[Python-checkins] Normalize jumps in compiler. All forward jumps to use JUMP_FORWARD. (GH-28755)

[markshannon]

https://github.com/python/cpython/commit/f6eafe18c004c55082de40d20cad084ef9dd3db7
commit: f6eafe18c004c55082de40d20cad084ef9dd3db7
branch: main
author: Mark Shannon <mark at hotpy.org>
committer: markshannon <mark at hotpy.org>
date: 2021-10-06T13:05:45+01:00
summary:

Normalize jumps in compiler. All forward jumps to use JUMP_FORWARD. (GH-28755)

files:
M Lib/test/test_dis.py
M Python/ceval.c
M Python/compile.c

diff --git a/Lib/test/test_dis.py b/Lib/test/test_dis.py
index 0899db66d8241..4755e3f3a1724 100644
--- a/Lib/test/test_dis.py
+++ b/Lib/test/test_dis.py
@@ -1046,7 +1046,7 @@ def _prepare_test_cases():
   Instruction(opname='COMPARE_OP', opcode=107, arg=4, argval='>', argrepr='>', offset=34, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='POP_JUMP_IF_FALSE', opcode=114, arg=21, argval=42, argrepr='to 42', offset=36, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='POP_TOP', opcode=1, arg=None, argval=None, argrepr='', offset=38, starts_line=8, is_jump_target=False, positions=None),
-  Instruction(opname='JUMP_ABSOLUTE', opcode=113, arg=26, argval=52, argrepr='to 52', offset=40, starts_line=None, is_jump_target=False, positions=None),
+  Instruction(opname='JUMP_FORWARD', opcode=110, arg=5, argval=52, argrepr='to 52', offset=40, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='JUMP_ABSOLUTE', opcode=113, arg=4, argval=8, argrepr='to 8', offset=42, starts_line=7, is_jump_target=True, positions=None),
   Instruction(opname='LOAD_GLOBAL', opcode=116, arg=1, argval='print', argrepr='print', offset=44, starts_line=10, is_jump_target=True, positions=None),
   Instruction(opname='LOAD_CONST', opcode=100, arg=4, argval='I can haz else clause?', argrepr="'I can haz else clause?'", offset=46, starts_line=None, is_jump_target=False, positions=None),
@@ -1071,7 +1071,7 @@ def _prepare_test_cases():
   Instruction(opname='LOAD_CONST', opcode=100, arg=2, argval=4, argrepr='4', offset=84, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='COMPARE_OP', opcode=107, arg=0, argval='<', argrepr='<', offset=86, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='POP_JUMP_IF_FALSE', opcode=114, arg=46, argval=92, argrepr='to 92', offset=88, starts_line=None, is_jump_target=False, positions=None),
-  Instruction(opname='JUMP_ABSOLUTE', opcode=113, arg=52, argval=104, argrepr='to 104', offset=90, starts_line=17, is_jump_target=False, positions=None),
+  Instruction(opname='JUMP_FORWARD', opcode=110, arg=6, argval=104, argrepr='to 104', offset=90, starts_line=17, is_jump_target=False, positions=None),
   Instruction(opname='LOAD_FAST', opcode=124, arg=0, argval='i', argrepr='i', offset=92, starts_line=11, is_jump_target=True, positions=None),
   Instruction(opname='POP_JUMP_IF_TRUE', opcode=115, arg=28, argval=56, argrepr='to 56', offset=94, starts_line=None, is_jump_target=False, positions=None),
   Instruction(opname='LOAD_GLOBAL', opcode=116, arg=1, argval='print', argrepr='print', offset=96, starts_line=19, is_jump_target=True, positions=None),
diff --git a/Python/ceval.c b/Python/ceval.c
index 2dbc291897c03..a3a173dfb7013 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -4051,19 +4051,18 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, InterpreterFrame *frame, int thr
 
         TARGET(JUMP_ABSOLUTE) {
             PREDICTED(JUMP_ABSOLUTE);
-            if (oparg < INSTR_OFFSET()) {
-                /* Increment the warmup counter and quicken if warm enough
-                * _Py_Quicken is idempotent so we don't worry about overflow */
-                if (!PyCodeObject_IsWarmedUp(co)) {
-                    PyCodeObject_IncrementWarmup(co);
-                    if (PyCodeObject_IsWarmedUp(co)) {
-                        if (_Py_Quicken(co)) {
-                            goto error;
-                        }
-                        int nexti = INSTR_OFFSET();
-                        first_instr = co->co_firstinstr;
-                        next_instr = first_instr + nexti;
+            assert(oparg < INSTR_OFFSET());
+            /* Increment the warmup counter and quicken if warm enough
+            * _Py_Quicken is idempotent so we don't worry about overflow */
+            if (!PyCodeObject_IsWarmedUp(co)) {
+                PyCodeObject_IncrementWarmup(co);
+                if (PyCodeObject_IsWarmedUp(co)) {
+                    if (_Py_Quicken(co)) {
+                        goto error;
                     }
+                    int nexti = INSTR_OFFSET();
+                    first_instr = co->co_firstinstr;
+                    next_instr = first_instr + nexti;
                 }
             }
             JUMPTO(oparg);
@@ -4072,6 +4071,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, InterpreterFrame *frame, int thr
         }
 
         TARGET(JUMP_ABSOLUTE_QUICK) {
+            assert(oparg < INSTR_OFFSET());
             JUMPTO(oparg);
             CHECK_EVAL_BREAKER();
             DISPATCH();
diff --git a/Python/compile.c b/Python/compile.c
index 694da29b771d0..2d82d6a1e5a91 100644
--- a/Python/compile.c
+++ b/Python/compile.c
@@ -7220,6 +7220,26 @@ assemble_emit(struct assembler *a, struct instr *i)
     return 1;
 }
 
+static void
+normalize_jumps(struct assembler *a)
+{
+    for (basicblock *b = a->a_entry; b != NULL; b = b->b_next) {
+        b->b_visited = 0;
+    }
+    for (basicblock *b = a->a_entry; b != NULL; b = b->b_next) {
+        b->b_visited = 1;
+        if (b->b_iused == 0) {
+            continue;
+        }
+        struct instr *last = &b->b_instr[b->b_iused-1];
+        if (last->i_opcode == JUMP_ABSOLUTE &&
+            last->i_target->b_visited == 0
+        ) {
+            last->i_opcode = JUMP_FORWARD;
+        }
+    }
+}
+
 static void
 assemble_jump_offsets(struct assembler *a, struct compiler *c)
 {
@@ -7897,6 +7917,9 @@ assemble(struct compiler *c, int addNone)
         clean_basic_block(b);
     }
 
+    /* Order of basic blocks must have been determined by now */
+    normalize_jumps(&a);
+
     /* Can't modify the bytecode after computing jump offsets. */
     assemble_jump_offsets(&a, c);