Why I am here? What's going on?

Ivan Butygin ivan.butygin@intel.com

Can someone merge this, please?

Updated: make a custom printers which won't print fastmath flags if they are equal to the default value, to avoid changes in unrelated tests

I mean the only change I need is to have tblgen to generate

if (fmf != <default value>)
  odsState.addAttribute("fmf", ::mlir::LLVM::FMFAttr::get(fmf, odsBuilder.getContext()));

instead of unconditional

odsState.addAttribute("fmf", ::mlir::LLVM::FMFAttr::get(fmf, odsBuilder.getContext()));

in build methods for DefaultValuedAttr.
In that case IR with default flags should look like exactly as before without OptionalAttr.
Does it make sense?
Is it worth an RFC?

let's make the attribute optional, so the absence thereof indicates there are no flags

Some more comments:

Updated. Went with custom attribute approach as there aren't any good decision how to handle build methods with bunch of unit attrs.

In D92485#2428754, @ftynse wrote:

Thanks for the patch. I would recommend sending an RFC to https://llvm.discourse.group/ to discuss the proposed design. Personally, I would expect us to have something like a set of unit attributes with appropriate names as opposed to array-of-str-enum. It is not clear if we want a trait or an interface for this purpose (note that you cannot dyn_cast to a trait class, and unit attribute accessors are generated automatically).

tests TBD