Add pass to lower Bitcast to nonstandard type instructions

[KornevNikita]

It is a pass to lower bitcast instructions to non-standard SPIR-V types.
It covers only known issues. At this moment there is only one pattern
that should be covered - use of vector with unsupported number of
elements. This pattern should be handled this way:

%0 = bitcast <3 x i64> addrspace(1)* @id to <6 x i32> addrspace(1)*
%1 = addrspacecast <6 x i32> addrspace(1)* %0 to <6 x i32> addrspace(4)*
%2 = load <6 x i32>, <6 x i32> addrspace(4)* %1, align 32
%conv = extractelement <6 x i32> %2, i32 4
%conv1 = sitofp i32 %conv to float

Must be replaced by:
%0 = addrspacecast <3 x i64> addrspace(1)* @id to <3 x i64> addrspace(4)*
%1 = load <3 x i64>, <3 x i64> addrspace(4)* %0, align 32
%2 = extractelement <3 x i64> %1, i32 0
%conv = trunc i64 %2 to i32
%conv1 = sitofp i32 %conv to float

It is assumed that the pass will be further developed as new patterns arise.

[KornevNikita]

@AlexeySachkov @AlexeySotkin @Fznamznon take a look please

[Fznamznon]

I didn't dig deeply into all loops, because for me recursive loop over users of faulty bitcast seems more convenient.
We also should remember that the pass like this, i.e. changing types can be very hard to maintain, since there is always more unhandled patterns coming.

[AlexeySotkin]

We also should remember that the pass like this, i.e. changing types can be very hard to maintain, since there is always more unhandled patterns coming.

I can't think of a better solution to handle such types. @Fznamznon, if you have any ideas, please share.

[AlexeySachkov]

What if there are several addressspace casts which use the same instruction?

[KornevNikita]

Added TODO comment in 587cc7f

[AlexeySotkin]

What if there are several addressspace casts which use the same instruction?

You mean we don't need break at the end of the if? If yes, I agree.

[mlychkov]

In this case whole function should be refactored to process several patterns "BitCast -> AScast -> Load -> ExtractElement" in an IR function.

[AlexeySachkov]

Yes, we can't just remove the break, because the code assumes that there is only one instruction which needs to be fixed

[AlexeySachkov]

I think that we can merge this as some kind of baseline, i.e. in general algorithm works, but in order for it to be useful in LLVM IR produced from real applications, we must ensure that we resolve TODO about multiple uses of non-standard vector types (from SPIRVLowerBitCastToNonStandardType.cpp:77)

[mlychkov]

Thank you very much for your patience and willingness to accept constructive remarks.
Now all looks good to me.

[KornevNikita]

Clang-tidy warning about recursion can be ignored, because the variable "RecursionDepth" is added to control the depth, so it cannot loop.