[Comgr] Add atomic noise flags to SPIR-V flag extraction

[lamb-j]

Summary

• Add -fatomic-fine-grained-memory, -fatomic-ignore-denormal-mode, and -fatomic-remote-memory to Comgr's ValidSpirvFlags allowlist so they are preserved through the SPIR-V → relocatable compilation pipeline
• Update the spirv-to-reloc.hip lit test to cover extraction of these flags

[z1-cciauto]

PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/11/builds/210

[chinmaydd]

I wonder if there's a better way to maintain this list of "allowed" flags...

[z1-cciauto]

PSDB Build Link: http://mlse-bdc-20dd129:8065/#/builders/11/builds/211

[AlexVlx]

We probably need to include the negations i.e. the -fno forms as well. Also, for legacy purposes, might be nice to include -munsafe-fp-atomics and -mnounsafe-fp-atomics

[AlexVlx]

I wonder if there's a better way to maintain this list of "allowed" flags...

The long-term plan is / was to start with all known / allowed -cc1 flags, and then mask out the few that are an issue.

[chinmaydd]

The long-term plan is / was to start with all known / allowed -cc1 flags, and then mask out the few that are an issue

That makes sense, thanks for the clarification.

[lamb-j]

We probably need to include the negations i.e. the -fno forms as well. Also, for legacy purposes, might be nice to include -munsafe-fp-atomics and -mnounsafe-fp-atomics

Doesn't look like the -fno forms or aliases will make it to the CC1 invocation (so not in @llvm.cmdline). They get converted to one of the above, or omitted (as they're default off)

Do we think the defaults for these will ever change? If so we could future-proof by adding the -fno versions I guess