diff mbox

Disable x87 inline functions for x86_64 and SSE [BZ #17262]

Message ID 20140812183008.GO9284@spoyarek.pnq.redhat.com
State Committed
Headers show

Commit Message

Siddhesh Poyarekar Aug. 12, 2014, 6:30 p.m. UTC
Hi,

Since:

commit 409e00bd69b8d8dd74d7327085351d26769ea6fc
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Wed Jan 29 07:51:41 2014 -0800

    Disable x87 inline functions for SSE2 math
    
    When i386 and x86-64 mathinline.h was merged into a single mathinline.h,
    "gcc -m32" enables x87 inline functions on x86-64 even when -mfpmath=sse
    and SSE2 is enabled.  It is a regression on x86-64.  We should check
    __SSE2_MATH__ instead of __x86_64__ when disabling x87 inline functions.


gcc-3.2 is unable to correctly compile x86_64 routines for llrint
since it gets redefined.  This is because gcc 3.2 does not set
__SSE2_MATH__ for x86_64, thus exposing the duplicate definition.

The correct fix ought to be to check for both __SSE2_MATH__ and
__x86_64__ and enable those bits only when neither are defined.

Tested fix with the reproducer for
409e00bd69b8d8dd74d7327085351d26769ea6fc as well as with gcc-3.2.

OK for 2.20?

Thanks,
Siddhesh

	[BZ #17262]
	* sysdeps/x86/fpu/bits/mathinline.h: Check both __SSE2_MATH__
	and __x86_64__ when disabling x87 inline functions.

Comments

Jakub Jelinek Aug. 12, 2014, 6:39 p.m. UTC | #1
On Wed, Aug 13, 2014 at 12:00:08AM +0530, Siddhesh Poyarekar wrote:
> --- a/sysdeps/x86/fpu/bits/mathinline.h
> +++ b/sysdeps/x86/fpu/bits/mathinline.h
> @@ -384,7 +384,10 @@ __END_NAMESPACE_C99
>  # endif
>  #endif
>  
> -#ifndef __SSE2_MATH__
> +/* Disable x87 inlines when -fpmath=sse is passed and also when we're building
> +   on x86_64.  Older gcc (gcc-3.2 for example) does not set __SSE2_MATH__

I'd use predefine or define instead of set, otherwise LGTM.  Thanks for
working on this.

> +   for x86_64.  */
> +#if !defined __SSE2_MATH__ && !defined __x86_64__
>  # if ((!defined __NO_MATH_INLINES || defined __LIBC_INTERNAL_MATH_INLINES) \
>       && defined __OPTIMIZE__)
>  
> @@ -970,4 +973,4 @@ __inline_mathcode2 (__ieee754_atan2, __y, __x,
>  		    return __value;)
>  # endif
>  
> -#endif /* !__SSE2_MATH__ */
> +#endif /* !__SSE2_MATH__ && !__x86_64__ */

	Jakub
H.J. Lu Aug. 12, 2014, 6:48 p.m. UTC | #2
On Tue, Aug 12, 2014 at 11:30 AM, Siddhesh Poyarekar
<siddhesh@redhat.com> wrote:
> Hi,
>
> Since:
>
> commit 409e00bd69b8d8dd74d7327085351d26769ea6fc
> Author: H.J. Lu <hjl.tools@gmail.com>
> Date:   Wed Jan 29 07:51:41 2014 -0800
>
>     Disable x87 inline functions for SSE2 math
>
>     When i386 and x86-64 mathinline.h was merged into a single mathinline.h,
>     "gcc -m32" enables x87 inline functions on x86-64 even when -mfpmath=sse
>     and SSE2 is enabled.  It is a regression on x86-64.  We should check
>     __SSE2_MATH__ instead of __x86_64__ when disabling x87 inline functions.
>
>
> gcc-3.2 is unable to correctly compile x86_64 routines for llrint
> since it gets redefined.  This is because gcc 3.2 does not set
> __SSE2_MATH__ for x86_64, thus exposing the duplicate definition.
>
> The correct fix ought to be to check for both __SSE2_MATH__ and
> __x86_64__ and enable those bits only when neither are defined.
>
> Tested fix with the reproducer for
> 409e00bd69b8d8dd74d7327085351d26769ea6fc as well as with gcc-3.2.
>
> OK for 2.20?
>
> Thanks,
> Siddhesh
>
>         [BZ #17262]
>         * sysdeps/x86/fpu/bits/mathinline.h: Check both __SSE2_MATH__
>         and __x86_64__ when disabling x87 inline functions.
>
> diff --git a/sysdeps/x86/fpu/bits/mathinline.h b/sysdeps/x86/fpu/bits/mathinline.h
> index 9c32e95..ee88b66 100644
> --- a/sysdeps/x86/fpu/bits/mathinline.h
> +++ b/sysdeps/x86/fpu/bits/mathinline.h
> @@ -384,7 +384,10 @@ __END_NAMESPACE_C99
>  # endif
>  #endif
>
> -#ifndef __SSE2_MATH__
> +/* Disable x87 inlines when -fpmath=sse is passed and also when we're building
> +   on x86_64.  Older gcc (gcc-3.2 for example) does not set __SSE2_MATH__
> +   for x86_64.  */
> +#if !defined __SSE2_MATH__ && !defined __x86_64__
>  # if ((!defined __NO_MATH_INLINES || defined __LIBC_INTERNAL_MATH_INLINES) \
>       && defined __OPTIMIZE__)
>
> @@ -970,4 +973,4 @@ __inline_mathcode2 (__ieee754_atan2, __y, __x,
>                     return __value;)
>  # endif
>
> -#endif /* !__SSE2_MATH__ */
> +#endif /* !__SSE2_MATH__ && !__x86_64__ */

Will this patch get x86 inline functions with -fpmath=sse -m32 for
GCC 3.2?
Jakub Jelinek Aug. 12, 2014, 7 p.m. UTC | #3
On Tue, Aug 12, 2014 at 11:48:44AM -0700, H.J. Lu wrote:
> > --- a/sysdeps/x86/fpu/bits/mathinline.h
> > +++ b/sysdeps/x86/fpu/bits/mathinline.h
> > @@ -384,7 +384,10 @@ __END_NAMESPACE_C99
> >  # endif
> >  #endif
> >
> > -#ifndef __SSE2_MATH__
> > +/* Disable x87 inlines when -fpmath=sse is passed and also when we're building
> > +   on x86_64.  Older gcc (gcc-3.2 for example) does not set __SSE2_MATH__
> > +   for x86_64.  */
> > +#if !defined __SSE2_MATH__ && !defined __x86_64__
> >  # if ((!defined __NO_MATH_INLINES || defined __LIBC_INTERNAL_MATH_INLINES) \
> >       && defined __OPTIMIZE__)
> >
> > @@ -970,4 +973,4 @@ __inline_mathcode2 (__ieee754_atan2, __y, __x,
> >                     return __value;)
> >  # endif
> >
> > -#endif /* !__SSE2_MATH__ */
> > +#endif /* !__SSE2_MATH__ && !__x86_64__ */
> 
> Will this patch get x86 inline functions with -fpmath=sse -m32 for
> GCC 3.2?

You'll get the i387 ones in that case, as GCC 3.2 and earlier just don't
predefine anything you could use for that.  Before this patch it didn't do
that either.  The bug in the current mathinline.h is that for gcc 3.2 and
earlier on x86_64 you get errors on the mathinline.h, because you have
multiple definitions of the inlines when __SSE2_MATH__ is not defined, but
__x86_64__ is.

	Jakub
Roland McGrath Aug. 12, 2014, 7:25 p.m. UTC | #4
Looks right to me.  It should go into 2.20 since it's a regression
introduced since 2.19.
Allan McRae Aug. 12, 2014, 11:30 p.m. UTC | #5
On 13/08/14 04:30, Siddhesh Poyarekar wrote:
> OK for 2.20?

Yes.

Allan
diff mbox

Patch

diff --git a/sysdeps/x86/fpu/bits/mathinline.h b/sysdeps/x86/fpu/bits/mathinline.h
index 9c32e95..ee88b66 100644
--- a/sysdeps/x86/fpu/bits/mathinline.h
+++ b/sysdeps/x86/fpu/bits/mathinline.h
@@ -384,7 +384,10 @@  __END_NAMESPACE_C99
 # endif
 #endif
 
-#ifndef __SSE2_MATH__
+/* Disable x87 inlines when -fpmath=sse is passed and also when we're building
+   on x86_64.  Older gcc (gcc-3.2 for example) does not set __SSE2_MATH__
+   for x86_64.  */
+#if !defined __SSE2_MATH__ && !defined __x86_64__
 # if ((!defined __NO_MATH_INLINES || defined __LIBC_INTERNAL_MATH_INLINES) \
      && defined __OPTIMIZE__)
 
@@ -970,4 +973,4 @@  __inline_mathcode2 (__ieee754_atan2, __y, __x,
 		    return __value;)
 # endif
 
-#endif /* !__SSE2_MATH__ */
+#endif /* !__SSE2_MATH__ && !__x86_64__ */