GH-105481: Mark more files as generated #107598
Conversation
brandtbucher
commented
Aug 3, 2023
•
brandtbucher
commented
- Issue: Generate opcode metadata from bytecodes.c instead of opcode.py #105481
| @@ -72,9 +72,11 @@ Doc/library/token-list.inc generated | |||
| Include/internal/pycore_ast.h generated | |||
| Include/internal/pycore_ast_state.h generated | |||
| Include/internal/pycore_opcode.h generated | |||
| Include/internal/pycore_opcode_metadata.h generated | |||
There was a problem hiding this comment.
I removed this from the list a couple of weeks ago to make it show up in the diff, to make sure people are aware when the instruction flags change.
There was a problem hiding this comment.
Agreed, let’s keep it out of the list for. This can catch accidental mistakes, or remind people of the consequences of intentional changes.
There was a problem hiding this comment.
Maybe we should leave it there commented out with an explanation so this won't come up again.
There was a problem hiding this comment.
Hm, really? I'm not sure I agree. This was prompted by the diff view of my recent PR, where it just adds noise.
We already hide other files with lists of tokens, keywords, and standard library module names that are arguably more important and more human-readable than this one. And people can still view them, GitHub just collapses them by default.
If we're this concerned about setting the correct flags, maybe we should just write some sort of test instead?
There was a problem hiding this comment.
Well, I'm not really sure what we're trying to protect against here. My understanding is that the flags are now generated automatically by analyzing the DSL's C code. If there's a bug somewhere, it's probably in the static analysis of the code, not the code itself (if I add a GETLOCAL to something, then I definitely want it to set the has "local" flag, ditto for JUMPBY and the "jump" flag).
So maybe tests for InstructionFlags.fromInstruction with some expected inputs and outputs? Or some error-checking in the generator that incompatible flag combinations (like "local" and "jump", or "const" without "oparg") don't happen?
There was a problem hiding this comment.
We could have a few of those. They would protect us from regressions when we change the static analysis code, or when we change the implementation of a bytecode that happens to be tested. But when you add a new bytecode, don't you want to look at that flags and see that they make sense?
There was a problem hiding this comment.
(I edited my comment above, adding that maybe the flags class could do some sanity checks on the flag combination.)
There was a problem hiding this comment.
But honestly I thought this would be uncontroversial. If you're getting value from the status quo, I can let it go. I just got kind of tired of scrolling past this file in PRs.
There was a problem hiding this comment.
Thinking more about this I think this PR is actually fine and we're just being overly paranoid. When changing the code generator I review all generated output anyways, and ditto when editing or adding bytecodes. But I'll leave the last word to @iritkatriel
|
Ok, we can try it this way. |
