Final project blog post: Mark Barbone #537
Conversation
sampsyo
left a comment
sampsyo
left a comment
There was a problem hiding this comment.
Really awesome work, Mark—this is fantastic. Thanks for the clear explanation of your MLIR-inspired IR and the overview of all the steps in the chain required to make the whole round trip work.
This seems like a fun foundation that someone might want to build on in the future. Would you be interested in contributing your machinery back to the Bril monorepo to perhaps help facilitate that?
| A jump to label `.l` is only allowed when `.l` is in scope (this is what makes | ||
| it structured!). Jumping to a `block` label is like a `break`, escaping to the | ||
| end of the block, while jumping to a `loop` label is like a `continue`, going | ||
| back to the loop header for the next iteration. |
There was a problem hiding this comment.
This is a clever way to repurpose Bril labels and jmp in a structured representation.
| flow, or around 65%. Considering just the relooper, interestingly, all 272 | ||
| functions in these combined test suites exhibit reducible control flow, and | ||
| successfully round-trip through just the structured control flow IR. |
There was a problem hiding this comment.
I agree that this is interesting! I would have expected some benchmark in there to do something weird with its control flow.
I would be interested in this! But I think it deserves some polishing before contributing it. Namely,
Let me know your thoughts -- if this seems reasonable, I'll do it sometime over the summer. |
|
Great! Sure, all of these sound quite reasonable. FWIW, I wouldn't bother trying to create a JSON representation for the static-control-flow form of Bril… that sounds great to have, but not necessary for an "MVP" that we could merge. Concocting a sensible serialization format for ISL sets seems like an "interesting" design problem that we could leave to future generations. |
2 participants
Closes #508 .