[libspirv][NFC] Fix build warnings when -Wall -Wextra is enabled

[wenju-he]

No description provided.

[Maetveis]

Suggested change

	int scope, global __CLC_GENTYPE *dst, const local __CLC_GENTYPE *src,
	size_t num_gentypes, size_t stride, event_t event) {
	(void)scope;
	int /*scope*/, global __CLC_GENTYPE *dst, const local __CLC_GENTYPE *src,
	size_t num_gentypes, size_t stride, event_t event) {

NIT: I'm not sure if libclc has an established existing practice for unused parameters, if it does it might be better to follow that, but I personally prefer this style. See also C++ Core Guidelines - Unused parameters should unnamed

Applies to all the places where there are unused parameters of course, I don't want to comment on all of them :)

[Maetveis]

If op is ever not one of the above values that is an error made by the developer of libclc, and never the user, right?
In that case I think this check should be some kind of build-time controlled assertion that compiles to only __builtin_unreachable in release builds (basically LLVM_UNREACHABLE).

I would also find a structure like this better, in which case we don't need unreachable:

if (op == Reduce) {                                                         \
    result = __clc__SubgroupShuffle(x, __spirv_BuiltInSubgroupSize() - 1);  \
    *carry = result;                                                        \
} /* For InclusiveScan, use results as computed */                          \
else if (op == InclusiveScan) {                                             \
  result = x;                                                               \
  *carry = result;                                                          \
} else {                                                                    \
  LIBCLC_ASSERT(op == ExclusiveScan && "Unexpected operation");             \
  /* For ExclusiveScan, shift and prepend identity */                       \
  *carry = x;                                                               \
  result = __clc__SubgroupShuffleUp(x, 1);                                  \
  if (sg_lid == 0) {                                                        \
    result = IDENTITY;                                                      \
  }                                                                         \
}

Also we're basically expecting the optimizer to fold these conditionals out, correct?
In this case I think just as OP is a macro, we could have a EPILOG_OP. Though honestly using some C++ templates and lambdas in libclc is sounding more and more attractive.

[Maetveis]

If you'd like to get this in quickly and return to LIBCLC_ASSERT or similar in a follow up then I would make the code like the above and only have the "assert" as a TODO comment.

[Maetveis]

Ditto here.

[Maetveis]

How did this ever work before? Good catch, ...I think :).