gh-111786: Use separate opcode vars for Tier 1 and Tier 2 (#112289)

This makes Windows about 3% faster on pyperformance benchmarks.

Author: mdboom

Python/ceval.c

@@ -689,7 +689,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #ifdef Py_STATS
     int lastopcode = 0;
 #endif
-    int opcode;        /* Current opcode */
+    uint8_t opcode;    /* Current opcode */
     int oparg;         /* Current opcode argument, if any */
 #ifdef LLTRACE
     int lltrace = 0;
@@ -776,9 +776,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     /* Start instructions */
 #if !USE_COMPUTED_GOTOS
     dispatch_opcode:
-        // Cast to an 8-bit value to improve the code generated by MSVC
-        // (in combination with the EXTRA_CASES macro).
-        switch ((uint8_t)opcode)
+        switch (opcode)
 #endif
         {
 
@@ -822,7 +820,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #if USE_COMPUTED_GOTOS
         _unknown_opcode:
 #else
-        EXTRA_CASES  // From pycore_opcode.h, a 'case' for each unused opcode
+        EXTRA_CASES  // From pycore_opcode_metadata.h, a 'case' for each unused opcode
 #endif
             /* Tell C compilers not to hold the opcode variable in the loop.
                next_instr points the current instruction without TARGET(). */
@@ -994,28 +992,29 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 
     OPT_STAT_INC(traces_executed);
     _PyUOpInstruction *next_uop = current_executor->trace;
+    uint16_t uopcode;
 #ifdef Py_STATS
     uint64_t trace_uop_execution_counter = 0;
 #endif
 
     for (;;) {
-        opcode = next_uop->opcode;
+        uopcode = next_uop->opcode;
         DPRINTF(3,
                 "%4d: uop %s, oparg %d, operand %" PRIu64 ", target %d, stack_level %d\n",
                 (int)(next_uop - current_executor->trace),
-                _PyUopName(opcode),
+                _PyUopName(uopcode),
                 next_uop->oparg,
                 next_uop->operand,
                 next_uop->target,
                 (int)(stack_pointer - _PyFrame_Stackbase(frame)));
         next_uop++;
         OPT_STAT_INC(uops_executed);
-        UOP_STAT_INC(opcode, execution_count);
+        UOP_STAT_INC(uopcode, execution_count);
 #ifdef Py_STATS
         trace_uop_execution_counter++;
 #endif
 
-        switch (opcode) {
+        switch (uopcode) {
 
 #include "executor_cases.c.h"
 
@@ -1053,7 +1052,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     STACK_SHRINK(1);
 error_tier_two:
     DPRINTF(2, "Error: [Uop %d (%s), oparg %d, operand %" PRIu64 ", target %d @ %d]\n",
-            opcode, _PyUopName(opcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
+            uopcode, _PyUopName(uopcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
             (int)(next_uop - current_executor->trace - 1));
     OPT_HIST(trace_uop_execution_counter, trace_run_length_hist);
     frame->return_offset = 0;  // Don't leave this random
@@ -1066,10 +1065,10 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     // On DEOPT_IF we just repeat the last instruction.
     // This presumes nothing was popped from the stack (nor pushed).
     DPRINTF(2, "DEOPT: [Uop %d (%s), oparg %d, operand %" PRIu64 ", target %d @ %d]\n",
-            opcode, _PyUopName(opcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
+            uopcode, _PyUopName(uopcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
             (int)(next_uop - current_executor->trace - 1));
     OPT_HIST(trace_uop_execution_counter, trace_run_length_hist);
-    UOP_STAT_INC(opcode, miss);
+    UOP_STAT_INC(uopcode, miss);
     frame->return_offset = 0;  // Dispatch to frame->instr_ptr
     _PyFrame_SetStackPointer(frame, stack_pointer);
     frame->instr_ptr = next_uop[-1].target + _PyCode_CODE(_PyFrame_GetCode(frame));