https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113035
Jovan Vukic <jovan.vu...@rt-rk.com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jovan.vu...@rt-rk.com
--- Comment #5 from Jovan Vukic <jovan.vu...@rt-rk.com> ---
Since we have many zicond tests failing here for the sifive-7-series
(currently 275 FAILs), I analyzed the issue this week.
In short, in the function riscv_expand_conditional_move (from riscv.cc),
the SFB is always preferred when both TARGET_SFB_ALU and
TARGET_ZICOND_LIKE are set to 1.
During the meeting this week, it was clarified to me that we generally steer
towards zicond. A simple change in the riscv_expand_conditional_move function
could help us achieve this here too and eliminate 120 FAILs.
- if (TARGET_SFB_ALU || TARGET_XTHEADCONDMOV)
+ if ((TARGET_SFB_ALU || TARGET_XTHEADCONDMOV)
+ && (!TARGET_ZICOND_LIKE || (code != EQ && code != NE)))
However, it was also noted that zicond + SFB cases are not currently
that relevant. So, even though I initially thought about turning this into
a patch, I assume it's better to leave this here as an observation.
