Improve recursion-depth checking a bit

Main difference is lowering the recursion-depth on unoptimized
builds, since we need more stack space per python frame.

src/codegen/ast_interpreter.cpp

@@ -452,6 +452,8 @@ Box* ASTInterpreter::executeInner(ASTInterpreter& interpreter, CFGBlock* start_b
 
 Box* ASTInterpreter::execute(ASTInterpreter& interpreter, CFGBlock* start_block, BST_stmt* start_at) {
     UNAVOIDABLE_STAT_TIMER(t0, "us_timer_in_interpreter");
+    RECURSIVE_BLOCK(CXX, " in function call");
+
     return executeInnerAndSetupFrame(interpreter, start_block, start_at);
 }
 

src/runtime/capi.cpp

@@ -628,7 +628,15 @@ extern "C" PyObject* PyExceptionInstance_Class(PyObject* o) noexcept {
     return PyInstance_Check(o) ? (Box*)static_cast<BoxedInstance*>(o)->inst_cls : o->cls;
 }
 
+// With optimizations turned off, we can handle much fewer Python stack frames
+// in the same amount of C-stack space.
+// Use NDEBUG as a proxy for whether optimizations are turned on.
+#ifndef NDEBUG
+#define Py_DEFAULT_RECURSION_LIMIT 300
+#else
 #define Py_DEFAULT_RECURSION_LIMIT 1000
+#endif
+
 static int recursion_limit = Py_DEFAULT_RECURSION_LIMIT;
 extern "C" {
 int _Py_CheckRecursionLimit = Py_DEFAULT_RECURSION_LIMIT;

src/runtime/objmodel.cpp

@@ -4747,7 +4747,7 @@ Box* callFunc(BoxedFunctionBase* func, CallRewriteArgs* rewrite_args, ArgPassSpe
 template <ExceptionStyle S>
 static Box* callChosenCF(CompiledFunction* chosen_cf, BoxedClosure* closure, BoxedGenerator* generator, Box* globals,
                          Box* oarg1, Box* oarg2, Box* oarg3, Box** oargs) noexcept(S == CAPI) {
-    // TODO: this should be done in the rewrite as well
+    // TODO: this should go into the emitted code
     RECURSIVE_BLOCK(S, " in function call");
 
     if (S != chosen_cf->exception_style) {

test/tests/stack_limits.py

-# skip-if: '-n' not in EXTRA_JIT_ARGS and '-O' not in EXTRA_JIT_ARGS
-# disable this test because the interpreter (with disabled bjit) currently uses too much stack
-
 # Make sure we can recurse at least 900 times on the three different types
 # of stacks that we have:
 
+import sys
+TEST_DEPTH = sys.getrecursionlimit() - 20
+
 def recurse(n):
-    print n
     if n > 0:
         return recurse(n - 1)
     return n
 
 print "Recursing on main thread..."
-recurse(900)
+recurse(TEST_DEPTH)
 
 print "Recursing in a generator..."
 def gen():
-    yield recurse(900)
+    yield recurse(TEST_DEPTH)
 print list(gen())
 
 print "Recursing in a thread..."
@@ -26,10 +25,17 @@ done = 0
 def thread_target():
     global done
 
-    recurse(900)
+    recurse(TEST_DEPTH)
 
     done = 1
 start_new_thread(thread_target, ())
 
 while not done:
     time.sleep(0.001)
+
+s = """
+if depth < TEST_DEPTH:
+    depth += 1
+    exec s
+"""
+exec s in {'depth': 0, 's': s, 'TEST_DEPTH': TEST_DEPTH}