On Sun, Aug 18, 2024 at 11:06 AM Jeff Law <jeffreya...@gmail.com> wrote:
>
>
>
> On 8/16/24 8:13 PM, Andrew Pinski wrote:
> > To start working on more with expressions with more than one operand,
> > converting
> > over to use gimple_match_op is needed.
> > The added side-effect here is factor_out_conditional_operation can now
> > support
> > builtins/internal calls that has one operand without any extra code added.
> >
> > Note on the changed testcases:
> > * pr87007-5.c: the test was testing testing for avoiding partial register
> > stalls
> > for the sqrt and making sure there is only one zero of the register before
> > the
> > branch, the phiopt would now merge the sqrt's so disable phiopt.
> >
> > Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> >
> > gcc/ChangeLog:
> >
> > * gimple-match-exports.cc
> > (gimple_match_op::operands_occurs_in_abnormal_phi):
> > New function.
> > * gimple-match.h (gimple_match_op): Add
> > operands_occurs_in_abnormal_phi.
> > * tree-ssa-phiopt.cc (factor_out_conditional_operation): Use
> > gimple_match_op
> > instead of manually extracting from/creating the gimple.
> >
> > gcc/testsuite/ChangeLog:
> >
> > * gcc.target/i386/pr87007-5.c: Disable phi-opt.
> >
>
> > diff --git a/gcc/testsuite/gcc.target/i386/pr87007-5.c
> > b/gcc/testsuite/gcc.target/i386/pr87007-5.c
> > index 8f2dc947f6c..1a240adef63 100644
> > --- a/gcc/testsuite/gcc.target/i386/pr87007-5.c
> > +++ b/gcc/testsuite/gcc.target/i386/pr87007-5.c
> > @@ -1,8 +1,11 @@
> > /* { dg-do compile } */
> > -/* { dg-options "-Ofast -march=skylake-avx512 -mfpmath=sse
> > -fno-tree-vectorize -fdump-tree-cddce3-details
> > -fdump-tree-lsplit-optimized" } */
> > +/* { dg-options "-Ofast -march=skylake-avx512 -mfpmath=sse
> > -fno-tree-vectorize -fdump-tree-cddce3-details -fdump-tree-lsplit-optimized
> > -fno-ssa-phiopt" } */
> > /* Load of d2/d3 is hoisted out, the loop is split, store of d1 and sqrt
> > are sunk out of the loop and the loop is elided. One vsqrtsd with
> > memory operand needs a xor to avoid partial dependence. */
> > +/* Phi-OPT needs to ne disabled otherwise, sqrt calls are merged which is
> > better
> > + but we are testing to make sure the partial register stall for SSE is
> > still avoided
> > + for sqrts. */
> Nit. s/to ne/to be/g
>
> OK with the nit fixed.
>
> Note this is getting closer to doing generalized sinking a common op
> through PHI nodes which is something we've wanted for a long time.
Yes that is the plan; I just want to do it in steps as I have a few
other projects in progress; and I don't know how much of each I will
be able to get done in time for GCC 15. I originally had this done
differently but I thought it would be better to reuse infrastructure
that was already there instead of creating new ones.
I had implemented this patch back in April and I didn't know if I
could get to the rest due to other projects going on so I submitted it
finally.
Thanks,
Andrew
>
> Jeff
>
