https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82636
--- Comment #1 from Michael Meissner <meissner at gcc dot gnu.org> ---
This looks like a reload bug. I see the same thing with automatically
generated fmas:
--> cat foo06c.c
__ieee128
__fmaf128_power9 (__ieee128 x, __ieee128 y, __ieee128 z)
{
return (x * y) + z;
}
After the LRA pass, the rtl dump is:
(insn 13 8 14 2 (set (reg/i:KF 79 2)
(fma:KF (reg:KF 79 2 [ x ])
(reg:KF 80 3 [ y ])
(reg:KF 81 4 [ z ]))) "foo06c.c":5 873 {fmakf4_hw}
(expr_list:REG_DEAD (reg:KF 81 4 [ z ])
(expr_list:REG_DEAD (reg:KF 80 3 [ y ])
(nil))))
(insn 14 13 16 2 (use (reg/i:KF 79 2)) "foo06c.c":5 -1
(nil))
However, reload inserts moves:
(insn 18 8 13 2 (set (reg:KF 77 0 [orig:126 z ] [126])
(reg:KF 81 4 [ z ])) "foo06c.c":5 1009 {*vsx_movkf_64bit}
(nil))
(insn 13 18 19 2 (set (reg:KF 77 0 [orig:126 z ] [126])
(fma:KF (reg:KF 79 2 [ x ])
(reg:KF 80 3 [ y ])
(reg:KF 77 0 [orig:126 z ] [126]))) "foo06c.c":5 873 {fmakf4_hw}
(nil))
(insn 19 13 14 2 (set (reg/i:KF 79 2)
(reg:KF 77 0 [orig:126 z ] [126])) "foo06c.c":5 1009 {*vsx_movkf_64bit}
(nil))
(insn 14 19 16 2 (use (reg/i:KF 79 2)) "foo06c.c":5 -1
(nil))
