[builtins] Fix floattitf.c etc. compilation on Solaris/SPARC

[rorth]

69660cc broke the Solaris/sparcv9 buildbot: compiler-rt/lib/builtins/int_to_fp.h unconditionally uses *int128_t which don't exist on 32-bit SPARC.

As suggested in #67540, this patch fixes this by moving the CRT_HAS_TF_MODE guard up which does the necessary checks.

Tested on sparcv9-sun-solaris2.11.

[arichardson]

LGTM, alternatively could also add an explicit check for INT128 availability to this #if rather than relying on CRT_HAS_TF_MODE only being defined when int128 also exists.

[asb]

This LGTM, thanks for fixing this and apologies my patch caused a regression for your target. Failing to guard int128_t was indeed an oversight.

Looking at the history for the linked builder, it seems pretty reliable other than this patch of course breaking it. Have you looked into moving it to the production buildmaster? That way, I would have had an email soon after landing the patch and could have reverted then fixed.

[rorth]

LGTM, alternatively could also add an explicit check for INT128 availability to this #if rather than relying on CRT_HAS_TF_MODE only being defined when int128 also exists.

I've committed the patch as is to quickly unbreak the buildbot. If we want to add more checks to make headers self-contained, I guess it's better to do this consistently rather than just in a single place.

[rorth]

Looking at the history for the linked builder, it seems pretty reliable other than this patch of course breaking it. Have you looked into moving it to the production buildmaster? That way, I would have had an email soon after landing the patch and could have reverted then fixed.

I have indeed, given that there seem to be no spurious failures on the Solaris/sparcv9 buildbot lately. I'd like to also move the corresponding Solaris/amd64 bot, but that one has one type of failure in compiler-rt once in a while which I couldn't reproduce outside of the bot yet, so I guess I'll be moving them separately.