[clang] [llvm] Reland "[HLSL] Implement the reflect HLSL function" (PR #125599)

[Farzon Lotfi via cfe-commits]

farzonl wrote:

>I am indeed asserting that many of the other uses of report_fatal_error in this file are also wrong. However, you're probably right that it makes sense to the existing pattern for now and clean this all up in a cohesive change.

Thanks for clarifying. Making a comprehensive change works for me.

> If the case is impossible why isn't this just an assert? In any case "failing immediately" and "causing the program to crash" are different things, and report_fatal_error is the latter.

Its impossible from clang, but anyone can invoke this intrinsic from opt\llc and so wanted a fail fast that we could test so no one thought they could use this intrinsic if the extention mode was opencl.

> IMO this is an argument for making a change that adds error handling to both functions before this PR, so that we don't end up in a state where the codebase is inconsistent.

I actually agree with this, but this funnction was marked `[[maybe_unused]]` so felt we need a use before making alteration. Plus we need an opencl counter part to test and we didn't have that until  #124045 was filed.



https://github.com/llvm/llvm-project/pull/125599