diff mbox series

[7/7] AArch64 Combine cmeq 0 + not into cmtst

Message ID 20210929162147.GA12194@arm.com
State New
Headers show
Series AArch64 Optimize truncation, shifts and bitmask comparisons | expand

Commit Message

Tamar Christina Sept. 29, 2021, 4:21 p.m. UTC
Hi All,

This turns a bitwise inverse of an equality comparison with 0 into a compare of
bitwise nonzero (cmtst).

We already have one pattern for cmsts, this adds an additional one which does
not require an additional bitwise and.

i.e.

#include <arm_neon.h>

uint8x8_t bar(int16x8_t abs_row0, int16x8_t row0) {
  uint16x8_t row0_diff =
    vreinterpretq_u16_s16(veorq_s16(abs_row0, vshrq_n_s16(row0, 15)));
  uint8x8_t abs_row0_gt0 =
    vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row0), vdupq_n_u16(0)));
  return abs_row0_gt0;
}

now generates:

bar:
        cmtst   v0.8h, v0.8h, v0.8h
        xtn     v0.8b, v0.8h
        ret

instead of:

bar:
        cmeq    v0.8h, v0.8h, #0
        not     v0.16b, v0.16b
        xtn     v0.8b, v0.8h
        ret

Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.

Ok for master?

Thanks,
Tamar

gcc/ChangeLog:

	* config/aarch64/aarch64-simd.md (*aarch64_cmtst_same_<mode>): New.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/mvn-cmeq0-1.c: New test.

--- inline copy of patch -- 
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 9d936428b438c95b56614c94081d7e2ebc47d89f..bce01c36386074bf475b8b7e5c69a1959a13fef3 100644


--

Comments

Kyrylo Tkachov Sept. 30, 2021, 9:35 a.m. UTC | #1
> -----Original Message-----
> From: Tamar Christina <Tamar.Christina@arm.com>
> Sent: Wednesday, September 29, 2021 5:22 PM
> To: gcc-patches@gcc.gnu.org
> Cc: nd <nd@arm.com>; Richard Earnshaw <Richard.Earnshaw@arm.com>;
> Marcus Shawcroft <Marcus.Shawcroft@arm.com>; Kyrylo Tkachov
> <Kyrylo.Tkachov@arm.com>; Richard Sandiford
> <Richard.Sandiford@arm.com>
> Subject: [PATCH 7/7]AArch64 Combine cmeq 0 + not into cmtst
> 
> Hi All,
> 
> This turns a bitwise inverse of an equality comparison with 0 into a compare
> of
> bitwise nonzero (cmtst).
> 
> We already have one pattern for cmsts, this adds an additional one which
> does
> not require an additional bitwise and.
> 
> i.e.
> 
> #include <arm_neon.h>
> 
> uint8x8_t bar(int16x8_t abs_row0, int16x8_t row0) {
>   uint16x8_t row0_diff =
>     vreinterpretq_u16_s16(veorq_s16(abs_row0, vshrq_n_s16(row0, 15)));
>   uint8x8_t abs_row0_gt0 =
>     vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row0),
> vdupq_n_u16(0)));
>   return abs_row0_gt0;
> }
> 
> now generates:
> 
> bar:
>         cmtst   v0.8h, v0.8h, v0.8h
>         xtn     v0.8b, v0.8h
>         ret
> 
> instead of:
> 
> bar:
>         cmeq    v0.8h, v0.8h, #0
>         not     v0.16b, v0.16b
>         xtn     v0.8b, v0.8h
>         ret
> 
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> 
> Ok for master?
> 
> Thanks,
> Tamar
> 
> gcc/ChangeLog:
> 
> 	* config/aarch64/aarch64-simd.md
> (*aarch64_cmtst_same_<mode>): New.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* gcc.target/aarch64/mvn-cmeq0-1.c: New test.
> 
> --- inline copy of patch --
> diff --git a/gcc/config/aarch64/aarch64-simd.md
> b/gcc/config/aarch64/aarch64-simd.md
> index
> 9d936428b438c95b56614c94081d7e2ebc47d89f..bce01c36386074bf475b8b7
> e5c69a1959a13fef3 100644
> --- a/gcc/config/aarch64/aarch64-simd.md
> +++ b/gcc/config/aarch64/aarch64-simd.md
> @@ -6585,6 +6585,23 @@ (define_insn "aarch64_cmtst<mode>"
>    [(set_attr "type" "neon_tst<q>")]
>  )
> 
> +;; One can also get a cmtsts by having to combine a
> +;; not (neq (eq x 0)) in which case you rewrite it to
> +;; a comparison against itself
> +
> +(define_insn "*aarch64_cmtst_same_<mode>"
> +  [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w")
> +	(plus:<V_INT_EQUIV>
> +	  (eq:<V_INT_EQUIV>
> +	    (match_operand:VDQ_I 1 "register_operand" "w")
> +	    (match_operand:VDQ_I 2 "aarch64_simd_imm_zero"))
> +	  (match_operand:<V_INT_EQUIV> 3
> "aarch64_simd_imm_minus_one")))
> +  ]
> +  "TARGET_SIMD"
> +  "cmtst\t%<v>0<Vmtype>, %<v>1<Vmtype>, %<v>1<Vmtype>"
> +  [(set_attr "type" "neon_tst<q>")]
> +)
> +
>  (define_insn_and_split "aarch64_cmtstdi"
>    [(set (match_operand:DI 0 "register_operand" "=w,r")
>  	(neg:DI
> diff --git a/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c
> b/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..59f3a230271c70d3bb51d03
> 38d9ec2613bd4394b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c
> @@ -0,0 +1,17 @@
> +/* { dg-do assemble } */
> +/* { dg-options "-O3 --save-temps --param=vect-epilogues-nomask=0" } */

I don't think we need the param here (or even anything higher than -O really).
Ok otherwise.
Thanks,
Kyrill


> +
> +#include <arm_neon.h>
> +
> +uint8x8_t bar(int16x8_t abs_row0, int16x8_t row0) {
> +  uint16x8_t row0_diff =
> +    vreinterpretq_u16_s16(veorq_s16(abs_row0, vshrq_n_s16(row0, 15)));
> +  uint8x8_t abs_row0_gt0 =
> +    vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row0),
> vdupq_n_u16(0)));
> +  return abs_row0_gt0;
> +}
> +
> +
> +/* { dg-final { scan-assembler-times {\tcmtst\t} 1 } } */
> +/* { dg-final { scan-assembler-not {\tcmeq\t} } } */
> +/* { dg-final { scan-assembler-not {\tnot\t} } } */
> 
> 
> --
diff mbox series

Patch

diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 9d936428b438c95b56614c94081d7e2ebc47d89f..bce01c36386074bf475b8b7e5c69a1959a13fef3 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -6585,6 +6585,23 @@  (define_insn "aarch64_cmtst<mode>"
   [(set_attr "type" "neon_tst<q>")]
 )
 
+;; One can also get a cmtsts by having to combine a
+;; not (neq (eq x 0)) in which case you rewrite it to
+;; a comparison against itself
+
+(define_insn "*aarch64_cmtst_same_<mode>"
+  [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w")
+	(plus:<V_INT_EQUIV>
+	  (eq:<V_INT_EQUIV>
+	    (match_operand:VDQ_I 1 "register_operand" "w")
+	    (match_operand:VDQ_I 2 "aarch64_simd_imm_zero"))
+	  (match_operand:<V_INT_EQUIV> 3 "aarch64_simd_imm_minus_one")))
+  ]
+  "TARGET_SIMD"
+  "cmtst\t%<v>0<Vmtype>, %<v>1<Vmtype>, %<v>1<Vmtype>"
+  [(set_attr "type" "neon_tst<q>")]
+)
+
 (define_insn_and_split "aarch64_cmtstdi"
   [(set (match_operand:DI 0 "register_operand" "=w,r")
 	(neg:DI
diff --git a/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c b/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c
new file mode 100644
index 0000000000000000000000000000000000000000..59f3a230271c70d3bb51d0338d9ec2613bd4394b
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/mvn-cmeq0-1.c
@@ -0,0 +1,17 @@ 
+/* { dg-do assemble } */
+/* { dg-options "-O3 --save-temps --param=vect-epilogues-nomask=0" } */
+
+#include <arm_neon.h>
+
+uint8x8_t bar(int16x8_t abs_row0, int16x8_t row0) {
+  uint16x8_t row0_diff =
+    vreinterpretq_u16_s16(veorq_s16(abs_row0, vshrq_n_s16(row0, 15)));
+  uint8x8_t abs_row0_gt0 =
+    vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row0), vdupq_n_u16(0)));
+  return abs_row0_gt0;
+}
+
+
+/* { dg-final { scan-assembler-times {\tcmtst\t} 1 } } */
+/* { dg-final { scan-assembler-not {\tcmeq\t} } } */
+/* { dg-final { scan-assembler-not {\tnot\t} } } */