Re: [PATCH] Struct-by-value support for x86-64 and ARM64

About a week before the end of the month. But changes to new code can come
in late.
On Thu, Jan 15, 2026 at 03:47 Jesse Bouwman <[email protected]>
wrote:

> I looked at win32, the struct classification is simpler than sysv, and
> other aspects look similar -- when does this release cycle close?
>
> > On Jan 14, 2026, at 9:20 AM, Stas Boukarev <[email protected]> wrote:
> >
> > Ok, I fixed that.
> >
> > On Wed, Jan 14, 2026 at 7:24 PM Stas Boukarev <[email protected]>
> wrote:
> >>
> >> It just segfaults.
> >>
> >> On Wed, Jan 14, 2026 at 7:14 PM Jesse Bouwman
> >> <[email protected]> wrote:
> >>>
> >>> Thank you for cleaning up the tests. I will take a look at win32.
> >>>
> >>> What result did you see from callback-struct-return-large? I reenabled
> it locally and it passed. Machine info:
> >>>
> >>> Linux wren 6.12.44 #1-NixOS SMP PREEMPT_DYNAMIC Thu Aug 28 14:31:16
> UTC 2025 x86_64 GNU/Linux
> >>>
> >>> On Jan 13, 2026, at 11:14 PM, Stas Boukarev <[email protected]>
> wrote:
> >>>
> >>> I merged this, but haven't reviewed it much. (I had to change it to
> >>> compile on unsupported targets).
> >>> And the callback-struct-return-large test fails on x86-64.
> >>> How hard would it be to add win32 support during this release cycle?
> >>>
> >>>
> >>> On Wed, Jan 14, 2026 at 2:58 AM Jesse Bouwman via Sbcl-devel
> >>> <[email protected]> wrote:
> >>>
> >>>
> >>> Hi,
> >>>
> >>> here is a patch that adds support for passing and returning structs by
> value on unix-like x86-64 and ARM64.
> >>>
> >>> Briefly, src/code/c-call.lisp dispatches to architecture-specific
> struct classifiers; classifiers determine whether a struct should be passed
> in registers or memory, per ABI rules: For x86-64, structs less than 16
> bytes are classified into "eightbytes" that map to int or SSE register
> classes. Larger structs are passed via hidden pointer in RDI, and returned
> in RAX. ARM64 supports HFA: structs of up to 4 same-type floats are
> passed/returned in floating-point registers. Non-HFA structs less than 16
> bytes use x0/x1; larger structs use a hidden pointer in x8. The
> alien-funcall IR1 transform handles both small struct returns (unpacking
> register to heap) and large struct returns (hidden pointer as synthetic
> first argument).
> >>>
> >>> There are no changes to the FFI interface.
> >>>
> >>> Other notes:
> >>>
> >>> Implementation owes a lot to the work done a year ago by Rongcui Dong
> and reviewers, documented on the launchpad thread at
> https://bugs.launchpad.net/sbcl/+bug/313202
> >>>
> >>> What I would deem pretty thorough testing is included: variously sized
> structs, nested structs, structs containing arrays, unions, etc. I've been
> debugging this with a binding to libclang, this has shaken out some
> serioous bugs, and I think it's now usable. I would welcome feedback about
> any aspects, especially relating to bad assumptions, missing functionality,
> or non-sbcl-idiomatic code.
> >>>
> >>> The branch is at:
> https://github.com/jbouwman/sbcl/tree/struct-by-value-final
> >>>
> >>> As the code is a result of bouncing between arm/darwin and x86/linux
> to try to get reasonably clean abstractions, it was squashed into a few big
> commits, but can be retroactively broken up in order to get something a
> little more bisectable for future debugging ergonomics.
> >>>
> >>> Windows is blocked by a non-SysV ABI which is not familiar to me, but
> looks pretty easy. Famous last words!
> >>>
> >>> Cheers,
> >>> Jesse
> >>>
> >>>
> >>> _______________________________________________
> >>> Sbcl-devel mailing list
> >>> [email protected]
> >>> https://lists.sourceforge.net/lists/listinfo/sbcl-devel
> >>>
> >>>
>
>
>

_______________________________________________
Sbcl-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-devel