[PATCH][committed] aarch64: Optimise ADDP with same source operands

[PATCH][committed] aarch64: Optimise ADDP with same source operands

[Kyrylo Tkachov]

[-- Attachment #1: Type: text/plain, Size: 817 bytes --]

Hi all,

We've been asked to optimise the testcase in this patch of a 64-bit ADDP with
the low and high halves of the same 128-bit vector. This can be done by a
single .4s ADDP followed by just reading the bottom 64 bits. A splitter for
this is quite straightforward now that all the vec_concat stuff is collapsed
by simplify-rtx.

With this patch we generate a single:
	addp	v0.4s, v0.4s, v0.4s
instead of:
        dup     d31, v0.d[1]
        addp    v0.2s, v0.2s, v31.2s
        ret

Bootstrapped and tested on aarch64-none-linux-gnu and aarch64_be-none-elf.
Pushing to trunk.
Thanks,
Kyrill

gcc/ChangeLog:

	* config/aarch64/aarch64-simd.md (*aarch64_addp_same_reg<mode>):
	New define_insn_and_split.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/simd/addp-same-low_1.c: New test.

[-- Attachment #2: addp-q.patch --]
[-- Type: application/octet-stream, Size: 2303 bytes --]

diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 5622d84cf3781b07e9a4f6780e56aa3f74c194a5..c17f21f688267394df34001c55027e4f1ecc56f3 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -7327,6 +7327,36 @@ (define_insn "aarch64_addp<mode><vczle><vczbe>_insn"
   [(set_attr "type" "neon_reduc_add<q>")]
 )
 
+;; A common usecase of 64-bit ADDP is to have both operands come from the same
+;; 128-bit vector and produce the pairwise addition results in the lower half.
+;; Split into the 128-bit ADDP form and extract the low half.
+(define_insn_and_split "*aarch64_addp_same_reg<mode>"
+  [(set (match_operand:<VHALF> 0 "register_operand" "=w")
+	(plus:<VHALF>
+	  (vec_select:<VHALF>
+	    (match_operand:VQ_I 1 "register_operand" "w")
+	    (match_operand:VQ_I 2 "vect_par_cnst_even_or_odd_half"))
+	  (vec_select:<VHALF>
+	    (match_dup 1)
+	    (match_operand:VQ_I 3 "vect_par_cnst_even_or_odd_half"))))]
+  "TARGET_SIMD && !rtx_equal_p (operands[2], operands[3])"
+  "#"
+  "&& 1"
+  [(const_int 0)]
+  {
+    rtx scratch;
+    if (can_create_pseudo_p ())
+      scratch = gen_reg_rtx (<MODE>mode);
+    else
+      scratch = lowpart_subreg (<MODE>mode, operands[0], <VHALF>mode);
+
+    emit_insn (gen_aarch64_addp<mode>_insn (scratch, operands[1], operands[1],
+					    operands[2], operands[3]));
+    emit_move_insn (operands[0], gen_lowpart (<VHALF>mode, scratch));
+    DONE;
+  }
+)
+
 (define_expand "aarch64_addp<mode>"
   [(match_operand:VDQ_I 0 "register_operand")
    (match_operand:VDQ_I 1 "register_operand")
diff --git a/gcc/testsuite/gcc.target/aarch64/simd/addp-same-low_1.c b/gcc/testsuite/gcc.target/aarch64/simd/addp-same-low_1.c
new file mode 100644
index 0000000000000000000000000000000000000000..42d8db77f1c91748f543507aa9f1316b242ffd8b
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/simd/addp-same-low_1.c
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-O" } */
+/* { dg-require-effective-target aarch64_little_endian } */
+/* { dg-final { check-function-bodies "**" "" "" } } */
+
+#include <arm_neon.h>
+
+
+/*
+** foo:
+**	addp	v0\.4s, v0\.4s, v0\.4s
+**	ret
+*/
+
+int32x2_t
+foo (int32x4_t a)
+{
+  return vpadd_s32 (vget_low_s32(a), vget_high_s32(a));
+}
+