libjpeg-turbo fails to compile on 32bit MSVC

[benjamin-otte]

Running the GTK compile with the MSVC 2019 32bit toolchain (apparently I was opening the wrong shortcut and got that one) failed with:

[1986/3743] Linking target subprojects/libjpeg-turbo-2.1.0/jpeg-8.2.2.dll
FAILED: subprojects/libjpeg-turbo-2.1.0/jpeg-8.2.2.dll subprojects/libjpeg-turbo
-2.1.0/jpeg-8.2.2.pdb
"link" @subprojects/libjpeg-turbo-2.1.0/jpeg-8.2.2.dll.rsp
   Creating library subprojects\libjpeg-turbo-2.1.0\jpeg.lib and object subproje
cts\libjpeg-turbo-2.1.0\jpeg.exp
jcphuff.c.obj : error LNK2019: unresolved external symbol ___builtin_ctzl refere
nced in function _count_zeroes
subprojects\libjpeg-turbo-2.1.0\jpeg-8.2.2.dll : fatal error LNK1120: 1 unresolv
ed externals
[1987/3743] Linking target subprojects...pixbuf/gdk-pixbuf/pixops/timescale.exe
ninja: build stopped: subcommand failed.

After trying to debug it a bit, it seems to be that there's a not just the check for ___builtin_ctzl but also one for sizeof(long)==sizeof(size_t) and that 2nd check fails on 64bit but not on 32bit, so ___builtin_ctzl gets disabled on the CI boxes because of that.

[eli-schwartz]

It should be checking that the __builtin_ctzl is functional:

wrapdb/subprojects/packagefiles/libjpeg-turbo/meson.build

Lines 68 to 69 in 0421165

	code = 'int main(int argc, char **argv) { unsigned long a = argc; return __builtin_ctzl(a); }'
	cdata.set('HAVE_BUILTIN_CTZL', cc.compiles(code))

Apparently on MSVC this does compile (both in the meson check, and in jcphuff.c.obj), but does not link? Is this somehow available on MSVC, but needs an additional library?

[tp-m]

I can confirm this fails to link on MSVC for x86. The only reason it "works" on x64 is that the check is bypassed entirely because of the size checks.

I think the best / correct thing to do here is to just do cc.has_function('__builtin_ctzl')? (Which works correctly).

I don't know why it compiles though, I'm not aware of these builtins being available on MSVC, they've got their own builtins for that (_BitScan*) as far as I know.

[xclaesse]

FWIW, that's a direct translation from cmake:

if(SIZE_T EQUAL UNSIGNED_LONG)
  check_c_source_compiles("int main(int argc, char **argv) { unsigned long a = argc;  return __builtin_ctzl(a); }"
    HAVE_BUILTIN_CTZL)
endif()

But... the cmake doc actually says:

Check that the source supplied in <code> can be compiled as a C source file and linked as an executable

So cc.links() is probably a better test here.

[tp-m]

Not convinced it's "better", but whatever works ;)

[xclaesse]

If someone can give it a try: #112

[xclaesse]

Not convinced it's "better", but whatever works ;)

I mean, closer to what cmake does, which is the reference here.

[eli-schwartz]

Does it need to be a reference to cmake? I agree that it makes more sense to use has_function, and it should be faster too.