Add __builtin_spirv_ internal builtins

[MrSidims]

Way they are implemented is described in:
#3221

The PR also adds SPV_EXT_float8 extension and packed conversions for
SPV_INTEL_int4

Currently only conversion instructions (and internal builtins) are
covered.

TODO: in the following PR Saturation decoration will be added.

Signed-off-by: Sidorov, Dmitry [email protected]

[MrSidims]

New LLVM intrinsic work is in progress, some draft implementation can be seen here: MrSidims/llvm-project@512a282 . But as I have said in https://discourse.llvm.org/t/rfc-spir-v-way-to-represent-float8-in-llvm-ir/87758 - we still need some builtins for the translator for older LLVM branches.

[MrSidims]

Applied most of the comments from @YuriPlyakhin . Thanks!

[YuriPlyakhin]

LGTM, there are just couple nit comments left.

[maarquitos14]

In general, LGTM, just a few details and nits here and there. A general comment, more of a nit actually, I think code is more readable when explicit types are used instead of auto (except if they are nasty, unreadable types such as iterators or similar). It's just a personal preference, so feel free to ignore.

[maarquitos14]

Nit: We repeat this double check a few times, might be worth considering a utility function to group them (e.g. IsTypeFP8).

[MrSidims]

There were some reasons, why I decided to drop such utility functions, but I don't remember particularly why. Lets push few other changes in the future and then decide if we want to add them.

[maarquitos14]

Applies to all tests: should we add a spirv-val check?

[MrSidims]

spirv-val conversions_scalar_vector.spv
error: line 7: Invalid capability operand: 4212

spirv-val is not yet ready

[maarquitos14]

Should we still add it with RUNx or add at least a TODO?

[maarquitos14]

The test doesn't actually do this: it checks if a function with __spirv_builtin placed in the middle of the name is translated correctly. It's subtly different, but I think what we want to do is what is described here, so the problem is in the function name and the checks, right?

[MrSidims]

Changed this a bit

[maarquitos14]

Sorry, I wasn't clear. The internal builtin prefix is __builtin_spirv, but the function name we are trying to call is _Z18boo__spirv_builtinfs. My understanding is that we would want to test _Z18boo__builtin_spirv_fs instead. Is that right?

[MrSidims]

Got you, updated the test

[maarquitos14]

Note that the order is still reversed: the internal builtin prefix is __builtin_spirv, but the function name contains __spirv_builtin. My understanding is that we want to check if having the internal builtin prefix in the middle of a function causes any trouble, and the function name does not contain the internal builtin prefix. Ping me offline if you need further clarification, I feel like I'm not being able to deliver my message correctly.

[MrSidims]

Ah, indeed, thanks!