diff mbox series

Expand __builtin_memcmp_eq with ptest for OI/TImode.

Message ID 20220505073707.95306-1-hongtao.liu@intel.com
State New
Headers show
Series Expand __builtin_memcmp_eq with ptest for OI/TImode. | expand

Commit Message

Liu, Hongtao May 5, 2022, 7:37 a.m. UTC 
Enable optimization for TImode only under 32-bit target, for 64-bit
target there could be extra ineteger <-> sse move regarding psABI,
not efficient.

Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
Ok for trunk?

gcc/ChangeLog:

	PR target/104610
	* config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
	for TI/QImode when code is EQ or NE.
	* config/i386/i386.md (SDWIM1248): New iterator.
	(cbranch<mode>4): Split TImode into a separate expander.
	(cbranchti4): New expander.
	* config/i386/predicates.md (timode_comparison_operator): New
	predicate.
	* config/i386/sse.md (cbranch<mode>4): Extend to OImode.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/pr104610.c: New test.
---
 gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
 gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
 gcc/config/i386/predicates.md            |  6 ++++++
 gcc/config/i386/sse.md                   | 10 +++++++--
 gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
 5 files changed, 74 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c

Comments

Hongtao Liu May 5, 2022, 7:37 a.m. UTC | #1 
On Thu, May 5, 2022 at 3:37 PM liuhongt <hongtao.liu@intel.com> wrote:
>
> Enable optimization for TImode only under 32-bit target, for 64-bit
> target there could be extra ineteger <-> sse move regarding psABI,
> not efficient.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for trunk?
>
> gcc/ChangeLog:
>
>         PR target/104610
>         * config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
>         for TI/QImode when code is EQ or NE.
>         * config/i386/i386.md (SDWIM1248): New iterator.
>         (cbranch<mode>4): Split TImode into a separate expander.
>         (cbranchti4): New expander.
>         * config/i386/predicates.md (timode_comparison_operator): New
>         predicate.
>         * config/i386/sse.md (cbranch<mode>4): Extend to OImode.
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.target/i386/pr104610.c: New test.
> ---
>  gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
>  gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
>  gcc/config/i386/predicates.md            |  6 ++++++
>  gcc/config/i386/sse.md                   | 10 +++++++--
>  gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
>  5 files changed, 74 insertions(+), 5 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c
>
> diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> index bc806ffa283..a2012a158ae 100644
> --- a/gcc/config/i386/i386-expand.cc
> +++ b/gcc/config/i386/i386-expand.cc
> @@ -2264,13 +2264,24 @@ ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
>  {
>    machine_mode mode = GET_MODE (op0);
>    rtx tmp;
> +  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> +  /* Using ptest for TImode only for 32-bit target since it's splitted into
> +     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
> +     move regarding psABI, not efficient.  */
> +  if ((code == EQ || code == NE)
> +       && ((mode == OImode && TARGET_AVX)
> +          || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
> +    {
> +      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
> +      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
> +      mode = p_mode;
> +    }
>
>    /* Handle special case - vector comparsion with boolean result, transform
>       it using ptest instruction.  */
>    if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
>      {
>        rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
> -      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
>
>        gcc_assert (code == EQ || code == NE);
>        /* Generate XOR since we can't check that one operand is zero vector.  */
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index b321cda1f22..f91325015c9 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -1069,6 +1069,10 @@ (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
>                              (HI "TARGET_HIMODE_MATH")
>                              SI DI (TI "TARGET_64BIT")])
>
> +(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
> +                                 (HI "TARGET_HIMODE_MATH")
> +                                 SI DI])
> +
>  ;; Math-dependant single word integer modes.
>  (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
>                             (HI "TARGET_HIMODE_MATH")
> @@ -1322,8 +1326,8 @@ (define_mode_iterator PTR
>
>  (define_expand "cbranch<mode>4"
>    [(set (reg:CC FLAGS_REG)
> -       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
> -                   (match_operand:SDWIM 2 "<general_operand>")))
> +       (compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
> +                   (match_operand:SDWIM1248 2 "<general_operand>")))
>     (set (pc) (if_then_else
>                (match_operator 0 "ordered_comparison_operator"
>                 [(reg:CC FLAGS_REG) (const_int 0)])
> @@ -1338,6 +1342,25 @@ (define_expand "cbranch<mode>4"
>    DONE;
>  })
>
> +(define_expand "cbranchti4"
> +  [(set (reg:CC FLAGS_REG)
> +       (compare:CC (match_operand:TI 1 "nonimmediate_operand")
> +                   (match_operand:TI 2 "x86_64_general_operand")))
> +   (set (pc) (if_then_else
> +              (match_operator 0 "timode_comparison_operator"
> +               [(reg:CC FLAGS_REG) (const_int 0)])
> +              (label_ref (match_operand 3))
> +              (pc)))]
> +  "TARGET_64BIT || TARGET_SSE4_1"
> +{
> +  if (MEM_P (operands[1]) && MEM_P (operands[2]))
> +    operands[1] = force_reg (TImode, operands[1]);
> +
> +  ix86_expand_branch (GET_CODE (operands[0]),
> +                     operands[1], operands[2], operands[3]);
> +  DONE;
> +})
> +
>  (define_expand "cstore<mode>4"
>    [(set (reg:CC FLAGS_REG)
>         (compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
> diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> index a8cc17a054d..fb3be3a262f 100644
> --- a/gcc/config/i386/predicates.md
> +++ b/gcc/config/i386/predicates.md
> @@ -1414,6 +1414,12 @@ (define_predicate "ix86_comparison_uns_operator"
>  (define_predicate "bt_comparison_operator"
>    (match_code "ne,eq"))
>
> +(define_predicate "timode_comparison_operator"
> +  (ior (and (match_test "TARGET_SSE4_1")
> +           (match_operand 0 "bt_comparison_operator"))
> +       (and (match_test "TARGET_64BIT")
> +           (match_operand 0 "ordered_comparison_operator"))))
> +
>  (define_predicate "shr_comparison_operator"
>    (match_code "gtu,leu"))
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index 7b791def542..0d194cee769 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -26034,10 +26034,14 @@ (define_expand "maskstore<mode><avx512fmaskmodelower>"
>           (match_operand:<avx512fmaskmode> 2 "register_operand")))]
>    "TARGET_AVX512BW")
>
> +(define_mode_iterator VI48_OI_AVX
> +  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
> +   V4SI V2DI])
> +
>  (define_expand "cbranch<mode>4"
>    [(set (reg:CC FLAGS_REG)
> -       (compare:CC (match_operand:VI48_AVX 1 "register_operand")
> -                   (match_operand:VI48_AVX 2 "nonimmediate_operand")))
> +       (compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
> +                   (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
>     (set (pc) (if_then_else
>                (match_operator 0 "bt_comparison_operator"
>                 [(reg:CC FLAGS_REG) (const_int 0)])
> @@ -26045,6 +26049,8 @@ (define_expand "cbranch<mode>4"
>                (pc)))]
>    "TARGET_SSE4_1"
>  {
> +  if (!vector_operand (operands[2], <MODE>mode))
> +    operands[2] = force_reg (<MODE>mode, operands[2]);
>    ix86_expand_branch (GET_CODE (operands[0]),
>                       operands[1], operands[2], operands[3]);
>    DONE;
> diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
> new file mode 100644
> index 00000000000..68f548594fd
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr104610.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
> +/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
> +/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
> +/* { dg-final { scan-assembler-times {sete} 2 } } */
> +/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
> +/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
> +
> +
> +#include<stdbool.h>
> +__attribute__((target("sse4.1")))
> +bool f128(char *a)
> +{
> +  char t[] = "012345678901234";
> +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> +}
> +
> +__attribute__((target("avx")))
> +bool f256(char *a)
> +{
> +  char t[] = "0123456789012345678901234567890";
> +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> +}
> --
> 2.18.1
>
Richard Biener May 5, 2022, 7:49 a.m. UTC | #2 
On Thu, May 5, 2022 at 9:37 AM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Enable optimization for TImode only under 32-bit target, for 64-bit
> target there could be extra ineteger <-> sse move regarding psABI,
> not efficient.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for trunk?

I wonder if this is better done in STV where we could assess this extra cost?

> gcc/ChangeLog:
>
>         PR target/104610
>         * config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
>         for TI/QImode when code is EQ or NE.
>         * config/i386/i386.md (SDWIM1248): New iterator.
>         (cbranch<mode>4): Split TImode into a separate expander.
>         (cbranchti4): New expander.
>         * config/i386/predicates.md (timode_comparison_operator): New
>         predicate.
>         * config/i386/sse.md (cbranch<mode>4): Extend to OImode.
>
> gcc/testsuite/ChangeLog:
>
>         * gcc.target/i386/pr104610.c: New test.
> ---
>  gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
>  gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
>  gcc/config/i386/predicates.md            |  6 ++++++
>  gcc/config/i386/sse.md                   | 10 +++++++--
>  gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
>  5 files changed, 74 insertions(+), 5 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c
>
> diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> index bc806ffa283..a2012a158ae 100644
> --- a/gcc/config/i386/i386-expand.cc
> +++ b/gcc/config/i386/i386-expand.cc
> @@ -2264,13 +2264,24 @@ ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
>  {
>    machine_mode mode = GET_MODE (op0);
>    rtx tmp;
> +  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> +  /* Using ptest for TImode only for 32-bit target since it's splitted into
> +     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
> +     move regarding psABI, not efficient.  */
> +  if ((code == EQ || code == NE)
> +       && ((mode == OImode && TARGET_AVX)
> +          || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
> +    {
> +      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
> +      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
> +      mode = p_mode;
> +    }
>
>    /* Handle special case - vector comparsion with boolean result, transform
>       it using ptest instruction.  */
>    if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
>      {
>        rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
> -      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
>
>        gcc_assert (code == EQ || code == NE);
>        /* Generate XOR since we can't check that one operand is zero vector.  */
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index b321cda1f22..f91325015c9 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -1069,6 +1069,10 @@ (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
>                              (HI "TARGET_HIMODE_MATH")
>                              SI DI (TI "TARGET_64BIT")])
>
> +(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
> +                                 (HI "TARGET_HIMODE_MATH")
> +                                 SI DI])
> +
>  ;; Math-dependant single word integer modes.
>  (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
>                             (HI "TARGET_HIMODE_MATH")
> @@ -1322,8 +1326,8 @@ (define_mode_iterator PTR
>
>  (define_expand "cbranch<mode>4"
>    [(set (reg:CC FLAGS_REG)
> -       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
> -                   (match_operand:SDWIM 2 "<general_operand>")))
> +       (compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
> +                   (match_operand:SDWIM1248 2 "<general_operand>")))
>     (set (pc) (if_then_else
>                (match_operator 0 "ordered_comparison_operator"
>                 [(reg:CC FLAGS_REG) (const_int 0)])
> @@ -1338,6 +1342,25 @@ (define_expand "cbranch<mode>4"
>    DONE;
>  })
>
> +(define_expand "cbranchti4"
> +  [(set (reg:CC FLAGS_REG)
> +       (compare:CC (match_operand:TI 1 "nonimmediate_operand")
> +                   (match_operand:TI 2 "x86_64_general_operand")))
> +   (set (pc) (if_then_else
> +              (match_operator 0 "timode_comparison_operator"
> +               [(reg:CC FLAGS_REG) (const_int 0)])
> +              (label_ref (match_operand 3))
> +              (pc)))]
> +  "TARGET_64BIT || TARGET_SSE4_1"
> +{
> +  if (MEM_P (operands[1]) && MEM_P (operands[2]))
> +    operands[1] = force_reg (TImode, operands[1]);
> +
> +  ix86_expand_branch (GET_CODE (operands[0]),
> +                     operands[1], operands[2], operands[3]);
> +  DONE;
> +})
> +
>  (define_expand "cstore<mode>4"
>    [(set (reg:CC FLAGS_REG)
>         (compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
> diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> index a8cc17a054d..fb3be3a262f 100644
> --- a/gcc/config/i386/predicates.md
> +++ b/gcc/config/i386/predicates.md
> @@ -1414,6 +1414,12 @@ (define_predicate "ix86_comparison_uns_operator"
>  (define_predicate "bt_comparison_operator"
>    (match_code "ne,eq"))
>
> +(define_predicate "timode_comparison_operator"
> +  (ior (and (match_test "TARGET_SSE4_1")
> +           (match_operand 0 "bt_comparison_operator"))
> +       (and (match_test "TARGET_64BIT")
> +           (match_operand 0 "ordered_comparison_operator"))))
> +
>  (define_predicate "shr_comparison_operator"
>    (match_code "gtu,leu"))
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index 7b791def542..0d194cee769 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -26034,10 +26034,14 @@ (define_expand "maskstore<mode><avx512fmaskmodelower>"
>           (match_operand:<avx512fmaskmode> 2 "register_operand")))]
>    "TARGET_AVX512BW")
>
> +(define_mode_iterator VI48_OI_AVX
> +  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
> +   V4SI V2DI])
> +
>  (define_expand "cbranch<mode>4"
>    [(set (reg:CC FLAGS_REG)
> -       (compare:CC (match_operand:VI48_AVX 1 "register_operand")
> -                   (match_operand:VI48_AVX 2 "nonimmediate_operand")))
> +       (compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
> +                   (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
>     (set (pc) (if_then_else
>                (match_operator 0 "bt_comparison_operator"
>                 [(reg:CC FLAGS_REG) (const_int 0)])
> @@ -26045,6 +26049,8 @@ (define_expand "cbranch<mode>4"
>                (pc)))]
>    "TARGET_SSE4_1"
>  {
> +  if (!vector_operand (operands[2], <MODE>mode))
> +    operands[2] = force_reg (<MODE>mode, operands[2]);
>    ix86_expand_branch (GET_CODE (operands[0]),
>                       operands[1], operands[2], operands[3]);
>    DONE;
> diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
> new file mode 100644
> index 00000000000..68f548594fd
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr104610.c
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
> +/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
> +/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
> +/* { dg-final { scan-assembler-times {sete} 2 } } */
> +/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
> +/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
> +
> +
> +#include<stdbool.h>
> +__attribute__((target("sse4.1")))
> +bool f128(char *a)
> +{
> +  char t[] = "012345678901234";
> +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> +}
> +
> +__attribute__((target("avx")))
> +bool f256(char *a)
> +{
> +  char t[] = "0123456789012345678901234567890";
> +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> +}
> --
> 2.18.1
>
Uros Bizjak May 5, 2022, 8:08 a.m. UTC | #3 
On Thu, May 5, 2022 at 9:50 AM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Thu, May 5, 2022 at 9:37 AM liuhongt via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > Enable optimization for TImode only under 32-bit target, for 64-bit
> > target there could be extra ineteger <-> sse move regarding psABI,
> > not efficient.
> >
> > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> > Ok for trunk?
>
> I wonder if this is better done in STV where we could assess this extra cost?

Yes, this should be handled via STV, Roger Sayle (CC'd) has proposed a
patch that does just that.


Uros.

> > gcc/ChangeLog:
> >
> >         PR target/104610
> >         * config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
> >         for TI/QImode when code is EQ or NE.
> >         * config/i386/i386.md (SDWIM1248): New iterator.
> >         (cbranch<mode>4): Split TImode into a separate expander.
> >         (cbranchti4): New expander.
> >         * config/i386/predicates.md (timode_comparison_operator): New
> >         predicate.
> >         * config/i386/sse.md (cbranch<mode>4): Extend to OImode.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.target/i386/pr104610.c: New test.
> > ---
> >  gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
> >  gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
> >  gcc/config/i386/predicates.md            |  6 ++++++
> >  gcc/config/i386/sse.md                   | 10 +++++++--
> >  gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
> >  5 files changed, 74 insertions(+), 5 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c
> >
> > diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> > index bc806ffa283..a2012a158ae 100644
> > --- a/gcc/config/i386/i386-expand.cc
> > +++ b/gcc/config/i386/i386-expand.cc
> > @@ -2264,13 +2264,24 @@ ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
> >  {
> >    machine_mode mode = GET_MODE (op0);
> >    rtx tmp;
> > +  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> > +  /* Using ptest for TImode only for 32-bit target since it's splitted into
> > +     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
> > +     move regarding psABI, not efficient.  */
> > +  if ((code == EQ || code == NE)
> > +       && ((mode == OImode && TARGET_AVX)
> > +          || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
> > +    {
> > +      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
> > +      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
> > +      mode = p_mode;
> > +    }
> >
> >    /* Handle special case - vector comparsion with boolean result, transform
> >       it using ptest instruction.  */
> >    if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
> >      {
> >        rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
> > -      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> >
> >        gcc_assert (code == EQ || code == NE);
> >        /* Generate XOR since we can't check that one operand is zero vector.  */
> > diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> > index b321cda1f22..f91325015c9 100644
> > --- a/gcc/config/i386/i386.md
> > +++ b/gcc/config/i386/i386.md
> > @@ -1069,6 +1069,10 @@ (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
> >                              (HI "TARGET_HIMODE_MATH")
> >                              SI DI (TI "TARGET_64BIT")])
> >
> > +(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
> > +                                 (HI "TARGET_HIMODE_MATH")
> > +                                 SI DI])
> > +
> >  ;; Math-dependant single word integer modes.
> >  (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
> >                             (HI "TARGET_HIMODE_MATH")
> > @@ -1322,8 +1326,8 @@ (define_mode_iterator PTR
> >
> >  (define_expand "cbranch<mode>4"
> >    [(set (reg:CC FLAGS_REG)
> > -       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
> > -                   (match_operand:SDWIM 2 "<general_operand>")))
> > +       (compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
> > +                   (match_operand:SDWIM1248 2 "<general_operand>")))
> >     (set (pc) (if_then_else
> >                (match_operator 0 "ordered_comparison_operator"
> >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > @@ -1338,6 +1342,25 @@ (define_expand "cbranch<mode>4"
> >    DONE;
> >  })
> >
> > +(define_expand "cbranchti4"
> > +  [(set (reg:CC FLAGS_REG)
> > +       (compare:CC (match_operand:TI 1 "nonimmediate_operand")
> > +                   (match_operand:TI 2 "x86_64_general_operand")))
> > +   (set (pc) (if_then_else
> > +              (match_operator 0 "timode_comparison_operator"
> > +               [(reg:CC FLAGS_REG) (const_int 0)])
> > +              (label_ref (match_operand 3))
> > +              (pc)))]
> > +  "TARGET_64BIT || TARGET_SSE4_1"
> > +{
> > +  if (MEM_P (operands[1]) && MEM_P (operands[2]))
> > +    operands[1] = force_reg (TImode, operands[1]);
> > +
> > +  ix86_expand_branch (GET_CODE (operands[0]),
> > +                     operands[1], operands[2], operands[3]);
> > +  DONE;
> > +})
> > +
> >  (define_expand "cstore<mode>4"
> >    [(set (reg:CC FLAGS_REG)
> >         (compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
> > diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> > index a8cc17a054d..fb3be3a262f 100644
> > --- a/gcc/config/i386/predicates.md
> > +++ b/gcc/config/i386/predicates.md
> > @@ -1414,6 +1414,12 @@ (define_predicate "ix86_comparison_uns_operator"
> >  (define_predicate "bt_comparison_operator"
> >    (match_code "ne,eq"))
> >
> > +(define_predicate "timode_comparison_operator"
> > +  (ior (and (match_test "TARGET_SSE4_1")
> > +           (match_operand 0 "bt_comparison_operator"))
> > +       (and (match_test "TARGET_64BIT")
> > +           (match_operand 0 "ordered_comparison_operator"))))
> > +
> >  (define_predicate "shr_comparison_operator"
> >    (match_code "gtu,leu"))
> >
> > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > index 7b791def542..0d194cee769 100644
> > --- a/gcc/config/i386/sse.md
> > +++ b/gcc/config/i386/sse.md
> > @@ -26034,10 +26034,14 @@ (define_expand "maskstore<mode><avx512fmaskmodelower>"
> >           (match_operand:<avx512fmaskmode> 2 "register_operand")))]
> >    "TARGET_AVX512BW")
> >
> > +(define_mode_iterator VI48_OI_AVX
> > +  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
> > +   V4SI V2DI])
> > +
> >  (define_expand "cbranch<mode>4"
> >    [(set (reg:CC FLAGS_REG)
> > -       (compare:CC (match_operand:VI48_AVX 1 "register_operand")
> > -                   (match_operand:VI48_AVX 2 "nonimmediate_operand")))
> > +       (compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
> > +                   (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
> >     (set (pc) (if_then_else
> >                (match_operator 0 "bt_comparison_operator"
> >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > @@ -26045,6 +26049,8 @@ (define_expand "cbranch<mode>4"
> >                (pc)))]
> >    "TARGET_SSE4_1"
> >  {
> > +  if (!vector_operand (operands[2], <MODE>mode))
> > +    operands[2] = force_reg (<MODE>mode, operands[2]);
> >    ix86_expand_branch (GET_CODE (operands[0]),
> >                       operands[1], operands[2], operands[3]);
> >    DONE;
> > diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
> > new file mode 100644
> > index 00000000000..68f548594fd
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr104610.c
> > @@ -0,0 +1,23 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
> > +/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
> > +/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
> > +/* { dg-final { scan-assembler-times {sete} 2 } } */
> > +/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
> > +/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
> > +
> > +
> > +#include<stdbool.h>
> > +__attribute__((target("sse4.1")))
> > +bool f128(char *a)
> > +{
> > +  char t[] = "012345678901234";
> > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > +}
> > +
> > +__attribute__((target("avx")))
> > +bool f256(char *a)
> > +{
> > +  char t[] = "0123456789012345678901234567890";
> > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > +}
> > --
> > 2.18.1
> >
Uros Bizjak May 5, 2022, 8:21 a.m. UTC | #4 
On Thu, May 5, 2022 at 10:08 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Thu, May 5, 2022 at 9:50 AM Richard Biener via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Thu, May 5, 2022 at 9:37 AM liuhongt via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Enable optimization for TImode only under 32-bit target, for 64-bit
> > > target there could be extra ineteger <-> sse move regarding psABI,
> > > not efficient.
> > >
> > > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> > > Ok for trunk?
> >
> > I wonder if this is better done in STV where we could assess this extra cost?
>
> Yes, this should be handled via STV, Roger Sayle (CC'd) has proposed a
> patch that does just that.

https://gcc.gnu.org/pipermail/gcc-patches/2022-April/593174.html
Uros.

>
> > > gcc/ChangeLog:
> > >
> > >         PR target/104610
> > >         * config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
> > >         for TI/QImode when code is EQ or NE.
> > >         * config/i386/i386.md (SDWIM1248): New iterator.
> > >         (cbranch<mode>4): Split TImode into a separate expander.
> > >         (cbranchti4): New expander.
> > >         * config/i386/predicates.md (timode_comparison_operator): New
> > >         predicate.
> > >         * config/i386/sse.md (cbranch<mode>4): Extend to OImode.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >         * gcc.target/i386/pr104610.c: New test.
> > > ---
> > >  gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
> > >  gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
> > >  gcc/config/i386/predicates.md            |  6 ++++++
> > >  gcc/config/i386/sse.md                   | 10 +++++++--
> > >  gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
> > >  5 files changed, 74 insertions(+), 5 deletions(-)
> > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c
> > >
> > > diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> > > index bc806ffa283..a2012a158ae 100644
> > > --- a/gcc/config/i386/i386-expand.cc
> > > +++ b/gcc/config/i386/i386-expand.cc
> > > @@ -2264,13 +2264,24 @@ ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
> > >  {
> > >    machine_mode mode = GET_MODE (op0);
> > >    rtx tmp;
> > > +  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> > > +  /* Using ptest for TImode only for 32-bit target since it's splitted into
> > > +     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
> > > +     move regarding psABI, not efficient.  */
> > > +  if ((code == EQ || code == NE)
> > > +       && ((mode == OImode && TARGET_AVX)
> > > +          || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
> > > +    {
> > > +      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
> > > +      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
> > > +      mode = p_mode;
> > > +    }
> > >
> > >    /* Handle special case - vector comparsion with boolean result, transform
> > >       it using ptest instruction.  */
> > >    if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
> > >      {
> > >        rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
> > > -      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> > >
> > >        gcc_assert (code == EQ || code == NE);
> > >        /* Generate XOR since we can't check that one operand is zero vector.  */
> > > diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> > > index b321cda1f22..f91325015c9 100644
> > > --- a/gcc/config/i386/i386.md
> > > +++ b/gcc/config/i386/i386.md
> > > @@ -1069,6 +1069,10 @@ (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
> > >                              (HI "TARGET_HIMODE_MATH")
> > >                              SI DI (TI "TARGET_64BIT")])
> > >
> > > +(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
> > > +                                 (HI "TARGET_HIMODE_MATH")
> > > +                                 SI DI])
> > > +
> > >  ;; Math-dependant single word integer modes.
> > >  (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
> > >                             (HI "TARGET_HIMODE_MATH")
> > > @@ -1322,8 +1326,8 @@ (define_mode_iterator PTR
> > >
> > >  (define_expand "cbranch<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > > -       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
> > > -                   (match_operand:SDWIM 2 "<general_operand>")))
> > > +       (compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
> > > +                   (match_operand:SDWIM1248 2 "<general_operand>")))
> > >     (set (pc) (if_then_else
> > >                (match_operator 0 "ordered_comparison_operator"
> > >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > > @@ -1338,6 +1342,25 @@ (define_expand "cbranch<mode>4"
> > >    DONE;
> > >  })
> > >
> > > +(define_expand "cbranchti4"
> > > +  [(set (reg:CC FLAGS_REG)
> > > +       (compare:CC (match_operand:TI 1 "nonimmediate_operand")
> > > +                   (match_operand:TI 2 "x86_64_general_operand")))
> > > +   (set (pc) (if_then_else
> > > +              (match_operator 0 "timode_comparison_operator"
> > > +               [(reg:CC FLAGS_REG) (const_int 0)])
> > > +              (label_ref (match_operand 3))
> > > +              (pc)))]
> > > +  "TARGET_64BIT || TARGET_SSE4_1"
> > > +{
> > > +  if (MEM_P (operands[1]) && MEM_P (operands[2]))
> > > +    operands[1] = force_reg (TImode, operands[1]);
> > > +
> > > +  ix86_expand_branch (GET_CODE (operands[0]),
> > > +                     operands[1], operands[2], operands[3]);
> > > +  DONE;
> > > +})
> > > +
> > >  (define_expand "cstore<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > >         (compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
> > > diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> > > index a8cc17a054d..fb3be3a262f 100644
> > > --- a/gcc/config/i386/predicates.md
> > > +++ b/gcc/config/i386/predicates.md
> > > @@ -1414,6 +1414,12 @@ (define_predicate "ix86_comparison_uns_operator"
> > >  (define_predicate "bt_comparison_operator"
> > >    (match_code "ne,eq"))
> > >
> > > +(define_predicate "timode_comparison_operator"
> > > +  (ior (and (match_test "TARGET_SSE4_1")
> > > +           (match_operand 0 "bt_comparison_operator"))
> > > +       (and (match_test "TARGET_64BIT")
> > > +           (match_operand 0 "ordered_comparison_operator"))))
> > > +
> > >  (define_predicate "shr_comparison_operator"
> > >    (match_code "gtu,leu"))
> > >
> > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > > index 7b791def542..0d194cee769 100644
> > > --- a/gcc/config/i386/sse.md
> > > +++ b/gcc/config/i386/sse.md
> > > @@ -26034,10 +26034,14 @@ (define_expand "maskstore<mode><avx512fmaskmodelower>"
> > >           (match_operand:<avx512fmaskmode> 2 "register_operand")))]
> > >    "TARGET_AVX512BW")
> > >
> > > +(define_mode_iterator VI48_OI_AVX
> > > +  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
> > > +   V4SI V2DI])
> > > +
> > >  (define_expand "cbranch<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > > -       (compare:CC (match_operand:VI48_AVX 1 "register_operand")
> > > -                   (match_operand:VI48_AVX 2 "nonimmediate_operand")))
> > > +       (compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
> > > +                   (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
> > >     (set (pc) (if_then_else
> > >                (match_operator 0 "bt_comparison_operator"
> > >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > > @@ -26045,6 +26049,8 @@ (define_expand "cbranch<mode>4"
> > >                (pc)))]
> > >    "TARGET_SSE4_1"
> > >  {
> > > +  if (!vector_operand (operands[2], <MODE>mode))
> > > +    operands[2] = force_reg (<MODE>mode, operands[2]);
> > >    ix86_expand_branch (GET_CODE (operands[0]),
> > >                       operands[1], operands[2], operands[3]);
> > >    DONE;
> > > diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
> > > new file mode 100644
> > > index 00000000000..68f548594fd
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/i386/pr104610.c
> > > @@ -0,0 +1,23 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
> > > +/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
> > > +/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
> > > +/* { dg-final { scan-assembler-times {sete} 2 } } */
> > > +/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
> > > +/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
> > > +
> > > +
> > > +#include<stdbool.h>
> > > +__attribute__((target("sse4.1")))
> > > +bool f128(char *a)
> > > +{
> > > +  char t[] = "012345678901234";
> > > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > > +}
> > > +
> > > +__attribute__((target("avx")))
> > > +bool f256(char *a)
> > > +{
> > > +  char t[] = "0123456789012345678901234567890";
> > > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > > +}
> > > --
> > > 2.18.1
> > >
Hongtao Liu May 5, 2022, 8:22 a.m. UTC | #5 
On Thu, May 5, 2022 at 4:09 PM Uros Bizjak via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Thu, May 5, 2022 at 9:50 AM Richard Biener via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Thu, May 5, 2022 at 9:37 AM liuhongt via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Enable optimization for TImode only under 32-bit target, for 64-bit
> > > target there could be extra ineteger <-> sse move regarding psABI,
> > > not efficient.
> > >
> > > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> > > Ok for trunk?
> >
> > I wonder if this is better done in STV where we could assess this extra cost?
>
> Yes, this should be handled via STV, Roger Sayle (CC'd) has proposed a
> patch that does just that.
>
My patch also handles OImode, I think that part could be a separate patch.
>
> Uros.
>
> > > gcc/ChangeLog:
> > >
> > >         PR target/104610
> > >         * config/i386/i386-expand.cc (ix86_expand_branch): Use ptest
> > >         for TI/QImode when code is EQ or NE.
> > >         * config/i386/i386.md (SDWIM1248): New iterator.
> > >         (cbranch<mode>4): Split TImode into a separate expander.
> > >         (cbranchti4): New expander.
> > >         * config/i386/predicates.md (timode_comparison_operator): New
> > >         predicate.
> > >         * config/i386/sse.md (cbranch<mode>4): Extend to OImode.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >         * gcc.target/i386/pr104610.c: New test.
> > > ---
> > >  gcc/config/i386/i386-expand.cc           | 13 +++++++++++-
> > >  gcc/config/i386/i386.md                  | 27 ++++++++++++++++++++++--
> > >  gcc/config/i386/predicates.md            |  6 ++++++
> > >  gcc/config/i386/sse.md                   | 10 +++++++--
> > >  gcc/testsuite/gcc.target/i386/pr104610.c | 23 ++++++++++++++++++++
> > >  5 files changed, 74 insertions(+), 5 deletions(-)
> > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr104610.c
> > >
> > > diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> > > index bc806ffa283..a2012a158ae 100644
> > > --- a/gcc/config/i386/i386-expand.cc
> > > +++ b/gcc/config/i386/i386-expand.cc
> > > @@ -2264,13 +2264,24 @@ ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
> > >  {
> > >    machine_mode mode = GET_MODE (op0);
> > >    rtx tmp;
> > > +  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> > > +  /* Using ptest for TImode only for 32-bit target since it's splitted into
> > > +     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
> > > +     move regarding psABI, not efficient.  */
> > > +  if ((code == EQ || code == NE)
> > > +       && ((mode == OImode && TARGET_AVX)
> > > +          || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
> > > +    {
> > > +      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
> > > +      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
> > > +      mode = p_mode;
> > > +    }
> > >
> > >    /* Handle special case - vector comparsion with boolean result, transform
> > >       it using ptest instruction.  */
> > >    if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
> > >      {
> > >        rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
> > > -      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
> > >
> > >        gcc_assert (code == EQ || code == NE);
> > >        /* Generate XOR since we can't check that one operand is zero vector.  */
> > > diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> > > index b321cda1f22..f91325015c9 100644
> > > --- a/gcc/config/i386/i386.md
> > > +++ b/gcc/config/i386/i386.md
> > > @@ -1069,6 +1069,10 @@ (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
> > >                              (HI "TARGET_HIMODE_MATH")
> > >                              SI DI (TI "TARGET_64BIT")])
> > >
> > > +(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
> > > +                                 (HI "TARGET_HIMODE_MATH")
> > > +                                 SI DI])
> > > +
> > >  ;; Math-dependant single word integer modes.
> > >  (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
> > >                             (HI "TARGET_HIMODE_MATH")
> > > @@ -1322,8 +1326,8 @@ (define_mode_iterator PTR
> > >
> > >  (define_expand "cbranch<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > > -       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
> > > -                   (match_operand:SDWIM 2 "<general_operand>")))
> > > +       (compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
> > > +                   (match_operand:SDWIM1248 2 "<general_operand>")))
> > >     (set (pc) (if_then_else
> > >                (match_operator 0 "ordered_comparison_operator"
> > >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > > @@ -1338,6 +1342,25 @@ (define_expand "cbranch<mode>4"
> > >    DONE;
> > >  })
> > >
> > > +(define_expand "cbranchti4"
> > > +  [(set (reg:CC FLAGS_REG)
> > > +       (compare:CC (match_operand:TI 1 "nonimmediate_operand")
> > > +                   (match_operand:TI 2 "x86_64_general_operand")))
> > > +   (set (pc) (if_then_else
> > > +              (match_operator 0 "timode_comparison_operator"
> > > +               [(reg:CC FLAGS_REG) (const_int 0)])
> > > +              (label_ref (match_operand 3))
> > > +              (pc)))]
> > > +  "TARGET_64BIT || TARGET_SSE4_1"
> > > +{
> > > +  if (MEM_P (operands[1]) && MEM_P (operands[2]))
> > > +    operands[1] = force_reg (TImode, operands[1]);
> > > +
> > > +  ix86_expand_branch (GET_CODE (operands[0]),
> > > +                     operands[1], operands[2], operands[3]);
> > > +  DONE;
> > > +})
> > > +
> > >  (define_expand "cstore<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > >         (compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
> > > diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
> > > index a8cc17a054d..fb3be3a262f 100644
> > > --- a/gcc/config/i386/predicates.md
> > > +++ b/gcc/config/i386/predicates.md
> > > @@ -1414,6 +1414,12 @@ (define_predicate "ix86_comparison_uns_operator"
> > >  (define_predicate "bt_comparison_operator"
> > >    (match_code "ne,eq"))
> > >
> > > +(define_predicate "timode_comparison_operator"
> > > +  (ior (and (match_test "TARGET_SSE4_1")
> > > +           (match_operand 0 "bt_comparison_operator"))
> > > +       (and (match_test "TARGET_64BIT")
> > > +           (match_operand 0 "ordered_comparison_operator"))))
> > > +
> > >  (define_predicate "shr_comparison_operator"
> > >    (match_code "gtu,leu"))
> > >
> > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > > index 7b791def542..0d194cee769 100644
> > > --- a/gcc/config/i386/sse.md
> > > +++ b/gcc/config/i386/sse.md
> > > @@ -26034,10 +26034,14 @@ (define_expand "maskstore<mode><avx512fmaskmodelower>"
> > >           (match_operand:<avx512fmaskmode> 2 "register_operand")))]
> > >    "TARGET_AVX512BW")
> > >
> > > +(define_mode_iterator VI48_OI_AVX
> > > +  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
> > > +   V4SI V2DI])
> > > +
> > >  (define_expand "cbranch<mode>4"
> > >    [(set (reg:CC FLAGS_REG)
> > > -       (compare:CC (match_operand:VI48_AVX 1 "register_operand")
> > > -                   (match_operand:VI48_AVX 2 "nonimmediate_operand")))
> > > +       (compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
> > > +                   (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
> > >     (set (pc) (if_then_else
> > >                (match_operator 0 "bt_comparison_operator"
> > >                 [(reg:CC FLAGS_REG) (const_int 0)])
> > > @@ -26045,6 +26049,8 @@ (define_expand "cbranch<mode>4"
> > >                (pc)))]
> > >    "TARGET_SSE4_1"
> > >  {
> > > +  if (!vector_operand (operands[2], <MODE>mode))
> > > +    operands[2] = force_reg (<MODE>mode, operands[2]);
> > >    ix86_expand_branch (GET_CODE (operands[0]),
> > >                       operands[1], operands[2], operands[3]);
> > >    DONE;
> > > diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
> > > new file mode 100644
> > > index 00000000000..68f548594fd
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/i386/pr104610.c
> > > @@ -0,0 +1,23 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
> > > +/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
> > > +/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
> > > +/* { dg-final { scan-assembler-times {sete} 2 } } */
> > > +/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
> > > +/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
> > > +
> > > +
> > > +#include<stdbool.h>
> > > +__attribute__((target("sse4.1")))
> > > +bool f128(char *a)
> > > +{
> > > +  char t[] = "012345678901234";
> > > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > > +}
> > > +
> > > +__attribute__((target("avx")))
> > > +bool f256(char *a)
> > > +{
> > > +  char t[] = "0123456789012345678901234567890";
> > > +  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
> > > +}
> > > --
> > > 2.18.1
> > >
Uros Bizjak May 5, 2022, 8:30 a.m. UTC | #6 
On Thu, May 5, 2022 at 10:23 AM Hongtao Liu <crazylht@gmail.com> wrote:
>
> On Thu, May 5, 2022 at 4:09 PM Uros Bizjak via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Thu, May 5, 2022 at 9:50 AM Richard Biener via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > On Thu, May 5, 2022 at 9:37 AM liuhongt via Gcc-patches
> > > <gcc-patches@gcc.gnu.org> wrote:
> > > >
> > > > Enable optimization for TImode only under 32-bit target, for 64-bit
> > > > target there could be extra ineteger <-> sse move regarding psABI,
> > > > not efficient.
> > > >
> > > > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> > > > Ok for trunk?
> > >
> > > I wonder if this is better done in STV where we could assess this extra cost?
> >
> > Yes, this should be handled via STV, Roger Sayle (CC'd) has proposed a
> > patch that does just that.
> >
> My patch also handles OImode, I think that part could be a separate patch.

Yes, OImode (and TImode on x86_32) can't be implemented using integer registers.

Uros,
diff mbox series

Patch

diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index bc806ffa283..a2012a158ae 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -2264,13 +2264,24 @@  ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label)
 {
   machine_mode mode = GET_MODE (op0);
   rtx tmp;
+  machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
+  /* Using ptest for TImode only for 32-bit target since it's splitted into
+     4 comparisons. For 64-bit target there could be extra ineteger <-> sse
+     move regarding psABI, not efficient.  */
+  if ((code == EQ || code == NE)
+       && ((mode == OImode && TARGET_AVX)
+	   || (mode == TImode && !TARGET_64BIT && TARGET_SSE4_1)))
+    {
+      op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode);
+      op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode);
+      mode = p_mode;
+    }
 
   /* Handle special case - vector comparsion with boolean result, transform
      it using ptest instruction.  */
   if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
     {
       rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG);
-      machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode;
 
       gcc_assert (code == EQ || code == NE);
       /* Generate XOR since we can't check that one operand is zero vector.  */
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index b321cda1f22..f91325015c9 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -1069,6 +1069,10 @@  (define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
 			     (HI "TARGET_HIMODE_MATH")
 			     SI DI (TI "TARGET_64BIT")])
 
+(define_mode_iterator SDWIM1248 [(QI "TARGET_QIMODE_MATH")
+				  (HI "TARGET_HIMODE_MATH")
+				  SI DI])
+
 ;; Math-dependant single word integer modes.
 (define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
 			    (HI "TARGET_HIMODE_MATH")
@@ -1322,8 +1326,8 @@  (define_mode_iterator PTR
 
 (define_expand "cbranch<mode>4"
   [(set (reg:CC FLAGS_REG)
-	(compare:CC (match_operand:SDWIM 1 "nonimmediate_operand")
-		    (match_operand:SDWIM 2 "<general_operand>")))
+	(compare:CC (match_operand:SDWIM1248 1 "nonimmediate_operand")
+		    (match_operand:SDWIM1248 2 "<general_operand>")))
    (set (pc) (if_then_else
 	       (match_operator 0 "ordered_comparison_operator"
 		[(reg:CC FLAGS_REG) (const_int 0)])
@@ -1338,6 +1342,25 @@  (define_expand "cbranch<mode>4"
   DONE;
 })
 
+(define_expand "cbranchti4"
+  [(set (reg:CC FLAGS_REG)
+	(compare:CC (match_operand:TI 1 "nonimmediate_operand")
+		    (match_operand:TI 2 "x86_64_general_operand")))
+   (set (pc) (if_then_else
+	       (match_operator 0 "timode_comparison_operator"
+		[(reg:CC FLAGS_REG) (const_int 0)])
+	       (label_ref (match_operand 3))
+	       (pc)))]
+  "TARGET_64BIT || TARGET_SSE4_1"
+{
+  if (MEM_P (operands[1]) && MEM_P (operands[2]))
+    operands[1] = force_reg (TImode, operands[1]);
+
+  ix86_expand_branch (GET_CODE (operands[0]),
+		      operands[1], operands[2], operands[3]);
+  DONE;
+})
+
 (define_expand "cstore<mode>4"
   [(set (reg:CC FLAGS_REG)
 	(compare:CC (match_operand:SWIM 2 "nonimmediate_operand")
diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
index a8cc17a054d..fb3be3a262f 100644
--- a/gcc/config/i386/predicates.md
+++ b/gcc/config/i386/predicates.md
@@ -1414,6 +1414,12 @@  (define_predicate "ix86_comparison_uns_operator"
 (define_predicate "bt_comparison_operator"
   (match_code "ne,eq"))
 
+(define_predicate "timode_comparison_operator"
+  (ior (and (match_test "TARGET_SSE4_1")
+	    (match_operand 0 "bt_comparison_operator"))
+       (and (match_test "TARGET_64BIT")
+	    (match_operand 0 "ordered_comparison_operator"))))
+
 (define_predicate "shr_comparison_operator"
   (match_code "gtu,leu"))
 
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 7b791def542..0d194cee769 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -26034,10 +26034,14 @@  (define_expand "maskstore<mode><avx512fmaskmodelower>"
 	  (match_operand:<avx512fmaskmode> 2 "register_operand")))]
   "TARGET_AVX512BW")
 
+(define_mode_iterator VI48_OI_AVX
+  [(V8SI "TARGET_AVX") (V4DI "TARGET_AVX") (OI "TARGET_AVX")
+   V4SI V2DI])
+
 (define_expand "cbranch<mode>4"
   [(set (reg:CC FLAGS_REG)
-	(compare:CC (match_operand:VI48_AVX 1 "register_operand")
-		    (match_operand:VI48_AVX 2 "nonimmediate_operand")))
+	(compare:CC (match_operand:VI48_OI_AVX 1 "register_operand")
+		    (match_operand:VI48_OI_AVX 2 "nonimmediate_operand")))
    (set (pc) (if_then_else
 	       (match_operator 0 "bt_comparison_operator"
 		[(reg:CC FLAGS_REG) (const_int 0)])
@@ -26045,6 +26049,8 @@  (define_expand "cbranch<mode>4"
 	       (pc)))]
   "TARGET_SSE4_1"
 {
+  if (!vector_operand (operands[2], <MODE>mode))
+    operands[2] = force_reg (<MODE>mode, operands[2]);
   ix86_expand_branch (GET_CODE (operands[0]),
 		      operands[1], operands[2], operands[3]);
   DONE;
diff --git a/gcc/testsuite/gcc.target/i386/pr104610.c b/gcc/testsuite/gcc.target/i386/pr104610.c
new file mode 100644
index 00000000000..68f548594fd
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr104610.c
@@ -0,0 +1,23 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -mmove-max=256 -mstore-max=256" } */
+/* { dg-final { scan-assembler-times {(?n)ptest.*xmm} 1 { target ia32 } } } */
+/* { dg-final { scan-assembler-times {(?n)vptest.*ymm} 1 } } */
+/* { dg-final { scan-assembler-times {sete} 2 } } */
+/* { dg-final { scan-assembler-not {(?n)je.*L[0-9]} } } */
+/* { dg-final { scan-assembler-not {(?n)jne.*L[0-9]} } } */
+
+
+#include<stdbool.h>
+__attribute__((target("sse4.1")))
+bool f128(char *a)
+{
+  char t[] = "012345678901234";
+  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
+}
+
+__attribute__((target("avx")))
+bool f256(char *a)
+{
+  char t[] = "0123456789012345678901234567890";
+  return __builtin_memcmp(a, &t[0], sizeof(t)) == 0;
+}