* Peter Maydell (peter.mayd...@linaro.org) wrote:
> On 6 June 2016 at 20:22, Dr. David Alan Gilbert <dgilb...@redhat.com> wrote:
> > Sigh I see; so the configure test is run with a different set of compiler
> > options than the main compiles? How is any configure test supposed to
> > check that it's got a sane configuration?
>
> I wasn't quite right -- we do run configure tests with ccache, we just
> force it to not read the cache. So I think your problem is just that
> your configure test doesn't fail when run under ccache, even though the
> main cutils.c compile does:
>
> $ cat /tmp/zz9.c
> #pragma GCC push_options
> #pragma GCC target("avx2")
> #include <cpuid.h>
> #include <immintrin.h>
>
> static int bar(void *a) {
> return _mm256_movemask_epi8(_mm256_cmpeq_epi8(*(__m256i *)a,
> (__m256i){0}));
> }
> static void *bar_ifunc(void) {return (void*) bar;}
> int foo(void *a) __attribute__((ifunc("bar_ifunc")));
> int main(int argc, char *argv[]) { return foo(argv[0]);}
> $ ccache gcc -g -Wall -o /tmp/zz9 /tmp/zz9.c
>
> This seems to be because to get the compile failure you have
> to be doing a .c to .o compile:
<snip>
> The difference is because ccache doesn't intercept .c-to-executable
> compiles, so they go straight to real gcc and don't have -save-temps,
> but it does intercept a .c-to-.o compile, which thus does get -save-temps
> and so hits the gcc bug.
>
> TLDR: if you make your configure test use compile_object it ought to
> correctly detect that AVX2 doesn't work for this gcc+ccache combo.
Does the following (on top of the existing patch) fix it for you:
diff --git a/configure b/configure
index 6449e65..65e603c 100755
--- a/configure
+++ b/configure
@@ -1793,8 +1793,8 @@ static void *bar_ifunc(void) {return (void*) bar;}
int foo(void *a) __attribute__((ifunc("bar_ifunc")));
int main(int argc, char *argv[]) { return foo(argv[0]);}
EOF
-if compile_prog "" ; then
- if readelf --syms $TMPE |grep "IFUNC.*foo" >/dev/null 2>&1; then
+if compile_object "" ; then
+ if readelf --syms $TMPO |grep "IFUNC.*foo" >/dev/null 2>&1; then
avx2_opt="yes"
fi
fi
Dave
>
> thanks
> -- PMM
--
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
