From: Lino Hsing-Yu Peng <[email protected]>
Tag Zvfofp8min conversion patterns with the new altfmt attribute so the
backend records which instructions require alt floating-point formats.
This enables later vsetvl decisions to preserve compatible settings.
gcc/ChangeLog:
* config/riscv/vector-float8.md (altfmt_kind): New attribute.
(@pred_extend_<altfmt_base>_to_<mode>): Set altfmt.
(@pred_trunc_<mode>_to_<altfmt_sat>,
@pred_quad_trunc_<mode>_to_<altfmt_sat>): Likewise.
---
gcc/config/riscv/vector-float8.md | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/gcc/config/riscv/vector-float8.md
b/gcc/config/riscv/vector-float8.md
index 20c656a948b..36806c4a9ed 100644
--- a/gcc/config/riscv/vector-float8.md
+++ b/gcc/config/riscv/vector-float8.md
@@ -64,6 +64,12 @@
(UNSPEC_F8E4M3_SAT "f8e4m3_sat")
(UNSPEC_F8E5M2_SAT "f8e5m2_sat")])
+(define_int_attr altfmt_kind
+ [(UNSPEC_F8E4M3 "none")
+ (UNSPEC_F8E5M2 "alt")
+ (UNSPEC_F8E4M3_SAT "none")
+ (UNSPEC_F8E5M2_SAT "alt")])
+
(define_int_attr sat
[(UNSPEC_F8E4M3 "")
(UNSPEC_F8E5M2 "")
@@ -90,7 +96,8 @@
"TARGET_VECTOR && TARGET_ZVFOFP8MIN && TARGET_ZVFBFMIN"
"vfwcvtbf16.f.f.v\t%0,%3%p1"
[(set_attr "type" "vfwcvtbf16")
- (set_attr "mode" "<VBF_DOUBLE_TRUNC>")])
+ (set_attr "mode" "<VBF_DOUBLE_TRUNC>")
+ (set_attr "altfmt" "<altfmt_kind>")])
;; Zvfofp8min extension: BF16 to FP8 narrowing conversions.
@@ -115,6 +122,7 @@
"vfncvtbf16<sat>.f.f.w\t%0,%3%p1"
[(set_attr "type" "vfncvtbf16")
(set_attr "mode" "<VBF_DOUBLE_TRUNC>")
+ (set_attr "altfmt" "<altfmt_kind>")
(set (attr "frm_mode")
(symbol_ref "riscv_vector::get_frm_mode (operands[8])"))])
@@ -141,5 +149,6 @@
"vfncvt<sat>.f.f.q\t%0,%3%p1"
[(set_attr "type" "vfncvtbf16")
(set_attr "mode" "<VF_QUAD_TRUNC>")
+ (set_attr "altfmt" "<altfmt_kind>")
(set (attr "frm_mode")
(symbol_ref "riscv_vector::get_frm_mode (operands[8])"))])
--
2.34.1
