Re: [RFC] trailing_wide_ints with runtime variable lengths

[Aldy Hernandez <aldyh@redhat.com>]

BTW, I don't know if it got lost in all my patches, but we already
have an irange allocator that given an irange, returns a chunk of
memory holding a clone of that irange squished into its minimum
representable pairs (see vrange_allocator and friends).  So we won't
ever be storing 255 or something equally absurd like I had proposed
years ago :).  We'll be storing the smallest representable range
inside a trailing_wide_int.

Aldy

On Fri, Jul 1, 2022 at 10:31 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>
> On Fri, Jul 1, 2022 at 8:53 PM Jakub Jelinek <jakub@redhat.com> wrote:
> >
> > On Fri, Jul 01, 2022 at 07:43:28PM +0200, Aldy Hernandez wrote:
> > > You can still say N=255 and things continue to work as they do now, since
> > > m_len[] is still statically determined. The only difference is that before,
> > > the size of the structure would be 2+1+255+sizeof(int) whereas now it would
> > > be 1 more because of the byte I'm using for num_elements.
> >
> > So, what N do you want to use for SSA_NAME_RANGE_INFO?
> > N=255 wouldn't be very space efficient especially if the common case is a
> > single range or two.
> > For such cases making e.g. m_len not an embedded array, but pointer to
> > somewhere after the HOST_WIDE_INT array in the same allocation would be
> > better.
>
> As I mentioned in my original post, 12.  This means that I'm taking
> the 4 bytes that are left over from the current padding plus 8
> (64-bits).  My trailing wide int structure for SSA_NAME_RANGE_INFO
> will be one word larger than what is currently there.  But we'll be
> able to store up to 5 pairs plus one for the nonzero bits plus one for
> future development (5*2 + 1 + 1 = 12), all without going over the 64
> bit alignment.
>
> This is a theoretical max, in reality as I mentioned, 99% of ranges
> calculated in infinite precision by the ranger fit into 3-4 pairs.
>
> Aldy