declare half @air.sincos.f16(half, half* nocapture writeonly)

vs

declare half @air.sincos.f16(half, i64)

So the problem is we aren't passing a pointer, but an i64. This can be fixed by using @typed_ccall from LLVM.jl instead of plain ccall.

It's also probably useful to add those annotations to

GPUCompiler.jl/src/metal.jl

Lines 1161 to 1203 in 47a4a8b

	function annotate_air_intrinsics!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
	changed = false
	for f in functions(mod)
	isdeclaration(f) || continue
	fn = LLVM.name(f)
	attrs = function_attributes(f)
	function add_attributes(names...)
	for name in names
	if LLVM.version() >= v"16" && name in ["argmemonly", "inaccessiblememonly",
	"inaccessiblemem_or_argmemonly",
	"readnone", "readonly", "writeonly"]
	# XXX: workaround for changes from https://reviews.llvm.org/D135780
	continue
	end
	push!(attrs, EnumAttribute(name, 0))
	end
	changed = true
	end
	# synchronization
	if fn == "air.wg.barrier" || fn == "air.simdgroup.barrier"
	add_attributes("nounwind", "convergent")
	# atomics
	elseif match(r"air.atomic.(local|global).load", fn) !== nothing
	# TODO: "memory(argmem: read)" on LLVM 16+
	add_attributes("argmemonly", "readonly", "nounwind")
	elseif match(r"air.atomic.(local|global).store", fn) !== nothing
	# TODO: "memory(argmem: write)" on LLVM 16+
	add_attributes("argmemonly", "writeonly", "nounwind")
	elseif match(r"air.atomic.(local|global).(xchg|cmpxchg)", fn) !== nothing
	# TODO: "memory(argmem: readwrite)" on LLVM 16+
	add_attributes("argmemonly", "nounwind")
	elseif match(r"^air.atomic.(local|global).(add|sub|min|max|and|or|xor)", fn) !== nothing
	# TODO: "memory(argmem: readwrite)" on LLVM 16+
	add_attributes("argmemonly", "nounwind")
	end
	end
	return changed
	end

Originally posted by @maleadt in #530