Hello,
On Tue, 30 Jun 2026, Richard Biener wrote:
> > The call chain for our testcase is the following:
> > gen_lowpart -> gen_lowpart_general -> gen_lowpart_common -> lowpart_subreg
> > -> simplify_gen_subreg -> validate_subreg.
> >
> > `validate_subreg` for OI and V4x4BF modes fails, leading to this fallback
> > in `gen_lowpart_general`:
> >
> > /* Handle SUBREGs and hard REGs that were rejected by
> > simplify_gen_subreg. */
> > else if (REG_P (x) || GET_CODE (x) == SUBREG)
> > {
> > result = gen_lowpart_common (mode, copy_to_reg (x));
> > gcc_assert (result != 0);
> > return result;
> > }
> >
> > `copy_to_reg (x)` creates the new reg.
>
> And what's 'x' here? That said, either gen_lowpart () isn't supposed to
> be used in lvalue context
This. It cannot just generally be used on the LHS, for exactly the reason
we see here: it may generate a new reg (on the unchecked but commented
assumption that the input was a hardreg that was invalid for the wanted
mode), which would require copy-back. FWIW, this was the original comment
for gen_lowpart_general() before 2004:
/* Assuming that X is an rtx (e.g., MEM, REG or SUBREG) for a value,
return an rtx (MEM, SUBREG, or CONST_INT) that refers to the
least-significant part of X. ... */
N.B: "value". Not LHS. That comment was unfortunately lost. All uses of
gen_lowpart (which translates to calling gen_lowpart_general) on the LHS
are dubious at best, but also pre-existing and some of them are _very_
old. That this was harmless indicates that most of the time, in the
circumstances that they are called (here for instance, with target mode
being an int_mode_for_mode), they aren't going the copy_to_reg path, and
hence return a proper normal subreg of the input rtx.
So, the question hence is: _why_ does validate_subreg fail on that mode
combination OI/V4x4BF, if this is supposed to be the int_mode_for_mode on
that target? I guess that should just be fixed.
The more correct change would be to generally just use
gen_lowpart_if_possible on LHSs, but then the callers need to be able to
deal with the fact that it may return NULL. But ... that will be larger
:)
Ciao,
Michael.
