aarch64: Use RTL builtins for vcvtx intrinsics

Rewrite vcvtx Neon intrinsics to use RTL builtins rather than inline
assembly code, allowing for better scheduling and optimization.

gcc/ChangeLog:

2021-02-18  Jonathan Wright  <jonathan.wright@arm.com>

	* config/aarch64/aarch64-simd-builtins.def: Add
	float_trunc_rodd builtin generator macros.
	* config/aarch64/aarch64-simd.md (aarch64_float_trunc_rodd_df):
	Define.
	(aarch64_float_trunc_rodd_lo_v2sf): Define.
	(aarch64_float_trunc_rodd_hi_v4sf_le): Define.
	(aarch64_float_trunc_rodd_hi_v4sf_be): Define.
	(aarch64_float_trunc_rodd_hi_v4sf): Define.
	* config/aarch64/arm_neon.h (vcvtx_f32_f64): Use RTL builtin
	instead of inline asm.
	(vcvtx_high_f32_f64): Likewise.
	(vcvtxd_f32_f64): Likewise.
	* config/aarch64/iterators.md: Add FCVTXN unspec.

4 files changed, 62 insertions, 18 deletions

diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def
index a7d4f2b..f01a1b4 100644
--- a/gcc/config/aarch64/aarch64-simd-builtins.def
+++ b/gcc/config/aarch64/aarch64-simd-builtins.def
@@ -632,6 +632,10 @@
   VAR1 (UNOP, float_extend_lo_,  0, FP, v4sf)
   BUILTIN_VDF (UNOP, float_truncate_lo_, 0, FP)
 
+  VAR1 (UNOP, float_trunc_rodd_, 0, FP, df)
+  VAR1 (UNOP, float_trunc_rodd_lo_, 0, FP, v2sf)
+  VAR1 (BINOP, float_trunc_rodd_hi_, 0, FP, v4sf)
+
   /* Implemented by aarch64_ld1<VALL_F16:mode>.  */
   BUILTIN_VALL_F16 (LOAD1, ld1, 0, LOAD)
   VAR1(STORE1P, ld1, 0, ALL, v2di)
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 299d911..72f429c 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -2981,6 +2981,60 @@
 
 ;; Float narrowing operations.
 
+(define_insn "aarch64_float_trunc_rodd_df"
+  [(set (match_operand:SF 0 "register_operand" "=w")
+	(unspec:SF [(match_operand:DF 1 "register_operand" "w")]
+		UNSPEC_FCVTXN))]
+  "TARGET_SIMD"
+  "fcvtxn\\t%s0, %d1"
+  [(set_attr "type" "neon_fp_cvt_narrow_d_q")]
+)
+
+(define_insn "aarch64_float_trunc_rodd_lo_v2sf"
+  [(set (match_operand:V2SF 0 "register_operand" "=w")
+	(unspec:V2SF [(match_operand:V2DF 1 "register_operand" "w")]
+		UNSPEC_FCVTXN))]
+  "TARGET_SIMD"
+  "fcvtxn\\t%0.2s, %1.2d"
+  [(set_attr "type" "neon_fp_cvt_narrow_d_q")]
+)
+
+(define_insn "aarch64_float_trunc_rodd_hi_v4sf_le"
+  [(set (match_operand:V4SF 0 "register_operand" "=w")
+	(vec_concat:V4SF
+	  (match_operand:V2SF 1 "register_operand" "0")
+	  (unspec:V2SF [(match_operand:V2DF 2 "register_operand" "w")]
+		UNSPEC_FCVTXN)))]
+  "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+  "fcvtxn2\\t%0.4s, %2.2d"
+  [(set_attr "type" "neon_fp_cvt_narrow_d_q")]
+)
+
+(define_insn "aarch64_float_trunc_rodd_hi_v4sf_be"
+  [(set (match_operand:V4SF 0 "register_operand" "=w")
+	(vec_concat:V4SF
+	  (unspec:V2SF [(match_operand:V2DF 2 "register_operand" "w")]
+		UNSPEC_FCVTXN)
+	  (match_operand:V2SF 1 "register_operand" "0")))]
+  "TARGET_SIMD && BYTES_BIG_ENDIAN"
+  "fcvtxn2\\t%0.4s, %2.2d"
+  [(set_attr "type" "neon_fp_cvt_narrow_d_q")]
+)
+
+(define_expand "aarch64_float_trunc_rodd_hi_v4sf"
+  [(match_operand:V4SF 0 "register_operand")
+   (match_operand:V2SF 1 "register_operand")
+   (match_operand:V2DF 2 "register_operand")]
+  "TARGET_SIMD"
+{
+  rtx (*gen) (rtx, rtx, rtx) = BYTES_BIG_ENDIAN
+			     ? gen_aarch64_float_trunc_rodd_hi_v4sf_be
+			     : gen_aarch64_float_trunc_rodd_hi_v4sf_le;
+  emit_insn (gen (operands[0], operands[1], operands[2]));
+  DONE;
+}
+)
+
 (define_insn "aarch64_float_truncate_lo_<mode>"
   [(set (match_operand:VDF 0 "register_operand" "=w")
       (float_truncate:VDF
diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h
index ead2bd0..4b8ec52 100644
--- a/gcc/config/aarch64/arm_neon.h
+++ b/gcc/config/aarch64/arm_neon.h
@@ -7014,36 +7014,21 @@ __extension__ extern __inline float32x2_t
 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 vcvtx_f32_f64 (float64x2_t __a)
 {
-  float32x2_t __result;
-  __asm__ ("fcvtxn %0.2s,%1.2d"
-           : "=w"(__result)
-           : "w"(__a)
-           : /* No clobbers */);
-  return __result;
+  return __builtin_aarch64_float_trunc_rodd_lo_v2sf (__a);
 }
 
 __extension__ extern __inline float32x4_t
 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 vcvtx_high_f32_f64 (float32x2_t __a, float64x2_t __b)
 {
-  float32x4_t __result;
-  __asm__ ("fcvtxn2 %0.4s,%1.2d"
-           : "=w"(__result)
-           : "w" (__b), "0"(__a)
-           : /* No clobbers */);
-  return __result;
+  return __builtin_aarch64_float_trunc_rodd_hi_v4sf (__a, __b);
 }
 
 __extension__ extern __inline float32_t
 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
 vcvtxd_f32_f64 (float64_t __a)
 {
-  float32_t __result;
-  __asm__ ("fcvtxn %s0,%d1"
-           : "=w"(__result)
-           : "w"(__a)
-           : /* No clobbers */);
-  return __result;
+  return __builtin_aarch64_float_trunc_rodd_df (__a);
 }
 
 __extension__ extern __inline float32x2_t
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index fe2c51c..3d66e63 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -861,6 +861,7 @@
     UNSPEC_BFCVTN      ; Used in aarch64-simd.md.
     UNSPEC_BFCVTN2     ; Used in aarch64-simd.md.
     UNSPEC_BFCVT       ; Used in aarch64-simd.md.
+    UNSPEC_FCVTXN	; Used in aarch64-simd.md.
 ])
 
 ;; ------------------------------------------------------------------