Why do doubles need special handling though, even on a 32-bit system? My suggestion is that the Double_tag be changed to Flat_tag, meaning that all non-pointer objects can reside in this tag. The only issue I've found so far is that polymorphic <, <=, > and >= would not work. However, these operators should not be allowed on a vector anyway since there is no natural ordering scheme for vectors. If there are other issues, please let me know.

I agree regarding the expansion of 246 constructors. This must have been kept for compatibility with 32 bit systems. I think what should happen in 32 bit systems is that one constructor should be reserved for having >246 constructors, in which case another word of memory could be utilized for the constructor code. In fact, you'd only need to use that extra word if the particular constructor exceeds 246. In 64 bit systems, the constructor count could easily be increased by a few bits, with the same backup mechanism for when you have more than X constructors (X being the maximum number of constructors).

Regards,
Yotam


On Mon, Sep 16, 2013 at 12:29 PM, Markus Mottl <markus.mottl@gmail.com> wrote:
On Mon, Sep 16, 2013 at 11:26 AM, Yotam Barnoy <yotambarnoy@gmail.com> wrote:
> Having looked through some of the ocaml runtime's code, I have a question
> regarding the Double_array block tag. Why not use a single tag for all block
> content that doesn't contain pointers instead? This would allow optimization
> of all cases where no pointers are involved, including float tuples, records
> with ints, bools and floats etc.

Floats are a little tricky, because they are always doubles (64 bits),
even on 32-bit platforms.  This requires some special identification.

Adding new tags by reducing the "No_scan_tag" might be a bad (and not
backward compatible) approach, too: the maximum number of non-constant
constructors is already pretty low at 246.  I think this number is too
small these days where 64-bit platforms are standard.  It's probably
hard to change this design decision now by reducing the overly
generous maximum "wosize".  Some automatically generated APIs can
easily blow the current limit on non-constant constructors and require
annoying, less efficient workarounds.

I guess it might be possible to allocate blocks that are known to be
all integers or atomic sum types by using the already available
Abstract_tag.  Large arrays would benefit most from that.  Doing this
in the compiler might break old marshaled data.  But if performance is
really critical, I could imagine trying that in a library with
well-hidden type representations.

Regards,
Markus

--
Markus Mottl        http://www.ocaml.info        markus.mottl@gmail.com