Conversation

GNU GCC does not round floating-point divisions to the nearest value

GNU GCC on x86 does not round floating-point divisions to the nearest value

I know that floating-point arithmetic is a bit crazy on modern computers. For example, floating-point numbers are not associative: 0.1+(0.2+0.3) == 0.599999999999999978 (0.1+0.2)+0.3 == 0.600000000...

Replying to

I assume this is 32b only, because the compiler is faithfully modeling the x87 extended-precision arithmetic (FLT_EVAL_METHOD=2) of the runtime environment.

Replying to

and

E.g. if you tell it to use SSE for fp, or compile for 64b, I would expect to see the "right" result.

Replying to

and

The difference with Clang is that it assumes SSE support by default and refuses to use x87 floating point if SSE is supported. If you pass -mfpmath=387 -mno-sse to Clang, it will work the same way as GCC does by default. If you don't pass -mno-sse, it refuses to use x87 support.

9:51 PM · Jun 26, 2020Twitter Web App

Replying to

and

I am still angry that we support -mno-sse in clang.

Replying to

and

So the line “fesetround(FE_TONEAREST);” is just for show, right? There is no guarantee at all that the compiler will honour it? So why is it not “trysetround(FE_TONEAREST);” or something?

Show replies