[PATCH] [ARM] Emit Tag_ABI_FP_denormal correctly in fast-math mode.

[Charlie Turner]

Hi Tim,

>>! In D6439#5, @t.p.northover wrote: 
> So where do we make these assumptions, or is this patch setting out a set of plans for the future? It's mostly curiosity, I don't remember seeing any code even mentioning denorms in my LLVM travels.

Which assumptions are you referring to? Perhaps the various "We default to IEEE 754 compliance" comments sprinkled in the test file? I assumed that's the default based on the original behaviour of LLVM (this tag is unconditionally set to 1 when we're not in unsafe maths mode). It's also what armcc does in ieee_full mode. Another assumption I'm making is that our default level of IEEE compliance is intended to mirror that of GCC. Whether that's currently the case is another matter, but it appears to be the goal.

The values for the unsafe maths mode are based off whether the target has an FPU, and if so what the behaviour of flush-to-zero is for the respective FP unit.

This work is only of interest to FP library selection.

Thanks a lot for taking a look!
Charlie.

http://reviews.llvm.org/D6439