https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118984
--- Comment #16 from Maxim Egorushkin <maxim.yegorushkin at gmail dot com> ---
(In reply to Maxim Egorushkin from comment #14)
> (In reply to Andrew Pinski from comment #6)
>
> > It happens more often with vector instructions/registers due to the
> > different "modes" of the registers that it can hold (subregs).
>
> ... Or, worse, allocating one constant variable to multiple registers, which
> gets never modified, for no apparent reason.
That happens most frequently with xmm registers storing one zero double value
versus a register storing two double zero values.
An xmm register storing one 0 double value is zeroed with a vxorpd instruction,
setting all its elements to 0. Yet, that fully zeroed xmm register isn't reused
in places where an xmm register with 2 double elements set to 0 is required.
gcc actually spill a register to stack in order to create another xmm register
with all its bits set to zero, because that existing xmm register with all its
bits set to 0 and never modified after that, has, apparently, an incompatible
type, and hence, cannot be reused where a wider xmm/ymm register filled with 0s
is required.
Fixing that xmm register spill to stack involved explicitly passing a double{0}
value into "v" asm parameter, instead of passing zero literal into it.
I wonder why gcc emits unnecessary register move instructions to treat ymm/xmm
registers as shorter vectors, unlike for GP registers being treated as
64/32/16/8-bit values without any register moves, please?
