3 ms·

Can you share any more information about the types of code patterns that can be superoptimized or how much speedup you see on benchmarks?

fenollp 9y ago
So if you look into [1] you'll find .hrl sources (actually .erl) and their .S counterpart in the subfolders. From these you will see that functions `a` and `b` look more and more like each other. Note that `b` is the literal version of `a`. How much speedup? Potentially infinite, at the price of compile time. In real world cases though I don't have any data yet. I want to be able to compile a whole project first and that fails for one last reason, described in the README. But I have faith :) The patterns that can be superoptimized are few for now. [2] describes how some of lists.erl functions work so that they can be supercompiled. The long-term idea is to have a local(/global if secured) cache of code so that inter-module calls can be optimized if the developer wants. My real world case/motivation is supercompilation of code such as [3] (goal is to turn `a` into `b` at the assembly level). This pattern is extremely common in my employer's software [4] (which is probably the biggest open source Erlang project AFAIK). A short example of a being super compiled into b: a() -> hd([get()]). b() -> get(). [1] https://github.com/fenollp/erlscp/tree/master/test/asm_data https://github.com/fenollp/erlscp/tree/master/test/asm_data [2] https://github.com/fenollp/erlscp/blob/master/src/erlang_supercompiler.erl#L29 https://github.com/fenollp/erlscp/blob/master/src/erlang_sup... [3] https://github.com/fenollp/erlscp/blob/master/test/asm_data/unfold3.hrl https://github.com/fenollp/erlscp/blob/master/test/asm_data/... [4] https://github.com/2600hz/kazoo https://github.com/2600hz/kazoo
- smaddox 9y ago
  From what I understand (on phone so no refs right now), "super compilation" is limited to a linear speedup. However, "distillation" can achieve super-linear speedup.

fenollp 9y ago

So if you look into [1] you'll find .hrl sources (actually .erl) and their .S counterpart in the subfolders. From these you will see that functions `a` and `b` look more and more like each other. Note that `b` is the literal version of `a`. How much speedup? Potentially infinite, at the price of compile time. In real world cases though I don't have any data yet. I want to be able to compile a whole project first and that fails for one last reason, described in the README. But I have faith :) The patterns that can be superoptimized are few for now. [2] describes how some of lists.erl functions work so that they can be supercompiled. The long-term idea is to have a local(/global if secured) cache of code so that inter-module calls can be optimized if the developer wants. My real world case/motivation is supercompilation of code such as [3] (goal is to turn `a` into `b` at the assembly level). This pattern is extremely common in my employer's software [4] (which is probably the biggest open source Erlang project AFAIK). A short example of a being super compiled into b: a() -> hd([get()]). b() -> get(). [1] https://github.com/fenollp/erlscp/tree/master/test/asm_data https://github.com/fenollp/erlscp/tree/master/test/asm_data [2] https://github.com/fenollp/erlscp/blob/master/src/erlang_supercompiler.erl#L29 https://github.com/fenollp/erlscp/blob/master/src/erlang_sup... [3] https://github.com/fenollp/erlscp/blob/master/test/asm_data/unfold3.hrl https://github.com/fenollp/erlscp/blob/master/test/asm_data/... [4] https://github.com/2600hz/kazoo https://github.com/2600hz/kazoo

smaddox 9y ago
From what I understand (on phone so no refs right now), "super compilation" is limited to a linear speedup. However, "distillation" can achieve super-linear speedup.