On Thu, Jan 14, 2021 at 7:16 PM Hongtao Liu <crazylht@gmail.com> wrote:
>
> ping.
>
> On Thu, Jan 7, 2021 at 1:22 PM Hongtao Liu <crazylht@gmail.com> wrote:
> >
> > On Wed, Jan 6, 2021 at 10:39 PM Jakub Jelinek <jakub@redhat.com> wrote:
> > >
> > > On Wed, Jan 06, 2021 at 02:49:13PM +0800, Hongtao Liu wrote:
> > > >   ix86_expand_fp_vec_cmp/ix86_expand_int_vec_cmp are used by vec_cmpmn
> > > > for vector comparison to vector mask, but ix86_expand_sse_cmp(which is
> > > > called in upper 2 functions.) may return integer mask whenever integer
> > > > mask is available, so convert integer mask back to vector mask if
> > > > needed.
> > > >
> > > > gcc/ChangeLog:
> > > >
> > > >         PR target/98537
> > > >         * config/i386/i386-expand.c (ix86_expand_fp_vec_cmp):
> > > >         When cmp is integer mask, convert it to vector.
> > > >         (ix86_expand_int_vec_cmp): Ditto.
> > > >
> > > > gcc/testsuite/ChangeLog:
> > > >
> > > >         PR target/98537
> > > >         * g++.target/i386/avx512bw-pr98537-1.C: New test.
> > > >         * g++.target/i386/avx512vl-pr98537-1.C: New test.
> > > >         * g++.target/i386/avx512vl-pr98537-2.C: New test.
> > >
> > > Do we optimize it then to an AVX/AVX2 comparison if possible?

  A new patch is proposed to solve a series of performance and
correctness regressions brought by r10-5250.
Integer mask comparison will only be used for 512-bit vectors, and
128/256-bit vcondmn(excluding the case where op_true/op_false is all
1s or 0s, it is actually a vec_cmpmn).

in ix86_expand_sse_cmp/ix86_expand_int_sse_cmp

-  if (ix86_valid_mask_cmp_mode (cmp_ops_mode))
+  if (GET_MODE_SIZE (mode) == 64
+      || (ix86_valid_mask_cmp_mode (cmp_ops_mode)
+         /* When op_true and op_false is NULL, vector dest is required.  */
+         && op_true && op_false
+         /* Gimple sometimes transforms vec_cmpmn to vcondmn with
+            op_true/op_false as constm1_rtx/const0_rtx.
+            Don't generate integer mask comparison then.  */
+         && !((vector_all_ones_operand (op_true, GET_MODE (op_true))
+               && CONST0_RTX (GET_MODE (op_false)) == op_false)
+              || (vector_all_ones_operand (op_false, GET_MODE (op_false))
+                  && CONST0_RTX (GET_MODE (op_true)) == op_true))))

Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
Ok for trunk?
Ok for backport to gcc10?

gcc/ChangeLog:

        PR target/98537
        * config/i386/i386-expand.c (ix86_expand_sse_cmp): Don't
        generate integer mask comparison for 128/256-bits vector when
        op_true/op_false is NULL_RTX or CONSTM1_RTX/CONST0_RTX. Also
        delete redundant !maskcmp condition.
        (ix86_expand_int_vec_cmp): Ditto but no redundant deletion
        here.
        (ix86_expand_sse_movcc): Delete definition of maskcmp, add the
        condition directly to if (maskcmp), add extra check for
        cmpmode, it should be MODE_INT.
        (ix86_expand_fp_vec_cmp): Pass NULL to ix86_expand_sse_cmp's
        parameters op_true/op_false.

gcc/testsuite/ChangeLog:

        PR target/98537
        * g++.target/i386/avx512bw-pr98537-1.C: New test.
        * g++.target/i386/avx512vl-pr98537-1.C: New test.
        * g++.target/i386/avx512vl-pr98537-2.C: New test.
        * gcc.target/i386/avx512vl-pr88547-1.c: Adjust testcase,
        integer mask comparison should not be generated.
        * gcc.target/i386/avx512vl-pr92686-vpcmp-1.c: This test is
        used to guard code generation of integer mask comparison, but
        for vector comparison to vector dest, integer mask comparison
        is disliked, so detele this useless test.
        * gcc.target/i386/avx512vl-pr92686-vpcmp-2.c: Ditto.
        * gcc.target/i386/avx512vl-pr92686-vpcmp-intelasm-1.c: Ditto.


-- 
BR,
Hongtao