[Qemu-devel] [PATCH] Improve avx2 configure test

[Qemu-devel] [PATCH] Improve avx2 configure test

[Dr. David Alan Gilbert (git)]

From: "Dr. David Alan Gilbert" <dgilbert@redhat.com>

When configured with --extra-cflags=-O2 gcc optimised out the test
and the readelf failed the check leaving avx2 disabled.

The fix for that is to remove 'static' on 'foo'.
At the same time, I thought it best to be more paranoid on
checking whether both the compiler and assembler could actually
use avx2, so make a couple of bogus calls to the the mm256
intrinsics.

(I also wondered about trying to replicate the gcc4.9 test that's
used in cutils.c - but I'd prefer to know the exact failure mode rather
than test for the version).

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
---
 configure | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/configure b/configure
index b5aab72..13fabc3 100755
--- a/configure
+++ b/configure
@@ -1781,10 +1781,17 @@ fi
 # avx2 optimization requirement check
 
 cat > $TMPC << EOF
-static void bar(void) {}
+#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;}
-static void foo(void) __attribute__((ifunc("bar_ifunc")));
-int main(void) { foo(); return 0; }
+int foo(void *a) __attribute__((ifunc("bar_ifunc")));
+int main(int argc, char *argv[]) { return foo(argv[0]);}
 EOF
 if compile_prog "-mavx2" "" ; then
     if readelf --syms $TMPE |grep "IFUNC.*foo" >/dev/null 2>&1; then
-- 
2.7.4

Re: [Qemu-devel] [PATCH] Improve avx2 configure test

[Dr. David Alan Gilbert]

Actually, skip this version, I've got a simple change that also
rolls in the gcc 4.9 check (just removing the "-mavx2") - I'll
post that after lunch.

Dave

* Dr. David Alan Gilbert (git) (dgilbert@redhat.com) wrote:
> From: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
> 
> When configured with --extra-cflags=-O2 gcc optimised out the test
> and the readelf failed the check leaving avx2 disabled.
> 
> The fix for that is to remove 'static' on 'foo'.
> At the same time, I thought it best to be more paranoid on
> checking whether both the compiler and assembler could actually
> use avx2, so make a couple of bogus calls to the the mm256
> intrinsics.
> 
> (I also wondered about trying to replicate the gcc4.9 test that's
> used in cutils.c - but I'd prefer to know the exact failure mode rather
> than test for the version).
> 
> Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> ---
>  configure | 13 ++++++++++---
>  1 file changed, 10 insertions(+), 3 deletions(-)
> 
> diff --git a/configure b/configure
> index b5aab72..13fabc3 100755
> --- a/configure
> +++ b/configure
> @@ -1781,10 +1781,17 @@ fi
>  # avx2 optimization requirement check
>  
>  cat > $TMPC << EOF
> -static void bar(void) {}
> +#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;}
> -static void foo(void) __attribute__((ifunc("bar_ifunc")));
> -int main(void) { foo(); return 0; }
> +int foo(void *a) __attribute__((ifunc("bar_ifunc")));
> +int main(int argc, char *argv[]) { return foo(argv[0]);}
>  EOF
>  if compile_prog "-mavx2" "" ; then
>      if readelf --syms $TMPE |grep "IFUNC.*foo" >/dev/null 2>&1; then
> -- 
> 2.7.4
> 
> 
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK