[ClR] Set optnone attribute for functions.

[htyu]

Summary:
Setting the Optnone attribute for CIR functions and progating it all the way down to LLVM IR for those not supposed to be optimized.

[bcardosolopes]

Great to see more attributes getting added to LLVM lowering.

[bcardosolopes]

Thanks for driving this, more comments inline!

[bcardosolopes]

There are also two spelling issues with the subject: [ClR] Set optnone attrbiute for functions

[htyu]

Reviving this as it's going to be used by the global ctor lowering work later.

[bcardosolopes]

Sweet, thanks for continuing this. Few last round of remarks

[bcardosolopes]

I see the output is:

extra( {inline = #cir.inline<no>, optnone = #cir.optnone} )

Given that the idea is to be unit-attr-like, it seems too verbose to have optnone = #cir.optnone, can we instead have (this builds on the assumption that nothing will show up if this CIRUnitAttr doesn't get created, which is how unit-attr usually works):

extra( {inline = #cir.inline<no>, #cir.optnone} )

?

[htyu]

Talked offline. It looks like MLIR has a hack that only works for the MLIR UnitAttr



clangir/mlir/lib/IR/AsmPrinter.cpp

Line 2584 in 8e37695

if (llvm::isa<UnitAttr>(attr.getValue()))

. 

I can think to create a subclass to override that method (will need to change it to virtual first). The parser (

clangir/mlir/lib/AsmParser/AttributeParser.cpp

Line 323 in 8e37695

attributes.push_back({*nameId, builder.getUnitAttr()});

) will need to be overriden too. Doesn’t sound trivial. 

The problem is that extra func attr is a list of named attributes, so no direct way to control how a CIR attribute is printed there. The name will always be printed first.