[jsc-dev] [webkit-dev] Can we drop supporting mixed Wasm::MemoryMode in one process?

[Saam Barati]

I would also expect bytecode interpreter compilation to be so much faster than JIT compilation that we could laziliy compile either:
- When we instantiate
- When functions are called for the first time

- Saam

> On Aug 28, 2018, at 12:17 PM, Yusuke Suzuki <yusukesuzuki at slowstart.org> wrote:
> 
> 
> 
> On Wed, Aug 29, 2018 at 4:10 Filip Pizlo <fpizlo at apple.com <mailto:fpizlo at apple.com>> wrote:
> 
>> On Aug 28, 2018, at 12:09 PM, Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>> 
>> 
>> 
>> On Wed, Aug 29, 2018 at 3:58 Filip Pizlo <fpizlo at apple.com <mailto:fpizlo at apple.com>> wrote:
>> 
>>> On Aug 28, 2018, at 11:56 AM, Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>>> 
>>> 
>>> 
>>> On Wed, Aug 29, 2018 at 3:49 Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>>> 
>>> 
>>> On Wed, Aug 29, 2018 at 3:27 Filip Pizlo <fpizlo at apple.com <mailto:fpizlo at apple.com>> wrote:
>>> 
>>>> On Aug 28, 2018, at 11:25 AM, Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>>>> 
>>>> Thanks!
>>>> 
>>>> On Wed, Aug 29, 2018 at 3:22 Filip Pizlo <fpizlo at apple.com <mailto:fpizlo at apple.com>> wrote:
>>>> I don’t like this proposal.
>>>> 
>>>> If we are running low on memory, we should switch to bounds checked memory.
>>>> 
>>>> How about using bound checking mode exclusively for low environment?
>>> 
>>> That would mean that, paradoxically, having a machine with a lot of memory means being able to spawn fewer wasm instances.
>>> 
>>> We want to support lightweight wasm instances because it wouldn’t be good to rule that out as a use case.
>>> 
>>> Hmmm, can we compile the BBQ phase / initial wasm code without knowing the attached memory’s mode? The current strategy basically defers compilation of wasm module until the memory mode is found.
>>> Because of this, WebAssembly.compile is not so meaningful in our implementation right now...
>>> And wasm ES6 module can import memory externally. This means that we cannot start wasm compilation after the memory mode of the impprted memory (described in the imported modulr) is downloaded, analyzed and found.
>>> 
>>> How about always compiling BBQ code with bound checking mode?
>>> It should work with signaling memory with small / no tweaks. And OMG code will be compiled based on the memory mode attached to the module.
>>> Since BBQ -> OMG function call should be linked, we need to call appropriate func for the running memory mode, but it would not introduce significant complexity.
>> 
>> What complexity are you trying to fix, specifically?
>> 
>> What I want is starting compilation before the memory is attached a.k.a. instantiated)
>> 
>> 
>> I think that what we really want is an interpreter as our baseline.  Then tier-up to BBQ or OMG from there.  In that world, I don’t think any of this matters.
>> 
>> Does this interpreter execute wasm binary directly? If so, we can skip compiling and all should work well!
>> 
>> Even if we want some own bytecode (like stack VM to register VM etc.), it is ok if the compilation result is not tied to the memory mode.
> 
> I don’t know if it will execute the wasm binary directly, but whatever bytecode it runs could be dissociated from memory mode.
> 
> Thanks, that sounds reasonable!
> If the bytecode compilation etc. is disassociated from the memory mode, the bytecode can be compiled before instantiated. It means that wasm module can be shared between workers (like postMessage the wasm module having compiled bytecode). And we can start compiling before the memory is attached (instantiated).
> 
> 
> 
> -Filip
> 
> 
>> 
>> If the compilation result is tied to the memory mode, then we still need to defer the compilation until the memory mode is attached.
>> 
>> 
>> -Filip
>> 
>> 
>>> 
>>> 
>>> 
>>> 
>>> -Filip
>>> 
>>> 
>>>> 
>>>> 
>>>> -Filip
>>>> 
>>>> 
>>>> 
>>>>> On Aug 28, 2018, at 11:21 AM, Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>>>>> 
>>>> 
>>>> 
>>>>> Posted this mail to webkit-dev mailing list too :)
>>>>> 
>>>>> On Wed, Aug 29, 2018 at 3:19 AM Yusuke Suzuki <yusukesuzuki at slowstart.org <mailto:yusukesuzuki at slowstart.org>> wrote:
>>>>> Hi JSC folks,
>>>>> 
>>>>> In Wasm supported environment, our MemoryMode is a bit dynamic.
>>>>> When we fail to allocate WasmMemory for signaling mode, we fall back to the bound checking memory instead.
>>>>> 
>>>>> But Wasm code compiled for signaling / bound checking is incompatible. If the code is compiled
>>>>> as signaling mode, and if we attach the memory for bound checking, we need to recompile the
>>>>> code for bound checking mode. This introduces significant complexity to our wasm compilation.
>>>>> And our WebAssembly.compile is not basically compiling: it is just validating.
>>>>> Actual compiling needs to be deferred until the memory is attached by instantiating.
>>>>> It is not good when we would like to share WasmModule among multiple wasm threads / workers in the future, since the "compiled" Wasm module is not actually compiled.
>>>>> 
>>>>> So, my proposal is, can we explore the way to exclusively support one of MemoryMode in a certain architecture?
>>>>> For example, in x64, enable signaling mode, and we report OOM errors if we fail to allocate WasmMemory with signaling mode.
>>>>> 
>>>>> Best regards,
>>>>> Yusuke Suzuki
>>>> 
>>>>> _______________________________________________
>>>>> webkit-dev mailing list
>>>>> webkit-dev at lists.webkit.org <mailto:webkit-dev at lists.webkit.org>
>>>>> https://lists.webkit.org/mailman/listinfo/webkit-dev <https://lists.webkit.org/mailman/listinfo/webkit-dev>
> 
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org
> https://lists.webkit.org/mailman/listinfo/webkit-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/jsc-dev/attachments/20180828/c93a3626/attachment-0001.html>