[PATCH] D24587: [RFC] Output optimization remarks in YAML

[Hal Finkel via llvm-commits]

hfinkel added a comment.

In https://reviews.llvm.org/D24587#548958, @anemet wrote:

> Ping.  @hfinkel, would be great if you could comment on the direction in this RFC.  I don't want to start updating the optimization before we settle on the API.  Thanks!


Thanks for working on this! I think this direction makes sense. I think that we'd like the remark classes to be serializable as YAML, but we probably don't want the file writing code to be in the backend. The diagnostic classes get passed to the handler, which is often in the frontend, and we should probably let the handler decide what to do with them (e.g. where/if to write them to a file).

> - string:  will not be inlined into


We don't need to fix this all at once, but we need to think carefully about how remarks are identified. I think we want it to be very clear how each diagnostic is identified, and we don't want strings that compose the human-readable summary message to subtly also form an interface contract with external tools. So the string here is fine, but we'll want each diagnostic to have some name that is designed to be machine consumed, and probably key-value pairs which likewise have machine-consumable names.


https://reviews.llvm.org/D24587