https://github.com/python/cpython/commit/c49dc3bd0f0ac57a038e9ba63e83cfa424e91b2a
commit: c49dc3bd0f0ac57a038e9ba63e83cfa424e91b2a
branch: main
author: Brandt Bucher <[email protected]>
committer: brandtbucher <[email protected]>
date: 2025-07-09T12:11:28-07:00
summary:
GH-115802: Optimize JIT stencils for size (GH-136393)
files:
M Tools/jit/_targets.py
diff --git a/Tools/jit/_targets.py b/Tools/jit/_targets.py
index ed10329d25d2f9..728f48128ce79c 100644
--- a/Tools/jit/_targets.py
+++ b/Tools/jit/_targets.py
@@ -137,7 +137,15 @@ async def _compile(
f"-I{CPYTHON / 'Include' / 'internal' / 'mimalloc'}",
f"-I{CPYTHON / 'Python'}",
f"-I{CPYTHON / 'Tools' / 'jit'}",
- "-O3",
+ # -O2 and -O3 include some optimizations that make sense for
+ # standalone functions, but not for snippets of code that are going
+ # to be laid out end-to-end (like ours)... common examples include
+ # passes like tail-duplication, or aligning jump targets with nops.
+ # -Os is equivalent to -O2 with many of these problematic passes
+ # disabled. Based on manual review, for *our* purposes it usually
+ # generates better code than -O2 (and -O2 usually generates better
+ # code than -O3). As a nice benefit, it uses less memory too:
+ "-Os",
"-S",
# Shorten full absolute file paths in the generated code (like the
# __FILE__ macro and assert failure messages) for reproducibility:
_______________________________________________
Python-checkins mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3//lists/python-checkins.python.org
Member address: [email protected]
