Re: For a cheaper ‘bytevector->pointer’

[Andy Wingo]

On Sun 24 Nov 2019 11:52, Ludovic Courtès <address@hidden> writes:

> A few days ago David was explaining on #guile how ‘bytevector->pointer’
> was generating too much garbage for his use case.  An idea we came up
> with was to embed the pointer object in the bytevector.
>
> The patch below does that but it leads to segfaults because I’m guessing
> there’s generated bytecode somewhere that still uses the wrong offset; I
> adjusted code that emits ‘pointer-ref/immediate’, what else did I
> miss?

The compiler :)  Bytevector literals are stored statically in the .go
files, so the assembler would need to change to emit the new layout.
Also, compiled access to bytevectors; see prepare-bytevector-access in
(language tree-il compile-cps).

> Also, since we disable internal pointers, we’d need to register an
> additional displacement, and I’m not sure if this is a good idea.
>
> Thoughts?

Honestly I would prefer not to do this.  If I understand correctly, the
problem is in FFI calls -- you have a bytevector and you want to pass it
as a pointer.  In that case the "right" optimization is to avoid the
scm_tc7_pointer altogether and instead having an unboxed raw pointer.
The idioms used in FFI are local enough that a compiler can do this.

More broadly -- the current FFI is an interpreter but it should be a
compiler.  When a call happens, the code interprets the description of
the ABI.  Instead, pointer->function should ideally *compile* a
trampoline.  In an ideal world this compilation can happen
ahead-of-time, when the .go file is compiled.

In the short term, what about allowing bytevectors as arguments
whereever a pointer is allowed?  Perhaps it's bad to expand the domain
of these functions but it may be the right trade-off.

Andy