Failing tests on i686 with GCC 13

[marxin]

As mentioned already here (#2975 (comment)), after the recent GCC change gcc-mirror/gcc@8020c9c, the following test-case fails now:

[   23s] /home/abuild/rpmbuild/BUILD/fmt-9.1.0/test/chrono-test.cc:571: Failure
[   23s] Expected equality of these values:
[   23s]   "44.000000000000"
[   23s]   fmt::format("{:%S}", std::chrono::duration<float, std::pico>( 1.54213895E+26))
[   23s]     Which is: "16.430007934570"
[   23s] [  FAILED  ] chrono_test.special_durations (0 ms)

A reduced test case can be seen here: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109085.

This is related to the excess precision provided by X87 and can be fixed with either -ffloat-store or -fexcess-precision=standard. Note that the default precision is going to change for GCC 13: https://gcc.gnu.org/gcc-13/porting_to.html#excess-precision.

One more comment about using -fexcess-precision=standard on i686, apparently, it breaks the following test-case:

[   43s] /home/abuild/rpmbuild/BUILD/fmt-9.1.0/test/format-test.cc:1395: Failure
[   43s] Expected equality of these values:
[   43s]   "0.10000000149011612"
[   43s]   fmt::format("{}", double(0.1f))
[   43s]     Which is: "0.1"
[   43s] [  FAILED  ] format_test.prettify_float (0 ms)

Can you please take a look, @vitaut?

[vitaut]

Please provide a godbolt repro.

[marxin]

https://godbolt.org/z/q8PPYKMM5

[vitaut]

Fixed the overspecified chrono test and removed the nonportable float test since it doesn't add much value (e475859). Thanks for reporting.