GH-91719: Make MSVC generate somewhat faster switch code (#91718)

gvanrossum · web-flow · commit f8dc6186d185 · 2022-04-21T11:53:57.000-07:00
Apparently a switch on an 8-bit quantity where all cases are present generates a more efficient jump (doing only one indexed memory load instead of two). So we make opcode and use_tracing uint8_t, and generate a macro full of extra `case NNN:` lines for all unused opcodes. See faster-cpython/ideas#321 (comment)
diff --git a/Include/cpython/pystate.h b/Include/cpython/pystate.h
@@ -46,7 +46,7 @@ typedef struct _PyCFrame {
      * discipline and make sure that instances of this struct cannot
      * accessed outside of their lifetime.
      */
-    int use_tracing;
+    uint8_t use_tracing;  // 0 or 255 (or'ed into opcode, hence 8-bit type)
     /* Pointer to the currently executing frame (it can be NULL) */
     struct _PyInterpreterFrame *current_frame;
     struct _PyCFrame *previous;
diff --git a/Include/opcode.h b/Include/opcode.h
diff --git a/Python/ceval.c b/Python/ceval.c
@@ -1662,7 +1662,9 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #ifdef Py_STATS
     int lastopcode = 0;
 #endif
-    int opcode;        /* Current opcode */
+    // opcode is an 8-bit value to improve the code generated by MSVC
+    // for the big switch below (in combination with the EXTRA_CASES macro).
+    uint8_t opcode;        /* Current opcode */
     int oparg;         /* Current opcode argument, if any */
     _Py_atomic_int * const eval_breaker = &tstate->interp->ceval.eval_breaker;
 
@@ -5645,7 +5647,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #if USE_COMPUTED_GOTOS
         _unknown_opcode:
 #else
-        default:
+        EXTRA_CASES  // From opcode.h, a 'case' for each unused opcode
 #endif
             fprintf(stderr, "XXX lineno: %d, opcode: %d\n",
                     _PyInterpreterFrame_GetLine(frame),  opcode);
diff --git a/Tools/scripts/generate_opcode_h.py b/Tools/scripts/generate_opcode_h.py
@@ -129,6 +129,13 @@ def main(opcode_py, outfile='Include/opcode.h'):
         fobj.write("};\n")
         fobj.write("#endif\n")
 
+        fobj.write("\n")
+        fobj.write("#define EXTRA_CASES \\\n")
+        for i, flag in enumerate(used):
+            if not flag:
+                fobj.write(f"    case {i}: \\\n")
+        fobj.write("        ;\n")
+
         fobj.write(footer)
 
 
