https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108764
Sinan <sinan.lin at linux dot alibaba.com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target| |riscv
--- Comment #1 from Sinan <sinan.lin at linux dot alibaba.com> ---
In the given test case, rv64gc generates better code than rv64gc_zba since asm
is less and simpler. In this case, we have an instruction chain like,
ashldi3: reg80=reg79 << 3;
adddi3: reg81=reg80+reg78;
movdi: reg82=reg80;
where ashldi3 has two uses, then combining ashldi3+adddi3 isn't really free.
I think one solution is to change the cost model of such complex instructions
to the sum of the cost for each part. E.g.
cost for shNadd = COSTS_N_INSNS (SINGLE_SHIFT_COST) + COSTS_N_INSNS (1) # cost
of addition
