Re: [PATCH] loading float member of parameter stored via int registers

[Segher Boessenkool <segher@kernel.crashing.org>]

On Fri, Dec 23, 2022 at 08:13:48PM +0100, Richard Biener wrote:
> > Am 23.12.2022 um 17:55 schrieb Segher Boessenkool <segher@kernel.crashing.org>:
> > There are at least six very different kinds of subreg:
> > 
> > 0) Lvalue subregs.  Most archs have no use for it, and it can be
> >   expressed much more clearly and cleanly always.
> > 1) Subregs of mem.  Do not use, deprecated.  When old reload goes away
> >   this will go away.
> > 2) Subregs of hard registers.  Do not use, there are much better ways to
> >   write subregs of a non-zero byte offset, and for zero offset this is
> >   non-canonical RTL.
> > 3) Bitcast subregs.  In principle they go from one mode to another mode
> >   of the same size (but read on).
> > 4) Paradoxical subregs.  A concept completely separate from the rest,
> >   different rules for everything, it has to be special cased almost
> >   everywhere, it would be better if it was a separate rtx_code imo.
> > 5) Finally, normal subregs, taking a contiguous span of bits from some
> >   value.
> > 
> > Now, it is invalid to have a subreg of a subreg, so a 3) of a 5) is
> > written as just one subreg, as you say.  And a 4) of a 5) is just
> > invalid afaics (and let's not talk about 0)..2) anymore :-) )
> > 
> >> Note whether targets actually support subreg operations needs to be queried and I’m not sure how subreg with offset validation should work there.
> > 
> > But 3) is always valid, no?  On pseudos.
> 
> Yes, but it will eventually result in a spill/reload which is undesirable when we created this from CSE from a load.  So I think for CSE we do want to know whether a spill will definitely not occur.

Does it cause reloads though?  On any sane backend?  If no movsf pattern
allows integer registers, can things work at all?

Anyway, the normal way to test if some RTL is valid is to just generate
it (using validate_change) and then do apply_change_group, which then
cancels the changes if they do not work.  CSE already does some of this.

(I am doubtful doing any of this in CSE is a good idea fwiw).


Segher