Copyright 2016-2017 Moddable Tech, Inc.
Revised: September 9, 2017
Warning: These notes are preliminary. Omissions and errors are likely. If you encounter problems, please ask for assistance.
A platform is a combination of hardware and system software. For each platform, XS requires an interface file, xsPlatform.h
, and an implementation file, xsPlatform.c
Historically, XS used one interface file, xsPlatform.h
splitting the implementation into two files: xsPlatform.c
and xsHost.c
. Many platforms shared the same interface and implementation files, based either on the KinomaJS platform abstraction, or on an adhoc platform abstraction for command line tools.
Further, an XS machine had many ways to find and load modules and programs: from JS files, from stand alone compiled XSB files with or without companion DLL or SO files, and from a linked XSA file with a companion DLL or SO file... The XS platform was in charge of providing such options.
When we started working on microcontrollers, the main inspiration for XS platforms was the adhoc platform abstraction for command line tools, which was the most complex version.
Today the XS runtime has been significantly streamlined, especially on microcontrollers. XS machines are always cloned from a read-only machine prepared by the XS linker. There are only modules, byte coded by the XS compiler. Modules are either preloaded or prepared to be loaded and unloaded at runtime.
Consequently, it is now much simpler to build an XS platform. This document describes the necessary interface and implementation files.
XS uses a few basic types that the interface file has to define.
#include <stdint.h>
typedef int8_t txS1;
typedef uint8_t txU1;
typedef int16_t txS2;
typedef uint16_t txU2;
typedef int32_t txS4;
typedef uint32_t txU4;
XS mostly relies on constants and functions from the C stantard library, accessed thru macros with C_
or c_
prefixes:
#include <math.h>
#define C_NAN NAN
//...
#include <stdlib.h>
#define c_free free
#define c_malloc malloc
//...
Such definitions, and the corresponding includes, are the most significant part of the interface file. The macros allows a platform to provide its own constants and functions. See any of the provided xsPlatform.h
for the list of macros to define.
The Xtensa instruction set and architecture, used most notably in microcontrollers by Espressif, requires special macros to locate certain constant data in ROM and to read that data. On other platforms these macros are trivially defined:
#define c_read8(POINTER) *((txU1 *)(POINTER))
#define c_read16(POINTER) *((txU2 *)(POINTER))
#define c_read32(POINTER) *((txU4 *)(POINTER))
#define ICACHE_FLASH_ATTR
#define ICACHE_STORE_ATTR
#define ICACHE_RODATA_ATTR
#define ICACHE_XS6RO_ATTR
#define ICACHE_XS6RO2_ATTR
#define ICACHE_XS6STRING_ATTR
#define mxGetKeySlotID(SLOT) (SLOT)->ID
#define mxGetKeySlotKind(SLOT) (SLOT)->kind
The platform can add fields to the machine record by defining the mxMachinePlatform
macro. Since the machine is passed to all functions that XS calls (as the ubiquitous the
), it is a convenient way for platforms to have their own context besides the application context.
For instance, on Mac, the mxMachinePlatform
macro adds references to a socket and a run loop source for the communication with xsbug, and another run loop source for promises.
#include <CoreServices/CoreServices.h>
#define mxMachinePlatfom \
CFSocketRef connection; \
CFRunLoopSourceRef connectionSource; \
CFRunLoopSourceRef promiseSource;
On Windows, the mxMachinePlatform
macro adds the socket and message window handles that are used for the same purposes.
#include <winsock2.h>
#define mxMachinePlatfom \
SOCKET connection; \
HWND window;
The implementation file first includes xsAll.h
, which contains the definitions of all XS macros and types, and the declarations of all XS extern functions. Then the platform has to implement the functions described here under.
XS machines do not support multiple threads, though platforms can support multiple threads, each with their own XS machines. All calls and callbacks described here must happen in the thread that created or cloned the machine.
The functions are grouped into meaningful sections. The xsPlatform.c file can also provide POSIX functions that the platform is missing.
--
void fxCreateMachinePlatform(txMachine* the)
fxCreateMachinePlatform
is called when creating and cloning an XS machine. The platform initializes the fields defined by its mxMachinePlatform
macro. By default all fields are zero.
--
void fxDeleteMachinePlatform(txMachine* the)
fxDeleteMachinePlatform
is called when deleting an XS machine. The platform must dispose or free here appropriate fields defined by its mxMachinePlatform
macro.
--
The functions in this section are only necessary for the debug version of XS. They can be condtionally defined within:
#ifdef mxDebug
// debug functions
#endif
If the platform does not support the communication with xsbug, functions in this section can be empty, except fxIsConnected
and fxIsReadable
, which must return 0
.
Communication between xsbug and the XS machine can be done over either a TCP/IP or serial connection. In the case of a TCP/IP connection, xsbug is the server and XS machines are clients. When using a serial connection, xsbug continues to communication over TCP/IP and a bridge running on the computer relays data between the serial and TCP connections. In the case of the ESP8266, this relay is performed by serial2xsbug.
Platforms must implement fxIsReadable
to allow XS machines to receive messages from xsbug while executing byte codes, i.e. when platforms are inside the fxRun
function. Most of the time, platforms are outside the fxRun
function. So they use a system event and fxDebugCommand
to tell XS about messages from xsbug.
For instance on Mac the platform uses CFSocketCreate
with a kCFSocketReadCallBack
:
void fxReadableCallback(CFSocketRef socketRef, CFSocketCallBackType cbType, CFDataRef addr, const void* data, void* context)
{
txMachine* the = context;
if (fxIsReadable(the))
fxDebugCommand(the);
}
On Windows the platform uses WSAAsyncSelect
with the WM_XSBUG
message:
LRESULT CALLBACK fxMessageWindowProc(HWND window, UINT message, WPARAM wParam, LPARAM lParam)
{
switch(message) {
#ifdef mxDebug
case WM_XSBUG: {
txMachine* the = (txMachine*)GetWindowLongPtr(window, 0);
if (fxIsReadable(the))
fxDebugCommand(the);
} break;
#endif
default:
return DefWindowProc(window, message, wParam, lParam);
}
return 0;
}
--
void fxConnect(txMachine* the)
XS calls fxConnect
to connect the
machine to xsbug.
For TCP/IP connections, platforms create a socket and connect it to xsbug. On Mac and Windows the address of xsbug is usually localhost
, on other platforms it is usually defined by an environment variable. The port of xsbug defaults to 5002
by convention.
Machines are connect to xsbug after being created, i.e. fxConnect
happens after fxCreateMachinePlatform
.
--
void fxDisconnect(txMachine* the)
XS calls fxDisconnect
to disconnect the
machine from xsbug.
For TCP/IP connections, platforms close the socket.
Machines are disconnected before being deleted, i.e. fxDisconnect
happens before fxDeleteMachinePlatform
.
--
txBoolean fxIsConnected(txMachine* the)
XS calls fxIsConnected
to know if the
machine is connected to xsbug.
--
txBoolean fxIsReadable(txMachine* the)
XS calls fxIsReadable
to know if the
machine received a message from xsbug. Platforms must return 1 or 0 depending on the availability of bytes to read.
The performance of the implementation of fxIsReadable
is important since XS calls fxIsReadable
at every LINE
byte code (e.g. for each line of JavaScript source code executed).
--
void fxReceive(txMachine* the)
XS calls fxReceive
to receive a message from xsbug. The implementation reads bytes into the->debugBuffer
and sets the->debugOffset
to the number of bytes received.
XS calls fxReceive
repeatedly until the entire message is received. The maximum number of bytes that can be read by fxReceive
is sizeof(the->debugBuffer) - 1
.
--
void fxSend(txMachine* the, txBoolean more)
XS calls fxSend
to send a message to xsbug. The implementation gets the number of bytes to send from the->echoOffset
and write bytes from the->echoBuffer
.
XS calls fxSend
repeatedly until the entire message is sent, more
equals 1
while the message is incomplete, 0
when the message is complete.
--
The standard eval
function, Function
constructor and Generator
constructor must transform source code into byte codes and keys.
XS lets the platform decides is such feature is worth the memory it takes.
--
txScript* fxParseScript(txMachine* the, void* stream, txGetter getter, txUnsigned flags)
XS calls fxParseScript
to transform source code into XS byte codes and keys. The stream
and getter
arguments allow the parser to access the source code. The flags
argument tells the parser the kind of source code: mxModuleCode
, mxProgramCode
or mxEvalCode
.
If the platform supports such feature, it must include xsScript.h
and implements fxParseScript
like:
#include "xsScript.h"
txScript* fxParseScript(txMachine* the, void* stream, txGetter getter, txUnsigned flags)
{
txParser _parser;
txParser* parser = &_parser;
txParserJump jump;
txScript* script = NULL;
fxInitializeParser(parser, the, 32*1024, 1993);
parser->firstJump = &jump;
if (c_setjmp(jump.jmp_buf) == 0) {
fxParserTree(parser, stream, getter, flags, NULL);
fxParserHoist(parser);
fxParserBind(parser);
script = fxParserCode(parser);
}
fxTerminateParser(parser);
return script;
}
The platform must also compile and link xsScript.c
, xsLexical.c
, xsSyntaxical.c
, xsTree.c
, xsSourceMap.c
, xsScope.c
and xsCode.c
.
If the platform does not support such feature, fxParseScript
must return NULL
and the C files here above do not have to be compiled and linked.
--
Keys are the names and symbols that XS uses to identify properties.
--
void fxBuildKeys(txMachine* the)
fxBuildKeys
is called only when creating an XS machine, in order to initialize the default keys used by the standard ECMAScript host functions.
On most platforms today, XS machines are cloned. The default keys are available and ready to be used in the read-only machine. So fxBuildKeys
is never called and can be empty.
If the platform supports the creation of XS machines from scratch, fxBuildKeys
must be implemented as:
void fxBuildKeys(txMachine* the)
{
int i;
for (i = 0; i < XS_SYMBOL_ID_COUNT; i++) {
txID id = the->keyIndex;
txSlot* description = fxNewSlot(the);
fxCopyStringC(the, description, gxIDStrings[i]);
the->keyArray[id] = description;
the->keyIndex++;
}
for (; i < XS_ID_COUNT; i++) {
fxID(the, gxIDStrings[i]);
}
}
--
XS machines use two heaps: the chunks heap and the slots heap.
Chunks are blocks of variable size that the garbage collector can move to compact memory. XS stores strings, buffers, arrays, etc into chunks. On microcontrollers without a dedicated memory management unit, chunks are also useful to store any kind of data. For instance Piu uses chunks to store its containment hierarchy.
Slots are blocks of fixed size (four times the size of a pointer) that never move. XS maintains a list of free slots, slots are allocated from the list and freed into the list by the garbage collector.
--
void* fxAllocateChunks(txMachine* the, txSize size)
XS calls fxAllocateChunks
to get a system memory block for chunks. Usually implemented as:
return c_malloc(size);
XS throws an exception if fxAllocateChunks
returns NULL.
XS checks if the result of fxAllocateChunks
is contiguous to the->firstBlock
so microcontrollers can grow the chunks heap without fragmenting system memory.
--
txSlot* fxAllocateSlots(txMachine* the, txSize count)
XS calls fxAllocateSlots
to get a system memory block for count
slots. Usually implemented as:
return (txSlot*)c_malloc(count * sizeof(txSlot));
XS throws an exception if fxAllocateSlots
returns NULL.
--
void fxFreeChunks(txMachine* the, void* chunks)
XS calls fxFreeChunks
to free the chunks
system memory block. Usually implemented as:
c_free(chunks);
--
void fxFreeSlots(txMachine* the, void* slots)
XS calls fxFreeSlots
to free the slots
system memory block. Usually implemented as:
c_free(slots);
--
void fxMarkHost(txMachine* the, txMarkRoot markRoot)
XS calls fxMarkHost
at the beginning of a garbage collection. The platform can use the markRoot
callback to mark its own roots:
(*markRoot)(the, &root);
--
void fxSweepHost(txMachine* the)
XS calls fxSweepHost
at the end of a garbage collection.
--
On platforms that support several ways to get modules, the implementation of fxFindModule
and fxLoadModule
can be complex. On microcontrollers, where all modules are prepared or preloaded, the implementation of fxFindModule
and fxLoadModule
can be simple enough, as demonstrated by the code snippets here under.
--
txID fxFindModule(txMachine* the, txID moduleID, txSlot* slot)
XS calls fxFindModule
to find an imported or required module.
The moduleID
argument is the importing or requiring module identifier. It is XS_NO_ID
when the machine itself requires a module.
The slot
argument is the imported or required module name. It is the module specifier of the import
syntactical construct or the module parameter of the require
host function.
If the module is found, fxFindModule
returns the module identifier, otherwise fxFindModule
returns XS_NO_ID
.
A module identifier is a unique txID
, but the platform defines the format of its corresponding key: it can be a path, a URL, a URI...
The platform defines also how the importing or requiring module identifier and the imported or required module name are merged. The usual convention is based on absolute (/*
), relative (./*
, ../*
) or search (*) paths.
Finding modules can involve looking for various kinds of files, using a set of preferred locations, etc. But on microcontrollers, all modules modules are prepared and ready to be found:
txID fxFindModule(txMachine* the, txID moduleID, txSlot* slot)
{
txPreparation* preparation = the->archive;
char name[PATH_MAX];
char path[PATH_MAX];
txBoolean absolute = 0, relative = 0, search = 0;
txInteger dot = 0;
txSlot *key;
txString slash;
txID id;
fxToStringBuffer(the, slot, name, sizeof(name));
if (!c_strncmp(name, "/", 1)) {
absolute = 1;
}
else if (!c_strncmp(name, "./", 2)) {
dot = 1;
relative = 1;
}
else if (!c_strncmp(name, "../", 3)) {
dot = 2;
relative = 1;
}
else {
relative = 1;
search = 1;
}
if (absolute) {
c_strcpy(path, preparation->base);
c_strcat(path, name + 1);
if (fxFindScript(the, path, &id))
return id;
}
if (relative && (moduleID != XS_NO_ID)) {
key = fxGetKey(the, moduleID);
c_strcpy(path, key->value.key.string);
slash = c_strrchr(path, '/');
if (!slash)
return XS_NO_ID;
if (dot == 0)
slash++;
else if (dot == 2) {
*slash = 0;
slash = c_strrchr(path, '/');
if (!slash)
return XS_NO_ID;
}
if (!c_strncmp(path, preparation->base, preparation->baseLength)) {
*slash = 0;
c_strcat(path, name + dot);
if (fxFindScript(the, path, &id))
return id;
}
}
if (search) {
c_strcpy(path, preparation->base);
c_strcat(path, name);
if (fxFindScript(the, path, &id))
return id;
}
return XS_NO_ID;
}
txBoolean fxFindScript(txMachine* the, txString path, txID* id)
{
txPreparation* preparation = the->archive;
txInteger c = preparation->scriptCount;
txScript* script = preparation->scripts;
path += preparation->baseLength;
c_strcat(path, ".xsb");
while (c > 0) {
if (!c_strcmp(path, script->path)) {
path -= preparation->baseLength;
*id = fxNewNameC(the, path);
return 1;
}
c--;
script++;
}
*id = XS_NO_ID;
return 0;
}
--
void fxLoadModule(txMachine* the, txID moduleID)
XS calls fxLoadModule
to tell the platform to prepare the byte codes, keys and host functions of a module. When ready, the platform must call fxResolveModule
with a txScript
structure that references the byte codes, keys and host functions of the module.
Preparing modules can involve reading and mapping files, parsing, scoping and byte coding scripts, loading dynamic libraries, etc. But on microcontrollers, all txScript
structures are available and ready to be used:
void fxLoadModule(txMachine* the, txID moduleID)
{
txString path = fxGetKeyName(the, moduleID);
txScript* script = fxLoadScript(the, path);
fxResolveModule(the, moduleID, script, C_NULL, C_NULL);
}
txScript* fxLoadScript(txMachine* the, txString path)
{
txPreparation* preparation = the->archive;
txInteger c = preparation->scriptCount;
txScript* script = preparation->scripts;
path += preparation->baseLength;
while (c > 0) {
if (!c_strcmp(path, script->path))
return script;
c--;
script++;
}
return C_NULL;
}
--
Promises are essentially asynchronous. The then
method of a Promise
object takes two arguments: a function to call when the promise is fulfilled and a function to call when the promise is rejected. Both functions have to be called by a Job:
A Job is an abstract operation that initiates an ECMAScript computation when no other ECMAScript computation is currently in progress. (ECMAScript® 2015 Language Specification, Section 8.4).
XS takes care queuing and running Jobs but relies on platforms for their scheduling.
--
void fxQueuePromiseJobs(txMachine* the)
XS calls fxQueuePromiseJobs
once when jobs have been queued. Platforms can use any mechanism to defer a call to fxRunPromiseJobs
.
For instance on Mac the platform uses a run loop source and CFRunLoopSourceSignal
:
void fxQueuePromiseJobsCallback(void *info)
{
txMachine* the = info;
fxRunPromiseJobs(the);
}
void fxQueuePromiseJobs(txMachine* the)
{
CFRunLoopSourceSignal(the->promiseSource);
}
On Windows the platform uses a message window and PostMessage
:
LRESULT CALLBACK fxMessageWindowProc(HWND window, UINT message, WPARAM wParam, LPARAM lParam)
{
switch(message) {
case WM_PROMISE: {
txMachine* the = (txMachine*)GetWindowLongPtr(window, 0);
fxRunPromiseJobs(the);
} break;
default:
return DefWindowProc(window, message, wParam, lParam);
}
return 0;
}
void fxQueuePromiseJobs(txMachine* the)
{
PostMessage(the->window, WM_PROMISE, 0, 0);
}
From XS point of view, SharedArrayBuffer
instances are host objects, i.e. instances with an internal host slot. The data of the host slot is a pointer to the data of a shared chunk. The destructor of the host slot is fxReleaseSharedChunk
.
What is a shared chunk is defined by the platform. XS atomically accesses 8-bit, 16-bit or 32-bit signed or unsigned integers inside the data of a shared chunk. XS accesses integers either thru GCC atomics, or between calls to fxLockSharedChunk
and fxUnlockSharedChunk
Since Atomics.wait
and Atomics.wake
require to synchonize the shared cluster of machines created or cloned by XS, platforms usually need a global synchronization mechanism, and synchronization related fields in every machine record, thru the mxMachinePlatform
macro explained here above.
void fxInitializeSharedCluster();
Applications that use Atomics.wait
and Atomics.wake
must call xsInitializeSharedCluster
before creating or cloning their first machine. xsInitializeSharedCluster
is the application programming interface, fxInitializeSharedCluster
is the platform implementation.
fxInitializeSharedCluster
allows the platform to setup its global synchronization mechanism.
The thread that calls fxInitializeSharedChunks
must be the thread that runs the user interface, usually the main thread. Atomics.wait
fails for all machines running in that thread.
void fxTerminateSharedCluster();
Applications that use Atomics.wait
and Atomics.wake
must call xsTerminateSharedCluster
after deleting their last machine. xsTerminateSharedCluster
is the application programming interface, fxTerminateSharedCluster
is the platform implementation.
fxTerminateSharedCluster
allows the platform to cleanup its global synchronization mechanism.
void* fxCreateSharedChunk(txInteger byteLength);
fxCreateSharedChunk
allocates a shared chunk, byteLength
is the size of its data, which must be initialised to zero.
Typically platforms use a reference count to track how many machines are referencing the shared chunk. fxCreateSharedChunk
must initialise the reference count to one.
fxCreateSharedChunk
returns a pointer to the data.
void fxLockSharedChunk(void* data);
fxLockSharedChunk
locks the shared chunk, data
is a pointer to the data of the shared chunk.
fxLockSharedChunk
is never called if the platform supports GCC atomics.
txInteger fxMeasureSharedChunk(void* data);
fxMeasureSharedChunk
returns the size of the data of the chunk, data
is a pointer to the data of the shared chunk.
void fxReleaseSharedChunk(void* data);
Machines call fxReleaseSharedChunk
when they do not reference the shared chunk anymore, data
is a pointer to the data of the shared chunk.
Typically platforms use an atomic operation to decrement the reference count of the shared chunk and free the shared chunk when the reference count is zero.
fxReleaseSharedChunk
is the destructor of the host slot.
void* fxRetainSharedChunk(void* data);
A machine calls fxRetainSharedChunk
when marshalling a shared chunk to another machine. data
is a pointer to the data of the shared chunk.
Typically platforms use an atomic operation to increment the reference count of the shared chunk.
void fxUnlockSharedChunk(void* data);
fxUnlockSharedChunk
unlocks the shared chunk. data
is a pointer to the data of the shared chunk.
fxUnlockSharedChunk
is never called if the platform supports GCC atomics.
txInteger fxWaitSharedChunk(txMachine* the, void* data, txInteger byteOffset, txInteger value, txNumber timeout);
If the application did not call fxInitializeSharedCluster
or if the current thread is the thread that called fxInitializeSharedCluster
, fxWaitSharedChunk
throws a TypeError
object.
If the 32-bit signed integer at byteOffset
in data
is not equal to value
, fxWaitSharedChunk
returns -1
immediately. Else fxWaitSharedChunk
suspends the current thread.
If a matching call to fxWakeSharedChunk
resumed the thread, fxWaitSharedChunk
returns 1
. A matching call is a call with the same data
and byteOffset
.
If timeout
expired, fxWaitSharedChunk
returns 0
. timeout
is a number between Date.now()
and C_INFINITY
.
txInteger fxWakeSharedChunk(txMachine* the, void* data, txInteger byteOffset, txInteger count);
If the application did not call fxInitializeSharedCluster
, fxWakeSharedChunk
returns 0
.
fxWakeSharedChunk
resumes at most count
threads that have been suspended by a matching call to fxWaitSharedChunk
. A matching call is a call with the same data
and byteOffset
.
fxWakeSharedChunk
returns the number of threads that resumed.
XS provides four default implementations of shared cluster and chunks:
- For systems with Linux futex and GCC atomics
- define
mxUseLinuxFutex
- define
mxUseGCCAtomics
- define
mxUseDefaultSharedChunks
- define
- For systems with POSIX threads, with or without GCC atomics
- define
mxUsePOSIXThreads
- define
mxUseGCCAtomics
if the tool chain supports GCC atomics - define
mxUseDefaultSharedChunks
- define
- For Windows
- define
mxUseDefaultSharedChunks
- define
- For systems with a single thread
- define
mxUseDefaultSharedChunks
- define
All default implementations use c_malloc
and c_free
to create and delete shared chunks.
On systems with POSIX threads and on Windows, to use the default implementation of shared cluster and chunks, the platform must define the mxMachinePlatform
macro with at least the following fields:
#define mxMachinePlatform \
void* waiterCondition; \
void* waiterData; \
txMachine* waiterLink;
Obviously, on systems with a single thread, Atomics.wait
always fails and Atomics.wake
always returns zero.