Hi,
Add support for the vmsumudm instruction and tie it into the vec_msum
built-in to support the variants of that built-in using vector
_int128 parameters.
vector _uint128_t vec_msum (vector unsigned long long,
vector unsigned long long,
vector _uint128_t);
vector _int128_t vec_msum (vector signed long long,
vector signed long long,
vector _int128_t);
[v2]
Corrected the define_insn and test requirements to be limited to P9.
Improve description to clarify the vmsum does
a widening multiply and horizontal addition.
Fresh regtests currently running on assorted powerpc targets.
OK for trunk?
Thanks,
-Will
[gcc]
2020-06-18 Will Schmidt <will_schm...@vnet.ibm.com>
* config/rs6000/altivec.h (vec_vmsumudm): New define.
* config/rs6000/altivec.md (UNSPEC_VMSUMUDM): New unspec.
(altivec_vmsumudm): New define_insn.
* config/rs6000/rs6000-builtin.def (altivec_vmsumudm): New
BU_ALTIVEC_3 entry. (vmsumudm): New BU_ALTIVEC_OVERLOAD_3
entry.
* config/rs6000/rs6000-call.c (altivec_overloaded_builtins):
Add entries for ALTIVEC_BUILTIN_VMSUMUDM variants of vec_msum.
[testsuite]
2020-06-18 Will Schmidt <will_schm...@vnet.ibm.com>
* gcc.target/powerpc/builtins-msum-runnable.c: New test.
* gcc.target/powerpc/vsx-builtin-msum.c: New test.
diff --git a/gcc/config/rs6000/altivec.h b/gcc/config/rs6000/altivec.h
index bb1524f4a679..0d199393556d 100644
--- a/gcc/config/rs6000/altivec.h
+++ b/gcc/config/rs6000/altivec.h
@@ -159,10 +159,11 @@
#define vec_vmsumuhm __builtin_vec_vmsumuhm
#define vec_vmsummbm __builtin_vec_vmsummbm
#define vec_vmsumubm __builtin_vec_vmsumubm
#define vec_vmsumshs __builtin_vec_vmsumshs
#define vec_vmsumuhs __builtin_vec_vmsumuhs
+#define vec_vmsumudm __builtin_vec_vmsumudm
#define vec_vmulesb __builtin_vec_vmulesb
#define vec_vmulesh __builtin_vec_vmulesh
#define vec_vmuleuh __builtin_vec_vmuleuh
#define vec_vmuleub __builtin_vec_vmuleub
#define vec_vmulosh __builtin_vec_vmulosh
diff --git a/gcc/config/rs6000/altivec.md b/gcc/config/rs6000/altivec.md
index 2ce9227c765a..84fee916d041 100644
--- a/gcc/config/rs6000/altivec.md
+++ b/gcc/config/rs6000/altivec.md
@@ -19,10 +19,11 @@
;; <http://www.gnu.org/licenses/>.
(define_c_enum "unspec"
[UNSPEC_VCMPBFP
UNSPEC_VMSUMU
+ UNSPEC_VMSUMUDM
UNSPEC_VMSUMM
UNSPEC_VMSUMSHM
UNSPEC_VMSUMUHS
UNSPEC_VMSUMSHS
UNSPEC_VMHADDSHS
@@ -970,10 +971,20 @@
UNSPEC_VMSUMU))]
"TARGET_ALTIVEC"
"vmsumu<VI_char>m %0,%1,%2,%3"
[(set_attr "type" "veccomplex")])
+(define_insn "altivec_vmsumudm"
+ [(set (match_operand:V1TI 0 "register_operand" "=v")
+ (unspec:V1TI [(match_operand:V2DI 1 "register_operand" "v")
+ (match_operand:V2DI 2 "register_operand" "v")
+ (match_operand:V1TI 3 "register_operand" "v")]
+ UNSPEC_VMSUMUDM))]
+ "TARGET_P9_VECTOR"
+ "vmsumudm %0,%1,%2,%3"
+ [(set_attr "type" "veccomplex")])
+
(define_insn "altivec_vmsumm<VI_char>m"
[(set (match_operand:V4SI 0 "register_operand" "=v")
(unspec:V4SI [(match_operand:VIshort 1 "register_operand" "v")
(match_operand:VIshort 2 "register_operand" "v")
(match_operand:V4SI 3 "register_operand" "v")]
diff --git a/gcc/config/rs6000/rs6000-builtin.def
b/gcc/config/rs6000/rs6000-builtin.def
index 363656ec05cc..ee0d787cfa22 100644
--- a/gcc/config/rs6000/rs6000-builtin.def
+++ b/gcc/config/rs6000/rs6000-builtin.def
@@ -1140,10 +1140,11 @@ BU_ALTIVEC_3 (VMHADDSHS, "vmhaddshs", SAT,
altivec_vmhaddshs)
BU_ALTIVEC_3 (VMHRADDSHS, "vmhraddshs", SAT,
altivec_vmhraddshs)
BU_ALTIVEC_3 (VMLADDUHM, "vmladduhm", CONST, fmav8hi4)
BU_ALTIVEC_3 (VMSUMUBM, "vmsumubm", CONST,
altivec_vmsumubm)
BU_ALTIVEC_3 (VMSUMMBM, "vmsummbm", CONST,
altivec_vmsummbm)
BU_ALTIVEC_3 (VMSUMUHM, "vmsumuhm", CONST,
altivec_vmsumuhm)
+BU_ALTIVEC_3 (VMSUMUDM, "vmsumudm", CONST,
altivec_vmsumudm)
BU_ALTIVEC_3 (VMSUMSHM, "vmsumshm", CONST,
altivec_vmsumshm)
BU_ALTIVEC_3 (VMSUMUHS, "vmsumuhs", SAT,
altivec_vmsumuhs)
BU_ALTIVEC_3 (VMSUMSHS, "vmsumshs", SAT,
altivec_vmsumshs)
BU_ALTIVEC_3 (VNMSUBFP, "vnmsubfp", FP, nfmsv4sf4)
BU_ALTIVEC_3 (VPERM_1TI, "vperm_1ti", CONST,
altivec_vperm_v1ti)
@@ -1497,10 +1498,11 @@ BU_ALTIVEC_OVERLOAD_3 (SEL, "sel")
BU_ALTIVEC_OVERLOAD_3 (VMSUMMBM, "vmsummbm")
BU_ALTIVEC_OVERLOAD_3 (VMSUMSHM, "vmsumshm")
BU_ALTIVEC_OVERLOAD_3 (VMSUMSHS, "vmsumshs")
BU_ALTIVEC_OVERLOAD_3 (VMSUMUBM, "vmsumubm")
BU_ALTIVEC_OVERLOAD_3 (VMSUMUHM, "vmsumuhm")
+BU_ALTIVEC_OVERLOAD_3 (VMSUMUDM, "vmsumudm")
BU_ALTIVEC_OVERLOAD_3 (VMSUMUHS, "vmsumuhs")
/* Altivec DST overloaded builtins. */
BU_ALTIVEC_OVERLOAD_D (DST, "dst")
BU_ALTIVEC_OVERLOAD_D (DSTT, "dstt")
diff --git a/gcc/config/rs6000/rs6000-call.c b/gcc/config/rs6000/rs6000-call.c
index d3cf2de88780..8e7bb54c73d1 100644
--- a/gcc/config/rs6000/rs6000-call.c
+++ b/gcc/config/rs6000/rs6000-call.c
@@ -3087,10 +3087,16 @@ const struct altivec_builtin_types
altivec_overloaded_builtins[] = {
RS6000_BTI_V4SI, RS6000_BTI_V16QI, RS6000_BTI_unsigned_V16QI,
RS6000_BTI_V4SI },
{ ALTIVEC_BUILTIN_VEC_MSUM, ALTIVEC_BUILTIN_VMSUMUHM,
RS6000_BTI_unsigned_V4SI, RS6000_BTI_unsigned_V8HI,
RS6000_BTI_unsigned_V8HI, RS6000_BTI_unsigned_V4SI },
{ ALTIVEC_BUILTIN_VEC_MSUM, ALTIVEC_BUILTIN_VMSUMSHM,
RS6000_BTI_V4SI, RS6000_BTI_V8HI, RS6000_BTI_V8HI, RS6000_BTI_V4SI },
+
+ { ALTIVEC_BUILTIN_VEC_MSUM, ALTIVEC_BUILTIN_VMSUMUDM,
+ RS6000_BTI_V1TI, RS6000_BTI_V2DI, RS6000_BTI_V2DI, RS6000_BTI_V1TI },
+ { ALTIVEC_BUILTIN_VEC_MSUM, ALTIVEC_BUILTIN_VMSUMUDM,
+ RS6000_BTI_unsigned_V1TI, RS6000_BTI_unsigned_V2DI,
RS6000_BTI_unsigned_V2DI, RS6000_BTI_unsigned_V1TI },
+
{ ALTIVEC_BUILTIN_VEC_VMSUMSHM, ALTIVEC_BUILTIN_VMSUMSHM,
RS6000_BTI_V4SI, RS6000_BTI_V8HI, RS6000_BTI_V8HI, RS6000_BTI_V4SI },
{ ALTIVEC_BUILTIN_VEC_VMSUMUHM, ALTIVEC_BUILTIN_VMSUMUHM,
RS6000_BTI_unsigned_V4SI, RS6000_BTI_unsigned_V8HI,
RS6000_BTI_unsigned_V8HI, RS6000_BTI_unsigned_V4SI },
{ ALTIVEC_BUILTIN_VEC_VMSUMMBM, ALTIVEC_BUILTIN_VMSUMMBM,
diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
index 5a57c1c49c56..8f75f60e7bc0 100644
--- a/gcc/doc/extend.texi
+++ b/gcc/doc/extend.texi
@@ -20211,10 +20211,17 @@ bool scalar_test_data_class (double source, const int
condition);
bool scalar_test_data_class (__ieee128 source, const int condition);
bool scalar_test_neg (float source);
bool scalar_test_neg (double source);
bool scalar_test_neg (__ieee128 source);
+
+vector _uint128_t vec_msum (vector unsigned long long,
+ vector unsigned long long,
+ vector _uint128_t);
+vector _int128_t vec_msum (vector signed long long,
+ vector signed long long,
+ vector _int128_t);
@end smallexample
The @code{scalar_extract_exp} and @code{scalar_extract_sig}
functions require a 64-bit environment supporting ISA 3.0 or later.
The @code{scalar_extract_exp} and @code{scalar_extract_sig} built-in
@@ -20230,10 +20237,15 @@ When supplied with a 128-bit @code{source} argument,
the
treated similarly.
Note that the sign of the significand is not represented in the result
returned from the @code{scalar_extract_sig} function. Use the
@code{scalar_test_neg} function to test the sign of its @code{double}
argument.
+The @code{vec_msum} functions perform a vector multiply-sum, returning
+the result of the widening multiply (arg1*arg2) which is then horizontally
+added to (arg3). ISA 3.0 adds support for the @code{vec_msum} builtin
+taking long long args for the multiply, an int128 args for the addition,
+and returning a vector int128 result.
The @code{scalar_insert_exp}
functions require a 64-bit environment supporting ISA 3.0 or later.
When supplied with a 64-bit first argument, the
@code{scalar_insert_exp} built-in function returns a double-precision
diff --git a/gcc/testsuite/gcc.target/powerpc/builtins-msum-runnable.c
b/gcc/testsuite/gcc.target/powerpc/builtins-msum-runnable.c
new file mode 100644
index 000000000000..0fa5c319b6de
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/builtins-msum-runnable.c
@@ -0,0 +1,74 @@
+/* { dg-do run { target { p9vector_hw } } } */
+/* { dg-options "-mdejagnu-cpu=power9 -O2" } */
+
+#include <altivec.h>
+
+#ifdef DEBUG
+#include <stdio.h>
+#endif
+
+void abort (void);
+
+int
+main()
+{
+ vector __uint128_t arg_uint128, result_uint128, expected_uint128;
+ vector __int128_t arg_int128, result_int128, expected_int128;
+
+ arg_uint128[0] = 0x1627384950617243;
+ arg_uint128[0] = arg_uint128[0] << 64;
+ arg_uint128[0] |= 0x9405182930415263;
+ expected_uint128[0] = 0x1627384950617243;
+ expected_uint128[0] = expected_uint128[0] << 64;
+ expected_uint128[0] |= 0xb6b07e42a570e5fe;
+ vector unsigned long long arg_vull2 = {0x12345678,0x44445555};
+ vector unsigned long long arg_vull3 = {0x6789abcd,0x66667777};
+ result_uint128 = vec_msum (arg_vull2, arg_vull3, arg_uint128);
+
+ if (result_uint128[0] != expected_uint128[0])
+ {
+#ifdef DEBUG
+ printf("result_uint128[0] doesn't match expected_u128[0]\n");
+ printf("arg_vull2 %llx %llx \n", arg_vull2[0], arg_vull2[1]);
+ printf("arg_vull3 %llx %llx \n", arg_vull3[0], arg_vull3[1]);
+ printf("arg_uint128[0] = %llx ", arg_uint128[0] >> 64);
+ printf(" %llx\n", arg_uint128[0] & 0xFFFFFFFFFFFFFFFF);
+
+ printf("result_uint128[0] = %llx ", result_uint128[0] >> 64);
+ printf(" %llx\n", result_uint128[0] & 0xFFFFFFFFFFFFFFFF);
+
+ printf("expected_uint128[0] = %llx ", expected_uint128[0] >> 64);
+ printf(" %llx\n", expected_uint128[0] & 0xFFFFFFFFFFFFFFFF);
+#else
+ abort();
+#endif
+ }
+
+ arg_int128[0] = 0x1627384950617283;
+ arg_int128[0] = arg_int128[0] << 64;
+ arg_int128[0] |= 0x9405182930415263;
+ expected_int128[0] = 0x1627384950617283;
+ expected_int128[0] = expected_int128[0] << 64;
+ expected_int128[0] |= 0xd99f35969c11cbfa;
+ vector signed long long arg_vll2 = { 0x567890ab, 0x1233456 };
+ vector signed long long arg_vll3 = { 0xcdef0123, 0x9873451 };
+ result_int128 = vec_msum (arg_vll2, arg_vll3, arg_int128);
+
+ if (result_int128[0] != expected_int128[0])
+ {
+#ifdef DEBUG
+ printf("result_int128[0] doesn't match expected128[0]\n");
+ printf("arg_int128[0] = %llx ", arg_int128[0] >> 64);
+ printf(" %llx\n", arg_int128[0] & 0xFFFFFFFFFFFFFFFF);
+
+ printf("result_int128[0] = %llx ", result_int128[0] >> 64);
+ printf(" %llx\n", result_int128[0] & 0xFFFFFFFFFFFFFFFF);
+
+ printf("expected_int128[0] = %llx ", expected_int128[0] >> 64);
+ printf(" %llx\n", expected_int128[0] & 0xFFFFFFFFFFFFFFFF);
+#else
+ abort();
+#endif
+ }
+}
+
diff --git a/gcc/testsuite/gcc.target/powerpc/vsx-builtin-msum.c
b/gcc/testsuite/gcc.target/powerpc/vsx-builtin-msum.c
new file mode 100644
index 000000000000..1974864de00c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vsx-builtin-msum.c
@@ -0,0 +1,25 @@
+/* Verify that overloaded built-ins for vec_msum with __int128
+ inputs generate the proper code. */
+
+/* { dg-do compile } */
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-mdejagnu-cpu=power9 -O3" } */
+
+#include <altivec.h>
+
+vector signed __int128
+test_msum_si (vector signed long long vsll_1, vector signed long long vsll_2,
+ vector signed __int128 vsi128)
+{
+ return vec_msum (vsll_1, vsll_2, vsi128);
+}
+
+vector unsigned __int128
+test_msum_ui (vector unsigned long long vull_1, vector unsigned long long
vull_2,
+ vector unsigned __int128 vui128)
+{
+ return vec_msum (vull_1, vull_2, vui128);
+}
+
+/* { dg_final { scan_assembler_times "vmsumudm" 2 } } */
+
