <t_arceri at yahoo.com.au> wrote:
> Nowadays GCC assumes stack pointer is 16-byte aligned even on 32-bits, but that is an assumption OpenGL drivers (or any dynamic library for that matter) can't afford to make as there are many closed- and open- source application binaries out there that only assume 4-byte stack alignment.
>
> https://bugs.freedesktop.org/show_bug.cgi?id=86788
> ---
>  The bug reporter has tested that this fixes the problem. But I haven't
>  tested the cross compile stuff. e.g. building 64-bit from 32-bit
>  not sure if anyone would acctaully be building that way.
>
>  Please note if this patch is ok it should also be applied to 10.4 with
>  the last hunk removed.
>
>  configure.ac               | 17 +++++++++++++++++
>  src/mesa/Makefile.am       |  4 ++++
>  src/mesa/main/sse_minmax.c |  3 ---
>  3 files changed, 21 insertions(+), 3 deletions(-)
>
> diff --git a/configure.ac b/configure.ac
> index b0df1bb..6bb43f8 100644
> --- a/configure.ac
> +++ b/configure.ac
> @@ -474,6 +474,21 @@ fi
>  dnl
>  dnl Arch/platform-specific settings
>  dnl
> +align_check_arch=""
> +if test "x$cross_compiling" = xno; then
> +    case "$host_cpu" in
> +        x86_64|amd64)
> +            align_check_arch=x86_64
> +    ;;
> +    esac
> +else
> +    case "$host_cpu" in
> +        i?86)
> +            align_check_arch=x86_64

This looks wrong.

> +    ;;

This is usually aligned with the code in the case.

> +    esac
> +fi
> +
>  AC_ARG_ENABLE([asm],
>      [AS_HELP_STRING([--disable-asm],
>          [disable assembly usage @<:@default=enabled on supported plaforms@:>@])],
> @@ -2087,6 +2102,8 @@ AM_CONDITIONAL(HAVE_X11_DRIVER, test "x$enable_xlib_glx" = xyes)
>  AM_CONDITIONAL(HAVE_OSMESA, test "x$enable_osmesa" = xyes)
>  AM_CONDITIONAL(HAVE_GALLIUM_OSMESA, test "x$enable_gallium_osmesa" = xyes)
>
> +AM_CONDITIONAL(HAVE_16BYTE_ALIGN, test "x$align_check_arch" = xx86_64)

We definitely need a better name. But, more importantly, isn't the
test wrong? You only want to do this on 32-bit.

Søren's patch to pixman does this check rather nicely, I think. See
http://lists.freedesktop.org/archives/pixman/2013-October/003077.html