duktape/doc/execution.rst

=========
Execution
=========

Overview
========

This document describes how Duktape manages its execution state.  Some details
are omitted but the goal is to give an overall picture how execution proceeds,
what state is involved, and what are the most important internal functions
involved.

The discussion is limited to a single Duktape heap as each Duktape heap is
independent of other Duktape heaps.  At any time, only one native thread may
be actively calling into a specific Duktape heap.

Execution states
================

There are three conceptual execution states for a Duktape heap:

* Idle

* Executing a Duktape/C function

* Executing an Ecmascript function

This conceptual model ignores details like heap initialization and
transitions from one state to another by "call handling".

Typical control flow
====================

Execution always begins from an idle state where no calls into Duktape are
active and user application has control.  User code may manipulate the value
stacks of Duktape contexts in this state without making any calls.  User code
may also call ``duk_debugger_cooperate()`` for integrating debugger into the
application event loop (or equivalent).

Eventually user code makes a call into either a Duktape/C function or an
Ecmascript function.  Such a call may be caused by an obvious API call like
``duk_pcall()``.  It may also be caused by a less obvious API call such as
``duk_get_prop()``, which may invoke a getter, or ``duk_to_string()`` which
may invoke a ``toString()`` coercion method.

The initial call into Duktape is always handled using ``duk_handle_call()``
which can handle a call from any state into any kind of target function.
Setting up a call involves a lot of state changes:

* A setjmp catchpoint is needed for protected calls.

* The call stack is resized if necessary, and an activation record
  (``duk_activation``) is set up for the new call.

* The value stack is resized if necessary, and a fresh value stack frame
  is established for the call.  The calling value stack frame and the target
  frame overlap for the call arguments, so that arguments on top of the
  calling stack are directly visible on the bottom of the target stack.

* An arguments object and an explicit environment record is created if
  necessary.

* Other small book-keeping (such as recursion depth tracking) is done.

When a call returns, the state changes are reversed before returning to
the caller.  If an error occurs during the call, a ``longjmp()`` will take
place and will be caught by the current (innermost) setjmp catchpoint
without tearing down the call state; the catchpoint will have to do that.

If the target function is a Duktape/C function, the corresponding C function
is looked up and called.  The C function now has access to a fresh value stack
frame it can operate on using the Duktape API.  It can make further calls which
get handled by ``duk_handle_call()``.

If the target function is an Ecmascript function, the value stack is resized
for the function register count (nregs) established by the compiler during
function compilation; unlike Duktape/C functions the value stack is mostly
static for the duration of bytecode execution.  Opcode handling may push
temporaries on the value stack but they must always be popped off before
proceeding to dispatch the next opcode.

The bytecode executor has its own setjmp catchpoint.  If bytecode makes a
call into a Duktape/C function it is handled normally using ``duk_handle_call()``;
such calls may happen also when the bytecode executor uses the value stack API
for various coercions etc.

If bytecode makes a function call into an Ecmascript function it is handled
specially by ``duk_handle_ecma_call_setup()``.  This call handler sets up a
new activation similarly to ``duk_handle_call()``, but instead of doing a
recursive call into the bytecode executor it returns to the bytecode executor
which restarts execution and starts executing the call target without
increasing C stack depth.  The call handler also supports tail calls where an
activation record is reused.

Both Duktape and user code may use ``duk_safe_call()`` to make protected
calls inside the current activation (or outside of any activations in the
idle state).  A safe call creates a new setjmp catchpoint but not a new
activation, so safe calls are not actual function calls.

Threading limitations
=====================

Only one native thread may call into a Duktape heap at any given time.
See ``threading.rst`` for more details.

Bytecode executor
=================

Basic functionality
-------------------

* Setjmp catchpoint which supports yield, resume, slow returns, try-catch, etc

* Opcode dispatch loop, central for performance

* Executor interrupt which facilitates script timeout and debugger integration

* Debugger support; breakpoint handling, checked and normal execution modes

Setjmp catchpoint
-----------------

The ``duk_handle_call()`` and ``duk_safe_call()`` catchpoints are only used to
handle ordinary error throws which propagate out of the calling function.  The
bytecode executor setjmp catchpoint handles a wider variety of longjmp call
types, and in many cases the longjmp may be handled without exiting the current
function:

* A slow break/continue uses a longjmp() so that if the break/continue crosses
  any finally clauses, they get executed as expected.  Similarly 'with' statement
  lexical environments are torn down, etc.

* A slow return uses a longjmp() so that any finally clauses, 'with' statement
  lexical environments, etc are handled appropriately.

* A coroutine resume is handled using longjmp(): the Duktape.Thread.resume()
  call adjusts the thread states (including their activations) and then uses
  this longjmp() type to restart execution in the target coroutine.

* A coroutine yield is handled using longjmp(): the Duktape.Thread.yield()
  call adjusts the states and uses this longjmp() type to restart execution
  in the target coroutine.

* An ordinary throw is handled as in ``duk_handle_call()`` with the difference
  that there are both 'try' and 'finally' sites.

Returns, coroutine yields, and throws may propagate out of the initial bytecode
executor entry and outwards to whatever code called into the executor.

Opcode dispatch loop and executor interrupt
-------------------------------------------

The opcode dispatch loop is a central performance critical part of the
executor.  The dispatch loop:

* Checks for an executor interrupt.  An interrupt can be taken for every
  opcode or for every N instructions; the interrupt handler provides e.g.
  script timeout and debugger integration.  This is performance critical
  because the check occurs for every opcode dispatch.  See separate section
  below on interrupt counter handling.

* Fetches an instruction from the topmost activation's "current PC",
  and increments the PC.  Managing the "current PC" is performance critical.
  See separate section below on current PC handling.

* Decodes and executes the opcode using a large switch-case.  The most
  important opcodes are in the main opcode space (64 opcodes); more rarely
  used opcodes are "extra" opcodes and need a double dispatch.

* Usually loops back to execute further opcodes.  May also (1) call another
  Duktape/C or Ecmascript function, (2) cause a longjmp, or (3) use
  ``goto restart_execution`` to restart the executor e.g. after call stack
  has been changed.

Debugger support
----------------

Debugger support relies on:

* Executor interrupt mechanism is needed to support debugging.

* A precheck in ``restart_execution`` where debugging status and breakpoints
  are checked.  Executor then either proceeds in "normal" or "checked"
  execution.  Checked execution means running one opcode at a time, and
  calling into the interrupt handler before each to see e.g. if a breakpoint
  has been triggered.

* There's some additional support outside the executor, e.g. call stack
  unwinding code handles the "step out" logic.

See ``debugger.rst`` for details.

Managing executor interrupt
===========================

The executor interrupt counter is currently tracked in
``thr->interrupt_counter``.  This seems to work well because ``thr`` is a
"hot" variable.

Another alternative would be to track the counter in an executor local
variable.  Error handling and other code paths jumping out of the executor
need to work similarly to how stack local ``curr_pc`` is handled.

Managing current PC
===================

Current approach
----------------

The current solution in Duktape 1.3 is to maintain a direct bytecode pointer
in each activation, and to keep a "cached copy" of the topmost activation's
bytecode pointer in a bytecode executor local variable ``curr_pc``.  A pointer
to the ``curr_pc`` in the stack frame (whose type is ``duk_instr_t **``) is
stored in ``thr->ptr_curr_pc`` so that when control exits the opcode dispatch
loop (e.g. when an error is thrown) the value in the stack frame can be read
and synced back into the topmost activation's ``act->curr_pc``.

Consistency depends on the compiler doing correct aliasing analysis, and
writing back the ``curr_pc`` value to the stack frame before any operation
that may potentially read it through ``thr->ptr_curr_pc``.  Using ``volatile``
would be safer but in practical testing it eliminates the performance benefit
entirely.

For the most part the bytecode executor can keep on dispatching opcodes
using ``curr_pc`` without copying the pointer back to the topmost activation.
However, the pointer needs to be synced (copied back) when:

* The current activation changes, i.e. a new function call is made.

* When an error is about to be thrown, to ensure any longjmp handlers
  will see correct PC values in activations.

* When the executor interrupt is entered; in particular, the debugger
  must see an up-to-date state in activations.

* When a ``goto restart_execution;`` occurs in bytecode dispatch, which
  happens for multiple opcodes.

Care must be taken *not* to sync when ``thr->ptr_curr_pc`` is no longer
pointing to the topmost activation and/or when the C stack frame pointed
to may no longer exist.  The current policy is to:

* Sync PC on function calls, also backup/restore ``thr->ptr_curr_pc`` on
  calls.

* Sync PC before a longjmp, often a bit earlier to ensure stacktraces
  come out right.

* Never sync or otherwise access ``thr->ptr_curr_pc`` in the setjmp
  catcher and unwind code paths.  This is to ensure we never dereference
  a ``thr->ptr_curr_pc`` no longer related to the topmost activation or
  pointing to an unwound C stack frame.  (The ``thr->ptr_curr_pc`` is
  not currently NULLed so it's intentionally dangling and must not be
  dereferenced incorrectly.)

Syncing the pointer back unnecessarily or multiple times is safe, however.

Function bytecode is behind a stable pointer, so there are no realloc or
other side effect concerns with using direct bytecode pointers.  Because
the function being executed is always reachable, a borrowed pointer can
be used.

This is a bit error prone, but it is worth the performance difference
of the alternatives.  This method of dispatch improves dispatch performance
by about 20-25% over Duktape 1.2.

Some alternatives
-----------------

* Duktape 1.3: maintain a direct bytecode pointer in each activation, and a
  "cached" copy of the topmost activation's bytecode pointer in a local
  variable of the executor.  Whenever something that might throw an error
  is executed, write the pointer back to the current activation using
  ``thr->ptr_curr_pc`` which points to the stack frame location containing
  ``curr_pc``.

* Duktape 1.2: maintain all PC values as numeric indices (not pointers and
  not pre-multiplied by bytecode opcode size).  The current PC is always
  looked up from the current activation.

* Same as Duktape 1.3 behavior but maintain a cached copy of the topmost
  activation's bytecode pointer in ``thr->curr_pc``.  The copy back operation
  is needed but doesn't need to peek into the bytecode executor stack frame.
  This works quite well because ``thr`` is a "hot" variable.  However, the
  stack local ``curr_pc`` used in Duktape 1.3 is faster.

* Use direct bytecode pointers in activations, keep a pointer to the current
  activation in the executor, and use ``act->curr_pc`` for dispatch.  There's
  no need for a copy back operation because activation states are always in
  sync.  This is faster than the Duktape 1.2 approach, but significantly
  slower than the ``thr->curr_pc`` or the Duktape 1.3 approach (part of that
  is probably because there's more register pressure).

Comparison between curr_pc alternatives
---------------------------------------

The current Duktape 1.3 approach is a bit error prone because of the need to
sync the executor local ``curr_pc`` back to ``act->curr_pc`` in multiple code
paths.  Another alternative would be to dispatch using ``act->curr_pc``
directly.  While that is faster than Duktape 1.2, it is significantly slower
than dispatching using executor local ``curr_pc`` (or ``thr->curr_pc``).

The measurements below are using ``gcc -O2`` on x64::

    # Duktape 1.3, dispatch using executor local variable curr_pc
    $ sudo nice -20 python util/time_multi.py --count 10 --mode all --verbose ./duk.O2.local_pc tests/perf/test-empty-loop.js
    Running: 2.180000 2.170000 2.180000 2.290000 2.180000 2.200000 2.190000 2.190000 2.220000 2.200000
    min=2.17, max=2.29, avg=2.20, count=10: [2.18, 2.17, 2.18, 2.29, 2.18, 2.2, 2.19, 2.19, 2.22, 2.2]

    # Duktape 1.2, dispatch using a numeric PC index
    $ sudo nice -20 python util/time_multi.py --count 10 --mode all --verbose ./duk.O2.123 tests/perf/test-empty-loop.js
    Running: 3.100000 3.100000 3.120000 3.120000 3.160000 3.300000 3.370000 3.410000 3.370000 3.390000
    min=3.10, max=3.41, avg=3.24, count=10: [3.1, 3.1, 3.12, 3.12, 3.16, 3.3, 3.37, 3.41, 3.37, 3.39]

    # Alternative; dispatch using thr->curr_pc
    $ sudo nice -20 python util/time_multi.py --count 10 --mode all --verbose ./duk.O2.thr_pc tests/perf/test-empty-loop.js
    Running: 2.310000 2.330000 2.310000 2.300000 2.400000 2.290000 2.310000 2.290000 2.300000 2.300000
    min=2.29, max=2.40, avg=2.31, count=10: [2.31, 2.33, 2.31, 2.3, 2.4, 2.29, 2.31, 2.29, 2.3, 2.3]

    # Alternative; dispatch using act->curr_pc
    $ sudo nice -20 python util/time_multi.py --count 10 --mode all --verbose ./duk.O2.act_pc tests/perf/test-empty-loop.js
    Running: 2.590000 2.580000 2.600000 2.600000 2.600000 2.660000 2.600000 2.640000 2.860000 2.860000
    min=2.58, max=2.86, avg=2.66, count=10: [2.59, 2.58, 2.6, 2.6, 2.6, 2.66, 2.6, 2.64, 2.86, 2.86]

Accessing constants
===================

The executor stores a copy of the ``duk_hcompiledfunction`` constant table
base address into a local variable ``consts``.  This reduces code footprint
and performs better compared to reading the consts base address on-the-fly
through the function reference.  Because the constants table has a stable
base address, this is easy and safe.

Accessing registers
===================

The executor currently accesses the stack frame base address (needed to read
registers) through ``thr`` as ``thr->valstack_bottom``.  This is reasonably
OK because ``thr`` is a "hot" variable.

The register base address could also be copied to a local variable as is done
for constants.  However, ``thr->valstack_bottom`` is not a stable address and
may be changed by any side effect (because any side effect can cause a value
stack resize, e.g. if a finalizer is invoked).

If a local variable were to be used, it would need to be updated when the
value stack is resized.  It's not certain if overall performance would be
improved.  This was postponed to Duktape 1.4:

* https://github.com/svaarala/duktape/issues/298