[gcc r14-266] aarch64: Reimplement RSHRN intrinsic patterns with standard RTL codes

[Kyrylo Tkachov <ktkachov@gcc.gnu.org>]

https://gcc.gnu.org/g:80afac3121778e509e1847be90a1999236cff03a

commit r14-266-g80afac3121778e509e1847be90a1999236cff03a
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date:   Wed Apr 26 15:10:18 2023 +0100

    aarch64: Reimplement RSHRN intrinsic patterns with standard RTL codes
    
    This patch reimplements the backend patterns for the rshrn intrinsics using standard RTL codes rather than UNSPECS.
    We already represent shrn as truncate of a shift. rshrn can be represented as truncate (src + (1 << (shft - 1)) >> shft),
    similar to how LLVM treats it.
    
    I have a follow-up patch to do the same for the rshrn2 pattern, which will allow us to remove the UNSPEC_RSHRN entirely.
    
    Bootstrapped and tested on aarch64-none-linux-gnu.
    
    gcc/ChangeLog:
    
            * config/aarch64/aarch64-simd.md (aarch64_rshrn<mode>_insn_le): Reimplement
            with standard RTL codes instead of an UNSPEC.
            (aarch64_rshrn<mode>_insn_be): Likewise.
            (aarch64_rshrn<mode>): Adjust for the above.
            * config/aarch64/predicates.md (aarch64_simd_rshrn_imm_vec): Define.

Diff:
---
 gcc/config/aarch64/aarch64-simd.md | 36 ++++++++++++++++++++++++------------
 gcc/config/aarch64/predicates.md   |  6 ++++++
 2 files changed, 30 insertions(+), 12 deletions(-)

diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index cb2223d29c2..f8913107aad 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -1990,11 +1990,15 @@
 (define_insn "aarch64_rshrn<mode>_insn_le"
   [(set (match_operand:<VNARROWQ2> 0 "register_operand" "=w")
 	(vec_concat:<VNARROWQ2>
-	  (unspec:<VNARROWQ> [(match_operand:VQN 1 "register_operand" "w")
-		(match_operand:VQN 2
-		  "aarch64_simd_shift_imm_vec_<vn_mode>")] UNSPEC_RSHRN)
-	  (match_operand:<VNARROWQ> 3 "aarch64_simd_or_scalar_imm_zero")))]
-  "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+	  (truncate:<VNARROWQ>
+	    (lshiftrt:VQN
+	      (plus:VQN (match_operand:VQN 1 "register_operand" "w")
+			(match_operand:VQN 3 "aarch64_simd_rshrn_imm_vec"))
+	      (match_operand:VQN 2 "aarch64_simd_shift_imm_vec_<vn_mode>")))
+	  (match_operand:<VNARROWQ> 4 "aarch64_simd_or_scalar_imm_zero")))]
+  "TARGET_SIMD && !BYTES_BIG_ENDIAN
+   && INTVAL (CONST_VECTOR_ELT (operands[3], 0))
+      == (HOST_WIDE_INT_1 << (INTVAL (CONST_VECTOR_ELT (operands[2], 0)) - 1))"
   "rshrn\\t%0.<Vntype>, %1.<Vtype>, %2"
   [(set_attr "type" "neon_shift_imm_narrow_q")]
 )
@@ -2002,11 +2006,15 @@
 (define_insn "aarch64_rshrn<mode>_insn_be"
   [(set (match_operand:<VNARROWQ2> 0 "register_operand" "=w")
 	(vec_concat:<VNARROWQ2>
-	  (match_operand:<VNARROWQ> 3 "aarch64_simd_or_scalar_imm_zero")
-	  (unspec:<VNARROWQ> [(match_operand:VQN 1 "register_operand" "w")
-		(match_operand:VQN 2 "aarch64_simd_shift_imm_vec_<vn_mode>")]
-		  UNSPEC_RSHRN)))]
-  "TARGET_SIMD && BYTES_BIG_ENDIAN"
+	  (match_operand:<VNARROWQ> 4 "aarch64_simd_or_scalar_imm_zero")
+	  (truncate:<VNARROWQ>
+	    (lshiftrt:VQN
+	      (plus:VQN (match_operand:VQN 1 "register_operand" "w")
+			(match_operand:VQN 3 "aarch64_simd_rshrn_imm_vec"))
+	      (match_operand:VQN 2 "aarch64_simd_shift_imm_vec_<vn_mode>")))))]
+  "TARGET_SIMD && BYTES_BIG_ENDIAN
+   && INTVAL (CONST_VECTOR_ELT (operands[3], 0))
+      == (HOST_WIDE_INT_1 << (INTVAL (CONST_VECTOR_ELT (operands[2], 0)) - 1))"
   "rshrn\\t%0.<Vntype>, %1.<Vtype>, %2"
   [(set_attr "type" "neon_shift_imm_narrow_q")]
 )
@@ -2024,18 +2032,22 @@
       }
     else
       {
+	rtx shft
+	  = aarch64_simd_gen_const_vector_dup (<MODE>mode,
+					       HOST_WIDE_INT_1U
+					        << (INTVAL (operands[2]) - 1));
 	rtx tmp = gen_reg_rtx (<VNARROWQ2>mode);
 	operands[2] = aarch64_simd_gen_const_vector_dup (<MODE>mode,
 						         INTVAL (operands[2]));
 	if (BYTES_BIG_ENDIAN)
 	  emit_insn (
 		gen_aarch64_rshrn<mode>_insn_be (tmp, operands[1],
-						 operands[2],
+						 operands[2], shft,
 						 CONST0_RTX (<VNARROWQ>mode)));
 	else
 	  emit_insn (
 		gen_aarch64_rshrn<mode>_insn_le (tmp, operands[1],
-						 operands[2],
+						 operands[2], shft,
 						 CONST0_RTX (<VNARROWQ>mode)));
 
 	/* The intrinsic expects a narrow result, so emit a subreg that will
diff --git a/gcc/config/aarch64/predicates.md b/gcc/config/aarch64/predicates.md
index 3f5f4df8c46..242f10aea1f 100644
--- a/gcc/config/aarch64/predicates.md
+++ b/gcc/config/aarch64/predicates.md
@@ -587,6 +587,12 @@
   (and (match_code "const_vector")
        (match_test "aarch64_const_vec_all_same_in_range_p (op, 1, 64)")))
 
+(define_predicate "aarch64_simd_rshrn_imm_vec"
+  (and (match_code "const_vector")
+       (match_test "aarch64_const_vec_all_same_in_range_p (op, 1,
+				HOST_WIDE_INT_1U
+				<< (GET_MODE_UNIT_BITSIZE  (mode) - 1))")))
+
 (define_predicate "aarch64_simd_shift_imm_bitsize_qi"
   (and (match_code "const_int")
        (match_test "IN_RANGE (INTVAL (op), 0, 8)")))