`LOAD_FAST_BORROW` not being used even when safe to do so, if value is live at BB end.

[markshannon]

This function

def f(x, y, c):
 return 1 + (x if c else y)

compiles to

 1 RESUME 0

 2 LOAD_SMALL_INT 1
 LOAD_FAST_BORROW 2 (c)
 TO_BOOL
 POP_JUMP_IF_FALSE 9 (to L1)
 NOT_TAKEN
 LOAD_FAST_BORROW 0 (x)
 BINARY_OP 0 (+)
 RETURN_VALUE
 L1: LOAD_FAST 1 (y)
 BINARY_OP 0 (+)
 RETURN_VALUE

Note that the load of y uses LOAD_FAST even though LOAD_FAST_BORROW is safe.

This becomes important with virtual iterators as the iterable for the loop is live at BB end.

Linked PRs

• gh-133672: Allow LOAD_FAST to be optimized to LOAD_FAST_BORROW #133721
• gh-133672: Allow LOAD_FAST to be optimized to LOAD_FAST_BORROW #148999

[markshannon]

@iritkatriel
@mpage

[ljfp]

@markshannon I've created this PR to modify the optimization condition in Python/flowgraph.c. The goal is to allow LOAD_FAST to be optimized to LOAD_FAST_BORROW even when the reference is still on the stack at the end of a basic block (i.e., the REF_UNCONSUMED flag is set).

If I got this right, this should be safe as long as the reference is not killed (SUPPORT_KILLED flag isn't set) and not stored as a local (STORED_AS_LOCAL flag isn't set).

With these changes, y in your original example now correctly uses LOAD_FAST_BORROW:

Python 3.15.0a0 (heads/fix-issue-133672:f038e7f472e, May 9 2025, 01:25:23) [GCC 15.1.1 20250425] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> def f(x, y, c):
... return 1 + (x if c else y)
... 
>>> import dis
>>> dis.dis(f)
 1 RESUME 0

 2 LOAD_SMALL_INT 1
 LOAD_FAST_BORROW 2 (c)
 TO_BOOL
 POP_JUMP_IF_FALSE 9 (to L1)
 NOT_TAKEN
 LOAD_FAST_BORROW 0 (x)
 BINARY_OP 0 (+)
 RETURN_VALUE
 L1: LOAD_FAST_BORROW 1 (y)
 BINARY_OP 0 (+)
 RETURN_VALUE

Could you take a look when you have a moment and confirm if this aligns with what you had in mind?

[ljfp]

@markshannon I opened a new PR for this in #148999.

I made a few mistakes with the previous PR, so I cleaned up the branch history and opened a replacement PR instead. The new PR should also address the failing tests from the earlier version.