compiler: fix llvmcall segfault with zero-sized types

[achaljhawar]

Fixes the segmentation fault that occurred when passing zero-sized types to llvmcall. The compiler now validates argument and return types and throws a clear error message instead of crashing.

Before

julia> f(x::T) where T = Base.llvmcall("ret i8 %0", Int8, Tuple{T}, x); f(nothing)
[26746] signal 11 (1): Segmentation fault

After

julia> f(x::T) where T = Base.llvmcall("ret i8 %0", Int8, Tuple{T}, x); f(nothing)
ERROR: llvmcall does not support zero-sized argument types

Changes

• Added validation in emit_llvmcall to detect zero-sized argument and return types
• Added tests to verify proper error handling

fixes: #60063

[LilithHafner]

Thanks! CI is failing with errors related to this PR, so maybe have a look at that. @vchuravy, would you be willing to give a closer review?

[vchuravy]

I would defer to @vtjnash for the ABI part

[JeffBezanson]

What about void functions?

[achaljhawar]

added a check for that

[vtjnash]

The semantics here of checking if the user made an egregious typo are still wrong. This is also well formed:

julia> mutable struct A end

julia> f() = Core.Intrinsics.llvmcall("ret ptr addrspace(10) null", A, Tuple{})

This sort of random checking for user error doesn't really make sense to me. The point of llvmcall is an unsafe escape hatch into unsafe territory. If you don't like that, just don't use it.