gh-111786: Optimize for space for _PyEval_EvalFrameDefault on MSVC for PGO (#111794)

In PGO mode, this function caused a compiler error in MSVC.
It turns out that optimizing for space only save the day, and is even faster.
However, without PGO, this is neither necessary nor slower.

diff --git a/PCbuild/pyproject.props b/PCbuild/pyproject.props
index fb12f2b0766d9b3248c29d3ae28dd34d93ab04a7..bb3555bd123089bcec281abe8dee9b7ff1c3a636 100644 (file)
--- a/PCbuild/pyproject.props
+++ b/PCbuild/pyproject.props
@@ -41,6 +41,7 @@
       <AdditionalIncludeDirectories>$(PySourcePath)Include;$(PySourcePath)Include\internal;$(PySourcePath)Include\internal\mimalloc;$(PySourcePath)PC;$(IntDir);%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <PreprocessorDefinitions>WIN32;$(_Py3NamePreprocessorDefinition);$(_PlatformPreprocessorDefinition)$(_DebugPreprocessorDefinition)$(_PydPreprocessorDefinition)%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <PreprocessorDefinitions Condition="'$(DisableGil)' == 'true'">Py_NOGIL=1;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions Condition="'$(SupportPGO)' and ($(Configuration) == 'PGInstrument' or $(Configuration) == 'PGUpdate')">_Py_USING_PGO=1;%(PreprocessorDefinitions)</PreprocessorDefinitions>
 
       <Optimization>MaxSpeed</Optimization>
       <IntrinsicFunctions>true</IntrinsicFunctions>

diff --git a/Python/ceval.c b/Python/ceval.c
index 68bed93eae3d86ae00a70ae266a23b2bbe656e7a..42100393f60ff721e7f07de248361a1c397abf31 100644 (file)
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -664,6 +664,15 @@ extern const struct _PyCode_DEF(8) _Py_InitCleanup;
  * so consume 3 units of C stack */
 #define PY_EVAL_C_STACK_UNITS 2
 
+#if defined(_MSC_VER) && defined(_Py_USING_PGO)
+/* gh-111786: _PyEval_EvalFrameDefault is too large to optimize for speed with
+   PGO on MSVC. Disable that optimization temporarily. If this is fixed
+   upstream, we should gate this on the version of MSVC.
+ */
+#  pragma optimize("t", off)
+/* This setting is reversed below following _PyEval_EvalFrameDefault */
+#endif
+
 PyObject* _Py_HOT_FUNCTION
 _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int throwflag)
 {
@@ -1080,6 +1089,7 @@ exit_trace:
 #  pragma GCC diagnostic pop
 #elif defined(_MSC_VER) /* MS_WINDOWS */
 #  pragma warning(pop)
+#  pragma optimize("", on)
 #endif
 
 static void