libdazzle @3.38.0_0: Static assert failure on PowerPC due to struct packing issue with different version of gcc

[evanmiller (Evan Miller)]

:info:build In file included from /opt/local/lib/glib-2.0/include/glibconfig.h:9:0,
:info:build                  from /opt/local/include/glib-2.0/glib/gtypes.h:32,
:info:build                  from /opt/local/include/glib-2.0/glib/galloca.h:32,
:info:build                  from /opt/local/include/glib-2.0/glib.h:30,
:info:build                  from /opt/local/include/glib-2.0/gobject/gbinding.h:28,
:info:build                  from /opt/local/include/glib-2.0/glib-object.h:23,
:info:build                  from ../libdazzle-3.38.0/src/search/dzl-fuzzy-mutable-index.h:22,
:info:build                  from ../libdazzle-3.38.0/src/search/dzl-fuzzy-mutable-index.c:24:
:info:build /opt/local/include/glib-2.0/glib/gmacros.h:738:31: error: static assertion failed: "Expression evaluates to false"
:info:build  #define G_STATIC_ASSERT(expr) _Static_assert (expr, "Expression evaluates to false")
:info:build                                ^
:info:build ../libdazzle-3.38.0/src/search/dzl-fuzzy-mutable-index.c:70:1: note: in expansion of macro 'G_STATIC_ASSERT'
:info:build  G_STATIC_ASSERT (sizeof(DzlFuzzyMutableIndexItem) == 6);
:info:build  ^~~~~~~~~~~~~~~
:info:build ninja: build stopped: subcommand failed.

Context:

#pragma pack(push, 1)
typedef struct
{
  guint64 id : 32;
  guint64 pos : 16;
} DzlFuzzyMutableIndexItem;
#pragma pack(pop)

In a test program, the above structure is 8 bytes but libdazzle is expecting it to be 6 bytes.

sysinfo

:debug:sysinfo Mac OS X 10.4.11 (darwin/8.11.0) arch powerpc
:debug:sysinfo MacPorts 2.7.1
:debug:sysinfo Xcode 2.5
:debug:sysinfo SDK 10.4
:debug:sysinfo MACOSX_DEPLOYMENT_TARGET: 10.4

Full log to follow.

[kencu (Ken)]

useful:

​http://personal.denison.edu/~bressoud/cs281-s07/MacOSXLowLevelABI.pdf

[evanmiller (Evan Miller)]

Thanks for the link. I'm still confused why there are 2 bytes of end padding with pack(push, 1) in place. Perhaps this is controlled by a separate pragma or flag.

[kencu (Ken)]

how does this work, exactly?

  guint64 pos : 16;

If a guint64 needs 4 bytes, how does adding the " : 16" somehow make it a two-byter?

I haven't seen this construct much before, or at least haven't noticed it...

[evanmiller (Evan Miller)]

These are bit fields ​https://docs.microsoft.com/en-us/cpp/c-language/c-bit-fields?view=msvc-160

[evanmiller (Evan Miller)]

Well, it looks like libdazzle has updated this code for Win32:

​https://gitlab.gnome.org/GNOME/libdazzle/-/blob/master/src/search/dzl-fuzzy-mutable-index.c#L62-75

Likely we'll need to include PPC in that ifdef or patch similarly

[evanmiller (Evan Miller)]

Actually, the Win32 code doesn't solve the problem on PPC, so further research is needed.

[kencu (Ken)]

interesting (Tiger PPC):

$ cat packcheck.c
#include <stdio.h>
#include <stdint.h>

int main(void) {

    #pragma option align=mac68k
    typedef struct
    {
    uint32_t id : 32;
    uint32_t  pos : 16;
    } KenStruct;
    
    KenStruct mystruct;
    
    printf("%i\n\n", sizeof(KenStruct));
    return 0;
}

$ gcc-mp-7 packcheck.c
$ ./a.out
8

$ gcc-apple-4.2 packcheck.c
$ ./a.out
8

$ gcc-apple-4.2 -malign-mac68k packcheck.c
$ ./a.out
6

$ gcc-mp-7 -malign-mac68k packcheck.c
gcc-mp-7: error: unrecognized argument in option '-malign-mac68k'
gcc-mp-7: note: valid arguments to '-malign-' are: natural power

[evanmiller (Evan Miller)]

Interestingly, just the regular pack directives work as expected with gcc-apple-4.2:

#include <stdint.h>

#pragma pack(push, 1)
typedef struct
{
  uint32_t id;
  uint16_t pos;
} DzlFuzzyMutableIndexItem;
#pragma pack(pop)

int main() {
    return sizeof(DzlFuzzyMutableIndexItem);
}

$ gcc-mp-7 test-pack.c && ./a.out
$ echo $?
8

$ gcc-apple-4.2 test-pack.c && ./a.out
$ echo $?
6

So the struct-padding algorithm changed at some point.

[kencu (Ken)]

This works properly:

$ cat packcheck.c
#include <stdio.h>
#include <stdint.h>

int main(void) {

    typedef struct __attribute__((packed))
    {
    uint32_t id : 32;
    uint16_t  pos : 16;
    } KenStruct;
    
    KenStruct mystruct;
    
    printf("%i\n\n", sizeof(KenStruct));
    return 0;
}

$ gcc-mp-7 packcheck.c
$ ./a.out
6

[kencu (Ken)]

and back to the original, looks pretty good:

$ cat packcheck.c
#include <stdio.h>
#include <glib.h>

int main(void) {

    typedef struct __attribute__((packed))
    {
    guint64 id : 32;
    guint64 pos : 16;
    } KenStruct;
    
    KenStruct mystruct;
    
    printf("%i\n\n", sizeof(KenStruct));
    return 0;

}

$ gcc-mp-7 `pkg-config --cflags glib-2.0` packcheck.c

$ ./a.out
6

so that looks like our solution.

[evanmiller (Evan Miller)]

Nice find! Wonder whether that's a GCC bug and whether GCC11 has the same behavior.

Anyway, we have a workaround now. I'll get a patch going, assuming everything else compiles OK.

[evanmiller (Evan Miller)]

In 4adcdabf66d166e5fc6dadc9f7f8a98fc3cc47b1/macports-ports (master):

libdazzle: Support for 10.4, 10.5, and PowerPC (​https://github.com/macports/macports-ports/pull/12250)

Closes: #63490

[barracuda156]

Replying to evanmiller:

Nice find! Wonder whether that's a GCC bug and whether GCC11 has the same behavior.

Anyway, we have a workaround now. I'll get a patch going, assuming everything else compiles OK.

Yes, as well as GCC12: ​https://github.com/macports/macports-ports/pull/18872#issuecomment-1568714259