public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] aarch64: remove useless pairs of rev instructions
@ 2023-07-18  9:50 Serval Martinot-Lagarde
  2023-07-18 10:53 ` Prathamesh Kulkarni
  0 siblings, 1 reply; 3+ messages in thread
From: Serval Martinot-Lagarde @ 2023-07-18  9:50 UTC (permalink / raw)
  To: gcc-patches; +Cc: Etienne Renault

SVE generates superflous rev instructions that can be replaced
by single mov instruction or a pair of (rev, mov) instructions

gcc/
        * config/aarch64/aarch64-sve.md: New peephole2.
        * testsuite/gcc.target/aarch64/sve/revrev.c: New dg test.

Signed-off-by: Serval Martinot-Lagarde <serval.martinot-lagarde@sipearl.com>
---
 gcc/config/aarch64/aarch64-sve.md             | 21 +++++++++++++++++++
 gcc/testsuite/gcc.target/aarch64/sve/revrev.c | 13 ++++++++++++
 2 files changed, 34 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/revrev.c

diff --git a/gcc/config/aarch64/aarch64-sve.md b/gcc/config/aarch64/aarch64-sve.md
index da5534c3e32..e5e0c7ddfc5 100644
--- a/gcc/config/aarch64/aarch64-sve.md
+++ b/gcc/config/aarch64/aarch64-sve.md
@@ -8836,6 +8836,27 @@
   "TARGET_SVE"
   "rev\t%0.<Vctype>, %1.<Vctype>")
 
+(define_peephole2
+  [(set (match_operand:SVE_ALL 0 "register_operand" "")
+        (unspec:SVE_ALL
+          [(match_operand:SVE_ALL 1 "register_operand" "")] UNSPEC_REV))
+   (set (match_operand:SVE_ALL 2 "register_operand" "")
+        (unspec:SVE_ALL
+          [(match_dup 0)] UNSPEC_REV))]
+  "TARGET_SVE"
+  [(const_int 0)]
+  {
+    if (REGNO (operands[2]) != REGNO (operands[0]))
+    {
+      emit_insn (gen_rtx_SET (operands[2], operands[1]));
+      rtx rev = gen_rtx_UNSPEC (<MODE>mode, gen_rtvec (1, operands[1]), UNSPEC_REV);
+      emit_insn (gen_rtx_SET (operands[0], rev));
+    }
+    else
+      emit_insn (gen_rtx_SET (operands[0], operands[1]));
+    DONE;
+  })
+
 ;; -------------------------------------------------------------------------
 ;; ---- [INT,FP] Special-purpose binary permutes
 ;; -------------------------------------------------------------------------
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/revrev.c b/gcc/testsuite/gcc.target/aarch64/sve/revrev.c
new file mode 100644
index 00000000000..04af6eed291
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/revrev.c
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-O3" } */
+
+#include <stdint.h>
+
+void
+test (uint8_t a[], uint8_t b[], uint64_t N)
+{
+  for (uint64_t i = N; i > 0; i--)
+    a[i - 1] = b[i - 1];
+}
+
+/* { dg-final { scan-assembler-not {\trev\t(z[0-9]+\.h), \1\n\trev\t\1, \1\n} } } */
-- 
2.21.0

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] aarch64: remove useless pairs of rev instructions
  2023-07-18  9:50 [PATCH] aarch64: remove useless pairs of rev instructions Serval Martinot-Lagarde
@ 2023-07-18 10:53 ` Prathamesh Kulkarni
  2023-07-18 13:38   ` Serval Martinot-Lagarde
  0 siblings, 1 reply; 3+ messages in thread
From: Prathamesh Kulkarni @ 2023-07-18 10:53 UTC (permalink / raw)
  To: Serval Martinot-Lagarde; +Cc: gcc-patches, Etienne Renault

On Tue, 18 Jul 2023 at 15:20, Serval Martinot-Lagarde via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> SVE generates superflous rev instructions that can be replaced
> by single mov instruction or a pair of (rev, mov) instructions
Hi Serval,
I had added a similar transform to remove pair of rev instructions in:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=f0eabc52c9a2d3da0bfc201da7a5c1658b76e9a4

which seems to be removing the superfluous VEC_PERM_EXPR for
your test:
For example, following pair of VEC_PERM_EXPR is input to forwprop4:
  vect__4.9_58 = VEC_PERM_EXPR <vect__4.8_57, vect__4.8_57, {
POLY_INT_CST [15, 16], POLY_INT_CST [30, 32], POLY_INT_CST [45, 48],
... }>;
  vect__4.12_63 = VEC_PERM_EXPR <vect__4.9_58, vect__4.9_58, {
POLY_INT_CST [15, 16], POLY_INT_CST [30, 32], POLY_INT_CST [45, 48],
... }>;

and forwprop4 dump shows:
Removing dead stmt vect__4.12_63 = vect__4.8_57;

Which shows that forwprop replaced the VEC_PERM_EXPR pair with
assignment vect__4.12_63 = vect__4.8_57 (which turned out to be
eventually dead).

Sorry if this sounds silly to ask but could you let me know how to
reproduce this issue on trunk ?
I tried using -O3 -mcpu=generic+sve for your test but grepping for rev
didn't return any results in code-gen.

Thanks,
Prathamesh
>
> gcc/
>         * config/aarch64/aarch64-sve.md: New peephole2.
>         * testsuite/gcc.target/aarch64/sve/revrev.c: New dg test.
>
> Signed-off-by: Serval Martinot-Lagarde <serval.martinot-lagarde@sipearl.com>
> ---
>  gcc/config/aarch64/aarch64-sve.md             | 21 +++++++++++++++++++
>  gcc/testsuite/gcc.target/aarch64/sve/revrev.c | 13 ++++++++++++
>  2 files changed, 34 insertions(+)
>  create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/revrev.c
>
> diff --git a/gcc/config/aarch64/aarch64-sve.md b/gcc/config/aarch64/aarch64-sve.md
> index da5534c3e32..e5e0c7ddfc5 100644
> --- a/gcc/config/aarch64/aarch64-sve.md
> +++ b/gcc/config/aarch64/aarch64-sve.md
> @@ -8836,6 +8836,27 @@
>    "TARGET_SVE"
>    "rev\t%0.<Vctype>, %1.<Vctype>")
>
> +(define_peephole2
> +  [(set (match_operand:SVE_ALL 0 "register_operand" "")
> +        (unspec:SVE_ALL
> +          [(match_operand:SVE_ALL 1 "register_operand" "")] UNSPEC_REV))
> +   (set (match_operand:SVE_ALL 2 "register_operand" "")
> +        (unspec:SVE_ALL
> +          [(match_dup 0)] UNSPEC_REV))]
> +  "TARGET_SVE"
> +  [(const_int 0)]
> +  {
> +    if (REGNO (operands[2]) != REGNO (operands[0]))
> +    {
> +      emit_insn (gen_rtx_SET (operands[2], operands[1]));
> +      rtx rev = gen_rtx_UNSPEC (<MODE>mode, gen_rtvec (1, operands[1]), UNSPEC_REV);
> +      emit_insn (gen_rtx_SET (operands[0], rev));
> +    }
> +    else
> +      emit_insn (gen_rtx_SET (operands[0], operands[1]));
> +    DONE;
> +  })
> +
>  ;; -------------------------------------------------------------------------
>  ;; ---- [INT,FP] Special-purpose binary permutes
>  ;; -------------------------------------------------------------------------
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/revrev.c b/gcc/testsuite/gcc.target/aarch64/sve/revrev.c
> new file mode 100644
> index 00000000000..04af6eed291
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/revrev.c
> @@ -0,0 +1,13 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O3" } */
> +
> +#include <stdint.h>
> +
> +void
> +test (uint8_t a[], uint8_t b[], uint64_t N)
> +{
> +  for (uint64_t i = N; i > 0; i--)
> +    a[i - 1] = b[i - 1];
> +}
> +
> +/* { dg-final { scan-assembler-not {\trev\t(z[0-9]+\.h), \1\n\trev\t\1, \1\n} } } */
> --
> 2.21.0

^ permalink raw reply	[flat|nested] 3+ messages in thread

* RE: [PATCH] aarch64: remove useless pairs of rev instructions
  2023-07-18 10:53 ` Prathamesh Kulkarni
@ 2023-07-18 13:38   ` Serval Martinot-Lagarde
  0 siblings, 0 replies; 3+ messages in thread
From: Serval Martinot-Lagarde @ 2023-07-18 13:38 UTC (permalink / raw)
  To: Prathamesh Kulkarni; +Cc: gcc-patches, Etienne Renault

Hi Prathamesh,

> On Tue, 18 Jul 2023 at 12:54, Prathamesh Kulkarni
> <prathamesh.kulkarni@linaro.org> wrote:
> I had added a similar transform to remove pair of rev instructions in:
> https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=f0eabc52c9a2d3da0bfc201da7a5
> c1658b76e9a4

I don't have much knowledge of GCC code, but your commit seems more
generic to me. Also, after a rebase on master, I realized that the
problem is already solved thanks to your work.

Thanks,

-- 
Serval Martinot-Lagarde

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2023-07-18 13:38 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-07-18  9:50 [PATCH] aarch64: remove useless pairs of rev instructions Serval Martinot-Lagarde
2023-07-18 10:53 ` Prathamesh Kulkarni
2023-07-18 13:38   ` Serval Martinot-Lagarde

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).