The pattern:
;; <su>q<addsub>
(define_insn "aarch64_<su_optab><optab><mode>"
[(set (match_operand:VSDQ_I 0 "register_operand" "=w")
(BINQOPS:VSDQ_I (match_operand:VSDQ_I 1 "register_operand" "w")
(match_operand:VSDQ_I 2 "register_operand" "w")))]
"TARGET_SIMD"
"<su_optab><optab>\\t%<v>0<Vmtype>, %<v>1<Vmtype>, %<v>2<Vmtype>"
[(set_attr "type" "neon_<optab><q>")]
)
meant that we overloaded "optab" to be "qadd" for both SQADD and
UQADD. Most other "optab" entries are instead the full optab name,
which for these patterns would be "ssadd" and "usadd" respectively.
(Unfortunately, the optabs don't extend to vectors yet, something
that would be good to fix in GCC 11.)
This patch therefore does what the comment implies and uses
q<addsub> to distinguish qadd and qsub instead.
Tested on aarch64-linux-gnu and applied as 280050.
Richard
2020-01-09 Richard Sandiford <richard.sandif...@arm.com>
gcc/
* config/aarch64/iterators.md (addsub): New code attribute.
* config/aarch64/aarch64-simd.md (aarch64_<su_optab><optab><mode>):
Re-express as...
(aarch64_<su_optab>q<addsub><mode>): ...this, making the same change
in the asm string and attributes. Fix indentation.
* config/aarch64/aarch64-sve.md (@aarch64_<su_optab><optab><mode>):
Re-express as...
(@aarch64_sve_<optab><mode>): ...this.
* config/aarch64/aarch64-sve-builtins.h
(function_expander::expand_signed_unpred_op): Delete.
* config/aarch64/aarch64-sve-builtins.cc
(function_expander::expand_signed_unpred_op): Likewise.
(function_expander::map_to_rtx_codes): If the optab isn't defined,
try using code_for_aarch64_sve instead.
* config/aarch64/aarch64-sve-builtins-base.cc (svqadd_impl): Delete.
(svqsub_impl): Likewise.
(svqadd, svqsub): Use rtx_code_function instead.
Index: gcc/config/aarch64/iterators.md
===================================================================
--- gcc/config/aarch64/iterators.md 2020-01-09 15:11:26.680035760 +0000
+++ gcc/config/aarch64/iterators.md 2020-01-09 15:14:22.178906205 +0000
@@ -1577,10 +1577,10 @@ (define_code_attr optab [(ashift "ashl")
(mult "mul")
(div "div")
(udiv "udiv")
- (ss_plus "qadd")
- (us_plus "qadd")
- (ss_minus "qsub")
- (us_minus "qsub")
+ (ss_plus "ssadd")
+ (us_plus "usadd")
+ (ss_minus "sssub")
+ (us_minus "ussub")
(ss_neg "qneg")
(ss_abs "qabs")
(smin "smin")
@@ -1599,6 +1599,11 @@ (define_code_attr optab [(ashift "ashl")
(gtu "gtu")
(abs "abs")])
+(define_code_attr addsub [(ss_plus "add")
+ (us_plus "add")
+ (ss_minus "sub")
+ (us_minus "sub")])
+
;; For comparison operators we use the FCM* and CM* instructions.
;; As there are no CMLE or CMLT instructions which act on 3 vector
;; operands, we must use CMGE or CMGT and swap the order of the
Index: gcc/config/aarch64/aarch64-simd.md
===================================================================
--- gcc/config/aarch64/aarch64-simd.md 2020-01-06 12:58:16.461794517 +0000
+++ gcc/config/aarch64/aarch64-simd.md 2020-01-09 15:14:22.174906229 +0000
@@ -3696,13 +3696,13 @@ (define_insn "*aarch64_vgetfmulx<mode>"
)
;; <su>q<addsub>
-(define_insn "aarch64_<su_optab><optab><mode>"
+(define_insn "aarch64_<su_optab>q<addsub><mode>"
[(set (match_operand:VSDQ_I 0 "register_operand" "=w")
(BINQOPS:VSDQ_I (match_operand:VSDQ_I 1 "register_operand" "w")
- (match_operand:VSDQ_I 2 "register_operand" "w")))]
+ (match_operand:VSDQ_I 2 "register_operand" "w")))]
"TARGET_SIMD"
- "<su_optab><optab>\\t%<v>0<Vmtype>, %<v>1<Vmtype>, %<v>2<Vmtype>"
- [(set_attr "type" "neon_<optab><q>")]
+ "<su_optab>q<addsub>\\t%<v>0<Vmtype>, %<v>1<Vmtype>, %<v>2<Vmtype>"
+ [(set_attr "type" "neon_q<addsub><q>")]
)
;; suqadd and usqadd
Index: gcc/config/aarch64/aarch64-sve.md
===================================================================
--- gcc/config/aarch64/aarch64-sve.md 2020-01-06 12:59:46.341203294 +0000
+++ gcc/config/aarch64/aarch64-sve.md 2020-01-09 15:14:22.178906205 +0000
@@ -3879,7 +3879,7 @@ (define_insn_and_rewrite "*aarch64_cond_
;; -------------------------------------------------------------------------
;; Unpredicated saturating signed addition and subtraction.
-(define_insn "@aarch64_<su_optab><optab><mode>"
+(define_insn "@aarch64_sve_<optab><mode>"
[(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, w, ?&w, ?&w, w")
(SBINQOPS:SVE_FULL_I
(match_operand:SVE_FULL_I 1 "register_operand" "0, 0, w, w, w")
@@ -3895,7 +3895,7 @@ (define_insn "@aarch64_<su_optab><optab>
)
;; Unpredicated saturating unsigned addition and subtraction.
-(define_insn "@aarch64_<su_optab><optab><mode>"
+(define_insn "@aarch64_sve_<optab><mode>"
[(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, ?&w, w")
(UBINQOPS:SVE_FULL_I
(match_operand:SVE_FULL_I 1 "register_operand" "0, w, w")
Index: gcc/config/aarch64/aarch64-sve-builtins.h
===================================================================
--- gcc/config/aarch64/aarch64-sve-builtins.h 2020-01-06 12:58:16.461794517
+0000
+++ gcc/config/aarch64/aarch64-sve-builtins.h 2020-01-09 15:14:22.174906229
+0000
@@ -554,7 +554,6 @@ class GTY((user)) function_instance
rtx map_to_rtx_codes (rtx_code, rtx_code, int,
unsigned int = DEFAULT_MERGE_ARGNO);
rtx map_to_unspecs (int, int, int, unsigned int = DEFAULT_MERGE_ARGNO);
- rtx expand_signed_unpred_op (rtx_code, rtx_code);
/* The function call expression. */
tree call_expr;
Index: gcc/config/aarch64/aarch64-sve-builtins.cc
===================================================================
--- gcc/config/aarch64/aarch64-sve-builtins.cc 2020-01-09 15:08:17.477253413
+0000
+++ gcc/config/aarch64/aarch64-sve-builtins.cc 2020-01-09 15:14:22.174906229
+0000
@@ -2871,7 +2871,10 @@ function_expander::use_contiguous_store_
(3) a normal unpredicated optab for PRED_none and PRED_x functions,
dropping the predicate in the latter case
- (4) "cond_<optab><mode>" otherwise
+ (4) an unpredicated "aarch64_sve_<code_optab><mode>" for PRED_none and
+ PRED_x functions, again dropping the predicate for PRED_x
+
+ (5) "cond_<optab><mode>" otherwise
where <optab> corresponds to:
@@ -2879,6 +2882,9 @@ function_expander::use_contiguous_store_
- CODE_FOR_UINT for unsigned integers
- UNSPEC_FOR_FP for floating-point values
+ and where <code_optab> is like <optab>, but uses CODE_FOR_SINT instead
+ of UNSPEC_FOR_FP for floating-point values.
+
MERGE_ARGNO is the argument that provides the values of inactive lanes for
_m functions, or DEFAULT_MERGE_ARGNO if we should apply the usual rules. */
rtx
@@ -2913,7 +2919,12 @@ function_expander::map_to_rtx_codes (rtx
/* Otherwise expand PRED_none and PRED_x operations without a predicate.
Floating-point operations conventionally use the signed rtx code. */
if (pred == PRED_none || pred == PRED_x)
- return use_unpred_insn (direct_optab_handler (code_to_optab (code), 0));
+ {
+ icode = direct_optab_handler (code_to_optab (code), 0);
+ if (icode == CODE_FOR_nothing)
+ icode = code_for_aarch64_sve (code, mode);
+ return use_unpred_insn (icode);
+ }
/* Don't use cond_*_optabs here, since not all codes have one yet. */
if (type_suffix (0).integer_p)
@@ -2964,22 +2975,6 @@ function_expander::map_to_unspecs (int u
return use_cond_insn (icode, merge_argno);
}
-/* Implement the call using an @aarch64 instruction and the
- instructions are parameterized by an rtx_code. CODE_FOR_SINT
- is the rtx_code for signed integer operations, CODE_FOR_UINT
- is the rtx_code for unsigned integer operations. */
-rtx
-function_expander::expand_signed_unpred_op (rtx_code code_for_sint,
- rtx_code code_for_uint)
-{
- insn_code icode;
- if (type_suffix (0).unsigned_p)
- icode = code_for_aarch64 (code_for_uint, code_for_uint, vector_mode (0));
- else
- icode = code_for_aarch64 (code_for_sint, code_for_sint, vector_mode (0));
- return use_unpred_insn (icode);
-}
-
/* Expand the call and return its lhs. */
rtx
function_expander::expand ()
Index: gcc/config/aarch64/aarch64-sve-builtins-base.cc
===================================================================
--- gcc/config/aarch64/aarch64-sve-builtins-base.cc 2020-01-06
12:58:16.457794543 +0000
+++ gcc/config/aarch64/aarch64-sve-builtins-base.cc 2020-01-09
15:14:22.174906229 +0000
@@ -1727,16 +1727,6 @@ public:
}
};
-class svqadd_impl : public function_base
-{
-public:
- rtx
- expand (function_expander &e) const OVERRIDE
- {
- return e.expand_signed_unpred_op (SS_PLUS, US_PLUS);
- }
-};
-
/* Implements svqdec[bhwd]{,_pat} and svqinc[bhwd]{,_pat}. */
class svqdec_svqinc_bhwd_impl : public function_base
{
@@ -1838,16 +1828,6 @@ public:
rtx_code m_code_for_uint;
};
-class svqsub_impl : public function_base
-{
-public:
- rtx
- expand (function_expander &e) const OVERRIDE
- {
- return e.expand_signed_unpred_op (SS_MINUS, US_MINUS);
- }
-};
-
class svrdffr_impl : public function_base
{
public:
@@ -2661,7 +2641,7 @@ FUNCTION (svptest_first, svptest_impl, (
FUNCTION (svptest_last, svptest_impl, (LTU))
FUNCTION (svptrue, svptrue_impl,)
FUNCTION (svptrue_pat, svptrue_pat_impl,)
-FUNCTION (svqadd, svqadd_impl,)
+FUNCTION (svqadd, rtx_code_function, (SS_PLUS, US_PLUS, -1))
FUNCTION (svqdecb, svqdec_bhwd_impl, (QImode))
FUNCTION (svqdecb_pat, svqdec_bhwd_impl, (QImode))
FUNCTION (svqdecd, svqdec_bhwd_impl, (DImode))
@@ -2680,7 +2660,7 @@ FUNCTION (svqinch_pat, svqinc_bhwd_impl,
FUNCTION (svqincp, svqdecp_svqincp_impl, (SS_PLUS, US_PLUS))
FUNCTION (svqincw, svqinc_bhwd_impl, (SImode))
FUNCTION (svqincw_pat, svqinc_bhwd_impl, (SImode))
-FUNCTION (svqsub, svqsub_impl,)
+FUNCTION (svqsub, rtx_code_function, (SS_MINUS, US_MINUS, -1))
FUNCTION (svrbit, unspec_based_function, (UNSPEC_RBIT, UNSPEC_RBIT, -1))
FUNCTION (svrdffr, svrdffr_impl,)
FUNCTION (svrecpe, unspec_based_function, (-1, -1, UNSPEC_FRECPE))
