Re: feature/jit-improved-type-punning 9d2a54bd8d: src/comp.c: Use libgccjit's bitcast API for type coercion, when available.

[Vibhav Pant]

On Wed, 2022-09-28 at 14:10 +0300, Eli Zaretskii wrote:
> > From: Po Lu <luangruo@yahoo.com>
> > Cc: Vibhav Pant <vibhavp@gmail.com>
> > Date: Wed, 28 Sep 2022 08:22:36 +0800
> > 
> > BTW, I don't see any code that loads this function dynamically on
> > MS
> > Windows.  Isn't that needed?
> 
> They were present in the patch that was posted.
> 
> FWIW, I'd like to see some rationale for these changes.  Perhaps
> libgccjit experts would see it immediately without any explanations,
> but I'm not such an expert.  TIA.

Type coercion in earlier versions of libgccjit was quite limited, as it
only supported casting between pointer types and integers <->
booleans/floats. Because working with lisp tagged pointers involved
converting from a Lisp_Object (which is usually a Lisp_X*) to an
integer, comp_t contained a union {uintptr_t; void *} which we would
use to emit two (cast_ptr_to_int, cast_int_to_ptr) functions before
JITting Lisp code, which would in turn initialize a union with the
input parameter, and return the required field. Combined with a 15 x 15
lookup table for every type used while compiling Lisp, all type casting
would therefore consist of emitting a call to one of these functions.

This branch uses libgccjit's new API for bitcasting between types,
which lets us skip emitting this good chunk of helper code as part of
the JIT process. Working with/extracting Lisp_Object tags in JITed code
is semantically the equivalent of a simple reinterpret_cast operation,
potentially improving optimization (although I imagine almost all calls
to these helper functions are inlined, unless native-comp-speed is 0).

Best,
Vibhav

-- 
Vibhav Pant
vibhavp@gmail.com

GPG: 7ED1 D48C 513C A024 BE3A 785F E3FB 28CB 6AB5 9598

signature.asc
Description: This is a digitally signed message part