pythongh-128563: Move GO_TO_INSTRUCTION and PREDICT to cases generator (pythonGH-129115)

Author: Fidget-Spinner

Lib/test/test_generated_cases.py

@@ -412,7 +412,7 @@ def test_predictions(self):
             frame->instr_ptr = next_instr;
             next_instr += 1;
             INSTRUCTION_STATS(OP1);
-            PREDICTED(OP1);
+            PREDICTED_OP1:;
             _PyStackRef res;
             res = Py_None;
             stack_pointer[-1] = res;
@@ -646,7 +646,7 @@ def test_macro_instruction(self):
             frame->instr_ptr = next_instr;
             next_instr += 6;
             INSTRUCTION_STATS(OP);
-            PREDICTED(OP);
+            PREDICTED_OP:;
             _Py_CODEUNIT* const this_instr = next_instr - 6;
             (void)this_instr;
             _PyStackRef left;

Python/ceval_macros.h

@@ -166,35 +166,6 @@ GETITEM(PyObject *v, Py_ssize_t i) {
 #define JUMPBY(x)       (next_instr += (x))
 #define SKIP_OVER(x)    (next_instr += (x))
 
-/* OpCode prediction macros
-    Some opcodes tend to come in pairs thus making it possible to
-    predict the second code when the first is run.  For example,
-    COMPARE_OP is often followed by POP_JUMP_IF_FALSE or POP_JUMP_IF_TRUE.
-
-    Verifying the prediction costs a single high-speed test of a register
-    variable against a constant.  If the pairing was good, then the
-    processor's own internal branch predication has a high likelihood of
-    success, resulting in a nearly zero-overhead transition to the
-    next opcode.  A successful prediction saves a trip through the eval-loop
-    including its unpredictable switch-case branch.  Combined with the
-    processor's internal branch prediction, a successful PREDICT has the
-    effect of making the two opcodes run as if they were a single new opcode
-    with the bodies combined.
-
-    If collecting opcode statistics, your choices are to either keep the
-    predictions turned-on and interpret the results as if some opcodes
-    had been combined or turn-off predictions so that the opcode frequency
-    counter updates for both opcodes.
-
-    Opcode prediction is disabled with threaded code, since the latter allows
-    the CPU to record separate branch prediction information for each
-    opcode.
-
-*/
-
-#define PREDICT_ID(op)          PRED_##op
-#define PREDICTED(op)           PREDICT_ID(op):
-
 
 /* Stack manipulation macros */
 
@@ -260,8 +231,6 @@ GETITEM(PyObject *v, Py_ssize_t i) {
                                      GETLOCAL(i) = value; \
                                      PyStackRef_XCLOSE(tmp); } while (0)
 
-#define GO_TO_INSTRUCTION(op) goto PREDICT_ID(op)
-
 #ifdef Py_STATS
 #define UPDATE_MISS_STATS(INSTNAME)                              \
     do {                                                         \
@@ -281,7 +250,7 @@ GETITEM(PyObject *v, Py_ssize_t i) {
         /* This is only a single jump on release builds! */ \
         UPDATE_MISS_STATS((INSTNAME));                      \
         assert(_PyOpcode_Deopt[opcode] == (INSTNAME));      \
-        GO_TO_INSTRUCTION(INSTNAME);                        \
+        goto PREDICTED_##INSTNAME;                          \
     }
 
 

Python/generated_cases.c.h

@@ -127,6 +127,7 @@ def __init__(self, out: CWriter):
             "DISPATCH": self.dispatch,
             "INSTRUCTION_SIZE": self.instruction_size,
             "POP_INPUT": self.pop_input,
+            "GO_TO_INSTRUCTION": self.go_to_instruction,
         }
         self.out = out
 
@@ -402,6 +403,23 @@ def sync_sp(
         self._print_storage(storage)
         return True
 
+    def go_to_instruction(
+        self,
+        tkn: Token,
+        tkn_iter: TokenIterator,
+        uop: Uop,
+        storage: Storage,
+        inst: Instruction | None,
+    ) -> bool:
+        next(tkn_iter)
+        name = next(tkn_iter)
+        next(tkn_iter)
+        next(tkn_iter)
+        assert name.kind == "IDENTIFIER"
+        self.emit("\n")
+        self.emit(f"goto PREDICTED_{name.text};\n")
+        return True
+
     def emit_save(self, storage: Storage) -> None:
         storage.save(self.out)
         self._print_storage(storage)

Tools/cases_generator/generators_common.py

@@ -158,7 +158,7 @@ def generate_tier1(
         out.emit(f"next_instr += {inst.size};\n")
         out.emit(f"INSTRUCTION_STATS({name});\n")
         if inst.is_target:
-            out.emit(f"PREDICTED({name});\n")
+            out.emit(f"PREDICTED_{name}:;\n")
             if needs_this:
                 out.emit(f"_Py_CODEUNIT* const this_instr = next_instr - {inst.size};\n")
                 out.emit(unused_guard)