[Bug rtl-optimization/94873] [8/9/10/11 Regression] wrong code with -O -fno-merge-constants -fno-split-wide-types -fno-tree-fre

["rsandifo at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94873

--- Comment #8 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
(In reply to Segher Boessenkool from comment #7)
> REG_EQ* is documented as only being allowed on insns that set only one
> register.  If you want to change that, you'll have to check *all* code
> that consumes this, see if they rely on that fact or not, and if so,
> change that.

But the point is that the word "set" is ambiguous.  Does it mean
set by a SET or set by any means?  I think it can be read both ways.
After all, a CLOBBER is a form of set too, but that's clearly meant
to be excluded.  

Either way will need auditing.  If we say that this kind of REG_EQUAL
is wrong, we'd in theory need to audit everything that adds REG_EQUAL
notes to make sure it has an appropriate check, or doesn't need one.

I'm also not sure if we really are concerned about multiple registers
in this particular case, or whether it's more the case that we don't
want REG_EQUAL notes on insns with side effects.