Add Julia routines for setting/getting "subnormals are zero" mode.

[ArchRobison]

This PR lets users more easily elect to treat subnormal numbers as zeros, which speeds up some programs on some x86 processors. Previously, doing such required ccall. The patch also lets users check the current treatment. See issue #12132 for discussion. Here is a demo program, which also demonstrates the "inject some noise" trick for avoiding subnormals.

The patch removes the undocumented routine jl_zero_subnormals and replaces it with jl_set_zero_subnormals. I renamed it because the semantics changed to make it similar to the Linux fesetround interface.

The implementation caches the result of CPUID inspection since CPUID is a serializing (slow) instruction.

I'll write the user documentation as a separate PR once we've agreed that this PR is the right way to go.

[yuyichao]

This is a question out of curiosity rather than an issue for this PR: How is cpuid handled in a multi CPU setup? Is it possible that different CPU's support different features? (or is this why cpuid is serializing?)

[ArchRobison]

In multi-CPU setups, some of the CPUID results can differ, because they tell you which hardware thread you are on. But for features like FZ/DAZ, they need to be the same across hardware threads for just about any time-slicing scheduler, since otherwise the end of a time slice (and rescheduling onto another hardware thread) would invalidate the result in a racy way.

I don't know why the designers made CPUID serializing in its introduction (reference: Section 18.8 of Pentium manual). It's regrettable since it's often needed when generating code that exploits version-specific features, and Pentium 4 introduced the mfence instruction that handles most serialization needs.

[yuyichao]

I see. Thanks.

It's regrettable since it's often needed when generating code that exploits version-specific features,

Of course Intel have never made any mistakes ;-P

[pao]

Thanks, Arch and Yu (...assuming your name here is in Western name order...) for digging this all back up and moving forward with a plan. While I regret not getting back to truly finish this up, I have no doubt that this is a more thorough, considered treatment than I could have given it.

[yuyichao]

Not a big deal but maybe it's better to be consistent with uint32_t and int32_t...

[ArchRobison]

Thanks for the sharp eyes. Now fixed.

[ArchRobison]

Documentation added to PR.

[yuyichao]

The doc added looks great! Thanks.

[ivarne]

I have no idea what subnormal numbers are, but are there any reason why we shouldn't support:

set_zero_subnormals() do
    heatflow(a,1000)
end

Apart from not thinking about the utility, or not bothering writing the code.

[yuyichao]

@ivarne See #12132 (comment) and my reply below. I think mainly because this is not needed that often (and anonymous is slow).

There's no reason we couldn't add that though.