On Wed, 2024-01-03 at 16:24 +0800, chenglulu wrote:
> LGTM!
> 
> Thanks!

Pushed r14-6890.

FWIW sometimes tree optimizer still fails to emit .reduc_f{max,min} or
it emits them sub-optimally.  I've commented in PR112457 but maybe I
should've created a new ticket...

> 在 2024/1/1 上午3:15, Xi Ruoyao 写道:
> > We already had smin/smax RTL pattern using vfmin/vfmax instructions.
> > But for smin/smax, it's unspecified what will happen if either operand
> > contains any NaN operands.  So we would not vectorize the loop with
> > -fno-finite-math-only (the default for all optimization levels expect
> > -Ofast).
> > 
> > But, LoongArch vfmin/vfmax instruction is IEEE-754-2008 conformant so we
> > can also use them and vectorize the loop.
> > 
> > gcc/ChangeLog:
> > 
> >     * config/loongarch/simd.md (fmax<mode>3): New define_insn.
> >     (fmin<mode>3): Likewise.
> >     (reduc_fmax_scal_<mode>3): New define_expand.
> >     (reduc_fmin_scal_<mode>3): Likewise.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> >     * gcc.target/loongarch/vfmax-vfmin.c: New test.
> > ---
> > 
> > Happy new year folks.  This is a follow-up of [1].
> > 
> > Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for trunk?
> > 
> > [1]:https://gcc.gnu.org/pipermail/gcc-patches/2023-December/641583.html
> > 
> >   gcc/config/loongarch/simd.md                  | 31 +++++++++++++++++++
> >   .../gcc.target/loongarch/vfmax-vfmin.c        | 31 +++++++++++++++++++
> >   2 files changed, 62 insertions(+)
> >   create mode 100644 gcc/testsuite/gcc.target/loongarch/vfmax-vfmin.c
> > 
> > diff --git a/gcc/config/loongarch/simd.md b/gcc/config/loongarch/simd.md
> > index 93fb39abcf5..8ac1d75a85c 100644
> > --- a/gcc/config/loongarch/simd.md
> > +++ b/gcc/config/loongarch/simd.md
> > @@ -426,6 +426,37 @@ (define_insn 
> > "<simd_isa>_<x>vfcmp_<fcond_unspec>_<simdfmt>"
> >     [(set_attr "type" "simd_fcmp")
> >      (set_attr "mode" "<MODE>")])
> >   
> > +; [x]vf{min/max} instructions are IEEE-754-2008 conforming, use them for
> > +; the corresponding IEEE-754-2008 operations.  We must use UNSPEC instead
> > +; of smin/smax though, see PR105414 and PR107013.
> > +
> > +(define_int_iterator UNSPEC_FMAXMIN [UNSPEC_FMAX UNSPEC_FMIN])
> > +(define_int_attr fmaxmin [(UNSPEC_FMAX "fmax") (UNSPEC_FMIN "fmin")])
> > +
> > +(define_insn "<fmaxmin><mode>3"
> > +  [(set (match_operand:FVEC 0 "register_operand" "=f")
> > +   (unspec:FVEC [(match_operand:FVEC 1 "register_operand" "f")
> > +                 (match_operand:FVEC 2 "register_operand" "f")]
> > +                UNSPEC_FMAXMIN))]
> > +  ""
> > +  "<x>v<fmaxmin>.<simdfmt>\t%<wu>0,%<wu>1,%<wu>2"
> > +  [(set_attr "type" "simd_fminmax")
> > +   (set_attr "mode" "<MODE>")])
> > +
> > +;; ... and also reduc operations.
> > +(define_expand "reduc_<fmaxmin>_scal_<mode>"
> > +  [(match_operand:<UNITMODE> 0 "register_operand")
> > +   (match_operand:FVEC 1 "register_operand")
> > +   (const_int UNSPEC_FMAXMIN)]
> > +  ""
> > +{
> > +  rtx tmp = gen_reg_rtx (<MODE>mode);
> > +  loongarch_expand_vector_reduc (gen_<fmaxmin><mode>3, tmp, operands[1]);
> > +  emit_insn (gen_vec_extract<mode><unitmode> (operands[0], tmp,
> > +                                         const0_rtx));
> > +  DONE;
> > +})
> > +
> >   ; The LoongArch SX Instructions.
> >   (include "lsx.md")
> >   
> > diff --git a/gcc/testsuite/gcc.target/loongarch/vfmax-vfmin.c 
> > b/gcc/testsuite/gcc.target/loongarch/vfmax-vfmin.c
> > new file mode 100644
> > index 00000000000..811fee361c3
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/loongarch/vfmax-vfmin.c
> > @@ -0,0 +1,31 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2 -mtune=la464 -mlasx" } */
> > +/* { dg-final { scan-assembler "\tvfmin\\.d" } } */
> > +/* { dg-final { scan-assembler "\tvfmax\\.d" } } */
> > +/* { dg-final { scan-assembler "\txvfmin\\.d" } } */
> > +/* { dg-final { scan-assembler "\txvfmax\\.d" } } */
> > +/* { dg-final { scan-assembler "\tvfmin\\.s" } } */
> > +/* { dg-final { scan-assembler "\tvfmax\\.s" } } */
> > +/* { dg-final { scan-assembler "\txvfmin\\.s" } } */
> > +/* { dg-final { scan-assembler "\txvfmax\\.s" } } */
> > +
> > +#define T(OP) __typeof__ (__builtin_##OP (0, 0))
> > +
> > +#define TEST(OP, LEN) \
> > +void \
> > +test_##OP##LEN (T (OP) *restrict dest, \
> > +           const T (OP) *restrict src1, \
> > +           const T (OP) *restrict src2) \
> > +{ \
> > +  for (int i = 0; i < LEN / sizeof (T(OP)); i++) \
> > +    dest[i] = __builtin_##OP (src1[i], src2[i]); \
> > +}
> > +
> > +TEST(fmin, 16)
> > +TEST(fmax, 16)
> > +TEST(fmin, 32)
> > +TEST(fmax, 32)
> > +TEST(fminf, 16)
> > +TEST(fmaxf, 16)
> > +TEST(fminf, 32)
> > +TEST(fmaxf, 32)
> 

-- 
Xi Ruoyao <xry...@xry111.site>
School of Aerospace Science and Technology, Xidian University

Reply via email to