Add support for SPV_INTEL_function_variants

[jzadnik]

This PR:

• Adds validation support for the SPV_INTEL_function_variants extension
• Augments the linker to be able to produce multitarget modules described in the extension spec. Details of how it works are documented in fnvar.h.

NOTE: This PR requires #6210 for the tests to pass because inline assembly is used to test conditional extensions.

The values to pass to OpSpecConstantArchitectureINTEL and OpSpecConstantTargetINTEL can be determined from targets registry.

[s-perron]

I've reviewed the spirv-opt changes. I'm actually surprized you do not need more changes. You add tests to spirv-opt. Mainly for the remove duplicates pass.

[s-perron]

I looked at some of the generic files. I believe you need a lot more testing. The way the extension was defined is that it created parallel opcode, but shifted the operands by 1. This means that every place that looks at the capabilities, extensions, ect. must make sure the operand indices are shift by one. I'm not sure that have been thoroughly checked or tests.

[jzadnik]

Thanks to the review. I fixed the issues you pointed out and added more tests.

The way I identified where the index should be shifted by one is that I searched for OpCapability/Extension/EntryPoint in the source. A more thorough check can be done, though I haven't found a systematic way to do it.

Could you also check the other PR I made (see this PR's description)? It's used to test conditional extensions and should be merged before this PR, otherwise the tests won't pass.

I'll be on vacation for the next three weeks, if more changes are required, I'll address them after that.

[alan-baker]

Validator changes are good.

[s-perron]

The changes to the optimizer look good. I still need to look at the code in the linker.

[s-perron]

The FunctionVariants.FAddAsm test seems to be failing. That is a linker test. Can you get that test fixed, and then someone will look at the linker code. Thanks.

[dneto0]

Partial review of the linking material.

I'm up to about line 200 in link/fnvar.cpp but have read the other parts of the linker changes.

[jzadnik]

I rebased this PR onto the latest main which has the inline assembly patch merged. This should fix the FunctionVariants.FAddAsm (all tests pass on my machine).

[dneto0]

Also, thanks for your patience.

Between the three reviewers so far I think we've approved all parts of this MR.

[jzadnik]

Thanks for the thorough review!

[dneto0]

This broke GN builds in downstream projects.
I'm looking into it.

E.g. vulkan-deps into dawn. https://dawn-review.googlesource.com/c/dawn/+/257354

(Not blaming anyone: GN is niche, and sometimes hard to get right)