[Bug target/57037] New: GCC does not generate non-temporal stores on i386 with SSE2+

[Bug target/57037] New: GCC does not generate non-temporal stores on i386 with SSE2+

[anlauf at gmx dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57037

             Bug #: 57037
           Summary: GCC does not generate non-temporal stores on i386 with
                    SSE2+
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: anlauf@gmx.de


Hello,

it appears that gcc does not generate non-temporal stores
available on i386 at least with SSE2.  This is an important
optimization for some memory-bandwidth limited codes.

Example: for the stream triad kernel,

subroutine stream_kernel_triad (a, b, c, n, s)
  integer         , intent(in) :: n
  double precision             :: a(*), b(*), c(*)
  double precision, intent(in) :: s

  integer :: j
  do j = 1,n
     a(j) = b(j) + s*c(j)
  end do
end subroutine stream_kernel_triad

the Intel compiler generates vectorized code with a
throughput that is 25% higher on my Core2 than when
disabling the generation of non-temporal stores
(i.e. compiling with "-opt-streaming-stores never").

gfortran (using -Ofast -fprefetch-loop-arrays) exactly
reproduces the performance of the Intel compiler without
temporal stores.  It appears that this is an important
optimization.

Harald

[Bug target/57037] GCC does not generate non-temporal stores on i386 with SSE2+

[anlauf at gmx dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57037

--- Comment #1 from Harald Anlauf <anlauf at gmx dot de> ---
(In reply to Harald Anlauf from comment #0)
> gfortran (using -Ofast -fprefetch-loop-arrays) exactly
> reproduces the performance of the Intel compiler without
> temporal stores.  It appears that this is an important
> optimization.

I tried a current snapshot from trunk (r219084) and found
that -fprefetch-loop-arrays now gives an additional boost,
matching Intel v15 for the above code, even without the
streaming stores.

[Bug target/57037] GCC does not generate non-temporal stores on i386 with SSE2+

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57037

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
GCC has a heustrics for deciding on non-temporal stores or not (when used with
-fprefetch-loop-arrays) but it is hard not hit it sometimes.

Anyways closing as invalid as we do have 2 testcases in the testsuite that
detects nontemporal stores, gcc.dg/tree-ssa/prefetch-{8,9}.c and they still
pass.