Linear IR broke llvmcall(something(...), ...)

[maleadt]

Not sure whether this is now unsupported, or a bug. Repro:

ptr = Ptr{Void}(1)
x() = Base.llvmcall(identity(ptr), Void, Tuple{})
x()

Obviously this should crash anyhow, but it does so because the ir argument to llvmcall is null, making the check for jl_is_tuple crash:

(gdb) break emit_llvmcall
Function "emit_llvmcall" not defined.
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (emit_llvmcall) pending.
(gdb) run

Thread 1 "julia" hit Breakpoint 1, emit_llvmcall (ctx=..., args=0x7fffe66ce4a0, nargs=3) at src/ccall.cpp:955
955         JL_NARGSV(llvmcall, 3);
(gdb) n
956         jl_value_t *rt = NULL, *at = NULL, *ir = NULL, *decl = NULL;
(gdb) 
957         JL_GC_PUSH4(&ir, &rt, &at, &decl);
(gdb) 
958         at = try_eval(ctx, args[3], "error statically evaluating llvmcall argument tuple", true);
(gdb) 
959         rt = try_eval(ctx, args[2], "error statically evaluating llvmcall return type", true);
(gdb) 
960         ir = try_eval(ctx, args[1], "error statically evaluating llvm IR argument", true);
(gdb) 
961         int i = 1;
(gdb) 
962         if (jl_is_tuple(ir)) {
(gdb) p ir
$1 = (jl_value_t *) 0x0
(gdb) n

Thread 1 "julia" received signal SIGSEGV, Segmentation fault.
0x00007ffff759f5fd in emit_llvmcall (ctx=..., args=0x7fffe66ce4a0, nargs=3) at src/ccall.cpp:962
962         if (jl_is_tuple(ir)) {

Breaks LLVM.jl, where the argument is LLVM.ref(fun::LLVM.Function)::Ptr{LLVM.Value}.

Ref #24113.

[vtjnash]

That's not really a bug, since the argument is required to be a literal.

[StefanKarpinski]

Can we avoid segfaulting?

[maleadt]

That's not really a bug, since the argument is required to be a literal.

Fair enough.