Maps, Unrolling, and Dynamic Frame Size

.codeclimate.yml

@@ -27,3 +27,4 @@ exclude_paths:
   - "repository/*"
   - "scripts/**/*"
   - "scripts/*"
+

rsqueakvm/constants.py

@@ -127,42 +127,42 @@
 SO_JIT_HOOK_RCVR = 59 # really selectorTrap
 
 constant_objects_in_special_object_table = {
-    "nil": (SO_NIL, "POINTERS"),
-    "true": (SO_TRUE, "POINTERS"),
-    "false": (SO_FALSE, "POINTERS"),
+    "nil": (SO_NIL, "FIXED"),
+    "true": (SO_TRUE, "FIXED"),
+    "false": (SO_FALSE, "FIXED"),
     "charactertable": (SO_CHARACTER_TABLE_ARRAY, "POINTERS"),
-    "schedulerassociationpointer": (SO_SCHEDULERASSOCIATIONPOINTER, "POINTERS"),
+    "schedulerassociationpointer": (SO_SCHEDULERASSOCIATIONPOINTER, "FIXED"),
     "special_selectors": (SO_SPECIAL_SELECTORS_ARRAY, "POINTERS"),
-    "smalltalkdict": (SO_SMALLTALK, "POINTERS"),
+    "smalltalkdict": (SO_SMALLTALK, "FIXED"),
     "doesNotUnderstand": (SO_DOES_NOT_UNDERSTAND, "BYTES"),
     "mustBeBoolean": (SO_MUST_BE_BOOLEAN, "BYTES"),
     "runWithIn": (SO_RUN_WITH_IN, "BYTES"),
     "cannotReturn": (SO_CANNOT_RETURN, "BYTES"),
     # classes
-    "Bitmap": (SO_BITMAP_CLASS, "POINTERS"),
-    "SmallInteger": (SO_SMALLINTEGER_CLASS, "POINTERS"),
-    "String": (SO_STRING_CLASS, "POINTERS"),
-    "Array": (SO_ARRAY_CLASS, "POINTERS"),
-    "Float": (SO_FLOAT_CLASS, "POINTERS"),
-    "MethodContext": (SO_METHODCONTEXT_CLASS, "POINTERS"),
-    "BlockContext": (SO_BLOCKCONTEXT_CLASS, "POINTERS"),
-    "BlockClosure": (SO_BLOCKCLOSURE_CLASS, "POINTERS"),
-    "Point": (SO_POINT_CLASS, "POINTERS"),
-    "LargePositiveInteger": (SO_LARGEPOSITIVEINTEGER_CLASS, "POINTERS"),
-    "Message": (SO_MESSAGE_CLASS, "POINTERS"),
-    "CompiledMethod": (SO_COMPILEDMETHOD_CLASS, "POINTERS"),
-    "Semaphore": (SO_SEMAPHORE_CLASS, "POINTERS"),
-    "Character": (SO_CHARACTER_CLASS, "POINTERS"),
-    "ByteArray": (SO_BYTEARRAY_CLASS, "POINTERS"),
-    "Process": (SO_PROCESS_CLASS, "POINTERS"),
-#    "PseudoContext" : (SO_PSEUDOCONTEXT_CLASS, "POINTERS"),
-#    "TranslatedMethod" : (SO_TRANSLATEDMETHOD_CLASS, "POINTERS"),
-    "LargeNegativeInteger" : (SO_LARGENEGATIVEINTEGER_CLASS, "POINTERS"),
+    "Bitmap": (SO_BITMAP_CLASS, "FIXED"),
+    "SmallInteger": (SO_SMALLINTEGER_CLASS, "FIXED"),
+    "String": (SO_STRING_CLASS, "FIXED"),
+    "Array": (SO_ARRAY_CLASS, "FIXED"),
+    "Float": (SO_FLOAT_CLASS, "FIXED"),
+    "MethodContext": (SO_METHODCONTEXT_CLASS, "FIXED"),
+    "BlockContext": (SO_BLOCKCONTEXT_CLASS, "FIXED"),
+    "BlockClosure": (SO_BLOCKCLOSURE_CLASS, "FIXED"),
+    "Point": (SO_POINT_CLASS, "FIXED"),
+    "LargePositiveInteger": (SO_LARGEPOSITIVEINTEGER_CLASS, "FIXED"),
+    "Message": (SO_MESSAGE_CLASS, "FIXED"),
+    "CompiledMethod": (SO_COMPILEDMETHOD_CLASS, "FIXED"),
+    "Semaphore": (SO_SEMAPHORE_CLASS, "FIXED"),
+    "Character": (SO_CHARACTER_CLASS, "FIXED"),
+    "ByteArray": (SO_BYTEARRAY_CLASS, "FIXED"),
+    "Process": (SO_PROCESS_CLASS, "FIXED"),
+#    "PseudoContext" : (SO_PSEUDOCONTEXT_CLASS, "FIXED"),
+#    "TranslatedMethod" : (SO_TRANSLATEDMETHOD_CLASS, "FIXED"),
+    "LargeNegativeInteger" : (SO_LARGENEGATIVEINTEGER_CLASS, "FIXED"),
     # ours, not in the table, but we'd like it to
-    "ClassBinding": (SPECIAL_OBJECTS_SIZE + 30, "POINTERS"),
-    "Metaclass": (SPECIAL_OBJECTS_SIZE + 31, "POINTERS"),
-    "Processor": (SPECIAL_OBJECTS_SIZE + 32, "POINTERS"),
-    "ByteSymbol": (SPECIAL_OBJECTS_SIZE + 33, "POINTERS")
+    "ClassBinding": (SPECIAL_OBJECTS_SIZE + 30, "FIXED"),
+    "Metaclass": (SPECIAL_OBJECTS_SIZE + 31, "FIXED"),
+    "Processor": (SPECIAL_OBJECTS_SIZE + 32, "FIXED"),
+    "ByteSymbol": (SPECIAL_OBJECTS_SIZE + 33, "FIXED")
 }
 
 variables_in_special_object_table = {

rsqueakvm/interpreter.py

@@ -126,9 +126,17 @@ def __init__(self, s_new_context, forced=False):
         ContextSwitchException.__init__(self, s_new_context)
         self.forced = forced
 
+class Optargs(object):
+    _attrs_ = ["max_squeak_unroll_count", "squeak_unroll_trace_limit"]
+    _immutable_fields_ = ["max_squeak_unroll_count", "squeak_unroll_trace_limit"]
+    def __init__(self):
+        self.max_squeak_unroll_count = 2
+        self.squeak_unroll_trace_limit = 32000
+
+
 UNROLLING_BYTECODE_RANGES = unroll.unrolling_iterable(interpreter_bytecodes.BYTECODE_RANGES)
 
-def get_printable_location(pc, self, method, w_class, blockmethod):
+def get_printable_location(pc, jump_back_pc, unrollings, frame_size, self, method, w_class, blockmethod):
     bc = ord(method.bytes[pc])
     name = method.safe_identifier_string()
     classname = "???"
@@ -143,8 +151,8 @@ def get_printable_location(pc, self, method, w_class, blockmethod):
         blockname = blockmethod.safe_identifier_string()
         return '%s(%s): (%s) [%d]: <%s>%s' % (classname, name, blockname, pc, hex(bc), interpreter_bytecodes.BYTECODE_NAMES[bc])
 
-def resume_get_printable_location(pc, self, method, w_class):
-    return "resume: %s" % get_printable_location(pc, self, method, w_class, None)
+def resume_get_printable_location(pc, frame_size, self, method, w_class):
+    return "resume: %s" % get_printable_location(pc, 0, 0, frame_size, self, method, w_class, None)
 
 # def confirm_enter_jit(pc, self, method, w_class, s_context):
 #     print get_printable_location(pc, self, method, w_class)
@@ -161,11 +169,13 @@ class Interpreter(object):
                           "evented",
                           "interrupts",
                           "trace_important",
-                          "trace"]
+                          "trace",
+                          "optargs"]
 
     jit_driver = jit.JitDriver(
         name=jit_driver_name,
-        greens=['pc', 'self', 'method', 'w_class', 'blockmethod'],
+        greens=['pc', 'jump_back_pc', 'unrollings', 'frame_size', 'self',
+                'method', 'w_class', 'blockmethod'],
         reds=['s_context'],
         virtualizables=['s_context'],
         get_printable_location=get_printable_location,
@@ -174,15 +184,15 @@ class Interpreter(object):
 
     resume_driver = jit.JitDriver(
         name=jit_driver_name + "_resume",
-        greens=['pc', 'self', 'method', 'w_class'],
+        greens=['pc', 'frame_size', 'self', 'method', 'w_class'],
         reds=['s_context'],
         # virtualizables=['s_context'],
         get_printable_location=resume_get_printable_location,
         is_recursive=True
     )
 
     def __init__(self, space, image=None, trace_important=False,
-                 trace=False, evented=True, interrupts=True):
+                 trace=False, evented=True, interrupts=True, optargs=None):
         # === Initialize immutable variables
         self.space = space
         self.image = image
@@ -199,6 +209,7 @@ def __init__(self, space, image=None, trace_important=False,
             self.interrupt_counter_size = constants.INTERRUPT_COUNTER_SIZE
         self.last_check = self.time_now()
         self.trace = trace
+        self.optargs = optargs or Optargs()
 
         # === Initialize mutable variables
         self.interrupt_check_counter = self.interrupt_counter_size
@@ -217,9 +228,11 @@ def loop(self, w_active_context):
         s_context = w_active_context.as_context_get_shadow(self.space)
         while True:
             method = s_context.w_method()
+            frame_size = method.frame_size()
             pc = s_context.pc()
             self.resume_driver.jit_merge_point(
                 pc=pc,
+                frame_size=frame_size,
                 self=self,
                 method=method,
                 w_class=self.getreceiverclass(s_context),
@@ -239,6 +252,7 @@ def loop(self, w_active_context):
                     self.resume_driver.can_enter_jit(
                         pc=pc,
                         self=self,
+                        frame_size=frame_size,
                         method=method,
                         w_class=self.getreceiverclass(s_context),
                         s_context=s_context)
@@ -304,27 +318,60 @@ def getreceiverclass(self, s_context):
         return s_context.w_receiver().safe_getclass(self.space)
 
     def getblockmethod(self, s_context):
-        return s_context.blockmethod
+        return s_context.blockmethod()
 
     def loop_bytecodes(self, s_context, may_context_switch=True):
         old_pc = 0
+        jump_back_pc = 0
+        unrollings = 0
         if not jit.we_are_jitted() and may_context_switch:
             self.quick_check_for_interrupt(s_context)
         method = s_context.w_method()
+        frame_size = method.frame_size()
         while True:
             pc = s_context.pc()
             if pc < old_pc:
-                if jit.we_are_jitted():
-                    # Do the interrupt-check at the end of a loop, don't interrupt loops midway.
-                    self.jitted_check_for_interrupt(s_context)
-                self.jit_driver.can_enter_jit(
-                    pc=pc, self=self, method=method,
-                    w_class=self.getreceiverclass(s_context),
-                    blockmethod=self.getblockmethod(s_context),
-                    s_context=s_context)
+                if jump_back_pc == old_pc:
+                    if (unrollings < self.optargs.max_squeak_unroll_count and
+                        jit.current_trace_length() < self.optargs.squeak_unroll_trace_limit):
+                        unrollings += 1
+                    else:
+                        if jit.we_are_jitted():
+                            # Do the interrupt-check at the end of a loop, don't
+                            # interrupt loops midway.
+                            self.jitted_check_for_interrupt(s_context)
+                        if not s_context.has_overflow_stack():
+                            self.jit_driver.can_enter_jit(
+                                pc=pc,
+                                jump_back_pc=jump_back_pc,
+                                unrollings=unrollings,
+                                frame_size=frame_size,
+                                self=self, method=method,
+                                w_class=self.getreceiverclass(s_context),
+                                blockmethod=self.getblockmethod(s_context),
+                                s_context=s_context)
+                else:
+                    jump_back_pc = old_pc
+                    unrollings = 1
+                    # we jumped back from the end of a loop. Instead of allowing
+                    # to enter the JIT here, we instead wait for the second time
+                    # this loop runs and call can_enter_jit only then
+                    # (effectively unrolling at least two iterations). This is
+                    # because the way the Squeak compiler generates loop
+                    # bytecodes: we get the branch condition in the header and a
+                    # conditional jump forward in case it is false. Then the
+                    # loop body and an unconditional jump back. In the case of 1
+                    # to: 1 do: or other loops that run for exactly one
+                    # iteration, we will still generate a call_assembler in that
+                    # case, which we work around with this. There are indeed a
+                    # few examples of loops that run exactly one iteration
             old_pc = pc
             self.jit_driver.jit_merge_point(
-                pc=pc, self=self, method=method,
+                pc=pc,
+                jump_back_pc=jump_back_pc,
+                unrollings=unrollings,
+                frame_size=frame_size,
+                self=self, method=method,
                 w_class=self.getreceiverclass(s_context),
                 blockmethod=self.getblockmethod(s_context),
                 s_context=s_context)

rsqueakvm/interpreter_bytecodes.py

@@ -365,7 +365,7 @@ def _sendSelfSelectorSpecial(self, selector, numargs, interp):
 
     @objectmodel.specialize.arg(3)
     def _sendSpecialSelector(self, interp, receiver, special_selector, w_args=[]):
-        space = jit.promote(self.space)
+        space = self.space
         w_special_selector = getattr(space, "w_" + special_selector)
         s_class = receiver.class_shadow(space)