2017-10-16 Tamar Christina <tamar.christina@arm.com>
+ * config/aarch64/aarch64-builtins.c
+ (aarch64_types_quadopu_lane_qualifiers): New.
+ (TYPES_QUADOPU_LANE): New.
+ * config/aarch64/aarch64-simd.md (aarch64_<sur>dot<vsi2qi>): New.
+ (<sur>dot_prod<vsi2qi>, aarch64_<sur>dot_lane<vsi2qi>): New.
+ (aarch64_<sur>dot_laneq<vsi2qi>): New.
+ * config/aarch64/aarch64-simd-builtins.def (sdot, udot): New.
+ (sdot_lane, udot_lane, sdot_laneq, udot_laneq): New.
+ * config/aarch64/iterators.md (sur): Add UNSPEC_SDOT, UNSPEC_UDOT.
+ (Vdottype, DOTPROD): New.
+ (sur): Add SDOT and UDOT.
+
+2017-10-16 Tamar Christina <tamar.christina@arm.com>
+
* config/aarch64/aarch64.h (AARCH64_FL_DOTPROD): New.
(AARCH64_ISA_DOTPROD, TARGET_DOTPROD): New.
* config/aarch64/aarch64-c.c (aarch64_update_cpp_builtins): Add TARGET_DOTPROD.
= { qualifier_none, qualifier_none, qualifier_none,
qualifier_none, qualifier_lane_index };
#define TYPES_QUADOP_LANE (aarch64_types_quadop_lane_qualifiers)
+static enum aarch64_type_qualifiers
+aarch64_types_quadopu_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS]
+ = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned,
+ qualifier_unsigned, qualifier_lane_index };
+#define TYPES_QUADOPU_LANE (aarch64_types_quadopu_lane_qualifiers)
static enum aarch64_type_qualifiers
aarch64_types_binop_imm_p_qualifiers[SIMD_MAX_BUILTIN_ARGS]
BUILTIN_VSDQ_I_DI (BINOP, srshl, 0)
BUILTIN_VSDQ_I_DI (BINOP_UUS, urshl, 0)
+ /* Implemented by aarch64_<sur><dotprod>{_lane}{q}<dot_mode>. */
+ BUILTIN_VB (TERNOP, sdot, 0)
+ BUILTIN_VB (TERNOPU, udot, 0)
+ BUILTIN_VB (QUADOP_LANE, sdot_lane, 0)
+ BUILTIN_VB (QUADOPU_LANE, udot_lane, 0)
+ BUILTIN_VB (QUADOP_LANE, sdot_laneq, 0)
+ BUILTIN_VB (QUADOPU_LANE, udot_laneq, 0)
+
BUILTIN_VDQ_I (SHIFTIMM, ashr, 3)
VAR1 (SHIFTIMM, ashr_simd, 0, di)
BUILTIN_VDQ_I (SHIFTIMM, lshr, 3)
}
)
+;; These instructions map to the __builtins for the Dot Product operations.
+(define_insn "aarch64_<sur>dot<vsi2qi>"
+ [(set (match_operand:VS 0 "register_operand" "=w")
+ (plus:VS (match_operand:VS 1 "register_operand" "0")
+ (unspec:VS [(match_operand:<VSI2QI> 2 "register_operand" "w")
+ (match_operand:<VSI2QI> 3 "register_operand" "w")]
+ DOTPROD)))]
+ "TARGET_DOTPROD"
+ "<sur>dot\\t%0.<Vtype>, %2.<Vdottype>, %3.<Vdottype>"
+ [(set_attr "type" "neon_dot")]
+)
+
+;; 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 "<sur>dot_prod<vsi2qi>"
+ [(set (match_operand:VS 0 "register_operand")
+ (plus:VS (unspec:VS [(match_operand:<VSI2QI> 1 "register_operand")
+ (match_operand:<VSI2QI> 2 "register_operand")]
+ DOTPROD)
+ (match_operand:VS 3 "register_operand")))]
+ "TARGET_DOTPROD"
+{
+ emit_insn (
+ gen_aarch64_<sur>dot<vsi2qi> (operands[3], operands[3], operands[1],
+ operands[2]));
+ emit_insn (gen_rtx_SET (operands[0], operands[3]));
+ DONE;
+})
+
+;; These instructions map to the __builtins for the Dot Product
+;; indexed operations.
+(define_insn "aarch64_<sur>dot_lane<vsi2qi>"
+ [(set (match_operand:VS 0 "register_operand" "=w")
+ (plus:VS (match_operand:VS 1 "register_operand" "0")
+ (unspec:VS [(match_operand:<VSI2QI> 2 "register_operand" "w")
+ (match_operand:V8QI 3 "register_operand" "<h_con>")
+ (match_operand:SI 4 "immediate_operand" "i")]
+ DOTPROD)))]
+ "TARGET_DOTPROD"
+ {
+ operands[4]
+ = GEN_INT (ENDIAN_LANE_N (V8QImode, INTVAL (operands[4])));
+ return "<sur>dot\\t%0.<Vtype>, %2.<Vdottype>, %3.4b[%4]";
+ }
+ [(set_attr "type" "neon_dot")]
+)
+
+(define_insn "aarch64_<sur>dot_laneq<vsi2qi>"
+ [(set (match_operand:VS 0 "register_operand" "=w")
+ (plus:VS (match_operand:VS 1 "register_operand" "0")
+ (unspec:VS [(match_operand:<VSI2QI> 2 "register_operand" "w")
+ (match_operand:V16QI 3 "register_operand" "<h_con>")
+ (match_operand:SI 4 "immediate_operand" "i")]
+ DOTPROD)))]
+ "TARGET_DOTPROD"
+ {
+ operands[4]
+ = GEN_INT (ENDIAN_LANE_N (V16QImode, INTVAL (operands[4])));
+ return "<sur>dot\\t%0.<Vtype>, %2.<Vdottype>, %3.4b[%4]";
+ }
+ [(set_attr "type" "neon_dot")]
+)
+
(define_expand "copysign<mode>3"
[(match_operand:VHSDF 0 "register_operand")
(match_operand:VHSDF 1 "register_operand")
UNSPEC_SQRDMLSH ; Used in aarch64-simd.md.
UNSPEC_FMAXNM ; Used in aarch64-simd.md.
UNSPEC_FMINNM ; Used in aarch64-simd.md.
+ UNSPEC_SDOT ; Used in aarch64-simd.md.
+ UNSPEC_UDOT ; Used in aarch64-simd.md.
])
;; ------------------------------------------------------------------
(define_mode_attr vsi2qi [(V2SI "v8qi") (V4SI "v16qi")])
(define_mode_attr VSI2QI [(V2SI "V8QI") (V4SI "V16QI")])
+
+;; Register suffix for DOTPROD input types from the return type.
+(define_mode_attr Vdottype [(V2SI "8b") (V4SI "16b")])
+
;; Sum of lengths of instructions needed to move vector registers of a mode.
(define_mode_attr insn_count [(OI "8") (CI "12") (XI "16")])
UNSPEC_SHSUB UNSPEC_UHSUB
UNSPEC_SRHSUB UNSPEC_URHSUB])
+(define_int_iterator DOTPROD [UNSPEC_SDOT UNSPEC_UDOT])
(define_int_iterator ADDSUBHN [UNSPEC_ADDHN UNSPEC_RADDHN
UNSPEC_SUBHN UNSPEC_RSUBHN])
(UNSPEC_USHLL "u") (UNSPEC_SSHLL "s")
(UNSPEC_URSHL "ur") (UNSPEC_SRSHL "sr")
(UNSPEC_UQRSHL "u") (UNSPEC_SQRSHL "s")
+ (UNSPEC_SDOT "s") (UNSPEC_UDOT "u")
])
(define_int_attr r [(UNSPEC_SQDMULH "") (UNSPEC_SQRDMULH "r")
projects / platform / upstream / gcc.git / commitdiff