Support typed LLVM getelementptr instructions

[HertzDevil]

This implements part of #12484 by allowing LLVM::Builder#gep and #inbounds_gep to accept an optional type argument. Given an LLVM instruction like:

%1 = getelementptr %T, %U* %0, i64 7

To maintain the same semantics before this change, %T and %U should remain identical types from the same context. There are now four possible scenarios:

• type is given, LLVM ≥ 8.0: %T is type, %U comes from value;
• type is omitted, LLVM ≥ 8.0: %T is value.type, %U comes from value;
• type is given, LLVM < 8.0: %T and %U come from value. The type argument is silently ignored;
• type is omitted, LLVM < 8.0: %T and %U come from value.

The compiler itself provides the type argument in all cases. The builder now always uses LLVMBuildGEP2 / LLVMBuildInBoundsGEP2 instead of the deprecated methods if LLVM 8.0 or above is detected. There are no breaking changes in this patch, and all the codegen specs have been locally tested on Nix shells for LLVM 7 and 14 on aarch64-apple-darwin.

Later a similar migration has to be done for #load / LLVMBuildLoad too.

[HertzDevil]

I have no idea what went wrong with the CI since I do not have an x86_64-darwin machine and LLVM 11.1 works locally.

[beta-ziliani]

I got the following error in LLVM 11.1.0 compiled with assertions:

Assertion failed: (PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()), function Create, file Instructions.h, line 941.

This is the assert failing: https://github.com/llvm/llvm-project/blob/1fdec59bffc11ae37eb51a1b9869f0696bfd5312/llvm/include/llvm/IR/Instructions.h#L939

[HertzDevil]

I have created a separate feature branch for the opaque pointer migrations. Let's keep them there until the whole set is complete

[beta-ziliani]

I'm quite certain that here is the problem. I looked into it because "unless struct get struct type" sounded weird. Locally I can build fine with every other place using gep2 but here. This said, I couldn't fix it and, weirdly, removing entirely this code (unless ... end) makes everything work. @HertzDevil if you have any clue, let me know, otherwise we call Ary.

[HertzDevil]

The llvm_type of any reference type is a pointer, llvm_struct_type is the pointee's layout:

class Foo
  @x = ""
end

# type                   # => Foo
# llvm_struct_type(type) # => %Foo = type { i32, %String* }
# llvm_type(type)        # => %Foo*
Foo.new

And indeed, @last's LLVM type is llvm_struct_type(type) due to allocate_aggregate, regardless of whether Foo is a value or reference.