garrettgu10 · garrettgu10 · May 19, 2024 · May 19, 2024 · May 20, 2024 · May 29, 2024
diff --git a/.devcontainer/devcontainer.json b/.devcontainer/devcontainer.json
@@ -15,14 +15,14 @@
     ],
     "updateContentCommand": {
         // Using the shell for `nproc` usage.
-        "python": "./configure --config-cache --with-pydebug && make -s -j `nproc`",
-        "docs": [
-            "make",
-            "--directory",
-            "Doc",
-            "venv",
-            "html"
-        ]
+        // "python": "./configure --config-cache --with-pydebug && make -s -j `nproc`",
+        // "docs": [
+        //     "make",
+        //     "--directory",
+        //     "Doc",
+        //     "venv",
+        //     "html"
+        // ]
     },
     "postCreateCommand": {
         // https://github.com/orgs/community/discussions/26026

diff --git a/Makefile.pre.in b/Makefile.pre.in
@@ -699,18 +699,7 @@ build_wasm: check-clean-src $(BUILDPYTHON) platform sharedmods \
 # Check that the source is clean when building out of source.
 .PHONY: check-clean-src
 check-clean-src:
-	@if test -n "$(VPATH)" -a \( \
-	    -f "$(srcdir)/$(BUILDPYTHON)" \
-	    -o -f "$(srcdir)/Programs/python.o" \
-	    -o -f "$(srcdir)/Python/frozen_modules/importlib._bootstrap.h" \
-	\); then \
-		echo "Error: The source directory ($(srcdir)) is not clean" ; \
-		echo "Building Python out of the source tree (in $(abs_builddir)) requires a clean source tree ($(abs_srcdir))" ; \
-		echo "Build artifacts such as .o files, executables, and Python/frozen_modules/*.h must not exist within $(srcdir)." ; \
-		echo "Try to run:" ; \
-		echo "  (cd \"$(srcdir)\" && make clean || git clean -fdx -e Doc/venv)" ; \
-		exit 1; \
-	fi
+	echo skipped clean check
 
 # Profile generation build must start from a clean tree.
 profile-clean-stamp:
@@ -1953,6 +1942,8 @@ Python/compile.o Python/assemble.o Python/flowgraph.o Python/instruction_sequenc
 
 Python/ceval.o: \
 		$(srcdir)/Python/ceval_macros.h \
+		$(srcdir)/Python/tail_call_funcs.h \
+		$(srcdir)/Python/opcode_funcs.h \
 		$(srcdir)/Python/condvar.h \
 		$(srcdir)/Python/generated_cases.c.h \
 		$(srcdir)/Python/executor_cases.c.h \

diff --git a/Python/bytecodes.c b/Python/bytecodes.c
@@ -806,7 +806,7 @@ dummy_func(
                 if (do_raise(tstate, exc, cause)) {
                     assert(oparg == 0);
                     monitor_reraise(tstate, frame, this_instr);
-                    goto exception_unwind;
+                    CEVAL_GOTO(exception_unwind);
                 }
                 break;
             default:
@@ -824,7 +824,7 @@ dummy_func(
             tstate->current_frame = frame->previous;
             assert(!_PyErr_Occurred(tstate));
             tstate->c_recursion_remaining += PY_EVAL_C_STACK_UNITS;
-            return retval;
+            return (uintptr_t) retval;
         }
 
         // The stack effect here is ambiguous.
@@ -868,7 +868,7 @@ dummy_func(
             _PyEval_FrameClearAndPop(tstate, dying);
             _PyFrame_StackPush(frame, retval);
             LOAD_IP(frame->return_offset);
-            goto resume_frame;
+            CEVAL_GOTO(resume_frame);
         }
 
         macro(RETURN_CONST) =
@@ -892,7 +892,7 @@ dummy_func(
             _PyEval_FrameClearAndPop(tstate, dying);
             _PyFrame_StackPush(frame, retval);
             LOAD_IP(frame->return_offset);
-            goto resume_frame;
+            CEVAL_GOTO(resume_frame);
         }
 
         inst(GET_AITER, (obj -- iter)) {
@@ -1095,7 +1095,7 @@ dummy_func(
             /* We don't know which of these is relevant here, so keep them equal */
             assert(INLINE_CACHE_ENTRIES_SEND == INLINE_CACHE_ENTRIES_FOR_ITER);
             LOAD_IP(1 + INLINE_CACHE_ENTRIES_SEND);
-            goto resume_frame;
+            CEVAL_GOTO(resume_frame);
         }
 
         inst(YIELD_VALUE, (retval -- value)) {
@@ -1157,7 +1157,7 @@ dummy_func(
             Py_INCREF(exc);
             _PyErr_SetRaisedException(tstate, exc);
             monitor_reraise(tstate, frame, this_instr);
-            goto exception_unwind;
+            CEVAL_GOTO(exception_unwind);
         }
 
         tier1 inst(END_ASYNC_FOR, (awaitable, exc -- )) {
@@ -1169,7 +1169,7 @@ dummy_func(
                 Py_INCREF(exc);
                 _PyErr_SetRaisedException(tstate, exc);
                 monitor_reraise(tstate, frame, this_instr);
-                goto exception_unwind;
+                CEVAL_GOTO(exception_unwind);
             }
         }
 
@@ -1184,7 +1184,7 @@ dummy_func(
             else {
                 _PyErr_SetRaisedException(tstate, Py_NewRef(exc_value));
                 monitor_reraise(tstate, frame, this_instr);
-                goto exception_unwind;
+                CEVAL_GOTO(exception_unwind);
             }
         }
 
@@ -3422,7 +3422,7 @@ dummy_func(
              * We don't check recursion depth here,
              * as it will be checked after start_frame */
             tstate->py_recursion_remaining--;
-            goto start_frame;
+            CEVAL_GOTO(start_frame);
         }
 
         inst(EXIT_INIT_CHECK, (should_be_none -- )) {
@@ -3615,7 +3615,7 @@ dummy_func(
             DEOPT_IF(!PyList_Check(self));
             STAT_INC(CALL, hit);
             if (_PyList_AppendTakeRef((PyListObject *)self, arg) < 0) {
-                goto pop_1_error;  // Since arg is DECREF'ed already
+                CEVAL_GOTO(pop_1_error);  // Since arg is DECREF'ed already
             }
             Py_DECREF(self);
             Py_DECREF(callable);

diff --git a/Python/ceval.c b/Python/ceval.c
@@ -2,6 +2,9 @@
 
 #define _PY_INTERPRETER
 
+#define USE_COMPUTED_GOTOS 0
+#define USE_TAIL_CALLS 1
+
 #include "Python.h"
 #include "pycore_abstract.h"      // _PyIndex_Check()
 #include "pycore_backoff.h"
@@ -684,6 +687,11 @@ extern void _PyUOpPrint(const _PyUOpInstruction *uop);
 /* This setting is reversed below following _PyEval_EvalFrameDefault */
 #endif
 
+#if USE_TAIL_CALLS
+#include "tail_call_funcs.h"
+#include "opcode_funcs.h"
+#endif
+
 PyObject* _Py_HOT_FUNCTION
 _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int throwflag)
 {
@@ -780,9 +788,14 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     assert(!_PyErr_Occurred(tstate));
 #endif
 
+#if USE_TAIL_CALLS
+    ret_state s;
+#endif
+
     DISPATCH();
 
     {
+#if !USE_TAIL_CALLS
     /* Start instructions */
 #if !USE_COMPUTED_GOTOS
     dispatch_opcode:
@@ -849,6 +862,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
             goto error;
 
         } /* End instructions */
+#endif
 
         /* This should never be reached. Every opcode should end with DISPATCH()
            or goto error. */

diff --git a/Python/ceval_macros.h b/Python/ceval_macros.h
@@ -70,9 +70,24 @@
 #define INSTRUCTION_STATS(op) ((void)0)
 #endif
 
+typedef struct {
+    PyThreadState *tstate;
+    _PyInterpreterFrame *frame;
+    PyObject **stack_pointer;
+    _Py_CODEUNIT *next_instr;
+    int opcode;
+    int oparg;
+} ret_state;
+
+typedef uintptr_t (*OPCODE_FUNC_PTR_TYPE)(PyThreadState *tstate, _PyInterpreterFrame *frame, PyObject **stack_pointer, _Py_CODEUNIT *next_instr, int opcode, int oparg, ret_state *state);
+static OPCODE_FUNC_PTR_TYPE opcode_funcs[256];
+
 #if USE_COMPUTED_GOTOS
 #  define TARGET(op) TARGET_##op:
 #  define DISPATCH_GOTO() goto *opcode_targets[opcode]
+#elif USE_TAIL_CALLS
+#  define DISPATCH_GOTO() OPCODE_FUNC_PTR_TYPE next = opcode_funcs[opcode]; \
+                          __attribute__((musttail)) return next(tstate, frame, stack_pointer, next_instr, opcode, oparg, state)
 #else
 #  define TARGET(op) case op: TARGET_##op:
 #  define DISPATCH_GOTO() goto dispatch_opcode
@@ -127,11 +142,11 @@ do { \
         (NEW_FRAME)->previous = frame;                  \
         frame = tstate->current_frame = (NEW_FRAME);     \
         CALL_STAT_INC(inlined_py_calls);                \
-        goto start_frame;                               \
+        CEVAL_GOTO(start_frame);                               \
     } while (0)
 
 // Use this instead of 'goto error' so Tier 2 can go to a different label
-#define GOTO_ERROR(LABEL) goto LABEL
+#define GOTO_ERROR(LABEL) CEVAL_GOTO(LABEL)
 
 #define CHECK_EVAL_BREAKER() \
     _Py_CHECK_EMSCRIPTEN_SIGNALS_PERIODICALLY(); \
@@ -266,7 +281,13 @@ GETITEM(PyObject *v, Py_ssize_t i) {
                                      GETLOCAL(i) = value; \
                                      Py_XDECREF(tmp); } while (0)
 
-#define GO_TO_INSTRUCTION(op) goto PREDICT_ID(op)
+#if USE_TAIL_CALLS
+#  define GO_TO_INSTRUCTION(OP, SUB) OPCODE_FUNC_PTR_TYPE next = opcode_funcs[OP]; \
+                                     next_instr -= SUB; \
+                                     __attribute__((musttail)) return next(tstate, frame, stack_pointer, next_instr, opcode, oparg, state)
+#else
+#  define GO_TO_INSTRUCTION(op) goto PREDICT_ID(op)
+#endif
 
 #ifdef Py_STATS
 #define UPDATE_MISS_STATS(INSTNAME)                              \
@@ -282,12 +303,12 @@ GETITEM(PyObject *v, Py_ssize_t i) {
 #define UPDATE_MISS_STATS(INSTNAME) ((void)0)
 #endif
 
-#define DEOPT_IF(COND, INSTNAME)                            \
+#define DEOPT_IF(COND, INSTNAME, SUB)                            \
     if ((COND)) {                                           \
         /* This is only a single jump on release builds! */ \
         UPDATE_MISS_STATS((INSTNAME));                      \
         assert(_PyOpcode_Deopt[opcode] == (INSTNAME));      \
-        GO_TO_INSTRUCTION(INSTNAME);                        \
+        GO_TO_INSTRUCTION(INSTNAME, SUB);                        \
     }
 
 
@@ -366,7 +387,7 @@ do { \
         stack_pointer = _PyFrame_GetStackPointer(frame); \
         if (next_instr == NULL) { \
             next_instr = (dest)+1; \
-            goto error; \
+            CEVAL_GOTO(error); \
         } \
     } \
 } while (0);