On Mon, Sep 6, 2021 at 8:35 PM H.J. Lu <hjl.to...@gmail.com> wrote:
>
> On Sun, Sep 5, 2021 at 10:34 PM Hongtao Liu <crazy...@gmail.com> wrote:
> >
> > On Sun, Sep 5, 2021 at 5:56 AM H.J. Lu via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Enable FMA in scalar/vector unsigned SI to SF expanders.
> > >
> > > gcc/
> > >
> > > PR target/85819
> > > * config/i386/i386-expand.c (ix86_expand_convert_uns_sisf_sse):
> > > Enable FMA.
> > > (ix86_expand_vector_convert_uns_vsivsf): Likewise.
> > >
> > > gcc/testsuite/
> > >
> > > PR target/85819
> > > * gcc.target/i386/pr85819-1.c: New test.
> > > * gcc.target/i386/pr85819-2a.c: Likewise.
> > > * gcc.target/i386/pr85819-2b.c: Likewise.
> > > * gcc.target/i386/pr85819-2c.c: Likewise.
> > > * gcc.target/i386/pr85819-3.c: Likewise.
> > > ---
> > > gcc/config/i386/i386-expand.c | 44 ++++++++++++++++------
> > > gcc/testsuite/gcc.target/i386/pr85819-1.c | 11 ++++++
> > > gcc/testsuite/gcc.target/i386/pr85819-2a.c | 17 +++++++++
> > > gcc/testsuite/gcc.target/i386/pr85819-2b.c | 6 +++
> > > gcc/testsuite/gcc.target/i386/pr85819-2c.c | 7 ++++
> > > gcc/testsuite/gcc.target/i386/pr85819-3.c | 18 +++++++++
> > > 6 files changed, 91 insertions(+), 12 deletions(-)
> > > create mode 100644 gcc/testsuite/gcc.target/i386/pr85819-1.c
> > > create mode 100644 gcc/testsuite/gcc.target/i386/pr85819-2a.c
> > > create mode 100644 gcc/testsuite/gcc.target/i386/pr85819-2b.c
> > > create mode 100644 gcc/testsuite/gcc.target/i386/pr85819-2c.c
> > > create mode 100644 gcc/testsuite/gcc.target/i386/pr85819-3.c
> > >
> > > diff --git a/gcc/config/i386/i386-expand.c b/gcc/config/i386/i386-expand.c
> > > index 2500dbfa7fb..26263bbe1af 100644
> > > --- a/gcc/config/i386/i386-expand.c
> > > +++ b/gcc/config/i386/i386-expand.c
> > > @@ -1851,12 +1851,21 @@ ix86_expand_convert_uns_sisf_sse (rtx target, rtx
> > > input)
> > > fp_lo = gen_reg_rtx (SFmode);
> > > emit_insn (gen_floatsisf2 (fp_hi, int_hi));
> > > emit_insn (gen_floatsisf2 (fp_lo, int_lo));
> > > - fp_hi = expand_simple_binop (SFmode, MULT, fp_hi, x, fp_hi,
> > > - 0, OPTAB_DIRECT);
> > > - fp_hi = expand_simple_binop (SFmode, PLUS, fp_hi, fp_lo, target,
> > > - 0, OPTAB_DIRECT);
> > > - if (!rtx_equal_p (target, fp_hi))
> > > - emit_move_insn (target, fp_hi);
> > > + if (TARGET_FMA || TARGET_AVX512F)
> > Looking at the expander floatunssi<mode>2, the || in the condition
> > should never be hit since we have direct vcvtsi2s[sd] instruction
> > under TARGET_AVX512F.
>
> Fixed.
>
> > > + {
> > > + x = validize_mem (force_const_mem (SFmode, x));
> > > + fp_hi = gen_rtx_FMA (SFmode, fp_hi, x, fp_lo);
> > > + emit_move_insn (target, fp_hi);
> > > + }
> > > + else
> > > + {
> > > + fp_hi = expand_simple_binop (SFmode, MULT, fp_hi, x, fp_hi,
> > > + 0, OPTAB_DIRECT);
> > > + fp_hi = expand_simple_binop (SFmode, PLUS, fp_hi, fp_lo, target,
> > > + 0, OPTAB_DIRECT);
> > > + if (!rtx_equal_p (target, fp_hi))
> > > + emit_move_insn (target, fp_hi);
> > > + }
> > > }
> > >
> > > /* floatunsv{4,8}siv{4,8}sf2 expander. Expand code to convert
> > > @@ -1888,12 +1897,23 @@ ix86_expand_vector_convert_uns_vsivsf (rtx
> > > target, rtx val)
> > > real_ldexp (&TWO16r, &dconst1, 16);
> > > tmp[5] = const_double_from_real_value (TWO16r, SFmode);
> > > tmp[5] = force_reg (fltmode, ix86_build_const_vector (fltmode, 1,
> > > tmp[5]));
> > > - tmp[6] = expand_simple_binop (fltmode, MULT, tmp[4], tmp[5], NULL_RTX,
> > > 1,
> > > - OPTAB_DIRECT);
> > > - tmp[7] = expand_simple_binop (fltmode, PLUS, tmp[3], tmp[6], target, 1,
> > > - OPTAB_DIRECT);
> > > - if (tmp[7] != target)
> > > - emit_move_insn (target, tmp[7]);
> > > + unsigned vector_size = GET_MODE_SIZE (fltmode);
> > > + if (TARGET_FMA
> > > + || (TARGET_AVX512F && vector_size == 64)
> > > + || (TARGET_AVX512VL && (vector_size == 32 || vector_size == 16)))
> > Similar here for the last two || in the condition.
>
> Fixed.
>
> > > + {
> > > + tmp[6] = gen_rtx_FMA (fltmode, tmp[4], tmp[5], tmp[3]);
> > > + emit_move_insn (target, tmp[6]);
> > > + }
> > > + else
> > > + {
> > > + tmp[6] = expand_simple_binop (fltmode, MULT, tmp[4], tmp[5],
> > > + NULL_RTX, 1, OPTAB_DIRECT);
> > > + tmp[7] = expand_simple_binop (fltmode, PLUS, tmp[3], tmp[6],
> > > + target, 1, OPTAB_DIRECT);
> > > + if (tmp[7] != target)
> > > + emit_move_insn (target, tmp[7]);
> > > + }
> > > }
> > >
> > > /* Adjust a V*SFmode/V*DFmode value VAL so that *sfix_trunc* resp.
> > > fix_trunc*
>
> Here is the v2 patch.
LGTM.
>
> --
> H.J.
--
BR,
Hongtao
