Improve compilation speed

[mneumann]

For an almost empty project, it takes about 32 seconds to compile.

> savi build --print-perf spec
     9 ms : manifests
   764 ms : load
     1 ms : populate_types
    14 ms : namespace
   336 ms : reparse
    15 ms : populate
    44 ms : macros
    35 ms : sugar
  1433 ms : refer_type
    21 ms : lambda
   158 ms : flow
   205 ms : classify
   116 ms : type_context
   236 ms : refer
   195 ms : jumps
   101 ms : inventory
   327 ms : pre_subtyping
  1989 ms : local
   515 ms : pre_infer
    84 ms : completeness
  2576 ms : infer_edge
  3387 ms : infer
  1637 ms : reach
   134 ms : verify
  1158 ms : privacy
   110 ms : paint
  1812 ms : type_check
  6306 ms : codegen
  4719 ms : binary

[jemc]

The biggest culprit is currently LLVM - I think we need to set up a two-stage build in the https://github.com/savi-lang/llvm-static repo to better optimize the LLVM build.

[mneumann]

Most likely I am using a debug build! Let me try again with gmake ci config=release!

[mneumann]

Or to have a savi check command, that just does type checking without code generation. Most of the time that'd help.

[jemc]

I'm planning to add a watch modes where the compiler can make use of its in-memory caching to speed up recompilation of the same program, whenever files change. I suspect this will help with the speed of the basic development workflow.

I'm definitely not opposed to having a mode that only runs compilation checks, though I personally wouldn't use it very often, because in my development workflow, whenever I get the program to compile i always want to run the tests next.

Let me try again with gmake ci config=release

That won't build the release-mode compiler - you want to run gmake build/savi-release.

[mneumann]

Just for reference, with savi-release it's 3x faster (10 seconds total instead of 32):

> time savi build --print-perf spec
     1 ms : manifests
   181 ms : load
     0 ms : populate_types
     1 ms : namespace
    20 ms : reparse
     2 ms : populate
    14 ms : macros
     7 ms : sugar
    79 ms : refer_type
     6 ms : lambda
    35 ms : flow
    35 ms : classify
    26 ms : type_context
    45 ms : refer
    22 ms : jumps
    12 ms : inventory
    25 ms : pre_subtyping
   337 ms : local
    92 ms : pre_infer
     7 ms : completeness
   451 ms : infer_edge
   619 ms : infer
   201 ms : reach
    13 ms : verify
   148 ms : privacy
    11 ms : paint
   361 ms : type_check
   806 ms : codegen
  4922 ms : binary
11.711u 0.346s 0:09.81 122.8%	93895+1492k 0+20io 38pf+0w

And 50% of that time is spent in LLVM.