Require C11

[devreal]

Error out if the compiler does not support C11 and limit the standard to C11 if the compiler accepts a standard flag. This limit prevents us from using features of newer standard versions.

Also removes support for __thread since we now rely on C11 _Thread_local.

Please check my m4 code, not an expert on this.

See #12658 (review) for notes from the RM discussion and the PR that initiated this.

There will likely be more changes over time replacing pre-C11 features with C11 features where possible but I wanted to keep this manageable.

@jsquyres where should I put a note about this for users?

[bosilca]

Not for 5.x

[devreal]

Mhh, CI fails with this error:

  LEX      keyval_lex.c
  CC       keyval_lex.lo
keyval_lex.c: In function ‘opal_util_keyval_yy_init_buffer’:
keyval_lex.c:1848:48: error: implicit declaration of function ‘fileno’ [-Werror=implicit-function-declaration]
 1848 |         b->yy_is_interactive = file ? (isatty( fileno(file) ) > 0) : 0;
      |                                                ^~~~~~
cc1: some warnings being treated as errors

I tried adding POSIX feature test macros but that doesn't seem to help. There seems to be a fix in flex from 7 years ago: westes/flex#263 Maybe that just hasn't trickled down to the distributions yet... I cannot reproduce this problem on my Mac, unfortunately.

[bosilca]

Add -D _POSIX_C_SOURCE=20240101L" to libopalutilkeyval_la_CFLAGS` in opal/util/keyval/Makefile.am ?

I guess you will need to so the same for all lex/yacc outputs

[jsquyres]

Do you need to put a check here to ensure that some flag was found and that C11 is, indeed, supported?

[devreal]

We do error out if C11 is not supported (the block above). This is just adding the C11 flag as a restriction (GCC will eventually default to C23 so we could accidentally sneak in C23 code that). If we don't find a flag for some reason we can just move on.

[jsquyres]

Ah -- so there's no case where an additional flag is needed, but we can't find that flag? If that's true, then this is good enough.

I ask because $opal_cv_c11_flag_required is set to true if we try to compile a C11 program (i.e., checking __STDC_VERSION__) with no additional flags and it fails.

[devreal]

The OPAL_PROG_CC_C11 macro sets the -std=c11 parameter if it's needed for C11 support and then tries to pick the right flag. If we don't find a suitable flag we error out. If C11 support is available without a flag then we get here and try to find one to restrict ourselves. So this block is somewhat optional and we I'm ok with removing it if we don't want to restrict ourselves to C11.

[jsquyres]

Ah, gotcha. Yes, if this is redundant code, let's ditch it.

Per your other point: I can't think of a reason to restrict ourselves to C11 -- can you? If neither of us can think of one, I think we're clear to remove this block.

[devreal]

Mhh, I was thinking we should protect ourselves from unintentionally introducing C23 code once compilers default to that but maybe that risk is low. I removed that part of configure.

[devreal]

Not sure why CI fails. All tests seem green but at the end it's marked as failed. Can someone please point me to the error?

[devreal]

Got it. Tests show green even if they failed in Jenkins... The culprit is our tests on RHEL 7.9, which still ships GCC 4.8. GCC added support for _Atomic, _Thread_local, and _Generic in 4.9 [1] and we are currently testing for that as part of our C11 test. We are not using _Atomic by default and _Generic is used in one place, optionally (the displacement/size wrappers for collectives). This PR makes _Thread_local the only option for TLS and removes the detection of __thread. All other features of C11 appear to be available in GCC 4.8, including __STDC_VERSION__ == 201112L.

I see three options:

1. Drop support for RHEL 7.9 and its prehistoric default compiler. According to [2] RHEL 7.9 is on extended support until 2028.
2. Separate out detection of _Atomic, _Generic, and _Thread_local, bring back the fallback to __thread, and require all other features of C11 by checking for __STDC_VERSION__ == 201112L (since GCC 4.7). That will give us access to some of the things C11 introduced without checking for all features explicitly, including _Static_assert, _Noreturn, stdalign.h, and anonymous structs.
3. Drop this PR and bring it back post 2028.

[1] https://gcc.gnu.org/wiki/C11Status
[2] https://en.wikipedia.org/wiki/Red_Hat_Enterprise_Linux

[wenduwan]

Drop support for RHEL 7.9 and its prehistoric default compiler.

The extended supported is a paid subscription IIUC so I don't think we can strictly secure the test environment. FYI EFA is dropping support for CentOS 7 and RHEL 7.

[devreal]

That's an important data point, thanks @wenduwan. I put this topic on the list for the next devel call.