This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH 1/8] x86: Add macros to implement ifunce selection in C
- From: "H.J. Lu" <hjl dot tools at gmail dot com>
- To: GNU C Library <libc-alpha at sourceware dot org>
- Date: Mon, 5 Jun 2017 07:25:42 -0700
- Subject: Re: [PATCH 1/8] x86: Add macros to implement ifunce selection in C
- Authentication-results: sourceware.org; auth=none
- References: <20170602194258.GA21335@lucon.org>
On Fri, Jun 2, 2017 at 12:42 PM, H.J. Lu <hongjiu.lu@intel.com> wrote:
> Should they go into include/libc-symbols.h so that they are available
> everywhere?
>
> H.J.
> ---
> These macros are used to implement ifunc selection in C. To implement
> an ifunc function, foo, which returns the address of __foo_sse2 or
> __foo_avx2:
>
> __foo_avx2:
>
> #define foo __redirect_foo
> #define __foo __redirect___foo
> #include <foo.h>
> #undef foo
> #undef __foo
> #define SYMBOL_NAME foo
> #include <init-arch.h>
>
> extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2) attribute_hidden;
> extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2) attribute_hidden;
>
> static inline void *
> foo_selector (void)
> {
> if (use AVX2)
> return OPTIMIZE (avx2);
>
> return OPTIMIZE (sse2);
> }
>
> libc_ifunc_redirected (__redirect_foo, foo, foo_selector ());
>
> 2017-06-02 Adhemerval Zanella <adhemerval.zanella@linaro.org>
> H.J. Lu <hongjiu.lu@intel.com>
>
> * sysdeps/x86/init-arch.h (PASTER1): New.
> (EVALUATOR1): Likewise.
> (PASTER2): Likewise.
> (EVALUATOR2): Likewise.
> (REDIRECT_NAME): Likewise.
> (OPTIMIZE): Likewise.
> (IFUNC_SELECTOR): Likewise.
> ---
> sysdeps/x86/init-arch.h | 40 ++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 40 insertions(+)
>
> diff --git a/sysdeps/x86/init-arch.h b/sysdeps/x86/init-arch.h
> index 6cbc16b..15d3f09 100644
> --- a/sysdeps/x86/init-arch.h
> +++ b/sysdeps/x86/init-arch.h
> @@ -21,6 +21,46 @@
> # include <ldsodefs.h>
> #endif
>
> +/* These macros are used to implement ifunc selection in C. To implement
> + an ifunc function, foo, which returns the address of __foo_sse2 or
> + __foo_avx2:
> +
> + #define foo __redirect_foo
> + #define __foo __redirect___foo
> + #include <foo.h>
> + #undef foo
> + #undef __foo
> + #define SYMBOL_NAME foo
> + #include <init-arch.h>
> +
> + extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2) attribute_hidden;
> + extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2) attribute_hidden;
> +
> + static inline void *
> + foo_selector (void)
> + {
> + if (use AVX2)
> + return OPTIMIZE (avx2);
> +
> + return OPTIMIZE (sse2);
> + }
> +
> + libc_ifunc_redirected (__redirect_foo, foo, foo_selector ());
> +
> +*/
> +
> +#define PASTER1(x,y) x##_##y
> +#define EVALUATOR1(x,y) PASTER1 (x,y)
> +#define PASTER2(x,y) __##x##_##y
> +#define EVALUATOR2(x,y) PASTER2 (x,y)
> +
> +/* Basically set '__redirect_<symbol>' to use as type definition,
> + '__<symbol>_<variant>' as the optimized implementation and
> + '<symbol>_ifunc_selector' as the IFUNC selector. */
> +#define REDIRECT_NAME EVALUATOR1 (__redirect, SYMBOL_NAME)
> +#define OPTIMIZE(name) EVALUATOR2 (SYMBOL_NAME, name)
> +#define IFUNC_SELECTOR EVALUATOR1 (SYMBOL_NAME, ifunc_selector)
> +
> #ifndef __x86_64__
> /* Due to the reordering and the other nifty extensions in i686, it is
> not really good to use heavily i586 optimized code on an i686. It's
> --
> 2.9.4
>
I will check it in today to unblock my AVX2 patches.
--
H.J.