[SYCL] Simplify arguments to computeModuleProperties

[sarnex]

Computing information about spec consts is difficult for callers when module properties generation and splitting are separate. Simplify it by storing information in module we can use invisibly to the caller.

This will be used for thinLTO, where we split early and compute module properties later.

This should be basically NFC.

[maksimsab]

Looks good to me.

[sarnex]

@AlexeySachkov Do you mind reviewing this too? Thanks

[sarnex]

@AlexeySachkov ping on this one when you get a sec

[AlexeySachkov]

I don't think that my comments are blocking, but to me they seem as a good cleanup which we can do

[AlexeySachkov]

Both collectSpecConstantMetadata and collectSpecConstantDefaultValuesMetadata perform this metadata check:

llvm/llvm/lib/SYCLLowerIR/SpecConstants.cpp

Lines 1035 to 1039 in 9723157

	bool SpecConstantsPass::collectSpecConstantMetadata(const Module &M,
	SpecIDMapTy &IDMap) {
	NamedMDNode *MD = M.getNamedMetadata(SPEC_CONST_MD_STRING);
	if (!MD)
	return false;

Therefore, I wonder if we just need to omit it all together and instead put PropSet.add under condition that TmpSpecIDMap is not empty.

[sarnex]

thanks, ill try this and make a follow-up PR if it works

[AlexeySachkov]

Conceptually, HandlingMode is both set outside of the pass and read outside of the pass (that property to indicate that a device image was produced with default values of spec constants), so it seems to me that that metadata is better to be set somewhere outside the pass.

[sarnex]

yeah i debated this, let me try putting it somewhere when we do the default split, will do in follow-up pr thx

[sarnex]

Follow-up PR #15346