GH-115457: Support splitting and replication of micro ops. (GH-115558)

1 files changed, 64 insertions, 3 deletions

diff --git a/Include/internal/pycore_uop_metadata.h b/Include/internal/pycore_uop_metadata.h
index 1e2dfec..c9def0e 100644
--- a/Include/internal/pycore_uop_metadata.h
+++ b/Include/internal/pycore_uop_metadata.h
@@ -12,6 +12,7 @@ extern "C" {
 #include <stdint.h>
 #include "pycore_uop_ids.h"
 extern const uint16_t _PyUop_Flags[MAX_UOP_ID+1];
+extern const uint8_t _PyUop_Replication[MAX_UOP_ID+1];
 extern const char * const _PyOpcode_uop_name[MAX_UOP_ID+1];
 
 #ifdef NEED_OPCODE_METADATA
@@ -19,10 +20,26 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_NOP] = HAS_PURE_FLAG,
     [_RESUME_CHECK] = HAS_DEOPT_FLAG,
     [_LOAD_FAST_CHECK] = HAS_ARG_FLAG | HAS_LOCAL_FLAG | HAS_ERROR_FLAG,
+    [_LOAD_FAST_0] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_1] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_2] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_3] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_4] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_5] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_6] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
+    [_LOAD_FAST_7] = HAS_LOCAL_FLAG | HAS_PURE_FLAG,
     [_LOAD_FAST] = HAS_ARG_FLAG | HAS_LOCAL_FLAG | HAS_PURE_FLAG,
     [_LOAD_FAST_AND_CLEAR] = HAS_ARG_FLAG | HAS_LOCAL_FLAG,
     [_LOAD_FAST_LOAD_FAST] = HAS_ARG_FLAG | HAS_LOCAL_FLAG,
     [_LOAD_CONST] = HAS_ARG_FLAG | HAS_CONST_FLAG | HAS_PURE_FLAG,
+    [_STORE_FAST_0] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_1] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_2] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_3] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_4] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_5] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_6] = HAS_LOCAL_FLAG,
+    [_STORE_FAST_7] = HAS_LOCAL_FLAG,
     [_STORE_FAST] = HAS_ARG_FLAG | HAS_LOCAL_FLAG,
     [_STORE_FAST_LOAD_FAST] = HAS_ARG_FLAG | HAS_LOCAL_FLAG,
     [_STORE_FAST_STORE_FAST] = HAS_ARG_FLAG | HAS_LOCAL_FLAG,
@@ -114,14 +131,20 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_LOAD_ATTR] = HAS_ARG_FLAG | HAS_NAME_FLAG | HAS_ERROR_FLAG | HAS_ESCAPES_FLAG,
     [_GUARD_TYPE_VERSION] = HAS_DEOPT_FLAG | HAS_EXIT_FLAG | HAS_PASSTHROUGH_FLAG,
     [_CHECK_MANAGED_OBJECT_HAS_VALUES] = HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
-    [_LOAD_ATTR_INSTANCE_VALUE] = HAS_ARG_FLAG | HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_INSTANCE_VALUE_0] = HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_INSTANCE_VALUE_1] = HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_INSTANCE_VALUE] = HAS_ARG_FLAG | HAS_DEOPT_FLAG | HAS_OPARG_AND_1_FLAG,
     [_CHECK_ATTR_MODULE] = HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
     [_LOAD_ATTR_MODULE] = HAS_ARG_FLAG | HAS_DEOPT_FLAG,
     [_CHECK_ATTR_WITH_HINT] = HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG | HAS_PASSTHROUGH_FLAG,
     [_LOAD_ATTR_WITH_HINT] = HAS_ARG_FLAG | HAS_NAME_FLAG | HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
-    [_LOAD_ATTR_SLOT] = HAS_ARG_FLAG | HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_SLOT_0] = HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_SLOT_1] = HAS_DEOPT_FLAG,
+    [_LOAD_ATTR_SLOT] = HAS_ARG_FLAG | HAS_DEOPT_FLAG | HAS_OPARG_AND_1_FLAG,
     [_CHECK_ATTR_CLASS] = HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
-    [_LOAD_ATTR_CLASS] = HAS_ARG_FLAG,
+    [_LOAD_ATTR_CLASS_0] = 0,
+    [_LOAD_ATTR_CLASS_1] = 0,
+    [_LOAD_ATTR_CLASS] = HAS_ARG_FLAG | HAS_OPARG_AND_1_FLAG,
     [_GUARD_DORV_VALUES] = HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
     [_STORE_ATTR_INSTANCE_VALUE] = HAS_ESCAPES_FLAG,
     [_STORE_ATTR_SLOT] = HAS_ESCAPES_FLAG,
@@ -168,6 +191,11 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_CHECK_PEP_523] = HAS_DEOPT_FLAG,
     [_CHECK_FUNCTION_EXACT_ARGS] = HAS_ARG_FLAG | HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
     [_CHECK_STACK_SPACE] = HAS_ARG_FLAG | HAS_DEOPT_FLAG | HAS_PASSTHROUGH_FLAG,
+    [_INIT_CALL_PY_EXACT_ARGS_0] = HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
+    [_INIT_CALL_PY_EXACT_ARGS_1] = HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
+    [_INIT_CALL_PY_EXACT_ARGS_2] = HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
+    [_INIT_CALL_PY_EXACT_ARGS_3] = HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
+    [_INIT_CALL_PY_EXACT_ARGS_4] = HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
     [_INIT_CALL_PY_EXACT_ARGS] = HAS_ARG_FLAG | HAS_ESCAPES_FLAG | HAS_PURE_FLAG,
     [_PUSH_FRAME] = HAS_ESCAPES_FLAG,
     [_CALL_TYPE_1] = HAS_ARG_FLAG | HAS_DEOPT_FLAG,
@@ -215,6 +243,12 @@ const uint16_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_CHECK_VALIDITY_AND_SET_IP] = HAS_DEOPT_FLAG,
 };
 
+const uint8_t _PyUop_Replication[MAX_UOP_ID+1] = {
+    [_LOAD_FAST] = 8,
+    [_STORE_FAST] = 8,
+    [_INIT_CALL_PY_EXACT_ARGS] = 5,
+};
+
 const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_BEFORE_ASYNC_WITH] = "_BEFORE_ASYNC_WITH",
     [_BEFORE_WITH] = "_BEFORE_WITH",
@@ -317,6 +351,11 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_GUARD_TYPE_VERSION] = "_GUARD_TYPE_VERSION",
     [_INIT_CALL_BOUND_METHOD_EXACT_ARGS] = "_INIT_CALL_BOUND_METHOD_EXACT_ARGS",
     [_INIT_CALL_PY_EXACT_ARGS] = "_INIT_CALL_PY_EXACT_ARGS",
+    [_INIT_CALL_PY_EXACT_ARGS_0] = "_INIT_CALL_PY_EXACT_ARGS_0",
+    [_INIT_CALL_PY_EXACT_ARGS_1] = "_INIT_CALL_PY_EXACT_ARGS_1",
+    [_INIT_CALL_PY_EXACT_ARGS_2] = "_INIT_CALL_PY_EXACT_ARGS_2",
+    [_INIT_CALL_PY_EXACT_ARGS_3] = "_INIT_CALL_PY_EXACT_ARGS_3",
+    [_INIT_CALL_PY_EXACT_ARGS_4] = "_INIT_CALL_PY_EXACT_ARGS_4",
     [_INTERNAL_INCREMENT_OPT_COUNTER] = "_INTERNAL_INCREMENT_OPT_COUNTER",
     [_IS_NONE] = "_IS_NONE",
     [_IS_OP] = "_IS_OP",
@@ -332,7 +371,11 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_LOAD_ASSERTION_ERROR] = "_LOAD_ASSERTION_ERROR",
     [_LOAD_ATTR] = "_LOAD_ATTR",
     [_LOAD_ATTR_CLASS] = "_LOAD_ATTR_CLASS",
+    [_LOAD_ATTR_CLASS_0] = "_LOAD_ATTR_CLASS_0",
+    [_LOAD_ATTR_CLASS_1] = "_LOAD_ATTR_CLASS_1",
     [_LOAD_ATTR_INSTANCE_VALUE] = "_LOAD_ATTR_INSTANCE_VALUE",
+    [_LOAD_ATTR_INSTANCE_VALUE_0] = "_LOAD_ATTR_INSTANCE_VALUE_0",
+    [_LOAD_ATTR_INSTANCE_VALUE_1] = "_LOAD_ATTR_INSTANCE_VALUE_1",
     [_LOAD_ATTR_METHOD_LAZY_DICT] = "_LOAD_ATTR_METHOD_LAZY_DICT",
     [_LOAD_ATTR_METHOD_NO_DICT] = "_LOAD_ATTR_METHOD_NO_DICT",
     [_LOAD_ATTR_METHOD_WITH_VALUES] = "_LOAD_ATTR_METHOD_WITH_VALUES",
@@ -340,6 +383,8 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_LOAD_ATTR_NONDESCRIPTOR_NO_DICT] = "_LOAD_ATTR_NONDESCRIPTOR_NO_DICT",
     [_LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES] = "_LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES",
     [_LOAD_ATTR_SLOT] = "_LOAD_ATTR_SLOT",
+    [_LOAD_ATTR_SLOT_0] = "_LOAD_ATTR_SLOT_0",
+    [_LOAD_ATTR_SLOT_1] = "_LOAD_ATTR_SLOT_1",
     [_LOAD_ATTR_WITH_HINT] = "_LOAD_ATTR_WITH_HINT",
     [_LOAD_BUILD_CLASS] = "_LOAD_BUILD_CLASS",
     [_LOAD_CONST] = "_LOAD_CONST",
@@ -349,6 +394,14 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_LOAD_CONST_INLINE_WITH_NULL] = "_LOAD_CONST_INLINE_WITH_NULL",
     [_LOAD_DEREF] = "_LOAD_DEREF",
     [_LOAD_FAST] = "_LOAD_FAST",
+    [_LOAD_FAST_0] = "_LOAD_FAST_0",
+    [_LOAD_FAST_1] = "_LOAD_FAST_1",
+    [_LOAD_FAST_2] = "_LOAD_FAST_2",
+    [_LOAD_FAST_3] = "_LOAD_FAST_3",
+    [_LOAD_FAST_4] = "_LOAD_FAST_4",
+    [_LOAD_FAST_5] = "_LOAD_FAST_5",
+    [_LOAD_FAST_6] = "_LOAD_FAST_6",
+    [_LOAD_FAST_7] = "_LOAD_FAST_7",
     [_LOAD_FAST_AND_CLEAR] = "_LOAD_FAST_AND_CLEAR",
     [_LOAD_FAST_CHECK] = "_LOAD_FAST_CHECK",
     [_LOAD_FAST_LOAD_FAST] = "_LOAD_FAST_LOAD_FAST",
@@ -388,6 +441,14 @@ const char *const _PyOpcode_uop_name[MAX_UOP_ID+1] = {
     [_STORE_ATTR_SLOT] = "_STORE_ATTR_SLOT",
     [_STORE_DEREF] = "_STORE_DEREF",
     [_STORE_FAST] = "_STORE_FAST",
+    [_STORE_FAST_0] = "_STORE_FAST_0",
+    [_STORE_FAST_1] = "_STORE_FAST_1",
+    [_STORE_FAST_2] = "_STORE_FAST_2",
+    [_STORE_FAST_3] = "_STORE_FAST_3",
+    [_STORE_FAST_4] = "_STORE_FAST_4",
+    [_STORE_FAST_5] = "_STORE_FAST_5",
+    [_STORE_FAST_6] = "_STORE_FAST_6",
+    [_STORE_FAST_7] = "_STORE_FAST_7",
     [_STORE_FAST_LOAD_FAST] = "_STORE_FAST_LOAD_FAST",
     [_STORE_FAST_STORE_FAST] = "_STORE_FAST_STORE_FAST",
     [_STORE_GLOBAL] = "_STORE_GLOBAL",