On Thu, Jul 1, 2021 at 12:18 PM Hongtao Liu <crazy...@gmail.com> wrote:
>
> On Thu, Jul 1, 2021 at 5:51 PM Richard Biener via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Thu, Jul 1, 2021 at 9:20 AM liuhongt via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> >
> > How does this look on GIMPLE and why's it not better handled there?
> Do you mean in match.pd, i'll try that.
>
> C++ FE doesn't support _FLoat16, and the place float/double are
> handled is in convert.c(which is GENERIC?), that's why I decided to do
> it in the backend.
>
> /* Disable until we figure out how to decide whether the functions are
> present in runtime. */
> /* Convert (float)sqrt((double)x) where x is float into sqrtf(x) */
> if (optimize
> && (TYPE_MODE (type) == TYPE_MODE (double_type_node)
> || TYPE_MODE (type) == TYPE_MODE (float_type_node)))
Yes, but we can easily add a pattern to match.pd, sth like
(for sq (SQRT)
(simplify
(convert (sq@1 (convert @0)))
(if (types_match (type, TREE_TYPE (@0))
&& TYPE_PRECISION (TREE_TYPE (@1)) > TYPE_PRECISION (TREE_TYPE (@0))
&& direct_internal_fn_supported_p (IFN_SQRT, type, OPTIMIZE_FOR_BOTH))
(IFN_SQRT @0)))
or so.
> >
> > Richard.
> >
> > > gcc/ChangeLog:
> > >
> > > * config/i386/i386.md (*sqrthf2): New define_insn.
> > > * config/i386/sse.md
> > > (*avx512fp16_vmsqrthf2<mask_scalar_name><round_scalar_name>):
> > > Ditto.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > > * gcc.target/i386/avx512fp16-builtin-sqrt-2.c: New test.
> > > ---
> > > gcc/config/i386/i386.md | 18 ++++++++++++++++++
> > > gcc/config/i386/sse.md | 18 ++++++++++++++++++
> > > .../i386/avx512fp16-builtin-sqrt-2.c | 18 ++++++++++++++++++
> > > 3 files changed, 54 insertions(+)
> > > create mode 100644
> > > gcc/testsuite/gcc.target/i386/avx512fp16-builtin-sqrt-2.c
> > >
> > > diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> > > index 5f45c4ff583..684b2080a93 100644
> > > --- a/gcc/config/i386/i386.md
> > > +++ b/gcc/config/i386/i386.md
> > > @@ -16583,6 +16583,24 @@ (define_insn "*sqrt<mode>2_sse"
> > > ]
> > > (symbol_ref "true")))])
> > >
> > > +/* Optimize for code like (_Float16) __builtin_sqrtf ((float) f16)
> > > + since it's not handled in frontend. */
> > > +(define_insn "*sqrthf2"
> > > + [(set (match_operand:HF 0 "register_operand" "=v,v")
> > > + (float_truncate:HF
> > > + (sqrt:MODEF
> > > + (float_extend:MODEF
> > > + (match_operand:HF 1 "nonimmediate_operand" "v,m")))))]
> > > + "TARGET_AVX512FP16"
> > > + "@
> > > + vsqrtsh\t{%d1, %0|%0, %d1}
> > > + vsqrtsh\t{%1, %d0|%d0, %1}"
> > > + [(set_attr "type" "sse")
> > > + (set_attr "atom_sse_attr" "sqrt")
> > > + (set_attr "prefix" "evex")
> > > + (set_attr "avx_partial_xmm_update" "false,true")
> > > + (set_attr "mode" "HF")])
> > > +
> > > (define_expand "sqrthf2"
> > > [(set (match_operand:HF 0 "register_operand")
> > > (sqrt:HF
> > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > > index a76c30c75cb..f87f6893835 100644
> > > --- a/gcc/config/i386/sse.md
> > > +++ b/gcc/config/i386/sse.md
> > > @@ -2407,6 +2407,24 @@ (define_insn
> > > "*<sse>_vmsqrt<mode>2<mask_scalar_name><round_scalar_name>"
> > > (set_attr "btver2_sse_attr" "sqrt")
> > > (set_attr "mode" "<ssescalarmode>")])
> > >
> > > +(define_insn "*avx512fp16_vmsqrthf2<mask_scalar_name><round_scalar_name>"
> > > + [(set (match_operand:V8HF 0 "register_operand" "=v")
> > > + (vec_merge:V8HF
> > > + (vec_duplicate:V8HF
> > > + (float_truncate:HF
> > > + (sqrt:MODEF
> > > + (float_extend:MODEF
> > > + (match_operand:HF 1 "nonimmediate_operand"
> > > "<round_scalar_constraint>")))))
> > > + (match_operand:VFH_128 2 "register_operand" "v")
> > > + (const_int 1)))]
> > > + "TARGET_AVX512FP16"
> > > + "vsqrtsh\t{<round_scalar_mask_op3>%1, %2,
> > > %0<mask_scalar_operand3>|%0<mask_scalar_operand3>, %2,
> > > %1<round_scalar_mask_op3>}"
> > > + [(set_attr "type" "sse")
> > > + (set_attr "atom_sse_attr" "sqrt")
> > > + (set_attr "prefix" "evex")
> > > + (set_attr "mode" "HF")])
> > > +
> > > +
> > > (define_expand "rsqrt<mode>2"
> > > [(set (match_operand:VF1_AVX512ER_128_256 0 "register_operand")
> > > (unspec:VF1_AVX512ER_128_256
> > > diff --git a/gcc/testsuite/gcc.target/i386/avx512fp16-builtin-sqrt-2.c
> > > b/gcc/testsuite/gcc.target/i386/avx512fp16-builtin-sqrt-2.c
> > > new file mode 100644
> > > index 00000000000..4fefee179af
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/i386/avx512fp16-builtin-sqrt-2.c
> > > @@ -0,0 +1,18 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-Ofast -mavx512fp16" } */
> > > +
> > > +#include<math.h>
> > > +_Float16
> > > +foo (_Float16 f16)
> > > +{
> > > + return sqrtf (f16);
> > > +}
> > > +
> > > +_Float16
> > > +foo1 (_Float16 f16)
> > > +{
> > > + return sqrt (f16);
> > > +}
> > > +
> > > +/* { dg-final { scan-assembler-not "vcvtsh2s\[sd\]" } } */
> > > +/* { dg-final { scan-assembler-times "vsqrtsh\[^\n\r\]*xmm\[0-9\]" 2 } }
> > > */
> > > --
> > > 2.18.1
> > >
>
>
>
> --
> BR,
> Hongtao
