GH-128682: Make PyStackRef_CLOSE escaping.
#129404
Conversation
| @@ -330,6 +337,16 @@ def convert_stack_item( | |||
| cond = replace_op_arg_1 | |||
| return StackItem(item.name, item.type, cond, item.size) | |||
|
|
|||
| def check_unused(stack: list[StackItem], input_names: dict[str, lexer.Token]): | |||
| "Unused items cannot be on the stack above used, non-peek items" | |||
There was a problem hiding this comment.
What's the rationale for this?
There was a problem hiding this comment.
We should probably remove any special treatment of variables named "unused" as they get skipped by the analysis, leading to some incorrect stacks. This is just avoiding one of those cases.
|
Benchmarking shows a 0.2% speedup on Linux (noise) and 1.2% slowdown on Windows. The slowdown on Windows includes an 8% slowdown on However, even with Looking at the generated code and the latest execution counts, the instructions changed account for 2-3% of the total instructions executed which I would expect to cause a slowdown of no more than 0.2%. |
Fidget-Spinner
added
the
🔨 test-with-refleak-buildbots
|
🤖 New build scheduled with the buildbot fleet by @Fidget-Spinner for commit 5dc7977 🤖 Results will be shown at: https://buildbot.python.org/all/#/grid?branch=refs%2Fpull%2F30617%2Fmerge If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again. |
bedevere-bot
removed
the
🔨 test-with-refleak-buildbots
|
Looks good, but worried about the close and stackref changes. Running with refleak buildbots. |
|
The buildbots are failing This fails on main: |
|
Wait did we address the slowdowns on Windows? |
Makes
PyStackRef_CLOSEescaping.PyStackRef_CLOSEkills the variable, and then escapes.This PR adds an
EscapingCalldataclass to the analysis, with a newkillattribute.PyStackRef_CLOSEandPyStackRef_XCLOSEset thekillattribute, so the code generator knows to kill the argument before spilling.Doing so, means that
pop_N_errorlabels no longer work as we need to adjust the stack pointer before allPyStackRef_CLOSErather than once at the jump to error.Performance numbers to come...