Patchwork [x86,AVX2] FP reassociation enabling for AVX2 targets

login
register
mail settings
Submitter Igor Zamyatin
Date Feb. 18, 2013, 6:44 p.m.
Message ID <CAKdSQZ=hffSFDQ2ZE8dpxS7tQPrY3iBmjmHCzX_xSfqLBXC-yw@mail.gmail.com>
Download mbox | patch
Permalink /patch/221446/
State New
Headers show

Comments

Igor Zamyatin - Feb. 18, 2013, 6:44 p.m.
Hi!

This small change enables FP reassociation for AVX2 processors. This
gives ~+1.5% in performance geomean for spec2006FP tests.

Bootstrapped/regtested on x86_64-linux, ok for trunk?


2013-02-14  Igor Zamyatin  <igor.zamyatin@intel.com>

        * config/i386/i386.c (initial_ix86_tune_features): Turn on fp
        reassociation for avx2 targets.



      regs instead of memory.  */
Uros Bizjak - Feb. 18, 2013, 7:34 p.m.
On Mon, Feb 18, 2013 at 7:44 PM, Igor Zamyatin <izamyatin@gmail.com> wrote:

> This small change enables FP reassociation for AVX2 processors. This
> gives ~+1.5% in performance geomean for spec2006FP tests.
>
> Bootstrapped/regtested on x86_64-linux, ok for trunk?
>
>
> 2013-02-14  Igor Zamyatin  <igor.zamyatin@intel.com>
>
>         * config/i386/i386.c (initial_ix86_tune_features): Turn on fp
>         reassociation for avx2 targets.

OK (it is a tuning patch).

Thanks,
Uros.
Kirill Yukhin - Feb. 20, 2013, 8:55 a.m.
> OK (it is a tuning patch).
>
Hi,
Checked in: http://gcc.gnu.org/ml/gcc-cvs/2013-02/msg00540.html

Thanks, K

Patch

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index caf4894..cb84866 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -2021,7 +2021,7 @@  static unsigned int
initial_ix86_tune_features[X86_TUNE_LAST] = {

   /* X86_TUNE_REASSOC_FP_TO_PARALLEL: Try to produce parallel computations
      during reassociation of fp computation.  */
-  m_ATOM,
+  m_ATOM | m_HASWELL,

   /* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE