return __builtin_neon_sdot_lanev16qi (__r, __a, __b, __index);
}
+__extension__ extern __inline uint32x2_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vdot_laneq_u32 (uint32x2_t __r, uint8x8_t __a, uint8x16_t __b, const int __index)
+{
+ return __builtin_neon_udot_laneqv8qi_uuuus (__r, __a, __b, __index);
+}
+
+__extension__ extern __inline uint32x4_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vdotq_laneq_u32 (uint32x4_t __r, uint8x16_t __a, uint8x16_t __b,
+ const int __index)
+{
+ return __builtin_neon_udot_laneqv16qi_uuuus (__r, __a, __b, __index);
+}
+
+__extension__ extern __inline int32x2_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vdot_laneq_s32 (int32x2_t __r, int8x8_t __a, int8x16_t __b, const int __index)
+{
+ return __builtin_neon_sdot_laneqv8qi (__r, __a, __b, __index);
+}
+
+__extension__ extern __inline int32x4_t
+__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
+vdotq_laneq_s32 (int32x4_t __r, int8x16_t __a, int8x16_t __b, const int __index)
+{
+ return __builtin_neon_sdot_laneqv16qi (__r, __a, __b, __index);
+}
+
#pragma GCC pop_options
#endif
VAR2 (UTERNOP, udot, v8qi, v16qi)
VAR2 (MAC_LANE, sdot_lane, v8qi, v16qi)
VAR2 (UMAC_LANE, udot_lane, v8qi, v16qi)
+VAR2 (MAC_LANE, sdot_laneq, v8qi, v16qi)
+VAR2 (UMAC_LANE, udot_laneq, v8qi, v16qi)
VAR1 (USTERNOP, usdot, v8qi)
VAR2 (USMAC_LANE_QUADTUP, usdot_lane, v8qi, v16qi)
})
-;; These instructions map to the __builtins for the Dot Product operations.
-(define_insn "neon_<sup>dot<vsi2qi>"
+;; These map to the auto-vectorizer Dot Product optab.
+;; The auto-vectorizer expects a dot product builtin that also does an
+;; accumulation into the provided register.
+;; Given the following pattern
+;;
+;; for (i=0; i<len; i++) {
+;; c = a[i] * b[i];
+;; r += c;
+;; }
+;; return result;
+;;
+;; This can be auto-vectorized to
+;; r = a[0]*b[0] + a[1]*b[1] + a[2]*b[2] + a[3]*b[3];
+;;
+;; given enough iterations. However the vectorizer can keep unrolling the loop
+;; r += a[4]*b[4] + a[5]*b[5] + a[6]*b[6] + a[7]*b[7];
+;; r += a[8]*b[8] + a[9]*b[9] + a[10]*b[10] + a[11]*b[11];
+;; ...
+;;
+;; and so the vectorizer provides r, in which the result has to be accumulated.
+(define_insn "<sup>dot_prod<vsi2qi>"
[(set (match_operand:VCVTI 0 "register_operand" "=w")
- (plus:VCVTI (match_operand:VCVTI 1 "register_operand" "0")
- (unspec:VCVTI [(match_operand:<VSI2QI> 2
- "register_operand" "w")
- (match_operand:<VSI2QI> 3
- "register_operand" "w")]
- DOTPROD)))]
+ (plus:VCVTI
+ (unspec:VCVTI [(match_operand:<VSI2QI> 1 "register_operand" "w")
+ (match_operand:<VSI2QI> 2 "register_operand" "w")]
+ DOTPROD)
+ (match_operand:VCVTI 3 "register_operand" "0")))]
"TARGET_DOTPROD"
- "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %<V_reg>3"
+ "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
[(set_attr "type" "neon_dot<q>")]
)
+;; These instructions map to the __builtins for the Dot Product operations
+(define_expand "neon_<sup>dot<vsi2qi>"
+ [(set (match_operand:VCVTI 0 "register_operand" "=w")
+ (plus:VCVTI
+ (unspec:VCVTI [(match_operand:<VSI2QI> 2 "register_operand")
+ (match_operand:<VSI2QI> 3 "register_operand")]
+ DOTPROD)
+ (match_operand:VCVTI 1 "register_operand")))]
+ "TARGET_DOTPROD"
+)
+
;; These instructions map to the __builtins for the Dot Product operations.
(define_insn "neon_usdot<vsi2qi>"
[(set (match_operand:VCVTI 0 "register_operand" "=w")
;; indexed operations.
(define_insn "neon_<sup>dot_lane<vsi2qi>"
[(set (match_operand:VCVTI 0 "register_operand" "=w")
- (plus:VCVTI (match_operand:VCVTI 1 "register_operand" "0")
- (unspec:VCVTI [(match_operand:<VSI2QI> 2
- "register_operand" "w")
- (match_operand:V8QI 3 "register_operand" "t")
- (match_operand:SI 4 "immediate_operand" "i")]
- DOTPROD)))]
+ (plus:VCVTI
+ (unspec:VCVTI [(match_operand:<VSI2QI> 2 "register_operand" "w")
+ (match_operand:V8QI 3 "register_operand" "t")
+ (match_operand:SI 4 "immediate_operand" "i")]
+ DOTPROD)
+ (match_operand:VCVTI 1 "register_operand" "0")))]
+ "TARGET_DOTPROD"
+ "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %P3[%c4]";
+ [(set_attr "type" "neon_dot<q>")]
+)
+
+;; These instructions map to the __builtins for the Dot Product
+;; indexed operations.
+(define_insn "neon_<sup>dot_laneq<vsi2qi>"
+ [(set (match_operand:VCVTI 0 "register_operand" "=w")
+ (plus:VCVTI
+ (unspec:VCVTI [(match_operand:<VSI2QI> 2 "register_operand" "w")
+ (match_operand:V16QI 3 "register_operand" "t")
+ (match_operand:SI 4 "immediate_operand" "i")]
+ DOTPROD)
+ (match_operand:VCVTI 1 "register_operand" "0")))]
"TARGET_DOTPROD"
{
- operands[4]
- = GEN_INT (NEON_ENDIAN_LANE_N (V8QImode, INTVAL (operands[4])));
- return "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %P3[%c4]";
+ int lane = INTVAL (operands[4]);
+ if (lane > GET_MODE_NUNITS (V2SImode) - 1)
+ {
+ operands[4] = GEN_INT (lane - GET_MODE_NUNITS (V2SImode));
+ return "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %f3[%c4]";
+ }
+ else
+ {
+ operands[4] = GEN_INT (lane);
+ return "v<sup>dot.<opsuffix>\\t%<V_reg>0, %<V_reg>2, %e3[%c4]";
+ }
}
[(set_attr "type" "neon_dot<q>")]
)
[(set_attr "type" "neon_dot<q>")]
)
-;; These expands map to the Dot Product optab the vectorizer checks for.
-;; The auto-vectorizer expects a dot product builtin that also does an
-;; accumulation into the provided register.
-;; Given the following pattern
-;;
-;; for (i=0; i<len; i++) {
-;; c = a[i] * b[i];
-;; r += c;
-;; }
-;; return result;
-;;
-;; This can be auto-vectorized to
-;; r = a[0]*b[0] + a[1]*b[1] + a[2]*b[2] + a[3]*b[3];
-;;
-;; given enough iterations. However the vectorizer can keep unrolling the loop
-;; r += a[4]*b[4] + a[5]*b[5] + a[6]*b[6] + a[7]*b[7];
-;; r += a[8]*b[8] + a[9]*b[9] + a[10]*b[10] + a[11]*b[11];
-;; ...
-;;
-;; and so the vectorizer provides r, in which the result has to be accumulated.
-(define_expand "<sup>dot_prod<vsi2qi>"
- [(set (match_operand:VCVTI 0 "register_operand")
- (plus:VCVTI (unspec:VCVTI [(match_operand:<VSI2QI> 1
- "register_operand")
- (match_operand:<VSI2QI> 2
- "register_operand")]
- DOTPROD)
- (match_operand:VCVTI 3 "register_operand")))]
- "TARGET_DOTPROD"
-{
- emit_insn (
- gen_neon_<sup>dot<vsi2qi> (operands[3], operands[3], operands[1],
- operands[2]));
- emit_insn (gen_rtx_SET (operands[0], operands[3]));
- DONE;
-})
-
;; Auto-vectorizer pattern for usdot
(define_expand "usdot_prod<vsi2qi>"
[(set (match_operand:VCVTI 0 "register_operand")
return vdotq_lane_s32 (r, x, y, 0);
}
-/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\td[0-9]+, d[0-9]+, d[0-9]+} 4 } } */
+int32x2_t sfoo_laneq1 (int32x2_t r, int8x8_t x, int8x16_t y)
+{
+ return vdot_laneq_s32 (r, x, y, 0);
+}
+
+int32x4_t sfooq_lane1 (int32x4_t r, int8x16_t x, int8x16_t y)
+{
+ return vdotq_laneq_s32 (r, x, y, 0);
+}
+
+int32x2_t sfoo_laneq2 (int32x2_t r, int8x8_t x, int8x16_t y)
+{
+ return vdot_laneq_s32 (r, x, y, 2);
+}
+
+int32x4_t sfooq_lane2 (int32x4_t r, int8x16_t x, int8x16_t y)
+{
+ return vdotq_laneq_s32 (r, x, y, 2);
+}
+
+/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\td[0-9]+, d[0-9]+, d[0-9]+} 6 } } */
/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\tq[0-9]+, q[0-9]+, q[0-9]+} 2 } } */
-/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\td[0-9]+, d[0-9]+, d[0-9]+\[#?[0-9]\]} 2 } } */
-/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\tq[0-9]+, q[0-9]+, d[0-9]+\[#?[0-9]\]} 2 } } */
+/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\td[0-9]+, d[0-9]+, d[0-9]+\[#?[0-9]\]} 4 } } */
+/* { dg-final { scan-assembler-times {v[us]dot\.[us]8\tq[0-9]+, q[0-9]+, d[0-9]+\[#?[0-9]\]} 4 } } */
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
# define ORDER(x, y) y
#else
-# define ORDER(x, y) x - y
+# define ORDER(x, y) (x - y)
#endif
#define P(n1,n2) n1,n1,n1,n1,n2,n2,n2,n2
t3 f##_##rx1 = {0}; \
f##_##rx1 = f (f##_##rx1, f##_##x, f##_##y, ORDER (1, 1)); \
if (f##_##rx1[0] != n3 || f##_##rx1[1] != n4) \
- abort (); \
+ abort ();
+
+#define P2(n1,n2) n1,n1,n1,n1,n2,n2,n2,n2,n1,n1,n1,n1,n2,n2,n2,n2
+#define TEST_LANEQ(t1, t2, t3, f, r1, r2, n1, n2, n3, n4) \
+ ARR(f, x, t1, r1); \
+ ARR(f, y, t2, r2); \
+ t3 f##_##rx = {0}; \
+ f##_##rx = f (f##_##rx, f##_##x, f##_##y, ORDER (3, 2)); \
+ if (f##_##rx[0] != n1 || f##_##rx[1] != n2) \
+ abort (); \
+ t3 f##_##rx1 = {0}; \
+ f##_##rx1 = f (f##_##rx1, f##_##x, f##_##y, ORDER (3, 3)); \
+ if (f##_##rx1[0] != n3 || f##_##rx1[1] != n4) \
+ abort ();
int
main()
TEST (int8x16_t, int8x16_t, int32x4_t, vdotq_s32, P(1,2), P(-2,-3), -8, -24);
TEST_LANE (uint8x8_t, uint8x8_t, uint32x2_t, vdot_lane_u32, P(1,2), P(2,3), 8, 16, 12, 24);
-
TEST_LANE (int8x8_t, int8x8_t, int32x2_t, vdot_lane_s32, P(1,2), P(-2,-3), -8, -16, -12, -24);
TEST_LANE (uint8x16_t, uint8x8_t, uint32x4_t, vdotq_lane_u32, P(1,2), P(2,3), 8, 16, 12, 24);
TEST_LANE (int8x16_t, int8x8_t, int32x4_t, vdotq_lane_s32, P(1,2), P(-2,-3), -8, -16, -12, -24);
+ TEST_LANEQ (uint8x8_t, uint8x16_t, uint32x2_t, vdot_laneq_u32, P(1,2), P2(2,3), 8, 16, 12, 24);
+ TEST_LANEQ (int8x8_t, int8x16_t, int32x2_t, vdot_laneq_s32, P(1,2), P2(-2,-3), -8, -16, -12, -24);
+
+ TEST_LANEQ (uint8x16_t, uint8x16_t, uint32x4_t, vdotq_laneq_u32, P2(1,2), P2(2,3), 8, 16, 12, 24);
+ TEST_LANEQ (int8x16_t, int8x16_t, int32x4_t, vdotq_laneq_s32, P2(1,2), P2(-2,-3), -8, -16, -12, -24);
+
return 0;
}