diff options
author | Tamar Christina <tamar.christina@arm.com> | 2017-10-16 09:56:41 +0000 |
---|---|---|
committer | Tamar Christina <tnfchris@gcc.gnu.org> | 2017-10-16 09:56:41 +0000 |
commit | 7a08d81307ef1a69bbdcb061e21881ab5def593a (patch) | |
tree | a592454f07b6d574bcdfdbe64b8d990dd6a614ee /gcc/config | |
parent | 1ddc47c05e80aed3c3ec5b636dd706c0d0264dbc (diff) | |
download | gcc-7a08d81307ef1a69bbdcb061e21881ab5def593a.zip gcc-7a08d81307ef1a69bbdcb061e21881ab5def593a.tar.gz gcc-7a08d81307ef1a69bbdcb061e21881ab5def593a.tar.bz2 |
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.
From-SVN: r253783
Diffstat (limited to 'gcc/config')
-rw-r--r-- | gcc/config/aarch64/aarch64-builtins.c | 5 | ||||
-rw-r--r-- | gcc/config/aarch64/aarch64-simd-builtins.def | 8 | ||||
-rw-r--r-- | gcc/config/aarch64/aarch64-simd.md | 81 | ||||
-rw-r--r-- | gcc/config/aarch64/iterators.md | 8 |
4 files changed, 102 insertions, 0 deletions
diff --git a/gcc/config/aarch64/aarch64-builtins.c b/gcc/config/aarch64/aarch64-builtins.c index 7edf75c..242b2e3 100644 --- a/gcc/config/aarch64/aarch64-builtins.c +++ b/gcc/config/aarch64/aarch64-builtins.c @@ -168,6 +168,11 @@ aarch64_types_quadop_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS] = { 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] diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def index d713d5d..52d0134 100644 --- a/gcc/config/aarch64/aarch64-simd-builtins.def +++ b/gcc/config/aarch64/aarch64-simd-builtins.def @@ -205,6 +205,14 @@ 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) diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 12da8be..49f615c 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -393,6 +393,87 @@ } ) +;; 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") diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md index 477dc35..48cedbe 100644 --- a/gcc/config/aarch64/iterators.md +++ b/gcc/config/aarch64/iterators.md @@ -354,6 +354,8 @@ 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. ]) ;; ------------------------------------------------------------------ @@ -800,6 +802,10 @@ (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")]) @@ -1029,6 +1035,7 @@ 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]) @@ -1166,6 +1173,7 @@ (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") |