From: Brandt Bucher <brandtbucher@microsoft.com>
Date: Wed, 9 Jul 2025 19:11:28 +0000 (-0700)
Subject: GH-115802: Optimize JIT stencils for size (GH-136393)
X-Git-Tag: v3.15.0a1~1036
X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=c49dc3bd0f0ac57a038e9ba63e83cfa424e91b2a;p=thirdparty%2FPython%2Fcpython.git

GH-115802: Optimize JIT stencils for size (GH-136393)
---

diff --git a/Tools/jit/_targets.py b/Tools/jit/_targets.py
index ed10329d25d2..728f48128ce7 100644
--- a/Tools/jit/_targets.py
+++ b/Tools/jit/_targets.py
@@ -137,7 +137,15 @@ class _Target(typing.Generic[_S, _R]):
             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: