meta: compiler performance

[charles-cooper]

this is a tracking issue to keep track of compiler performance.

the following numbers are for a fairly large, single file vyper contract (2k loc, near 24kb limit).

some hotspots:

• vyper initialization. costs 200ms just to run vyper --help

  • some of this probably comes from dataclass initialization
  • e.g. calling dataclass() on a simple class costs 180us

• parse_to_ast has hotspots in tokenize() and ASTTokens.mark_tokens() - total 300ms

• semantic analysis is relatively cheap, 240ms

• translation to IRnodes, 600ms

• biggest hotspot is in legacy IR optimizer, due to all the calls to IRnode.from_list() - 1.7s (1700ms)

• venom (optional) - 1.1s (1100ms), although this changes fairly rapidly depending what we are doing with the passes

while the frontend looks cheaper, note that it scales with the size of all imports (source code), while code generation scales with contract size.

[harkal]

As we move away from the legacy IR optimizer, we are going to get a lot of extra margin for venom, which I am confident we will be able to make better use of. I feel confident about this, as for venom I never optimized for speed up until this point. My only goal is the efficiency of the generated code.

p.s. so no rewrites in zig just yet ::D