[iree][codegen] Add #iree_codegen.denormal_fp_math to set denormals behavior

[fabianmcg]

Adds support for configuring denormal floating point math mode for f32 operations
on AMD GPUs. This includes:

• DenormalFpMath enum with None, PreserveSign, and PositiveZero modes
• Command line flag --iree-hip-denormal-fp-math-f32 to control the behavior
• Attribute propagation through to LLVM kernels

I based the options on https://llvm.org/doxygen/structllvm_1_1DenormalMode.html#a29b26e3ae30f3f6ec4106ff181282893

[kuhar]

Having this attribute at the level of an executable target makes sense to me but I wonder if we may end up requiring a finer granularity in the future

(This is a non-blocking review, I need to think about this more...)

[kuhar]

@benvanik do you want to give this a quick look?

[fabianmcg]

Having this attribute at the level of an executable target makes sense to me but I wonder if we may end up requiring a finer granularity in the future

I talked with @Groverkss about this, as he expressed similar concerns. We definitely need a finer-grained mechanism that has the info per dispatch. However, having a global value is still useful, as any dispatches without a more specialized value can use the global.

I'd also argue that the current mechanism around target attributes, flags, ..., is not very extensible. It likely should be revisited, and we should likely consider using the work Renato and Rolf are doing.

[krzysz00]

I'd also like to flag that fp16 and fp64 denormal handling is a separate flag ... but also, in many cases, the same flag

[kuhar]

LGTM but please wait for a review from Ben before merging

[qedawkins]

Just a couple comments about naming and code location, otherwise LGTM. Thanks for making it work in IR and not just behind a flag!

[MaheshRavishankar]

Is the flag added here for testing purpose? This is going to flush all the denormals right? I thought the point was that we only do this for dispatches that we know this is OK to do?

[fabianmcg]

This is to have a fallback option to set a default value, similar to global behavior flags in clang. The other patch, will allow to set it for specific dispatches. If that's not wanted I can drop this patch.

[MaheshRavishankar]

This looks fine to me. Thanks!