Re: Immediate doubles (up to 2^256) and rationals coming to Guile 3

[Mark H Weaver <mhw@netris.org>]

Hi Ludovic,

Ludovic Courtès <ludo@gnu.org> writes:
> Though an immediate, like a fixnum or an iflo, is still something
> different from a tagged heap object like a pair, right?  So I would
> expect SCM_THOB_P to be a different test, not a drop-in replacement for
> SCM_NIMP, is that correct?

That's right.  It's not possible to create a drop-in replacement for
SCM_NIMP, because it is being used to answer two different questions
which used to be effectively equivalent, but no longer are:

(1) Is X a pointer to a heap object with a heap tag in the first word?
(2) Is X a reference to a heap object?

Test (1) needs to be done before checking the heap tag, to implement
type predicates for heap objects.  Test (2) is needed in relatively few
places, e.g. to decide whether to register disappearing links when
adding an entry to a weak hash table.

Actually, in my current branch I've removed the SCM_IMP and SCM_NIMP
macros outright, because it seems to me they are likely to be misused.

SCM_THOB_P implements test (1) and SCM_HEAP_OBJECT_P implements test (2).

>> (2) Our existing VM instructions almost invariably specify offsets with
>>     a granularity of whole words.  To support tagged pair pointers with
>>     good performance, I think we need a few new instructions that
>>     specify byte offsets, to avoid the expensive extra step of removing
>>     the tag before accessing the CAR or CDR of a pair.
>
> So instead of a pointer dereference, SCM_CAR becomes mask + dereference,
> right?

There's no masking involved.  Rather, it is subtracted from the pointer,
which allows the tag to be fused with the field offset.  For example, on
x86-64, whereas CAR and CDR were previously:

     1c0:   48 8b 07                mov    (%rdi),%rax      ;old car
and:
     1d0:   48 8b 47 08             mov    0x8(%rdi),%rax   ;old cdr

Now they become:

     1e0:   48 8b 47 fa             mov    -0x6(%rdi),%rax  ;new car
and:
     1f0:   48 8b 47 02             mov    0x2(%rdi),%rax   ;new cdr

In the VM, I've added four new instructions:

  make-tagged-non-immediate dst:12 tag:12 offset:32
  tagged-allocate-words/immediate dst:8 count:8 tag:8
  tagged-scm-ref/immediate dst:8 obj:8 byte-offset:8
  tagged-scm-set!/immediate obj:8 byte-offset:8 val:8

The last two instructions above are like 'scm-ref/immediate' and
'scm-set!/immediate' except that they accept byte offsets instead of
word offsets.  CAR and CDR become:

     (tagged-scm-ref/immediate DST SRC -6)    ;CAR
and:
     (tagged-scm-ref/immediate DST SRC 2)     ;CDR

(although at present the -6 prints as 250 in the disassembler).

> I think we disable GC “interior pointer” scanning.

I agree.

> With this scheme, an SCM for a pair would actually point in the middle
> of a pair; could this be an issue for GC?

It is already the case that Guile has tagged pointers in the first word
of every struct.  The first word of a struct contains a pointer to the
vtable, with scm_tc3_struct added.

Fortunately, BDW-GC provides GC_REGISTER_DISPLACEMENT, which allows us
to register a small offset K, such that BDW-GC should recognize pointers
that point K bytes into a heap block.  We've been using this in both 2.0
and 2.2 from scm_init_struct (), and in 'master' it's done in
scm_storage_prehistory ().

This new approach entails registering one additional displacement.

In 2.0 and 2.2, we register displacements 0, 16, and 17.  The last two
are for struct vtables, which point 2 words into a heap block, even
before adding scm_tc3_struct (which is 1).

In current master, we register 0 and 1 (scm_tc3_struct).  With this new
approach, we would register 0, 1, and 6 (scm_pair_tag).

What do you think?

      Thanks,
        Mark