gh-133672: Allow LOAD_FAST to be optimized to LOAD_FAST_BORROW

[ljfp]

The LOAD_FAST_BORROW instruction (op code 86) loads a borrowed reference onto the operand stack, which is a performance optimization that avoids unnecessary reference counting operations.

Previously, we were only applying this optimization when the reference was consumed within the same basic block. If the value was still on the stack at the end of a basic block (indicated by the REF_UNCONSUMED flag), we wouldn't perform the optimization.

However, (if I understood this correctly) there are cases where it's perfectly safe to use LOAD_FAST_BORROW even when the value is still on the stack at the end of a basic block. The optimization is safe as long as:

1. The supporting reference in the frame isn't killed before the borrowed reference is consumed (checked by the SUPPORT_KILLED flag)
2. The borrowed reference isn't stored as a local (checked by the STORED_AS_LOCAL flag)

This fix allows us to optimize more cases, which is seems to be particularly important for the virtual iterators implementation (PR #132555) where the iterable for a loop is often live at basic block end.

• Issue: LOAD_FAST_BORROW not being used even when safe to do so, if value is live at BB end. #133672

[bedevere-app]

Most changes to Python require a NEWS entry. Add one using the blurb_it web app or the blurb command-line tool.

If this change has little impact on Python users, wait for a maintainer to apply the skip news label instead.

[mpage]

Thanks for working on this! This isn't safe to do unconditionally and will require an analysis that operates on the entire CFG, rather than per basic block. It's only safe to optimize a LOAD_FAST instruction that leaves a value on the stack at the end of a basic block if the other two conditions hold along all paths between the end of the basic block and when the value is popped from the stack. For example, it's not safe to optimize the first LOAD_FAST in bb0 below because the local is overwritten before the value is consumes from the stack in bb1:

bb0:
  LOAD_FAST 0
  LOAD_FAST 1
  TO_BOOL.
   POP_JUMP_IF_FALSE <bb2>

bb1:
  LOAD_CONST 0
  STORE_FAST 0
  RETURN_VALUE

bb2:
  RETURN_VALUE

[bedevere-app]

A Python core developer has requested some changes be made to your pull request before we can consider merging it. If you could please address their requests along with any other requests in other reviews from core developers that would be appreciated.

Once you have made the requested changes, please leave a comment on this pull request containing the phrase I have made the requested changes; please review again. I will then notify any core developers who have left a review that you're ready for them to take another look at this pull request.

[ljfp]

@mpage Thanks for your feedback! I've pushed an update that should address cases like the one you mentioned.

The main change involves doing, as suggested, a cfg analysis (using dfs via is_borrow_safe and check_borrow_safety_globally functions). This analysis now:

• Tracks the borrowed value across basic block boundaries if it's live at the end of a block.
• Checks along all execution paths if the original local variable backing the borrow is overwritten before the borrowed value is consumed.
• Verifies that the borrowed value itself isn't stored into another local variable.
• Detects and handles cycles in the CFG to ensure termination and safety.

I think this way we are properly handling the scenarios like the one you provided and ensuring the LOAD_FAST_BORROW optimization is applied safely.

Could you please take another look when you have a moment?

To be honest, this dived deeper into flowgraph.c than I initially anticipated, and it's a bit outside my usual comfort zone, so any further guidance would be much appreciated if I've missed something or if there are better ways to approach this.

[Fidget-Spinner]

Could you please check out the test failures? I'll try to give this PR a look if Matt doesn't get to it first :).

[python-cla-bot]

The following commit authors need to sign the Contributor License Agreement:

• 35541026+Prometheus3375@users.noreply.github.com

[ljfp]

Why was this PR closed? I was just synching my branch with the latest changes from main, after receiving an email saying my PR was marked "stale".

[ZeroIntensity]

Your rebase messed something up, which ended up tagging every codeowner, so everyone is going to get pings in their inbox for all activity on this PR. Please open a new PR.