Enzyme v0.11.11+ much slower to load

[staticfloat]

It appears that Enzyme v0.11.11 introduced a pretty serious load time regression in Enzyme.Compiler.__init__():

(tmp.QXqGKVm6jS) pkg> add [email protected]
   Resolving package versions...             
    Updating `/private/var/folders/b1/lrtym5_j0zl_8r96ydwfrgy40000gp/T/tmp.QXqGKVm6jS/Project.toml`
⌃ [7da242da] + Enzyme v0.11.10
julia> @time_imports using Enzyme
      1.6 ms  EnzymeCore
      1.5 ms  CEnum
      9.5 ms  Preferences
      0.5 ms  LazyArtifacts
      0.5 ms  JLLWrappers
               ┌ 3.2 ms LLVMExtra_jll.__init__() 70.46% compilation time
     19.1 ms  LLVMExtra_jll 88.50% compilation time
               ┌ 0.1 ms LLVM.__init__() 
     25.2 ms  LLVM
               ┌ 0.7 ms Enzyme_jll.__init__() 
      1.7 ms  Enzyme_jll
      0.4 ms  ExprTools
               ┌ 0.1 ms TimerOutputs.__init__() 
     11.2 ms  TimerOutputs
      0.4 ms  Scratch
               ┌ 1.7 ms GPUCompiler.__init__() 
     70.0 ms  GPUCompiler 0.04% compilation time
      0.4 ms  Reexport
      0.4 ms  StructIO
               ┌ 0.0 ms ObjectFile.__init__() 
      6.0 ms  ObjectFile
               ┌ 1.1 ms Enzyme.API.__init__() 
               ├ 0.2 ms Enzyme.Compiler.JIT.__init__() 
               ├ 0.0 ms Enzyme.Compiler.FFI.BLASSupport.__init__() 
               ├ 1.8 ms Enzyme.Compiler.FFI.__init__() 
               ├ 0.1 ms Enzyme.Compiler.__init__() 
    110.6 ms  Enzyme 0.03% compilation time

Versus:

(tmp.QXqGKVm6jS) pkg> add [email protected]
   Resolving package versions...             
    Updating `/private/var/folders/b1/lrtym5_j0zl_8r96ydwfrgy40000gp/T/tmp.QXqGKVm6jS/Project.toml`
⌃ [7da242da] + Enzyme v0.11.11
julia> @time_imports using Enzyme
      1.6 ms  EnzymeCore
      1.5 ms  CEnum
     40.9 ms  Preferences
      0.4 ms  LazyArtifacts
      0.5 ms  JLLWrappers
               ┌ 2.6 ms LLVMExtra_jll.__init__() 68.41% compilation time
     15.2 ms  LLVMExtra_jll 89.35% compilation time
               ┌ 0.1 ms LLVM.__init__() 
     21.0 ms  LLVM
               ┌ 0.8 ms Enzyme_jll.__init__() 
      1.6 ms  Enzyme_jll
      0.4 ms  ExprTools
               ┌ 0.0 ms TimerOutputs.__init__() 
     10.8 ms  TimerOutputs
      0.4 ms  Scratch
               ┌ 1.8 ms GPUCompiler.__init__() 
     65.0 ms  GPUCompiler 0.05% compilation time
      0.4 ms  Reexport
      0.3 ms  StructIO
               ┌ 0.0 ms ObjectFile.__init__() 
      5.4 ms  ObjectFile
               ┌ 6.4 ms Enzyme.API.__init__() 
               ├ 0.2 ms Enzyme.Compiler.JIT.__init__() 
               ├ 0.0 ms Enzyme.Compiler.FFI.BLASSupport.__init__() 
               ├ 1.9 ms Enzyme.Compiler.FFI.__init__() 
               ├ 3086.9 ms Enzyme.Compiler.__init__() 99.99% compilation time
   3202.9 ms  Enzyme 96.37% compilation time

[staticfloat]

This also apparently can be worsened if you've loaded many other packages, adding just SciMLSensitivity to the environment and then timing that load shows the time spent growing to absurd numbers:

...
10.8 ms  ObjectFile
               ┌ 5.9 ms Enzyme.API.__init__() 
               ├ 0.2 ms Enzyme.Compiler.JIT.__init__() 
               ├ 0.0 ms Enzyme.Compiler.FFI.BLASSupport.__init__() 
               ├ 2.1 ms Enzyme.Compiler.FFI.__init__() 
               ├ 10521.1 ms Enzyme.Compiler.__init__() 100.00% compilation time
  10673.4 ms  Enzyme 98.57% compilation time
...

This is just with a fresh environment on Julia v1.10:

(tmp.QXqGKVm6jS) pkg> st
Status `/private/var/folders/b1/lrtym5_j0zl_8r96ydwfrgy40000gp/T/tmp.QXqGKVm6jS/Project.toml`
⌃ [7da242da] Enzyme v0.11.11
  [1ed8b502] SciMLSensitivity v7.52.0
Info Packages marked with ⌃ have new versions available and may be upgradable.

Using Enzyme v0.11.10 eliminates this time, and lowers the overall load time of SciMLSensitivity from 16.7s -> 6.1s.

[wsmoses]

I think duplicate of #776, so let's move this info/conversation to that issue.

But regardless my tldr understanding is that the Julia compiler itself needs improvements here