From b12677052e69b67310c1d63360db2793354414cb Mon Sep 17 00:00:00 2001
From: Jiong Wang <jiong.wang@arm.com>
Date: Tue, 7 Jun 2016 17:01:22 +0100
Subject: [PATCH 01/14] [1/14] ARMv8.2 FP16 data processing intrinsics
---
gcc/config/aarch64/aarch64-simd.md | 22 +--
gcc/config/aarch64/aarch64.c | 16 +++
gcc/config/aarch64/arm_neon.h | 275 ++++++++++++++++++++++++++++++++++++-
3 files changed, 298 insertions(+), 15 deletions(-)
@@ -5161,10 +5161,10 @@
)
(define_insn "aarch64_<PERMUTE:perm_insn><PERMUTE:perm_hilo><mode>"
- [(set (match_operand:VALL 0 "register_operand" "=w")
- (unspec:VALL [(match_operand:VALL 1 "register_operand" "w")
- (match_operand:VALL 2 "register_operand" "w")]
- PERMUTE))]
+ [(set (match_operand:VALL_F16 0 "register_operand" "=w")
+ (unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")
+ (match_operand:VALL_F16 2 "register_operand" "w")]
+ PERMUTE))]
"TARGET_SIMD"
"<PERMUTE:perm_insn><PERMUTE:perm_hilo>\\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>"
[(set_attr "type" "neon_permute<q>")]
@@ -5172,11 +5172,11 @@
;; Note immediate (third) operand is lane index not byte index.
(define_insn "aarch64_ext<mode>"
- [(set (match_operand:VALL 0 "register_operand" "=w")
- (unspec:VALL [(match_operand:VALL 1 "register_operand" "w")
- (match_operand:VALL 2 "register_operand" "w")
- (match_operand:SI 3 "immediate_operand" "i")]
- UNSPEC_EXT))]
+ [(set (match_operand:VALL_F16 0 "register_operand" "=w")
+ (unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")
+ (match_operand:VALL_F16 2 "register_operand" "w")
+ (match_operand:SI 3 "immediate_operand" "i")]
+ UNSPEC_EXT))]
"TARGET_SIMD"
{
operands[3] = GEN_INT (INTVAL (operands[3])
@@ -5187,8 +5187,8 @@
)
(define_insn "aarch64_rev<REVERSE:rev_op><mode>"
- [(set (match_operand:VALL 0 "register_operand" "=w")
- (unspec:VALL [(match_operand:VALL 1 "register_operand" "w")]
+ [(set (match_operand:VALL_F16 0 "register_operand" "=w")
+ (unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")]
REVERSE))]
"TARGET_SIMD"
"rev<REVERSE:rev_op>\\t%0.<Vtype>, %1.<Vtype>"
@@ -12053,6 +12053,8 @@ aarch64_evpc_trn (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_trn2v4si; break;
case V2SImode: gen = gen_aarch64_trn2v2si; break;
case V2DImode: gen = gen_aarch64_trn2v2di; break;
+ case V4HFmode: gen = gen_aarch64_trn2v4hf; break;
+ case V8HFmode: gen = gen_aarch64_trn2v8hf; break;
case V4SFmode: gen = gen_aarch64_trn2v4sf; break;
case V2SFmode: gen = gen_aarch64_trn2v2sf; break;
case V2DFmode: gen = gen_aarch64_trn2v2df; break;
@@ -12071,6 +12073,8 @@ aarch64_evpc_trn (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_trn1v4si; break;
case V2SImode: gen = gen_aarch64_trn1v2si; break;
case V2DImode: gen = gen_aarch64_trn1v2di; break;
+ case V4HFmode: gen = gen_aarch64_trn1v4hf; break;
+ case V8HFmode: gen = gen_aarch64_trn1v8hf; break;
case V4SFmode: gen = gen_aarch64_trn1v4sf; break;
case V2SFmode: gen = gen_aarch64_trn1v2sf; break;
case V2DFmode: gen = gen_aarch64_trn1v2df; break;
@@ -12136,6 +12140,8 @@ aarch64_evpc_uzp (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_uzp2v4si; break;
case V2SImode: gen = gen_aarch64_uzp2v2si; break;
case V2DImode: gen = gen_aarch64_uzp2v2di; break;
+ case V4HFmode: gen = gen_aarch64_uzp2v4hf; break;
+ case V8HFmode: gen = gen_aarch64_uzp2v8hf; break;
case V4SFmode: gen = gen_aarch64_uzp2v4sf; break;
case V2SFmode: gen = gen_aarch64_uzp2v2sf; break;
case V2DFmode: gen = gen_aarch64_uzp2v2df; break;
@@ -12154,6 +12160,8 @@ aarch64_evpc_uzp (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_uzp1v4si; break;
case V2SImode: gen = gen_aarch64_uzp1v2si; break;
case V2DImode: gen = gen_aarch64_uzp1v2di; break;
+ case V4HFmode: gen = gen_aarch64_uzp1v4hf; break;
+ case V8HFmode: gen = gen_aarch64_uzp1v8hf; break;
case V4SFmode: gen = gen_aarch64_uzp1v4sf; break;
case V2SFmode: gen = gen_aarch64_uzp1v2sf; break;
case V2DFmode: gen = gen_aarch64_uzp1v2df; break;
@@ -12224,6 +12232,8 @@ aarch64_evpc_zip (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_zip2v4si; break;
case V2SImode: gen = gen_aarch64_zip2v2si; break;
case V2DImode: gen = gen_aarch64_zip2v2di; break;
+ case V4HFmode: gen = gen_aarch64_zip2v4hf; break;
+ case V8HFmode: gen = gen_aarch64_zip2v8hf; break;
case V4SFmode: gen = gen_aarch64_zip2v4sf; break;
case V2SFmode: gen = gen_aarch64_zip2v2sf; break;
case V2DFmode: gen = gen_aarch64_zip2v2df; break;
@@ -12242,6 +12252,8 @@ aarch64_evpc_zip (struct expand_vec_perm_d *d)
case V4SImode: gen = gen_aarch64_zip1v4si; break;
case V2SImode: gen = gen_aarch64_zip1v2si; break;
case V2DImode: gen = gen_aarch64_zip1v2di; break;
+ case V4HFmode: gen = gen_aarch64_zip1v4hf; break;
+ case V8HFmode: gen = gen_aarch64_zip1v8hf; break;
case V4SFmode: gen = gen_aarch64_zip1v4sf; break;
case V2SFmode: gen = gen_aarch64_zip1v2sf; break;
case V2DFmode: gen = gen_aarch64_zip1v2df; break;
@@ -12286,6 +12298,8 @@ aarch64_evpc_ext (struct expand_vec_perm_d *d)
case V8HImode: gen = gen_aarch64_extv8hi; break;
case V2SImode: gen = gen_aarch64_extv2si; break;
case V4SImode: gen = gen_aarch64_extv4si; break;
+ case V4HFmode: gen = gen_aarch64_extv4hf; break;
+ case V8HFmode: gen = gen_aarch64_extv8hf; break;
case V2SFmode: gen = gen_aarch64_extv2sf; break;
case V4SFmode: gen = gen_aarch64_extv4sf; break;
case V2DImode: gen = gen_aarch64_extv2di; break;
@@ -12361,6 +12375,8 @@ aarch64_evpc_rev (struct expand_vec_perm_d *d)
case V2SImode: gen = gen_aarch64_rev64v2si; break;
case V4SFmode: gen = gen_aarch64_rev64v4sf; break;
case V2SFmode: gen = gen_aarch64_rev64v2sf; break;
+ case V8HFmode: gen = gen_aarch64_rev64v8hf; break;
+ case V4HFmode: gen = gen_aarch64_rev64v4hf; break;
default:
return false;
}
@@ -466,6 +466,8 @@ typedef struct poly16x8x4_t
#define __aarch64_vdup_lane_any(__size, __q, __a, __b) \
vdup##__q##_n_##__size (__aarch64_vget_lane_any (__a, __b))
+#define __aarch64_vdup_lane_f16(__a, __b) \
+ __aarch64_vdup_lane_any (f16, , __a, __b)
#define __aarch64_vdup_lane_f32(__a, __b) \
__aarch64_vdup_lane_any (f32, , __a, __b)
#define __aarch64_vdup_lane_f64(__a, __b) \
@@ -492,6 +494,8 @@ typedef struct poly16x8x4_t
__aarch64_vdup_lane_any (u64, , __a, __b)
/* __aarch64_vdup_laneq internal macros. */
+#define __aarch64_vdup_laneq_f16(__a, __b) \
+ __aarch64_vdup_lane_any (f16, , __a, __b)
#define __aarch64_vdup_laneq_f32(__a, __b) \
__aarch64_vdup_lane_any (f32, , __a, __b)
#define __aarch64_vdup_laneq_f64(__a, __b) \
@@ -518,6 +522,8 @@ typedef struct poly16x8x4_t
__aarch64_vdup_lane_any (u64, , __a, __b)
/* __aarch64_vdupq_lane internal macros. */
+#define __aarch64_vdupq_lane_f16(__a, __b) \
+ __aarch64_vdup_lane_any (f16, q, __a, __b)
#define __aarch64_vdupq_lane_f32(__a, __b) \
__aarch64_vdup_lane_any (f32, q, __a, __b)
#define __aarch64_vdupq_lane_f64(__a, __b) \
@@ -544,6 +550,8 @@ typedef struct poly16x8x4_t
__aarch64_vdup_lane_any (u64, q, __a, __b)
/* __aarch64_vdupq_laneq internal macros. */
+#define __aarch64_vdupq_laneq_f16(__a, __b) \
+ __aarch64_vdup_lane_any (f16, q, __a, __b)
#define __aarch64_vdupq_laneq_f32(__a, __b) \
__aarch64_vdup_lane_any (f32, q, __a, __b)
#define __aarch64_vdupq_laneq_f64(__a, __b) \
@@ -10369,6 +10377,12 @@ vaddvq_f64 (float64x2_t __a)
/* vbsl */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vbsl_f16 (uint16x4_t __a, float16x4_t __b, float16x4_t __c)
+{
+ return __builtin_aarch64_simd_bslv4hf_suss (__a, __b, __c);
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vbsl_f32 (uint32x2_t __a, float32x2_t __b, float32x2_t __c)
{
@@ -10444,6 +10458,12 @@ vbsl_u64 (uint64x1_t __a, uint64x1_t __b, uint64x1_t __c)
{__builtin_aarch64_simd_bsldi_uuuu (__a[0], __b[0], __c[0])};
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vbslq_f16 (uint16x8_t __a, float16x8_t __b, float16x8_t __c)
+{
+ return __builtin_aarch64_simd_bslv8hf_suss (__a, __b, __c);
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vbslq_f32 (uint32x4_t __a, float32x4_t __b, float32x4_t __c)
{
@@ -12967,6 +12987,12 @@ vcvtpq_u64_f64 (float64x2_t __a)
/* vdup_n */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vdup_n_f16 (float16_t __a)
+{
+ return (float16x4_t) {__a, __a, __a, __a};
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vdup_n_f32 (float32_t __a)
{
@@ -13041,6 +13067,12 @@ vdup_n_u64 (uint64_t __a)
/* vdupq_n */
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vdupq_n_f16 (float16_t __a)
+{
+ return (float16x8_t) {__a, __a, __a, __a, __a, __a, __a, __a};
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vdupq_n_f32 (float32_t __a)
{
@@ -13118,6 +13150,12 @@ vdupq_n_u64 (uint64_t __a)
/* vdup_lane */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vdup_lane_f16 (float16x4_t __a, const int __b)
+{
+ return __aarch64_vdup_lane_f16 (__a, __b);
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vdup_lane_f32 (float32x2_t __a, const int __b)
{
@@ -13192,6 +13230,12 @@ vdup_lane_u64 (uint64x1_t __a, const int __b)
/* vdup_laneq */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vdup_laneq_f16 (float16x8_t __a, const int __b)
+{
+ return __aarch64_vdup_laneq_f16 (__a, __b);
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vdup_laneq_f32 (float32x4_t __a, const int __b)
{
@@ -13265,6 +13309,13 @@ vdup_laneq_u64 (uint64x2_t __a, const int __b)
}
/* vdupq_lane */
+
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vdupq_lane_f16 (float16x4_t __a, const int __b)
+{
+ return __aarch64_vdupq_lane_f16 (__a, __b);
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vdupq_lane_f32 (float32x2_t __a, const int __b)
{
@@ -13338,6 +13389,13 @@ vdupq_lane_u64 (uint64x1_t __a, const int __b)
}
/* vdupq_laneq */
+
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vdupq_laneq_f16 (float16x8_t __a, const int __b)
+{
+ return __aarch64_vdupq_laneq_f16 (__a, __b);
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vdupq_laneq_f32 (float32x4_t __a, const int __b)
{
@@ -13430,6 +13488,13 @@ vdupb_lane_u8 (uint8x8_t __a, const int __b)
}
/* vduph_lane */
+
+__extension__ static __inline float16_t __attribute__ ((__always_inline__))
+vduph_lane_f16 (float16x4_t __a, const int __b)
+{
+ return __aarch64_vget_lane_any (__a, __b);
+}
+
__extension__ static __inline poly16_t __attribute__ ((__always_inline__))
vduph_lane_p16 (poly16x4_t __a, const int __b)
{
@@ -13449,6 +13514,7 @@ vduph_lane_u16 (uint16x4_t __a, const int __b)
}
/* vdups_lane */
+
__extension__ static __inline float32_t __attribute__ ((__always_inline__))
vdups_lane_f32 (float32x2_t __a, const int __b)
{
@@ -13509,6 +13575,13 @@ vdupb_laneq_u8 (uint8x16_t __a, const int __b)
}
/* vduph_laneq */
+
+__extension__ static __inline float16_t __attribute__ ((__always_inline__))
+vduph_laneq_f16 (float16x8_t __a, const int __b)
+{
+ return __aarch64_vget_lane_any (__a, __b);
+}
+
__extension__ static __inline poly16_t __attribute__ ((__always_inline__))
vduph_laneq_p16 (poly16x8_t __a, const int __b)
{
@@ -13528,6 +13601,7 @@ vduph_laneq_u16 (uint16x8_t __a, const int __b)
}
/* vdups_laneq */
+
__extension__ static __inline float32_t __attribute__ ((__always_inline__))
vdups_laneq_f32 (float32x4_t __a, const int __b)
{
@@ -13567,6 +13641,19 @@ vdupd_laneq_u64 (uint64x2_t __a, const int __b)
/* vext */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vext_f16 (float16x4_t __a, float16x4_t __b, __const int __c)
+{
+ __AARCH64_LANE_CHECK (__a, __c);
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__b, __a,
+ (uint16x4_t) {4 - __c, 5 - __c, 6 - __c, 7 - __c});
+#else
+ return __builtin_shuffle (__a, __b,
+ (uint16x4_t) {__c, __c + 1, __c + 2, __c + 3});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vext_f32 (float32x2_t __a, float32x2_t __b, __const int __c)
{
@@ -13698,6 +13785,22 @@ vext_u64 (uint64x1_t __a, uint64x1_t __b, __const int __c)
return __a;
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vextq_f16 (float16x8_t __a, float16x8_t __b, __const int __c)
+{
+ __AARCH64_LANE_CHECK (__a, __c);
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__b, __a,
+ (uint16x8_t) {8 - __c, 9 - __c, 10 - __c, 11 - __c,
+ 12 - __c, 13 - __c, 14 - __c,
+ 15 - __c});
+#else
+ return __builtin_shuffle (__a, __b,
+ (uint16x8_t) {__c, __c + 1, __c + 2, __c + 3,
+ __c + 4, __c + 5, __c + 6, __c + 7});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vextq_f32 (float32x4_t __a, float32x4_t __b, __const int __c)
{
@@ -14333,8 +14436,7 @@ vld1q_u64 (const uint64_t *a)
__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
vld1_dup_f16 (const float16_t* __a)
{
- float16_t __f = *__a;
- return (float16x4_t) { __f, __f, __f, __f };
+ return vdup_n_f16 (*__a);
}
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
@@ -14414,8 +14516,7 @@ vld1_dup_u64 (const uint64_t* __a)
__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
vld1q_dup_f16 (const float16_t* __a)
{
- float16_t __f = *__a;
- return (float16x8_t) { __f, __f, __f, __f, __f, __f, __f, __f };
+ return vdupq_n_f16 (*__a);
}
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
@@ -18018,6 +18119,12 @@ vmlsq_laneq_u32 (uint32x4_t __a, uint32x4_t __b,
/* vmov_n_ */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vmov_n_f16 (float16_t __a)
+{
+ return vdup_n_f16 (__a);
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vmov_n_f32 (float32_t __a)
{
@@ -18090,6 +18197,12 @@ vmov_n_u64 (uint64_t __a)
return (uint64x1_t) {__a};
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vmovq_n_f16 (float16_t __a)
+{
+ return vdupq_n_f16 (__a);
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vmovq_n_f32 (float32_t __a)
{
@@ -20834,6 +20947,12 @@ vrev32q_u16 (uint16x8_t a)
return __builtin_shuffle (a, (uint16x8_t) { 1, 0, 3, 2, 5, 4, 7, 6 });
}
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vrev64_f16 (float16x4_t __a)
+{
+ return __builtin_shuffle (__a, (uint16x4_t) { 3, 2, 1, 0 });
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vrev64_f32 (float32x2_t a)
{
@@ -20888,6 +21007,12 @@ vrev64_u32 (uint32x2_t a)
return __builtin_shuffle (a, (uint32x2_t) { 1, 0 });
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vrev64q_f16 (float16x8_t __a)
+{
+ return __builtin_shuffle (__a, (uint16x8_t) { 3, 2, 1, 0, 7, 6, 5, 4 });
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vrev64q_f32 (float32x4_t a)
{
@@ -23840,6 +23965,16 @@ vtbx4_p8 (poly8x8_t __r, poly8x8x4_t __tab, uint8x8_t __idx)
/* vtrn */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vtrn1_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 1, 7, 3});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 2, 6});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vtrn1_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -23930,6 +24065,16 @@ vtrn1_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vtrn1q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 1, 11, 3, 13, 5, 15, 7});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 8, 2, 10, 4, 12, 6, 14});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vtrn1q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -24056,6 +24201,16 @@ vtrn1q_u64 (uint64x2_t __a, uint64x2_t __b)
#endif
}
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vtrn2_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 6, 2});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 5, 3, 7});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vtrn2_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -24146,6 +24301,16 @@ vtrn2_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vtrn2q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 0, 10, 2, 12, 4, 14, 6});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 9, 3, 11, 5, 13, 7, 15});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vtrn2q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -24272,6 +24437,12 @@ vtrn2q_u64 (uint64x2_t __a, uint64x2_t __b)
#endif
}
+__extension__ static __inline float16x4x2_t __attribute__ ((__always_inline__))
+vtrn_f16 (float16x4_t __a, float16x4_t __b)
+{
+ return (float16x4x2_t) {vtrn1_f16 (__a, __b), vtrn2_f16 (__a, __b)};
+}
+
__extension__ static __inline float32x2x2_t __attribute__ ((__always_inline__))
vtrn_f32 (float32x2_t a, float32x2_t b)
{
@@ -24326,6 +24497,12 @@ vtrn_u32 (uint32x2_t a, uint32x2_t b)
return (uint32x2x2_t) {vtrn1_u32 (a, b), vtrn2_u32 (a, b)};
}
+__extension__ static __inline float16x8x2_t __attribute__ ((__always_inline__))
+vtrnq_f16 (float16x8_t __a, float16x8_t __b)
+{
+ return (float16x8x2_t) {vtrn1q_f16 (__a, __b), vtrn2q_f16 (__a, __b)};
+}
+
__extension__ static __inline float32x4x2_t __attribute__ ((__always_inline__))
vtrnq_f32 (float32x4_t a, float32x4_t b)
{
@@ -24574,6 +24751,7 @@ vuqaddd_s64 (int64_t __a, uint64_t __b)
}
#define __INTERLEAVE_LIST(op) \
+ __DEFINTERLEAVE (op, float16x4x2_t, float16x4_t, f16,) \
__DEFINTERLEAVE (op, float32x2x2_t, float32x2_t, f32,) \
__DEFINTERLEAVE (op, poly8x8x2_t, poly8x8_t, p8,) \
__DEFINTERLEAVE (op, poly16x4x2_t, poly16x4_t, p16,) \
@@ -24583,6 +24761,7 @@ vuqaddd_s64 (int64_t __a, uint64_t __b)
__DEFINTERLEAVE (op, uint8x8x2_t, uint8x8_t, u8,) \
__DEFINTERLEAVE (op, uint16x4x2_t, uint16x4_t, u16,) \
__DEFINTERLEAVE (op, uint32x2x2_t, uint32x2_t, u32,) \
+ __DEFINTERLEAVE (op, float16x8x2_t, float16x8_t, f16, q) \
__DEFINTERLEAVE (op, float32x4x2_t, float32x4_t, f32, q) \
__DEFINTERLEAVE (op, poly8x16x2_t, poly8x16_t, p8, q) \
__DEFINTERLEAVE (op, poly16x8x2_t, poly16x8_t, p16, q) \
@@ -24595,6 +24774,16 @@ vuqaddd_s64 (int64_t __a, uint64_t __b)
/* vuzp */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vuzp1_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {5, 7, 1, 3});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 2, 4, 6});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vuzp1_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -24685,6 +24874,16 @@ vuzp1_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vuzp1q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {9, 11, 13, 15, 1, 3, 5, 7});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {0, 2, 4, 6, 8, 10, 12, 14});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vuzp1q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -24811,6 +25010,16 @@ vuzp1q_u64 (uint64x2_t __a, uint64x2_t __b)
#endif
}
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vuzp2_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 6, 0, 2});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {1, 3, 5, 7});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vuzp2_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -24901,6 +25110,16 @@ vuzp2_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vuzp2q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {8, 10, 12, 14, 0, 2, 4, 6});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x8_t) {1, 3, 5, 7, 9, 11, 13, 15});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vuzp2q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -25031,6 +25250,16 @@ __INTERLEAVE_LIST (uzp)
/* vzip */
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vzip1_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {6, 2, 7, 3});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {0, 4, 1, 5});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vzip1_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -25121,6 +25350,18 @@ vzip1_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vzip1q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b,
+ (uint16x8_t) {12, 4, 13, 5, 14, 6, 15, 7});
+#else
+ return __builtin_shuffle (__a, __b,
+ (uint16x8_t) {0, 8, 1, 9, 2, 10, 3, 11});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vzip1q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -25250,6 +25491,16 @@ vzip1q_u64 (uint64x2_t __a, uint64x2_t __b)
#endif
}
+__extension__ static __inline float16x4_t __attribute__ ((__always_inline__))
+vzip2_f16 (float16x4_t __a, float16x4_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {4, 0, 5, 1});
+#else
+ return __builtin_shuffle (__a, __b, (uint16x4_t) {2, 6, 3, 7});
+#endif
+}
+
__extension__ static __inline float32x2_t __attribute__ ((__always_inline__))
vzip2_f32 (float32x2_t __a, float32x2_t __b)
{
@@ -25340,6 +25591,18 @@ vzip2_u32 (uint32x2_t __a, uint32x2_t __b)
#endif
}
+__extension__ static __inline float16x8_t __attribute__ ((__always_inline__))
+vzip2q_f16 (float16x8_t __a, float16x8_t __b)
+{
+#ifdef __AARCH64EB__
+ return __builtin_shuffle (__a, __b,
+ (uint16x8_t) {8, 0, 9, 1, 10, 2, 11, 3});
+#else
+ return __builtin_shuffle (__a, __b,
+ (uint16x8_t) {4, 12, 5, 13, 6, 14, 7, 15});
+#endif
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vzip2q_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -25479,6 +25742,7 @@ __INTERLEAVE_LIST (zip)
#undef __aarch64_vget_lane_any
#undef __aarch64_vdup_lane_any
+#undef __aarch64_vdup_lane_f16
#undef __aarch64_vdup_lane_f32
#undef __aarch64_vdup_lane_f64
#undef __aarch64_vdup_lane_p8
@@ -25491,6 +25755,7 @@ __INTERLEAVE_LIST (zip)
#undef __aarch64_vdup_lane_u16
#undef __aarch64_vdup_lane_u32
#undef __aarch64_vdup_lane_u64
+#undef __aarch64_vdup_laneq_f16
#undef __aarch64_vdup_laneq_f32
#undef __aarch64_vdup_laneq_f64
#undef __aarch64_vdup_laneq_p8
@@ -25503,6 +25768,7 @@ __INTERLEAVE_LIST (zip)
#undef __aarch64_vdup_laneq_u16
#undef __aarch64_vdup_laneq_u32
#undef __aarch64_vdup_laneq_u64
+#undef __aarch64_vdupq_lane_f16
#undef __aarch64_vdupq_lane_f32
#undef __aarch64_vdupq_lane_f64
#undef __aarch64_vdupq_lane_p8
@@ -25515,6 +25781,7 @@ __INTERLEAVE_LIST (zip)
#undef __aarch64_vdupq_lane_u16
#undef __aarch64_vdupq_lane_u32
#undef __aarch64_vdupq_lane_u64
+#undef __aarch64_vdupq_laneq_f16
#undef __aarch64_vdupq_laneq_f32
#undef __aarch64_vdupq_laneq_f64
#undef __aarch64_vdupq_laneq_p8
--
2.5.0
Several data-processing instructions are agnostic to the type of their operands. This patch add the mapping between them and those bit- and lane-manipulation instructions. No ARMv8.2-A FP16 extension hardware support is required for these intrinsics. gcc/ 2016-07-07 Jiong Wang <jiong.wang@arm.com> * config/aarch64/aarch64-simd.md (aarch64_<PERMUTE:perm_insn><PERMUTE:perm_hilo><mode>): Use VALL_F16. (aarch64_ext<mode>): Likewise. (aarch64_rev<REVERSE:rev_op><mode>): Likewise. * config/aarch64/aarch64.c (aarch64_evpc_trn): Support V4HFmode and V8HFmode. (aarch64_evpc_uzp): Likewise. (aarch64_evpc_zip): Likewise. (aarch64_evpc_ext): Likewise. (aarch64_evpc_rev): Likewise. * config/aarch64/arm_neon.h (__aarch64_vdup_lane_f16): New. (__aarch64_vdup_laneq_f16): New.. (__aarch64_vdupq_lane_f16): New. (__aarch64_vdupq_laneq_f16): New. (vbsl_f16): New. (vbslq_f16): New. (vdup_n_f16): New. (vdupq_n_f16): New. (vdup_lane_f16): New. (vdup_laneq_f16): New. (vdupq_lane_f16): New. (vdupq_laneq_f16): New. (vduph_lane_f16): New. (vduph_laneq_f16): New. (vext_f16): New. (vextq_f16): New. (vmov_n_f16): New. (vmovq_n_f16): New. (vrev64_f16): New. (vrev64q_f16): New. (vtrn1_f16): New. (vtrn1q_f16): New. (vtrn2_f16): New. (vtrn2q_f16): New. (vtrn_f16): New. (vtrnq_f16): New. (__INTERLEAVE_LIST): Support float16x4_t, float16x8_t. (vuzp1_f16): New. (vuzp1q_f16): New. (vuzp2_f16): New. (vuzp2q_f16): New. (vzip1_f16): New. (vzip2q_f16): New. (vmov_n_f16): Reimplement using vdup_n_f16. (vmovq_n_f16): Reimplement using vdupq_n_f16..