Daniel Micay on Twitter: "@lemire @stephentyrone You seem to be talking about 2 distinct issues: 1) GCC uses x87 floating point by default on x86 (pass -mfpmath=sse), unlike Clang 2) Neither Clang or GCC supports setting a non-default floating point rounding mode (which you'll notice more with optimization)" / Twitter

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...

2

11

26

Replying to

Set -mfpmath=sse or avoid the legacy architecture. The only difference with Clang is that the default is SSE. If you pass -mfpmath=387 -mno-sse, Clang will act the same way as GCC does by default. So, either avoid the legacy architecture or configure it the way you want instead.

1

Replying to

Ok. I added -mfpmath=sse (see my script on GitHub). It does ***not*** help. See screenshot. cc

@stephentyrone

2

1

Replying to

and

It works fine here. Which GCC version are you using? I'm testing with GCC 10.1.0 and cannot reproduce your results.

1

Replying to

and

It reproduces with trunk on Godbolt.

2

Replying to

and

I have a docker script, so you should be able to reproduce it. Just install docker and run my dockerscript.sh. github.com/lemire/Code-us

1

Replying to

and

You seem to be talking about 2 distinct issues: 1) GCC uses x87 floating point by default on x86 (pass -mfpmath=sse), unlike Clang 2) Neither Clang or GCC supports setting a non-default floating point rounding mode (which you'll notice more with optimization)

9:59 PM · Jun 26, 2020Twitter Web App

Replying to

and

Rounding direction doesn't enter into what Daniel is observing at all. It's a complete red herring.

1

Replying to

and

I think twitter.com/DanielMicay/st is a full explanation of everything being observed. It's 2 different things being noticed. llvm.org/docs/LangRef.h documents that LLVM does not support the floating point environment without using special intrinsics and this applies to Clang.

Quote Tweet

Daniel Micay

@DanielMicay

Jun 26, 2020

Replying to @lemire and @stephentyrone

You seem to be talking about 2 distinct issues: 1) GCC uses x87 floating point by default on x86 (pass -mfpmath=sse), unlike Clang 2) Neither Clang or GCC supports setting a non-default floating point rounding mode (which you'll notice more with optimization)

1

Show replies

Conversation