Re: Modes on CONST_INTs

[Geoff Keating <geoffk at geoffk dot org>]

Bernd Schmidt <bernds@redhat.com> writes:

> It seems that every now and then, we run into bugs involving a lack of mode
> information on CONST_INTs.  So far everyone I've asked has agreed that we
> really should have a mode for constants, but no one ever bothered because the
> change is going to be huge.
> 
> Even if the change is going to be huge, I don't see how it can be any more
> painful than the recurring problems we have because of the lack of mode
> information.  So, I'm looking for viable ways to gradually transform the
> compiler towards using modes on constants.
> 
> Here's what I have in mind so far:
> 1. Create a new macro GENMINT (MODE, INTVAL).  This macro will at first
>    ignore the MODE argument.
>    Replace all uses of GEN_INT with GENMINT.  Once done, remove the old
>    GEN_INT and rename GENMINT to GEN_INT.  (I'm open to suggestions for
>    a better name than GENMINT, but it should have 7 characters to avoid
>    reformatting madness).

This sounds like a good idea.  We too often end up in places in the
compiler where we know we need to do PLUS on (const_int A) and
(const_int B) but have somehow lost the mode we need to do it in.

Don't forget you also need to put modes on CONST_DOUBLEs.  

I believe that most uses of GEN_INT in the middle-end are actually
wrong, and should be using immed_double_const anyway.

> 2. In a similar way, replace uses of const0_rtx etc. with CONST0_RTX(MODE).
>    We'll also need a few new macros, e.g. CONSTM1_RTX(MODE) etc.
> 
> Then we'll probably have to fix some code that recognizes constants by
> the fact that their mode is VOIDmode and other problems likely to show up.

There's much code that distinguishes an integer CONST_DOUBLE from a FP
CONST_DOUBLE by looking to see whether it's VOIDmode; each instance is
easily fixed, but there are lots of them.

> I haven't quite figured out yet what mode should be used for variable shift
> counts.

Presumably the same mode that would be used if the shift count was a
register.

-- 
- Geoffrey Keating <geoffk@geoffk.org>