Similar to other languages that use linear memory, all data in AssemblyScript becomes stored at a specific memory offset so other parts of the program can read and modify it.
The WebAssembly.Memory instance used by your program can be imported from the host by using the
--importMemory flag on the command line. The module will then expect an import named
memory within the
env module. Likewise, a module always exports its memory as
There is one special case to mention when it comes to accessing memory while top-level statements are still executing. Instantiation must return first so one can get a hold of the exported memory instance, hence it is not possible to print a string externally before this happens, unless memory has been imported or an explicit start function has been exported using the
--explicitStart option essentially delays all top-level statements until
__start is called externally, which must happen before calling any other exports, so one can get a hold of the memory early, before top-level statements run, even if it has not been imported.
Internally, there are two regions of memory the compiler is aware of:
Memory starts with static data, like strings and arrays (of constant values) the compiler encountered while translating the program. Unlike in other languages, there is no concept of a stack in AssemblyScript and it instead relies on WebAssembly's execution stack exclusively.
A custom region of memory can be reserved using the
--memoryBase option. For example, if one needs an image buffer of exactly N bytes, instead of allocating it one could reserve that space, telling the compiler to place its own static data afterwards, partitioning memory in this order:
As specified with
Starting right after reserved memory
Starting right after static memory
Dynamic memory, commonly known as the heap, is managed by the AssemblyScript runtime, at runtime. When a chunk of memory is requested by the program, the runtime's memory manager reserves a suitable region and returns a pointer to it to the program. The lifetime of objects allocated from the heap is then tracked by the runtime's garbage collector, and once it is not needed anymore, the object's chunk of memory is returned to the memory manager for reuse.
When writing high level AssemblyScript, the above is pretty much everything one needs to know. But there is a low level perspective of course.
When implementing low-level code, the end of static memory respectively the start of dynamic memory can be obtained by reading the
__heap_base global. Memory managers for example do this to decide where to place their bookkeeping structures before partitioning the heap.
The memory manager guarantees an alignment of 16 bytes, so an
ArrayBuffer, which can be the backing buffer of multiple views, always fits up to
v128 values with native alignment.
Objects in AssemblyScript have a common hidden header used by the runtime to keep track of them. The header includes information about the block used by the memory manager, state information used by the garbage collector, a unique id per concrete class and the data's actual size. The length of a
String (id = 1) is computed from that size for example. The header is "hidden" in that the reference to an object points right after it, at the first byte of the object's actual data.
General block information
Reference count etc.
Unique id of the concrete class
Size of the data following the header
The most basic objects using the common header are
String. These do not have any fields but are just data:
The first byte
The second byte
The N-th byte
The first character
The second character
N << 1
The N-th character
Set use one or multiple
ArrayBuffers to store their data, but the backing buffer is exposed by the typed array views only because other collections can grow automatically, which would otherwise lead to no longer valid references sticking around.
This is not an actual class but a helper to simplify the implementation of the typed array views. Each typed array view like
Uint8Array is like a subclass of
ArrayBufferView with a static value type attached.
Backing buffer reference
Start of the data within .data
Length of the data from .dataStart
Mutable length of the data the user is interested in
The layouts of
Map are a bit more sophisticated and therefore not mentioned here, but feel free to take a look at their sources.
One obvious shortcoming of this layout alone is that accessing the values of an array must go through one level of indirection. For example, an
arr[i] must first obtain a reference to the backing buffer and only then is able to read the actual value. One possible way to solve this is to make use of the Multi-value proposal for WebAssembly eventually, lowering typed arrays for example to multiple values that are passed around directly, so the reference to the backing buffer is already on the execution stack. Another is to provide more knowledge about specific structures to Binaryen, so it can cache the reference to the backing buffer as long as the array does not grow, but it is unclear how feasible this is.
A more general issue is that AssemblyScript is essentially a compiler on top of another compiler (here. Binaryen), with Binaryen already providing all the utility to do optimizations that we do not want to duplicate. Thus, it is still to decide how to get rid of unnecessary range checks for example. For now, there is the famous
unchecked(expression) built-in that explicitly request no bounds checks on indexed access, but this certainly isn't forever.