-mfpmath=sse,387 is dangerous

[scmasaru]

I used -mfpmath=sse,387 to compile LAME, an mp3 encoder, and it turned out that it not only slowed the thing down but also produced different mp3 (using cmp).

I would suggest avoid this option like plague.

icc also has the similar problem.

[neuron]

hmmm... interesting /me 's always had it on.

[bsolar]

What are the whole CFLAGS and what's the CPU.

Also can you clarify what do you mean with "different" if possible?
_________________
I may not agree with what you say, but I'll defend to the death your right to say it.

[Lovechild]

actually what did you expect - you are telling your CPU to use another floating point unit, which might I add is less effective than the default 387 one.

Of course it will do some math different.

if you have an Athlon CPU do not use that flag, it won't speed up anything, rather it will slow things down.

[scmasaru]

[Megaptera]

I've read where -mfpmath=sse,387 and Pentium 3s and 4s don't get along. Saw it in this thread , which you may find interesting.
_________________
It is not like the world will end if I take the day off from eating worlds.

[Exner]

[zhenlin]

sse is for parallelized computations. 387 is for single floating point operations.

sse can be convinced to do the job of a FPU, but it's a waste of computing power.

Using sse,387 is a sure way to get into trouble. Such programs may try to access a 387 register when the data they are looking for is in a SSE register...

Are they audibly different? If not, does it matter? It is probably a small computational difference, like 0.00000000000000001 and 0.00000000000000002.

[Exner]