QuoteUnpicker.unpickleExpr and QuoteUnpicker.unpickleType are slow

[deusaquilus]

Compiler version

3.0.0-RC3

I am getting ~700ms range bottlenecks from unpickleExpr and unpickleType (blue and light-blue in the chart below) for the Proto-Quill parser and expr-accumulation (for a simple query with 20 parameters). I can't accommodate large nested queries with this kind of latency.

Are there any tricks I can use to optimize unpickling?
Are there any compiler args I can use to make unpickling faster or to disable it for certain parts of the code?

Also, if you want more detailed information in the flame graph, let me know which packages you'd like me to instrument and filter for (since I can't instrument the entire Dotty).

[nicolasstucki]

Is it possible to get a wider flame graph or an interactive one? It is impossible to see the method names.

Can you create a small self-contained example that shows the same issue?

[nicolasstucki]

It would also be quite useful to see the timings of the code within unpickleTerm https://github.com/lampepfl/dotty/blob/9632f6b68ea726cf2b219f55995e0bd80abd66c1/compiler/src/dotty/tools/dotc/quoted/PickledQuotes.scala#L56-L63

We should know what are the proportions of unpickle, spliceTypes and spliceTerms are for this method.

[deusaquilus]

I'll try to get some more detailed information about what happening is inside of unpickleExpr/unpickleType.
It's hard to make a self-contained example because this flame-graph is basically the entire ProtoQuill quotation subsystem. I could publish the ProtoQuill library to maven-central and then a small code-sample that uses it if that helps.

[deusaquilus]

Also, do you have JProfiler? If you do, I can send you the saved JProfile session that you can interactively explore.
If not, you can get a trial version. They also offer academic licenses.

[deusaquilus]

Here is a bit more info. I have colorized unpickle (pink), unpickleTerm (red), and spliceTerms (purple). I didn't find spliceTypes anywhere in the chart. It seems QuoteUnpickler.unpickle is the one that takes the most time.
Color legend is below.

I can send you the profiling session separately if you'd like.

[deusaquilus]

Here is even more info with additional colorizations. You can see DottyUnpicker.tree (light green) slow things down in many places, also RefTree.isTerm() and NamedType.denot() (both are light blue in different shades).

[deusaquilus]

Also, it is worth noting that TreeAccumulator.foldOverTree also appears in many places (in yellow). It might also be useful to optimize.

[deusaquilus]

Hopefully, this provides some additional useful information. Please let me know if you need something else.

[deusaquilus]

Here's one more chart with the info about "hot-spots" in the dotty.tools.dotc call stack.

Here is the call-tree of one instance where I use unpickleTerm.

Here's the flame-graph of that unpickleTerm call.

Hope this helps.

[odersky]

It seems to read a very large number of names. We should analyze which names are serialized when we pickle quotes and whether we can cut the number somehow.

[nicolasstucki]

One thing we should be able to do is cache the result of PickleQuotes.unpickle.