This patch adds the following intrinsics (all __arm_streaming only) along with
asm tests for them.
BFCVT, FCVT Convert to packed 8-bit floating-point format:
- svmfloat8_t svcvt_mf8[_f16_x2]_fpm(svfloat16x2_t zn, fpm_t fpm)
- svmfloat8_t svcvt_mf8[_bf16_x2]_fpm(svbfloat16x2_t zn, fpm_t fpm)
- svmfloat8_t svcvt_mf8[_f32_x4]_fpm(svfloat32x4_t zn, fpm_t fpm)
FCVTN Convert to interleaved 8-bit floating-point format.
- svmfloat8_t svcvtn_mf8[_f32_x4]_fpm(svfloat32x4_t zn, fpm_t fpm)
gcc/
* config/aarch64/aarch64-sve-builtins-base.cc (svcvt_impl): Update to
handle fp8 cases.
* config/aarch64/aarch64-sve-builtins-sve2.def (svcvt, svcvtn): Added
DEF_SVE_FUNCTION_GS_FPM instances.
* config/aarch64/aarch64-sve2.md
(@aarch64_sve2_fp8_cvtn<mode>): Updated define_insn for additional case.
(@aarch64_sme2_fp8_cvt<mode>): Added new define_insn.
* config/aarch64/iterators.md (VNx16F_NARROW): Added new iterator to
handle narrowing SVE floating point operations.
(UNSPEC_FCVT): Added new unspec.
gcc/testsuite/
* gcc.target/aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c: Added test file.
* gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c: Likewise.
* gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c: Likewise.
* gcc.target/aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c: Likewise.
* gcc.target/aarch64/sve/acle/asm/test_sve_acle.h
(TEST_X2_NARROW): Added fpm0 argument for intrinsics.
(TEST_X4_NARROW): Likewise.
---
.../aarch64/aarch64-sve-builtins-base.cc | 26 ++++---
.../aarch64/aarch64-sve-builtins-sve2.def | 3 +
gcc/config/aarch64/aarch64-sve2.md | 14 +++-
gcc/config/aarch64/iterators.md | 4 ++
.../aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c | 56 +++++++++++++++
.../aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c | 56 +++++++++++++++
.../aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c | 72 +++++++++++++++++++
.../aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c | 72 +++++++++++++++++++
.../aarch64/sve/acle/asm/test_sve_acle.h | 2 +
9 files changed, 293 insertions(+), 12 deletions(-)
create mode 100644
gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c
create mode 100644
gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c
create mode 100644
gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c
create mode 100644
gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c
diff --git a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
index ecc06877cac..622485effb3 100644
--- a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
+++ b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
@@ -779,17 +779,23 @@ public:
{
machine_mode mode0 = e.result_mode ();
machine_mode mode1 = GET_MODE (e.args[0]);
- convert_optab optab;
- if (e.type_suffix (0).integer_p)
- optab = e.type_suffix (0).unsigned_p ? ufix_optab : sfix_optab;
- else if (e.type_suffix (1).integer_p)
- optab = e.type_suffix (1).unsigned_p ? ufloat_optab : sfloat_optab;
- else if (e.type_suffix (0).element_bits
- < e.type_suffix (1).element_bits)
- optab = trunc_optab;
+ if (e.fpm_mode == aarch64_sve::FPM_set)
+ icode = code_for_aarch64_sme2_fp8_cvt (mode1);
else
- optab = sext_optab;
- icode = convert_optab_handler (optab, mode0, mode1);
+ {
+ convert_optab optab;
+ if (e.type_suffix (0).integer_p)
+ optab = e.type_suffix (0).unsigned_p ? ufix_optab : sfix_optab;
+ else if (e.type_suffix (1).integer_p)
+ optab = e.type_suffix (1).unsigned_p ? ufloat_optab
+ : sfloat_optab;
+ else if (e.type_suffix (0).element_bits
+ < e.type_suffix (1).element_bits)
+ optab = trunc_optab;
+ else
+ optab = sext_optab;
+ icode = convert_optab_handler (optab, mode0, mode1);
+ }
gcc_assert (icode != CODE_FOR_nothing);
return e.use_exact_insn (icode);
}
diff --git a/gcc/config/aarch64/aarch64-sve-builtins-sve2.def
b/gcc/config/aarch64/aarch64-sve-builtins-sve2.def
index c271b97de87..869e006ffde 100644
--- a/gcc/config/aarch64/aarch64-sve-builtins-sve2.def
+++ b/gcc/config/aarch64/aarch64-sve-builtins-sve2.def
@@ -421,6 +421,9 @@ DEF_SVE_FUNCTION_GS_FPM (svdot_lane,
ternary_mfloat8_lane_group_selection, h_flo
#define REQUIRED_EXTENSIONS \
streaming_only (AARCH64_FL_SME2 | AARCH64_FL_FP8)
+DEF_SVE_FUNCTION_GS_FPM (svcvt, unary_convertxn_narrow, cvtn_mf8, x2, none,
set)
+DEF_SVE_FUNCTION_GS_FPM (svcvt, unary_convertxn_narrow, cvtnx_mf8, x4, none,
set)
+DEF_SVE_FUNCTION_GS_FPM (svcvtn, unary_convertxn_narrow, cvtnx_mf8, x4, none,
set)
DEF_SVE_FUNCTION_GS_FPM (svcvt1, unary_convert, cvt_mf8, x2, none, set)
DEF_SVE_FUNCTION_GS_FPM (svcvt2, unary_convert, cvt_mf8, x2, none, set)
DEF_SVE_FUNCTION_GS_FPM (svcvtl1, unary_convert, cvt_mf8, x2, none, set)
diff --git a/gcc/config/aarch64/aarch64-sve2.md
b/gcc/config/aarch64/aarch64-sve2.md
index ab8098d3327..da7a7a3c23c 100644
--- a/gcc/config/aarch64/aarch64-sve2.md
+++ b/gcc/config/aarch64/aarch64-sve2.md
@@ -3635,10 +3635,10 @@ (define_insn "@aarch64_sve_cvtn<mode>"
(define_insn "@aarch64_sve2_fp8_cvtn<mode>"
[(set (match_operand:VNx16QI 0 "register_operand" "=w")
(unspec:VNx16QI
- [(match_operand:SVE_FULL_HFx2 1 "aligned_register_operand" "Uw2")
+ [(match_operand:VNx16F_NARROW 1 "aligned_register_operand"
"Uw<vector_count>")
(reg:DI FPM_REGNUM)]
UNSPEC_FP8FCVTN))]
- "TARGET_SSVE_FP8"
+ "<MODE>mode == VNx16SFmode ? TARGET_SSME2_FP8 : TARGET_STREAMING_SME2"
"<b>fcvtn\t%0.b, %1"
[(set_attr "sve_type" "sve_fp_cvt")]
)
@@ -3666,6 +3666,16 @@ (define_insn "@aarch64_sve_cvtnt<mode>"
[(set_attr "sve_type" "sve_fp_cvt")]
)
+(define_insn "@aarch64_sme2_fp8_cvt<mode>"
+ [(set (match_operand:VNx16QI 0 "register_operand" "=w")
+ (unspec:VNx16QI
+ [(match_operand:VNx16F_NARROW 1 "aligned_register_operand"
"Uw<vector_count>")
+ (reg:DI FPM_REGNUM)]
+ UNSPEC_FCVT))]
+ "TARGET_SSME2_FP8"
+ "<b>fcvt\t%0.b, %1"
+)
+
;; -------------------------------------------------------------------------
;; ---- [FP<-INT] Multi-vector conversions
;; -------------------------------------------------------------------------
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 026c3101e38..e6f59d22d0c 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -712,6 +712,9 @@ (define_mode_iterator VNx4_WIDE [VNx4SI])
(define_mode_iterator VNx2_NARROW [VNx2QI VNx2HI VNx2SI])
(define_mode_iterator VNx2_WIDE [VNx2DI])
+;; Used for narrowing SVE floating point operations.
+(define_mode_iterator VNx16F_NARROW [VNx16BF VNx16HF VNx16SF])
+
;; All SVE predicate modes.
(define_mode_iterator PRED_ALL [VNx16BI VNx8BI VNx4BI VNx2BI])
@@ -1061,6 +1064,7 @@ (define_c_enum "unspec"
UNSPEC_F2CVTL ; Used in aarch64-sve2.md.
UNSPEC_F2CVTLT ; Used in aarch64-sve2.md.
UNSPEC_FADDP ; Used in aarch64-sve2.md.
+ UNSPEC_FCVT ; Used in aarch64-sve2.md.
UNSPEC_FCVTNB ; Used in aarch64-sve2.md.
UNSPEC_FCVTNT ; Used in aarch64-sve2.md.
UNSPEC_FMAXNMP ; Used in aarch64-sve2.md.
diff --git a/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c
b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c
new file mode 100644
index 00000000000..bdda0fd36d6
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_bf16_x2.c
@@ -0,0 +1,56 @@
+/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
+
+#pragma GCC target "+fp8+bf16"
+#include "test_sme2_acle.h"
+
+/*
+** cvt_z0_z0:
+** msr fpmr, x0
+** bfcvt z0\.b, {z0\.h - z1\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z0, svbfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_bf16_x2_fpm (z0, fpm0),
+ z0_res = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z0_z6:
+** msr fpmr, x0
+** bfcvt z0\.b, {z6\.h - z7\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z6, svbfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_bf16_x2_fpm (z6, fpm0),
+ z0_res = svcvt_mf8_fpm (z6, fpm0))
+
+/*
+** cvt_z0_z29:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** bfcvt z0\.b, [^\n]+
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z29, svbfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_bf16_x2_fpm (z29, fpm0),
+ z0_res = svcvt_mf8_fpm (z29, fpm0))
+
+/*
+** cvt_z5_z0:
+** msr fpmr, x0
+** bfcvt z5\.b, {z0\.h - z1\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z5_z0, svbfloat16x2_t, svmfloat8_t,
+ z5 = svcvt_mf8_bf16_x2_fpm (z0, fpm0),
+ z5 = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z22_z16:
+** msr fpmr, x0
+** bfcvt z22\.b, {z16\.h - z17\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z22_z16, svbfloat16x2_t, svmfloat8_t,
+ z22 = svcvt_mf8_bf16_x2_fpm (z16, fpm0),
+ z22 = svcvt_mf8_fpm (z16, fpm0))
diff --git a/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c
b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c
new file mode 100644
index 00000000000..93792e90945
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f16_x2.c
@@ -0,0 +1,56 @@
+/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
+
+#include "test_sme2_acle.h"
+#pragma GCC target "+fp8"
+
+/*
+** cvt_z0_z0:
+** msr fpmr, x0
+** fcvt z0\.b, {z0\.h - z1\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z0, svfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f16_x2_fpm (z0, fpm0),
+ z0_res = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z0_z6:
+** msr fpmr, x0
+** fcvt z0\.b, {z6\.h - z7\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z6, svfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f16_x2_fpm (z6, fpm0),
+ z0_res = svcvt_mf8_fpm (z6, fpm0))
+
+/*
+** cvt_z0_z29:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** fcvt z0\.b, {z0\.h - z1\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z0_z29, svfloat16x2_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f16_x2_fpm (z29, fpm0),
+ z0_res = svcvt_mf8_fpm (z29, fpm0))
+
+/*
+** cvt_z5_z0:
+** msr fpmr, x0
+** fcvt z5\.b, {z0\.h - z1\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z5_z0, svfloat16x2_t, svmfloat8_t,
+ z5 = svcvt_mf8_f16_x2_fpm (z0, fpm0),
+ z5 = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z22_z16:
+** msr fpmr, x0
+** fcvt z22\.b, {z16\.h - z17\.h}
+** ret
+*/
+TEST_X2_NARROW (cvt_z22_z16, svfloat16x2_t, svmfloat8_t,
+ z22 = svcvt_mf8_f16_x2_fpm (z16, fpm0),
+ z22 = svcvt_mf8_fpm (z16, fpm0))
diff --git a/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c
b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c
new file mode 100644
index 00000000000..a9ee10de0df
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvt_mf8_f32_x4.c
@@ -0,0 +1,72 @@
+/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
+
+#include "test_sme2_acle.h"
+#pragma GCC target "+fp8"
+
+/*
+** cvt_z0_z0:
+** msr fpmr, x0
+** fcvt z0\.b, {z0\.s - z3\.s}
+** ret
+*/
+TEST_X4_NARROW (cvt_z0_z0, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f32_x4_fpm (z0, fpm0),
+ z0_res = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z0_z4:
+** msr fpmr, x0
+** fcvt z0\.b, {z4\.s - z7\.s}
+** ret
+*/
+TEST_X4_NARROW (cvt_z0_z4, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f32_x4_fpm (z4, fpm0),
+ z0_res = svcvt_mf8_fpm (z4, fpm0))
+
+/*
+** cvt_z0_z21:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** fcvt z0\.b, [^\n]+
+** ret
+*/
+TEST_X4_NARROW (cvt_z0_z21, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvt_mf8_f32_x4_fpm (z21, fpm0),
+ z0_res = svcvt_mf8_fpm (z21, fpm0))
+
+/*
+** cvt_z25_z26:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** fcvt z25\.b, {z28\.s - z31\.s}
+** ret
+*/
+TEST_X4_NARROW (cvt_z25_z26, svfloat32x4_t, svmfloat8_t,
+ z25 = svcvt_mf8_f32_x4_fpm (z26, fpm0),
+ z25 = svcvt_mf8_fpm (z26, fpm0))
+
+/*
+** cvt_z25_z0:
+** msr fpmr, x0
+** fcvt z25\.b, {z0\.s - z3\.s}
+** ret
+*/
+TEST_X4_NARROW (cvt_z25_z0, svfloat32x4_t, svmfloat8_t,
+ z25 = svcvt_mf8_f32_x4_fpm (z0, fpm0),
+ z25 = svcvt_mf8_fpm (z0, fpm0))
+
+/*
+** cvt_z22_z16:
+** msr fpmr, x0
+** fcvt z22\.b, {z16\.s - z19\.s}
+** ret
+*/
+TEST_X4_NARROW (cvt_z22_z16, svfloat32x4_t, svmfloat8_t,
+ z22_res = svcvt_mf8_f32_x4_fpm (z16, fpm0),
+ z22_res = svcvt_mf8_fpm (z16, fpm0))
diff --git a/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c
b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c
new file mode 100644
index 00000000000..2d0bd7eda0e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sme2/acle-asm/cvtn_mf8_f32_x4.c
@@ -0,0 +1,72 @@
+/* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
+
+#include "test_sme2_acle.h"
+#pragma GCC target "+fp8"
+
+/*
+** cvtn_z0_z0:
+** msr fpmr, x0
+** fcvtn z0\.b, {z0\.s - z3\.s}
+** ret
+*/
+TEST_X4_NARROW (cvtn_z0_z0, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvtn_mf8_f32_x4_fpm (z0, fpm0),
+ z0_res = svcvtn_mf8_fpm (z0, fpm0))
+
+/*
+** cvtn_z0_z4:
+** msr fpmr, x0
+** fcvtn z0\.b, {z4\.s - z7\.s}
+** ret
+*/
+TEST_X4_NARROW (cvtn_z0_z4, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvtn_mf8_f32_x4_fpm (z4, fpm0),
+ z0_res = svcvtn_mf8_fpm (z4, fpm0))
+
+/*
+** cvtn_z0_z21:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** fcvtn z0\.b, [^\n]+
+** ret
+*/
+TEST_X4_NARROW (cvtn_z0_z21, svfloat32x4_t, svmfloat8_t,
+ z0_res = svcvtn_mf8_f32_x4_fpm (z21, fpm0),
+ z0_res = svcvtn_mf8_fpm (z21, fpm0))
+
+/*
+** cvtn_z25_z26:
+** msr fpmr, x0
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** mov [^\n]+
+** fcvtn z25\.b, {z28\.s - z31\.s}
+** ret
+*/
+TEST_X4_NARROW (cvtn_z25_z26, svfloat32x4_t, svmfloat8_t,
+ z25 = svcvtn_mf8_f32_x4_fpm (z26, fpm0),
+ z25 = svcvtn_mf8_fpm (z26, fpm0))
+
+/*
+** cvtn_z25_z0:
+** msr fpmr, x0
+** fcvtn z25\.b, {z0\.s - z3\.s}
+** ret
+*/
+TEST_X4_NARROW (cvtn_z25_z0, svfloat32x4_t, svmfloat8_t,
+ z25 = svcvtn_mf8_f32_x4_fpm (z0, fpm0),
+ z25 = svcvtn_mf8_fpm (z0, fpm0))
+
+/*
+** cvtn_z22_z16:
+** msr fpmr, x0
+** fcvtn z22\.b, {z16\.s - z19\.s}
+** ret
+*/
+TEST_X4_NARROW (cvtn_z22_z16, svfloat32x4_t, svmfloat8_t,
+ z22_res = svcvtn_mf8_f32_x4_fpm (z16, fpm0),
+ z22_res = svcvtn_mf8_fpm (z16, fpm0))
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/acle/asm/test_sve_acle.h
b/gcc/testsuite/gcc.target/aarch64/sve/acle/asm/test_sve_acle.h
index 7c156c4cf2a..8d4ed537c87 100644
--- a/gcc/testsuite/gcc.target/aarch64/sve/acle/asm/test_sve_acle.h
+++ b/gcc/testsuite/gcc.target/aarch64/sve/acle/asm/test_sve_acle.h
@@ -733,6 +733,7 @@
#define TEST_X2_NARROW(NAME, TTYPE, ZTYPE, CODE1, CODE2) \
PROTO (NAME, void, ()) \
{ \
+ register fpm_t fpm0 __asm ("x0"); \
register TTYPE z0 __asm ("z0"); \
register ZTYPE z5 __asm ("z5"); \
register TTYPE z6 __asm ("z6"); \
@@ -749,6 +750,7 @@
#define TEST_X4_NARROW(NAME, TTYPE, ZTYPE, CODE1, CODE2) \
PROTO (NAME, void, ()) \
{ \
+ register fpm_t fpm0 __asm ("x0"); \
register TTYPE z0 __asm ("z0"); \
register TTYPE z4 __asm ("z4"); \
register TTYPE z16 __asm ("z16"); \
--
2.51.0
