Linus Torvalds' Latest Commentary Against -O3'ing The Linux Kernel

[chithanh]

No that is allowed, as long as the result is more precise than IEEE minimum, which it is with contractions. This is also what allowed doubles to be internally 80bits on x87. Though perhaps you can only do it if you assume no signaling NaN, as they could signal independently on each of the statements, but -fno-signaling-nans is default in gcc.

To quote the C standard 6.3.1.8:

That is not what I meant, see gcc manpage for the -ffp-contract=... parameter and the associated comments in gcc code. gcc defaults to -ffp-contract=fast which means it even contracts when crossing expression boundaries, which is forbidden by ISO C. There was a proposal a while back to make gcc conformant but it didn't go anywhere I think.

https://gcc.gnu.org/legacy-ml/gcc-patches/2013-11/msg00602.html

[kcrudup]

FWIW, I've been running with -O3, -march/-mtune=icelake-client and LTO=thin for a couple of years now with no ill effects, and I always run Linus' latest master.

I don't get the kvetching here about it, just enable it yourself (all I had to do was remove the conditional that turned it off for all but the "ARC" processors).

[eltomito]

Where they I should have waited for the first FreeBSD release and never had made the Linux kernel?

Maybe but there was a swear word between every two non-swear words.