Repository navigation
Optimize reference tracking and eliminate branching during returns and yields #144540
Description
Activity
- added a commit that references this issue
on Mar 11, 2026 @cocolato I suspect we can do more with _MAKE_HEAP_SAFE, e.g. if we have the following (return of a return)
_MAKE_HEAP_SAFE _RETURN_VALUE _MAKE_HEAP_SAFE <----- this can be eliminated! _RETURN_VALUEThis shows up commonly in the
bm_generatorsbenchmark in pyperformance.Reacted by Hai ZhuThis shows up commonly in the
bm_generatorsbenchmark in pyperformance.@Sacul0457 would you like to work on this? It's just adding this conditional in _MAKE_HEAP_SAFE:
+ if (sym_is_immortal(PyJitRef_Unwrap(value)) || !PyJitRef_IsBorrowed(value)) {And adding a test.
@cocolato if you don't mind, I'm sensing you're a little busier now, so I will let Sacul take over for this one. I have some very interesting problems for you in the future :). Thanks a lot for all the work you've done so far!
Reacted by Hai Zhu@cocolato if you don't mind, I'm sensing you're a little busier now, so I will let Sacul take over for this one. I have some very interesting problems for you in the future :). Thanks a lot for all the work you've done so far!
Never mind. I’ve been little busy because I’m in the middle of a job handover, but I should have plenty of time starting in early April. I’ll also look into some of the existing issues then. If there's anything I can do to help, please let me know!
Reacted by Ken Jin and SaculThank you both! I'll look into this in a few hours!
Reacted by Hai Zhu
There are a few optimizations we can make to returns and yields, but they are somewhat related so I'm grouping them into a single issue.
During a
return, the VM needs to make the returned reference "heap safe" (convert any borrowed references to strong references) and then pop and clear the frame, which can involve quite a lot of decrefs.During a
yield, the VM needs to make the returned reference heap safe and then pop the frame, but not clear it.We want to minimize the amount of refcounting operations that we do.
Before we can do much else, we should split
RETURN_VALUEandYIELD_VALUEinto micro-ops:so that we can optimize the uops independently
Eliminate making heap safe if we know that reference already is
If TOS is a strong reference
_MAKE_HEAP_SAFE->_NOPAvoid the branch in
_RETURN_VALUEby splitting into_RETURN_VALUE_GENand_RETURN_VALUE_FUNCTo avoid using up opcodes, we'll need to do this in the JIT, not the interpreter.
Track borrows/immortals to optimize clearing frames
E.g. if a frame has four local variables,
a,b,c,dbut we can tell thataanddare immortal, or borrowed, then instead of looping over all the variables, we could emit code just to decrefbandc.This might end up bloating the code, as we would need to inline the decrefs, but it could be quite a lot faster.
Linked PRs
_MAKE_HEAP_FUNCTIONfor owned references #146320