[clang] Align -ffp-model=fast denormal handling with -ffast-math (PR #89477)

[Joshua Cranmer via cfe-commits]

jcranmer-intel wrote:

> I'm not sure what the correct behavior is across all platforms. My perspective on this is heavily X86-biased, so I'd really like some guidance from people who work on other targets about their expectations. I think I saw at least one target that sets denormal-fp-math=preservesign as their default even without any fast-math options enabled. It wouldn't surprise me if there are targets that don't want to set denormal-fp-math=preservesign even with fast-math.

Some of the GPU targets, IIRC, want daz/ftz by default. Not all targets have DAZ/FTZ bits that can be set; I think RISC-V is in this category, although to be honest, trying to track down all the ISA extensions to make sure is a bit beyond my ken.

I kinda do long for the day when daz/ftz can be consigned to the same sort of dustbin of history as, say, EBCDIC.

> For some(?) Linux targets, we've got the mess I'm touching here where the driver is looking for crtfastmath.o and trying to inspect the command-line for fast-math settings outside the other FP option rendering, and then the FP option rendering, which is meant to be target-independent, is kind of making a mess of it, especially with inconsistent handling of "denormal-fp-math" and "denormal-fp32-math".

The issue here I think is that of the Linux targets, only the x86 people cared enough to get the denormal-fp-math set correctly, but the logic actually applies across targets. It's a real mess since some targets flush to positive zero and some targets flush to signed zero and some manuals I couldn't make heads or tails out of what they actually did.

https://github.com/llvm/llvm-project/pull/89477