Re: MPS image cache

["Gerd Möllmann" <gerd.moellmann@gmail.com>]

Eli Zaretskii <eliz@gnu.org> writes:

>> Pushed that, but not yet using it. We can always remove that later...
>
> It breaks the 32-bit build:
>
>   In file included from lisp.h:42,
> 		   from igc.c:32:
>   igc.c: In function 'fix_ptr_vec':
>   ../lib/verify.h:213:41: error: static assertion failed: "verify (sizeof *h == sizeof *v)"
>     213 | # define _GL_VERIFY(R, DIAGNOSTIC, ...) _Static_assert (R, DIAGNOSTIC)
> 	|                                         ^~~~~~~~~~~~~~
>   ../lib/verify.h:330:20: note: in expansion of macro '_GL_VERIFY'
>     330 | # define verify(R) _GL_VERIFY (R, "verify (" #R ")", -)
> 	|                    ^~~~~~~~~~
>   igc.c:107:30: note: in expansion of macro 'verify'
>     107 | #define igc_static_assert(x) verify (x)
> 	|                              ^~~~~~
>   igc.c:1085:5: note: in expansion of macro 'igc_static_assert'
>    1085 |     igc_static_assert (sizeof *h == sizeof *v);
> 	|     ^~~~~~~~~~~~~~~~~
>   Makefile:453: recipe for target `igc.o' failed
>
> And I truly don't understand how it was supposed to work:
>
>   static mps_res_t
>   fix_ptr_vec (mps_ss_t ss, void *client)
>   {
>     MPS_SCAN_BEGIN (ss)
>     {
>       struct igc_header *h = client_to_base (client);
>       void **v = client;
>       igc_static_assert (sizeof *h == sizeof *v);  <<<<<<<<<<<<<<<<<
>
> Here '*v' is a pointer, whereas '*h' is 'struct igc_header', which is
> a structure whose size is larger than 32 bits:

I guess this comes from igc_header being 1 word on 64 bit, and 2 on
32-bit. Thing is we need to know how large the vector is to scan it.
We have the info the igc_header::nwords.

>   enum
>   {
>     IGC_TYPE_BITS = 5,
>     IGC_PVEC_BITS = 6,
>     IGC_HASH_BITS = 21,
>     IGC_SIZE_BITS = 32
>   };
>
>   struct igc_header
>   {
>     enum igc_obj_type obj_type : IGC_TYPE_BITS;
>     enum pvec_type pvec_type : IGC_PVEC_BITS;
>     mps_word_t hash : IGC_HASH_BITS;
>     mps_word_t nwords : IGC_SIZE_BITS;
>   };
>
> So this can only work on a 64-bit architecture, and even then only if
> the compiler packs the structure (which is entirely not guaranteed
> AFAIK).  Or what am I missing here?

Apparently it only works as is on 64-bits and so on. What do you expect
me to do? A thing I promise not to do is to develop and check on any
platform except mine.

> And wasn't it supposed to say
>
>       void **v = &client;
>
> instead?

The address of a local variable? I think it's correct as is.