Add support for floating point exceptions

[simonbyrne]

Fixes #27705.

Still more to do, just want to see if it builds

[simonbyrne]

@c42f you expressed interest in this.

One issue is that the floating point environment is tied to a thread: for our purposes, it would probably make more sense to tie it to a task. this would require saving and restoring the floating point environment whenever we switch tasks. I assume we can do this by adding a fenv_t field to the jl_task_t struct, and calling fegetenv/fesetenv. Which functions would need the hooks?

[Keno]

This doesn't really make sense without hooking up the constrained fp intrinsics from the LLVM side, because otherwise LLVM will just be free to move things around and you'll get unexpected results.

[simonbyrne]

I disagree: while the results might be a bit unreliable, it's still helpful for figuring out where NaNs are coming from

[simonbyrne]

It's also a bit of a chicken and egg problem: we need one to motivate the other

[simonbyrne]

It's also a bit of a chicken and egg problem: we need one to motivate the other

[Keno]

You're breaking things like the compiler's nothrow modeling though, so you'll definitely get memory corruption and undefined behavior by turning this on.

[c42f]

breaking things like the compiler's nothrow modeling

So, this would be similar to the way that force throwing InterruptException breaks the compiler's assumptions and can lead to unexpected results? Though we have that escape hatch because it's so darn useful because it mostly "kind of works" for interactive use.

Regarding InexactFloatingPointException I can see you'd add that for completeness. Though I found when I played with this ages ago, it was almost impossible to enable. For example, entering the compiler as part of a user's generic call can result in inexact FP operations as part of compilation happening in the same task. Which then easily result in an exception.

[gbaraldi]

What if you make a macro that wraps a function call, or something, that from julia itself inspects the floating point exception flags, and in the macro you throw an exception?