Re: [PATCH] c-family, middle-end: Add __builtin_c[lt]zg (arg, 0ULL) exception

[Richard Biener <rguenther@suse.de>]

On Mon, 20 Nov 2023, Jakub Jelinek wrote:

> On Mon, Nov 20, 2023 at 08:37:55AM +0000, Richard Biener wrote:
> > > I'm not sure about that, it would be nice for them to be usable there,
> > 
> > Btw, I think that {( .. )} should be made usable in sizeof () and
> > possibly even in at least C++ constant expressions (not sure about C).
> 
> I believe the problkem is having new VAR_DECLs in those which actually
> aren't file scope/namespace scope variables but there is no function
> DECL_CONTEXT to attach to them.  So, it probably wouldn't be one afternoon
> change to allow that.
> 
> > > but I think e.g. none of Joseph's implementation of those macros
> > > made them usable there (except inside of sizeof/typeof/typeof_unquall)
> > > and I don't see a requirement in the C23 standard that they must be usable
> > > in constant expressions.
> > > The versions I've posted on Thursday were usable there except for
> > > stdc_has_single_bit (but that actually can be implemented that way too)
> > > and stdc_bit_floor.  And the version I haven't posted that used the 3
> > > patches posted on Saturday would have all functions usable when the
> > > argument to those macros is a constant expression.
> > > 
> > > BTW, if we go route of implementing all of the stdc_ type-generic macros
> > > as builtins, we could as well not implement that way the following 4
> > > # define stdc_first_leading_one(x) (__builtin_clzg (x, -1) + 1U)
> > > # define stdc_first_trailing_one(x) (__builtin_ctzg (x, -1) + 1U)
> > > # define stdc_count_ones(x) ((unsigned int) __builtin_popcountg (x))
> > > # define stdc_has_single_bit(x) ((_Bool) (__builtin_popcountg (x) == 1))
> > > which are implementable without any new extensions.
> > 
> > I'd rather do all of those necessary as builtins instead of hacking
> > around limitations.  If we don't want to solve those limitations in
> > a more generic way.
> 
> Ok, I can prepare a patch for that, shouldn't be that hard.
> Do you want all 14, or just the 10 and leave the above 4 with the
> above definitions?

I'd say all of them for consistency, we can parse/gimplify them to
the open-coded variants then.

> > And of course nobody would write
> > 
> > const int x = sizeof (stdc_first_leading_one (5));
> > 
> > that's just stupid ... (but oh well).
> 
> Well, standard testsuite needs to include that at least.
> But of course, if it is usable in constant expressions,
> unsigned a = stdc_bit_width ((unsigned _BitInt(824)) 435987349856735489657489657468954768954674589674598uwb * 49876558967549867548967548967548967549867548967456uwb);
> etc. can be useful in constant expressions.
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)