Explain configurations that get faster with types

[bennn]

Such configurations exist in

• [x] futen
• [x] http2
• [x] call*
• [x] spectralnorm
• [x] Espionage (maybe nothing)

[bennn]

also explain slowSHA's untyped config.

EDIT: the histogram for slowSHA has a typo, it had listed 2 configurations with 0 types. I just went through the other data & checked their "bucketing" was all right. It was. Just slowSHA was the problem.

[bennn]

Spent a little time on this. Probably too much, but still I don't have great answers.

• futen : maybe due to tuple types ... adding a tuple annotation speeds up the config

• call_method : definitely because of replacing Dyn with an object type

• Espionage : adds a String type, which replaces one check with a check_type_string. The old check (slow check?) opens a try/except block and checks whether the given value has the split attribute.

• http2 : not sure ... maybe the same issue as Espionage where adding a type annotation leads to a cheaper check. When I looked at this earlier today, it seemed like the "faster" config would have MORE runtime checks and retic -p confirmed, but type-checking the whole program is different from retic -p on individual files.

• • -

Here's my reference for "what configs got faster with types"? The file is a:

• list of
  • pairs, where the first element is a benchmark name
  • and the second element is a list of pairs of strings
    • the 1st element of each pair is a configuration
    • the 2nd element is a configuration with more type annotations that runs faster than the first

faster-with-types.txt

(I mean it's a simple data format just long to explain)