8.19.0 breaks compilation with ancient GCC (eg GCC 4.2)

[fds242]

I did this

3 source files in 8.19.0 got added

#pragma GCC diagnostic push
#pragma GCC diagnostic ignored …

directives, only wrapping it in a basic check for __GNUC__. This breaks compilation in ancient versions of GCC with an error of #pragma GCC diagnostic not allowed inside functions.
It isn't clear to me from past issues like #16062 or #16152 on what was ultimately decided on supporting such old compilers. I'm filing this issue then since:

• those previous issues were corrected
• curl is a stickler for requiring nothing newer than C89, and even GCC 4.2 was a C99-capable compiler
• simple enough, if wordy, to extend the check for the correct version:
  (defined(__GNUC__) && ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))

The newly affected files in 8.19.0:
lib/mprintf.c inside function out_double()
src/tool_cb_prg.c inside function tool_progress_cb()
src/tool_writeout.c inside function outtime()

I expected the following

No response

curl/libcurl version

curl 8.19.0

operating system

n/a

[bagder]

Why do we care about ancient gcc?

[vszakats]

Earlier attempts were driven by Windows CE with gcc 4.4 its latest. With
it gone, the oldest version in CI is 4.8.1 (paired with mingw-w64 3) and
6.3.0 on old linux. Things without CI tests tend to decay.

Keeping compatibility is more complicated than extending guards: these
warnings then need to be disabled globally in these cases, in autotools
and cmake. Possibly more, because these ancient versions also do not
support push/pop.

It's not impossible to support ancient gcc (also present in tests via
CURL_GNUC_DIAG, and constraining an example), but unfun to spend
effort on, esp. without tests. Drawing the line at 4.8.1 (2013-03-22)
may be a reasonable compromise IMO.

[bagder]

Drawing the line at 4.8.1 (2013-03-22) may be a reasonable compromise IMO.

Sounds good to me. We could add that in the INTERNALS.md list.

[fds242]

Documenting a minimum version of GCC, even a much more recent one than 4.8.1, is of course eminently reasonable, and will be perfectly sufficient for me to know not to file related issues, and keep my patches local. There's only the funny dichotomy to state that curl, while only requiring a C89 compiler, in practice requires at least GCC 4.8, a C11 compiler. (A C11 compiler already is required by important dependencies such as ngtcp2.)

I must say though that the average user does not wear CI-colored glasses or compile with -Werror enabled, and as such, consider compiler warnings quite harmless, and easily ignored. In that sense then, it really is sufficient not to squelch the warning, and simply extend the ifdef guard. A modicum of support – warnings but no errors – for the ancient compiler would be more than enough! It really is way too much effort to devise a way to actually silence the warning everywhere.

[My interest here is retro computing, and long abandoned PowerPC Macs. The very last compiler Apple provided that could target these systems happens to be GCC 4.2.1. Mainline GCC, including the current GCC 15.2, also continues to support these systems. There remains however a significant practical divide between software that is compilable with the final Apple-provided GCC and those that absolutely need a modern compiler. Particularly for various package managers that attempt to locally build everything, it is very desirable to have some basic, foundational packages that could be built before having new tools.]

[bagder]

There's only the funny dichotomy to state that curl, while only requiring a C89 compiler, in practice requires at least GCC 4.8, a C11 compiler

curl builds fine with lots of other compilers as well.