RFC: serialize pointer- and padding-free objects in one write

[JeffBezanson]

This is a hack to try to address #14106. Any objects that contain no references and no padding can be sent as a single binary blob, instead of iterating over the fields which entails expensive boxing.

Before:

    From worker 2:  (300000,)
  2.356665 seconds (20.25 M allocations: 1.398 GB, 3.00% gc time)
    From worker 3:  (300000,)
  2.218371 seconds (20.10 M allocations: 1.391 GB, 2.71% gc time)

after:

    From worker 2:  (300000,)
  0.649856 seconds (2.25 M allocations: 165.977 MB, 3.34% gc time)
    From worker 3:  (300000,)
  0.501824 seconds (2.10 M allocations: 159.813 MB, 1.42% gc time)

[JeffBezanson]

This is regrettable but there are no read! methods that accept a pointer.

[vtjnash]

i've thought they should be unsafe_read! and implemented consistently instead of the current ad-hoc implementations

[JeffBezanson]

Sounds good to me.

[JeffBezanson]

cc @amitmurthy @abieler

[JeffBezanson]

Oh bloody hell this ends up routing around the code that serializes Ptr as NULL. Gotta love those tests that check that something doesn't work. Does anybody else think this isn't worth a 4x slowdown and 10x increase in allocation? If we want both, I guess we could add a hasptr flag to DataType.

[StefanKarpinski]

Le sigh. Should probably just add the hasptr flag.

[JeffBezanson]

But then, why is Ptr the only bits type that's allowed to have a non-obvious serialize method? I guess for this to really work we need to limit it to a fixed set of known types.

[yuyichao]

What about Cstring and Cwstring. These are also basically used as alias of the corresponding Ptr types although they are probably not currently used too much in composite types.

[JeffBezanson]

Those types currently can't be serialized at all. And for the sake of not making people's parallel code slow as hell, let's hope they never can be.

[vtjnash]

the pointer returned here may be to a box that is not gc-rooted. in general, pointer_from_objref should never be called on something that is immutable and pointerfree (I think we should make it an error)

i'm looking into possible alternatives that would provide the same benefit, but doesn't require calling pointer_from_objref