ekuber, 22 days ago

@jhpratt looking forward to the repro and either finding out that it is a 2 line fix or requires a complete reimplementation of 2 stages of the compiler.

predrag, 22 days ago

@ekuber @jhpratt ah yes, the two bug flavors...

Hoping it's the first one! 🤞

jhpratt, 22 days ago

@predrag @ekuber Me too! I'm going to use the self-profile feature to try to determine where time is being spent. Creating a minimal reproduction may be tricky, as the underlying library is relatively intertwined.

jhpratt, 22 days ago

@predrag @ekuber Oh, I forgot to mention. This happens with cargo check, so it is unquestionably rustc and not LLVM.

jhpratt, 21 days ago

@predrag @ekuber

Discarding the smallest checks as insignificant, it's certainly exponential. Best fit is an exponent of 1.27749 × the number of methods chained; correlation is 0.992.

A chain of 11 methods took 7m51s to type check, and I didn't dare to try 12 (which was my original code).

typeck: 28.941%
normalize_canonicalized_projection_ty 26.938%
type_op_prove_predicate 25.097%
codegen_select_candidate 12.080%
LLVM_passes 5.534%

Any tips on creating a minimal example for an issue?

ekuber, 21 days ago

@jhpratt @predrag replacing bodies with loop {} is the first step I do, I'm guessing this is on some where bound or impl Trait return?

jhpratt, 21 days ago

@ekuber @predrag Very much so — every part of the API is based around trait impls. Stubbing bodies is a good starting place.

ekuber, 21 days ago

@jhpratt @predrag I'm sure @errs will be looking forward to the report 😀