[libclc] CLC (ab)uses ASM labels to emit llvm intrinsics instead of using Clang Builtins

[Maetveis]

LibCLC uses the Asm labels feature to directly call LLVM intrinsic functions from OpenCL code.
For an example overloads of __clc_round

llvm-project/libclc/generic/lib/math/round.cl

Lines 3 to 6 in 1b6340d

	// Map the llvm intrinsic to an OpenCL function.
	#define __CLC_FUNCTION __clc_round
	#define __CLC_INTRINSIC "llvm.round"
	#include <clc/math/unary_intrin.inc>

llvm-project/libclc/clc/include/clc/math/unary_intrin.inc

Line 29 in 1b6340d

_CLC_OVERLOAD half8 __CLC_FUNCTION(half8 d) __asm(__CLC_INTRINSIC ".v8f16");

would look roughly like the following after preprocessing:

__attribute__((overloadable)) half8 __clc_round(half8 d) __asm("llvm.round.v8f16");

For many (all?) intrinsics that are called like this Clang has builtins that lower to the same llvm intrinsics. For the above example __builtin_round or __builtin_elementwise_round could be used instead.

This might be problem because Clang will emit different IR when it knows it's calling an intrinsic and not a real function. Such unusual IR might need to be worked around in the LLVM backends of targets.

As an example for the SPIR-V target function calls receive the spir_func calling convention by default, while intrinsic calls will not. Demonstration on Compiler Explorer

EDIT: Technically the IR generated by clang even has UB according to LangRef:

LLVM functions, calls and invokes can all have an optional calling convention specified for the call. The calling convention of any pair of dynamic caller/callee must match, or the behavior of the program is undefined.

[frasercrmck]

Agreed! I've been getting rid of them from the codebase as I refactor (see e.g., 06789cc). As you note, there are still some remaining. I hope that soon all of the users of the *_intrin.inc "helpers" will be gone.

There are also some AMDGPU-specific intrinsics being used. I haven't looked but I bet those have target-specific builtins too.

[Maetveis]

Agreed! I've been getting rid of them from the codebase as I refactor (see e.g., 06789cc). As you note, there are still some remaining. I hope that soon all of the users of the *_intrin.inc "helpers" will be gone.

Nice, thank you! Feel free to use this issue to track progress or close it when the work is complete.