use C++ LLVM API rather than C API for array types to support 64 bits

[andrewrk]

See #1422

fn recur() [4294967397]u8 {
    return recur();
}

export fn entry() void {
    var loophole = recur();
}

const builtin = @import("builtin");
pub fn panic(msg: []const u8, error_return_trace: ?*builtin.StackTrace) noreturn {
    while (true) {}
}

the LLVM IR for this is:

; Function Attrs: nobuiltin nounwind
define internal fastcc void @recur([101 x i8]* nonnull sret) unnamed_addr #2 !dbg !53 {
Entry:

The 101 is because it wrapped around 32 bits. This is because the LLVM C API uses unsigned but the C++ API uses uint64_t: static ArrayType *get(Type *ElementType, uint64_t NumElements);

So if we switch to the C++ API to make LLVM array types it fixes this bug.

[BarabasGitHub]

I seriously hope nobody will ever need a fixed size array larger than 4GB.

[thejoshwolfe]

I seriously hope nobody will ever need a fixed size array larger than 4GB.

It's definitely possible, especially when you're not running in an OS.

[nektro]

triaging upstream as well llvm/llvm-project#56496

[nektro]

I submitted a patch to enable the C api to take advantage of u64 array lengths which has now landed as llvm/llvm-project@35276f1 and will likely make it as part of LLVM 17

[nektro]

completed in 6a07b70

[andrewrk]

Nice work @nektro!