Re: regression on zOS between 5.4.8 to 5.5.0 on zOS

On Mon, Feb 2, 2026 at 2:19 PM Roberto Ierusalimschy
<[email protected]> wrote:
>
> > The same SW stack and automated test works for me on AIX (which is
> > also Big Endian), Windows, Linux/x86_64, Linux/s390x, and
> > Linux/ppc64le
> > The previous stack, with 5.4.8 or even 5.2.2. worked OK on the above + zOS
> >
> > My suspicion is that something about the structs, unions, and
> > shifting/masking around the representation of the stack is not as
> > portable as 5.4.8.
> > But I didn't get very far trying various de-optimizations in lobject.h
> > and lopcodes.h
>
> That is possible. Does zOS has something unusual related to alignment,
> words, sizes for basic objects, or other memory issues?

Big endian is the obvious one but it shares that with AIX which is unaffected.
I have also seen warnings that "strict aliasing" is rather
strict/aggressive on zOS / xlc.


> In particular, can you try the following?
>
> -------------------------------------------------------------------------
> --- a/lstate.c
> +++ b/lstate.c
> @@ -345,2 +345,3 @@ LUA_API lua_State *lua_newstate (lua_Alloc f, void *ud, unsigned seed) {
>                         (*f)(ud, NULL, LUA_TTHREAD, sizeof(global_State)));
> +printf("%ld %ld\n", sizeof(StackValue), sizeof(TValue));
>    if (g == NULL) return NULL;
> -------------------------------------------------------------------------
>
> StackValue is a union than contains a TValue, and there is a lot of
> casts from one to the other. If we want to get the next element, we
> must add 1 to a pointer to the first type, before the cast.
>
> In most machines, these two objects have the same size, so a bug of
> adding 1 after the cast won't be detected.

Both 16 bytes.  I have gone a few rounds of unsuccessful AI-assisted
debugging but I do have some more verbose debug related to the various
offsets and macros, in lvm.c and lstate.c

I have collected it here:

https://gist.github.com/covener/4412c8be7762916335bb05e1d8171511

-- 
You received this message because you are subscribed to the Google Groups "lua-l" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
To view this discussion visit https://groups.google.com/d/msgid/lua-l/CALK%3DYjNAfPGXYdBK-Qh9qDCUzfM%3DSVtuDKUFZGRpSFhT8%2BVVfg%40mail.gmail.com.