From: Uros Bizjak <ubizjak@gmail.com>
To: liuhongt <hongtao.liu@intel.com>
Cc: gcc-patches@gcc.gnu.org, crazylht@gmail.com, hjl.tools@gmail.com
Subject: Re: [PATCH V2] [x86] Fix incorrect digit constraint
Date: Mon, 31 Oct 2022 10:22:05 +0100 [thread overview]
Message-ID: <CAFULd4a60FTTXvYjLWVzv2PEDnK+hfo665f3dY3EerC3omUxAg@mail.gmail.com> (raw)
In-Reply-To: <20221031011036.1158443-1-hongtao.liu@intel.com>
On Mon, Oct 31, 2022 at 2:10 AM liuhongt <hongtao.liu@intel.com> wrote:
>
> >You have a couple of other patterns where operand 1 is matched to
> >produce vmovddup insn. These are *avx512f_unpcklpd512<mask_name> and
> >avx_unpcklpd256<mask_name>. You can also remove expander in both
> >cases.
>
> Yes, changed in V2 patch.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
>
>
> Matching constraints are used in these circumstances. More precisely,
> the two operands that match must include one input-only operand and
> one output-only operand. Moreover, the digit must be a smaller number
> than the number of the operand that uses it in the constraint.
>
> In pr107057, the 2 operands in the pattern are both input operands.
>
> gcc/ChangeLog:
>
> PR target/107057
> * config/i386/sse.md (*vec_interleave_highv2df): Remove
> constraint 1.
> (*vec_interleave_lowv2df): Ditto.
> (vec_concatv2df): Ditto.
> (*avx512f_unpcklpd512<mask_name>): Ditto and renamed to ..
> (avx512f_unpcklpd512<mask_name>): .. this.
> (avx512f_movddup512<mask_name>): Change to define_insn.
> (avx_movddup256<mask_name>): Ditto.
> (*avx_unpcklpd256<mask_name>): Remove constraint 1 and renamed
> to ..
> (avx_unpcklpd256<mask_name>): .. this.
> * config/i386/i386.cc (ix86_vec_interleave_v2df_operator_ok):
> Disallow MEM_P (op1) && MEM_P (op2).
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr107057.c: New test.
OK with two small adjustments.
Thanks,
Uros.
> ---
> gcc/config/i386/i386.cc | 2 +-
> gcc/config/i386/sse.md | 140 +++++++++--------------
> gcc/testsuite/gcc.target/i386/pr107057.c | 19 +++
> 3 files changed, 77 insertions(+), 84 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr107057.c
>
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index aeea26ef4be..e3b7bea0d68 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -15652,7 +15652,7 @@ ix86_vec_interleave_v2df_operator_ok (rtx operands[3], bool high)
> if (MEM_P (operands[0]))
> return rtx_equal_p (operands[0], operands[1 + high]);
> if (MEM_P (operands[1]) && MEM_P (operands[2]))
> - return TARGET_SSE3 && rtx_equal_p (operands[1], operands[2]);
> + return false;
> return true;
> }
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index f4b5506703f..b7922521734 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -12170,107 +12170,88 @@ (define_expand "vec_interleave_highv2df"
> })
>
> (define_insn "*vec_interleave_highv2df"
> - [(set (match_operand:V2DF 0 "nonimmediate_operand" "=x,v,v,x,v,m")
> + [(set (match_operand:V2DF 0 "nonimmediate_operand" "=x,v,x,v,m")
> (vec_select:V2DF
> (vec_concat:V4DF
> - (match_operand:V2DF 1 "nonimmediate_operand" " 0,v,o,o,o,v")
> - (match_operand:V2DF 2 "nonimmediate_operand" " x,v,1,0,v,0"))
> + (match_operand:V2DF 1 "nonimmediate_operand" " 0,v,o,o,v")
> + (match_operand:V2DF 2 "nonimmediate_operand" " x,v,0,v,0"))
> (parallel [(const_int 1)
> (const_int 3)])))]
> "TARGET_SSE2 && ix86_vec_interleave_v2df_operator_ok (operands, 1)"
> "@
> unpckhpd\t{%2, %0|%0, %2}
> vunpckhpd\t{%2, %1, %0|%0, %1, %2}
> - %vmovddup\t{%H1, %0|%0, %H1}
> movlpd\t{%H1, %0|%0, %H1}
> vmovlpd\t{%H1, %2, %0|%0, %2, %H1}
> %vmovhpd\t{%1, %0|%q0, %1}"
> - [(set_attr "isa" "noavx,avx,sse3,noavx,avx,*")
> - (set_attr "type" "sselog,sselog,sselog,ssemov,ssemov,ssemov")
> + [(set_attr "isa" "noavx,avx,noavx,avx,*")
> + (set_attr "type" "sselog,sselog,ssemov,ssemov,ssemov")
> (set (attr "prefix_data16")
> - (if_then_else (eq_attr "alternative" "3,5")
> + (if_then_else (eq_attr "alternative" "2,4")
> (const_string "1")
> (const_string "*")))
> - (set_attr "prefix" "orig,maybe_evex,maybe_vex,orig,maybe_evex,maybe_vex")
> - (set_attr "mode" "V2DF,V2DF,DF,V1DF,V1DF,V1DF")])
> + (set_attr "prefix" "orig,maybe_evex,orig,maybe_evex,maybe_vex")
> + (set_attr "mode" "V2DF,V2DF,V1DF,V1DF,V1DF")])
>
> -(define_expand "avx512f_movddup512<mask_name>"
> - [(set (match_operand:V8DF 0 "register_operand")
> +(define_insn "avx512f_movddup512<mask_name>"
> + [(set (match_operand:V8DF 0 "register_operand" "=v")
> (vec_select:V8DF
> (vec_concat:V16DF
> - (match_operand:V8DF 1 "nonimmediate_operand")
> + (match_operand:V8DF 1 "memory_operand" "m")
I think you should leave nonimmediate_operand here with "m" predicate.
Reload is able to move the register to the memory, and it is
beneficial to allow registers for possible combine opportunities.
> (match_dup 1))
> (parallel [(const_int 0) (const_int 8)
> (const_int 2) (const_int 10)
> (const_int 4) (const_int 12)
> (const_int 6) (const_int 14)])))]
> - "TARGET_AVX512F")
> -
> -(define_expand "avx512f_unpcklpd512<mask_name>"
> - [(set (match_operand:V8DF 0 "register_operand")
> - (vec_select:V8DF
> - (vec_concat:V16DF
> - (match_operand:V8DF 1 "register_operand")
> - (match_operand:V8DF 2 "nonimmediate_operand"))
> - (parallel [(const_int 0) (const_int 8)
> - (const_int 2) (const_int 10)
> - (const_int 4) (const_int 12)
> - (const_int 6) (const_int 14)])))]
> - "TARGET_AVX512F")
> + "TARGET_AVX512F"
> + "vmovddup\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "sselog")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "V8DF")])
>
> -(define_insn "*avx512f_unpcklpd512<mask_name>"
> - [(set (match_operand:V8DF 0 "register_operand" "=v,v")
> +(define_insn "avx512f_unpcklpd512<mask_name>"
> + [(set (match_operand:V8DF 0 "register_operand" "=v")
> (vec_select:V8DF
> (vec_concat:V16DF
> - (match_operand:V8DF 1 "nonimmediate_operand" "vm, v")
> - (match_operand:V8DF 2 "nonimmediate_operand" "1 ,vm"))
> + (match_operand:V8DF 1 "register_operand" "v")
> + (match_operand:V8DF 2 "nonimmediate_operand" "vm"))
> (parallel [(const_int 0) (const_int 8)
> (const_int 2) (const_int 10)
> (const_int 4) (const_int 12)
> (const_int 6) (const_int 14)])))]
> "TARGET_AVX512F"
> - "@
> - vmovddup\t{%1, %0<mask_operand3>|%0<mask_operand3>, %1}
> - vunpcklpd\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
> + "vunpcklpd\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
> [(set_attr "type" "sselog")
> (set_attr "prefix" "evex")
> (set_attr "mode" "V8DF")])
>
> ;; Recall that the 256-bit unpck insns only shuffle within their lanes.
> -(define_expand "avx_movddup256<mask_name>"
> - [(set (match_operand:V4DF 0 "register_operand")
> +(define_insn "avx_movddup256<mask_name>"
> + [(set (match_operand:V4DF 0 "register_operand" "=v")
> (vec_select:V4DF
> (vec_concat:V8DF
> - (match_operand:V4DF 1 "nonimmediate_operand")
> + (match_operand:V4DF 1 "memory_operand" "m")
Also here, leave nonimmediate_operand.
> (match_dup 1))
> (parallel [(const_int 0) (const_int 4)
> (const_int 2) (const_int 6)])))]
> - "TARGET_AVX && <mask_avx512vl_condition>")
> -
> -(define_expand "avx_unpcklpd256<mask_name>"
> - [(set (match_operand:V4DF 0 "register_operand")
> - (vec_select:V4DF
> - (vec_concat:V8DF
> - (match_operand:V4DF 1 "register_operand")
> - (match_operand:V4DF 2 "nonimmediate_operand"))
> - (parallel [(const_int 0) (const_int 4)
> - (const_int 2) (const_int 6)])))]
> - "TARGET_AVX && <mask_avx512vl_condition>")
> + "TARGET_AVX && <mask_avx512vl_condition>"
> + "vmovddup\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "sselog")
> + (set_attr "prefix" "<mask_prefix>")
> + (set_attr "mode" "V4DF")])
>
> -(define_insn "*avx_unpcklpd256<mask_name>"
> - [(set (match_operand:V4DF 0 "register_operand" "=v,v")
> +(define_insn "avx_unpcklpd256<mask_name>"
> + [(set (match_operand:V4DF 0 "register_operand" "=v")
> (vec_select:V4DF
> (vec_concat:V8DF
> - (match_operand:V4DF 1 "nonimmediate_operand" " v,m")
> - (match_operand:V4DF 2 "nonimmediate_operand" "vm,1"))
> + (match_operand:V4DF 1 "register_operand" " v")
> + (match_operand:V4DF 2 "nonimmediate_operand" "vm"))
> (parallel [(const_int 0) (const_int 4)
> (const_int 2) (const_int 6)])))]
> "TARGET_AVX && <mask_avx512vl_condition>"
> - "@
> - vunpcklpd\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}
> - vmovddup\t{%1, %0<mask_operand3>|%0<mask_operand3>, %1}"
> + "vunpcklpd\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
> [(set_attr "type" "sselog")
> - (set_attr "prefix" "vex")
> + (set_attr "prefix" "<mask_prefix>")
> (set_attr "mode" "V4DF")])
>
> (define_expand "vec_interleave_lowv4df"
> @@ -12332,29 +12313,28 @@ (define_expand "vec_interleave_lowv2df"
> })
>
> (define_insn "*vec_interleave_lowv2df"
> - [(set (match_operand:V2DF 0 "nonimmediate_operand" "=x,v,v,x,v,o")
> + [(set (match_operand:V2DF 0 "nonimmediate_operand" "=x,v,x,v,o")
> (vec_select:V2DF
> (vec_concat:V4DF
> - (match_operand:V2DF 1 "nonimmediate_operand" " 0,v,m,0,v,0")
> - (match_operand:V2DF 2 "nonimmediate_operand" " x,v,1,m,m,v"))
> + (match_operand:V2DF 1 "nonimmediate_operand" " 0,v,0,v,0")
> + (match_operand:V2DF 2 "nonimmediate_operand" " x,v,m,m,v"))
> (parallel [(const_int 0)
> (const_int 2)])))]
> "TARGET_SSE2 && ix86_vec_interleave_v2df_operator_ok (operands, 0)"
> "@
> unpcklpd\t{%2, %0|%0, %2}
> vunpcklpd\t{%2, %1, %0|%0, %1, %2}
> - %vmovddup\t{%1, %0|%0, %q1}
> movhpd\t{%2, %0|%0, %q2}
> vmovhpd\t{%2, %1, %0|%0, %1, %q2}
> %vmovlpd\t{%2, %H0|%H0, %2}"
> - [(set_attr "isa" "noavx,avx,sse3,noavx,avx,*")
> - (set_attr "type" "sselog,sselog,sselog,ssemov,ssemov,ssemov")
> + [(set_attr "isa" "noavx,avx,noavx,avx,*")
> + (set_attr "type" "sselog,sselog,ssemov,ssemov,ssemov")
> (set (attr "prefix_data16")
> - (if_then_else (eq_attr "alternative" "3,5")
> + (if_then_else (eq_attr "alternative" "2,4")
> (const_string "1")
> (const_string "*")))
> - (set_attr "prefix" "orig,maybe_evex,maybe_vex,orig,maybe_evex,maybe_vex")
> - (set_attr "mode" "V2DF,V2DF,DF,V1DF,V1DF,V1DF")])
> + (set_attr "prefix" "orig,maybe_evex,orig,maybe_evex,maybe_vex")
> + (set_attr "mode" "V2DF,V2DF,V1DF,V1DF,V1DF")])
>
> (define_split
> [(set (match_operand:V2DF 0 "memory_operand")
> @@ -13560,56 +13540,50 @@ (define_insn "vec_dupv2df<mask_name>"
> (set_attr "mode" "V2DF,DF,DF")])
>
> (define_insn "vec_concatv2df"
> - [(set (match_operand:V2DF 0 "register_operand" "=x,x,v,x,v,x,x, v,x,x")
> + [(set (match_operand:V2DF 0 "register_operand" "=x,x,v,x,x, v,x,x")
> (vec_concat:V2DF
> - (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,m,m,0,x,vm,0,0")
> - (match_operand:DF 2 "nonimm_or_0_operand" " x,x,v,1,1,m,m, C,x,m")))]
> - "TARGET_SSE
> - && (!(MEM_P (operands[1]) && MEM_P (operands[2]))
> - || (TARGET_SSE3 && rtx_equal_p (operands[1], operands[2])))"
> + (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,0,x,vm,0,0")
> + (match_operand:DF 2 "nonimm_or_0_operand" " x,x,v,m,m, C,x,m")))]
> + "TARGET_SSE && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
> "@
> unpcklpd\t{%2, %0|%0, %2}
> vunpcklpd\t{%2, %1, %0|%0, %1, %2}
> vunpcklpd\t{%2, %1, %0|%0, %1, %2}
> - %vmovddup\t{%1, %0|%0, %1}
> - vmovddup\t{%1, %0|%0, %1}
> movhpd\t{%2, %0|%0, %2}
> vmovhpd\t{%2, %1, %0|%0, %1, %2}
> %vmovq\t{%1, %0|%0, %1}
> movlhps\t{%2, %0|%0, %2}
> movhps\t{%2, %0|%0, %2}"
> [(set (attr "isa")
> - (cond [(eq_attr "alternative" "0,5")
> + (cond [(eq_attr "alternative" "0,3")
> (const_string "sse2_noavx")
> - (eq_attr "alternative" "1,6")
> + (eq_attr "alternative" "1,4")
> (const_string "avx")
> - (eq_attr "alternative" "2,4")
> + (eq_attr "alternative" "2")
> (const_string "avx512vl")
> - (eq_attr "alternative" "3")
> - (const_string "sse3")
> - (eq_attr "alternative" "7")
> + (eq_attr "alternative" "5")
> (const_string "sse2")
> ]
> (const_string "noavx")))
> (set (attr "type")
> (if_then_else
> - (eq_attr "alternative" "0,1,2,3,4")
> + (eq_attr "alternative" "0,1,2")
> (const_string "sselog")
> (const_string "ssemov")))
> (set (attr "prefix_data16")
> - (if_then_else (eq_attr "alternative" "5")
> + (if_then_else (eq_attr "alternative" "3")
> (const_string "1")
> (const_string "*")))
> (set (attr "prefix")
> - (cond [(eq_attr "alternative" "1,6")
> + (cond [(eq_attr "alternative" "1,4")
> (const_string "vex")
> - (eq_attr "alternative" "2,4")
> + (eq_attr "alternative" "2")
> (const_string "evex")
> - (eq_attr "alternative" "3,7")
> + (eq_attr "alternative" "5")
> (const_string "maybe_vex")
> ]
> (const_string "orig")))
> - (set_attr "mode" "V2DF,V2DF,V2DF, DF, DF, V1DF,V1DF,DF,V4SF,V2SF")])
> + (set_attr "mode" "V2DF,V2DF,V2DF,V1DF,V1DF,DF,V4SF,V2SF")])
>
> ;; vmovq clears also the higher bits.
> (define_insn "vec_set<mode>_0"
> diff --git a/gcc/testsuite/gcc.target/i386/pr107057.c b/gcc/testsuite/gcc.target/i386/pr107057.c
> new file mode 100644
> index 00000000000..40b49ac21ec
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr107057.c
> @@ -0,0 +1,19 @@
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-mavx -mcmodel=large -O3" } */
> +
> +typedef double v2df __attribute__ ((vector_size (16)));
> +v2df f (double a, double b)
> +{
> + v2df v;
> + double *c = (double *)&v;
> + *c = a;
> + *(c+1) = b;
> + return v;
> +}
> +void g ()
> +{
> + v2df x = f (1.0, 1.0);
> + v2df y = f (2.0, 2.0);
> + for (;*(double *)&x<=8; x+=y)
> + g ();
> +}
> --
> 2.27.0
>
next prev parent reply other threads:[~2022-10-31 9:22 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-10-27 10:53 [PATCH] " liuhongt
2022-10-27 12:29 ` Uros Bizjak
2022-10-31 1:10 ` [PATCH V2] " liuhongt
2022-10-31 9:22 ` Uros Bizjak [this message]
2022-11-01 8:39 ` Hongtao Liu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAFULd4a60FTTXvYjLWVzv2PEDnK+hfo665f3dY3EerC3omUxAg@mail.gmail.com \
--to=ubizjak@gmail.com \
--cc=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--cc=hongtao.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).