bug#25791: address argument to atomic operation must be a pointer to _Atomic type with 2.1.7 on macOS

bug#25791: address argument to atomic operation must be a pointer to _Atomic type with 2.1.7 on macOS

[ilove zfs]

[-- Attachment #1: Type: text/plain, Size: 1500 bytes --]

On macOS 10.11 and 10.12 building with Xcode 8, guile 2.1.7 fails to compile with several errors saying "address argument to atomic operation must be a pointer to _Atomic type." I can avoid this by setting ac_cv_header_stdatomic_h=no.



Build logs here:
https://gist.github.com/ilovezfs/b18b7e6160e0a4c5da297cc430961f63
https://gist.github.com/57e1513e03ede26dc97c0f63eaad3a98

Build failure is
```
  CC       libguile_2.2_la-arrays.lo
  CC       libguile_2.2_la-async.lo
In file included from async.c:27:
../libguile/atomics-internal.h:37:10: error: address argument to atomic operation must be a pointer to _Atomic type ('gl_uint32_t *' (aka 'unsigned int *') invalid)
  return atomic_fetch_sub (loc, arg);
         ^                 ~~~
/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/../lib/clang/8.0.0/include/stdatomic.h:149:43: note: expanded from macro 'atomic_fetch_sub'
#define atomic_fetch_sub(object, operand) __c11_atomic_fetch_sub(object, operand, __ATOMIC_SEQ_CST)
```
and several similar errors.

Previously (as of guile 2.1.4) I had been able to work around the issue without setting ac_cv_header_stdatomic_h=no with this patch:
https://raw.githubusercontent.com/ilovezfs/formula-patches/d2798a468346a7a28fbcd1c8aa3c1ccd5627b03f/guile/guile-atomic-type.patch

However, the patch is no longer sufficient since now void** is one the types needing _Atomic, which isn't allowed.



[-- Attachment #2.1: Type: text/html, Size: 2165 bytes --]

bug#25791: atomics on macOS

[Matt Wette]

[-- Attachment #1: Type: text/plain, Size: 421 bytes --]

I saw this too and gave up trying to fix it.  I think this needs to be worked bottom up: I believe the C99 standard is specific about the type qualifier _Atomic being present for the first argument.  The workaround for now IMO, is to change atomics-internal.h from this:

	#ifdef HAVE_STDATOMIC_H

to this:

	#if (defined HAVE_STDATOMIC_H) && (!defined __clang__)

By the way, gcc is forgiving here.

Matt



[-- Attachment #2: Type: text/html, Size: 2369 bytes --]

bug#25791: address argument to atomic operation must be a pointer to _Atomic type with 2.1.7 on macOS

[Andy Wingo]

On Sun 19 Feb 2017 09:41, ilove zfs <ilovezfs@icloud.com> writes:

> On macOS 10.11 and 10.12 building with Xcode 8, guile 2.1.7 fails to compile with several errors saying "address argument to atomic operation must be a pointer to _Atomic type." I
> can avoid this by setting ac_cv_header_stdatomic_h=no.
>
> Build logs here:
> https://gist.github.com/ilovezfs/b18b7e6160e0a4c5da297cc430961f63
> https://gist.github.com/57e1513e03ede26dc97c0f63eaad3a98
>
> Build failure is
> ```
>   CC       libguile_2.2_la-arrays.lo
>   CC       libguile_2.2_la-async.lo
> In file included from async.c:27:
> ../libguile/atomics-internal.h:37:10: error: address argument to atomic operation must be a pointer to _Atomic type ('gl_uint32_t *' (aka 'unsigned int *') invalid)
>   return atomic_fetch_sub (loc, arg);
>          ^                 ~~~
> /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/../lib/clang/8.0.0/include/stdatomic.h:149:43: note: expanded from macro 'atomic_fetch_sub'
> #define atomic_fetch_sub(object, operand) __c11_atomic_fetch_sub(object, operand, __ATOMIC_SEQ_CST)
> ```
> and several similar errors.
>
> Previously (as of guile 2.1.4) I had been able to work around the issue without setting ac_cv_header_stdatomic_h=no with this patch:
> https://raw.githubusercontent.com/ilovezfs/formula-patches/d2798a468346a7a28fbcd1c8aa3c1ccd5627b03f/guile/guile-atomic-type.patch
>
> However, the patch is no longer sufficient since now void** is one the types needing _Atomic, which isn't allowed.

Any ideas here?  Sure would be nice to release a Guile with atomics that
compiles with clang.

Andy

bug#25791: address argument to atomic operation must be a pointer to _Atomic type with 2.1.7 on macOS

[Andy Wingo]

On Wed 01 Mar 2017 17:43, Andy Wingo <wingo@pobox.com> writes:

> On Sun 19 Feb 2017 09:41, ilove zfs <ilovezfs@icloud.com> writes:
>
>> On macOS 10.11 and 10.12 building with Xcode 8, guile 2.1.7 fails to
>> compile with several errors saying "address argument to atomic
>> operation must be a pointer to _Atomic type." I
>> can avoid this by setting ac_cv_header_stdatomic_h=no.
>>
>> Build logs here:
>> https://gist.github.com/ilovezfs/b18b7e6160e0a4c5da297cc430961f63
>> https://gist.github.com/57e1513e03ede26dc97c0f63eaad3a98
>>
>> Build failure is
>> ```
>>   CC       libguile_2.2_la-arrays.lo
>>   CC       libguile_2.2_la-async.lo
>> In file included from async.c:27:
>> ../libguile/atomics-internal.h:37:10: error: address argument to
>> atomic operation must be a pointer to _Atomic type ('gl_uint32_t *'
>> (aka 'unsigned int *') invalid)
>>   return atomic_fetch_sub (loc, arg);
>>          ^                 ~~~
>> /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/../lib/clang/8.0.0/include/stdatomic.h:149:43:
>> note: expanded from macro 'atomic_fetch_sub'
>> #define atomic_fetch_sub(object, operand) __c11_atomic_fetch_sub(object, operand, __ATOMIC_SEQ_CST)
>> ```
>> and several similar errors.
>>
>> Previously (as of guile 2.1.4) I had been able to work around the issue without setting ac_cv_header_stdatomic_h=no with this patch:
>> https://raw.githubusercontent.com/ilovezfs/formula-patches/d2798a468346a7a28fbcd1c8aa3c1ccd5627b03f/guile/guile-atomic-type.patch
>>
>> However, the patch is no longer sufficient since now void** is one the types needing _Atomic, which isn't allowed.
>
> Any ideas here?  Sure would be nice to release a Guile with atomics that
> compiles with clang.

Fixed in 2.1.8 I think.

Andy