x86: Re-enable partial_reg_dependency and movx for Haswell
commitac899f83accb916a9068d64b286c3a7a268d1286
authorhjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Thu, 31 May 2018 15:37:22 +0000 (31 15:37 +0000)
committerhjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Thu, 31 May 2018 15:37:22 +0000 (31 15:37 +0000)
tree711df1b4cb3a1740a15ade8d61067cd35139f4f8
parenta209c8ccc0cd0e7d90db28dc81d87777f612fe27
x86: Re-enable partial_reg_dependency and movx for Haswell

r254152 disabled partial_reg_dependency and movx for Haswell and newer
Intel processors.  r258972 restored them for skylake-avx512.  For Haswell,
movx improves performance.  But partial_reg_stall may be better than
partial_reg_dependency in theory.  We will investigate performance impact
of partial_reg_stall vs partial_reg_dependency on Haswell for GCC 9.  In
the meantime, this patch restores both partial_reg_dependency and mox for
Haswell in GCC 8.

On Haswell, improvements for EEMBC benchmarks with

-mtune-ctrl=movx,partial_reg_dependency -Ofast -march=haswell

vs

-Ofast -mtune=haswell

are

automotive
=========
  aifftr01 (default) - goodperf: Runtime improvement of   2.6% (time).
  aiifft01 (default) - goodperf: Runtime improvement of   2.2% (time).

networking
=========
  ip_pktcheckb1m (default) - goodperf: Runtime improvement of   3.8% (time).
  ip_pktcheckb2m (default) - goodperf: Runtime improvement of   5.2% (time).
  ip_pktcheckb4m (default) - goodperf: Runtime improvement of   4.4% (time).
  ip_pktcheckb512k (default) - goodperf: Runtime improvement of   4.2% (time).

telecom
=========
  fft00data_1 (default) - goodperf: Runtime improvement of   8.4% (time).
  fft00data_2 (default) - goodperf: Runtime improvement of   8.6% (time).
  fft00data_3 (default) - goodperf: Runtime improvement of   9.0% (time).

PR target/85829
* config/i386/x86-tune.def: Re-enable partial_reg_dependency
and movx for Haswell.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@261028 138bc75d-0d04-0410-961f-82ee72b054a4
gcc/ChangeLog
gcc/config/i386/x86-tune.def