Use VarULE baked storage in PackedPatterns

[sffc]

Depends on #7310

This should have been done after #5230 but got dropped.

[sffc]

Build time measurements of the icu_datetime crate with default features:

Before:

$ /usr/bin/time -f "%es, %M kB" cargo build --release
   Compiling icu_datetime v2.1.1
    Finished `release` profile [optimized] target(s) in 4.11s
4.17s, 632548 kB

$ /usr/bin/time -f "%e sec, %M kB" cargo build --release
   Compiling icu_datetime v2.1.1
    Finished `release` profile [optimized] target(s) in 4.01s
4.07 sec, 634132 kB

After:

$ /usr/bin/time -f "%e sec, %M kB" cargo build --release
   Compiling icu_datetime v2.1.1
    Finished `release` profile [optimized] target(s) in 3.27s
3.33 sec, 586932 kB

$ /usr/bin/time -f "%e sec, %M kB" cargo build --release
   Compiling icu_datetime v2.1.1
    Finished `release` profile [optimized] target(s) in 3.22s
3.28 sec, 587620 kB

[sffc]

Benchmark measurements:

Benchmarks that say "format only" are unchanged. Ones that say "construct and format" have a regression of somewhere between 1% and 9% depending on how my computer is feeling; the consensus seems to be somewhere around 5-6%. To reproduce, run: cargo bench --all-features construct_and_format

[sffc]

If I mark the ZeroFrom impl as inline, the results are still noisy but seem to go down to about 3-4% regression.

[sffc]

Finally, code size.

Before:

Details

-rwxr-xr-x 1 sffc primarygroup   52750 Dec 11 17:57 and_list.wasm
-rwxr-xr-x 1 sffc primarygroup   41593 Dec 11 17:57 casemapping.wasm
-rwxr-xr-x 1 sffc primarygroup  276517 Dec 11 17:57 chrome_transliterators.wasm
-rwxr-xr-x 1 sffc primarygroup 3409535 Dec 11 17:57 chrono_jiff.wasm
-rwxr-xr-x 1 sffc primarygroup   36041 Dec 11 17:57 code_line_diff.wasm
-rwxr-xr-x 1 sffc primarygroup   47052 Dec 11 17:57 date_try_from_fields.wasm
-rwxr-xr-x 1 sffc primarygroup    9306 Dec 11 17:57 derives.wasm
-rwxr-xr-x 1 sffc primarygroup   22222 Dec 11 17:57 elevator_floors.wasm
-rwxr-xr-x 1 sffc primarygroup     415 Dec 11 17:57 experimental_segmenter.wasm
-rwxr-xr-x 1 sffc primarygroup   14863 Dec 11 17:57 filter_langids.wasm
-rwxr-xr-x 1 sffc primarygroup   12318 Dec 11 17:57 first_weekday_for_region.wasm
-rwxr-xr-x 1 sffc primarygroup    1492 Dec 11 17:57 iso_date_manipulations.wasm
-rwxr-xr-x 1 sffc primarygroup   10836 Dec 11 17:57 language_names_hash_map.wasm
-rwxr-xr-x 1 sffc primarygroup   16187 Dec 11 17:57 language_names_lite_map.wasm
-rwxr-xr-x 1 sffc primarygroup   28387 Dec 11 17:57 litemap_bincode.wasm
-rwxr-xr-x 1 sffc primarygroup   16921 Dec 11 17:57 litemap_postcard.wasm
-rwxr-xr-x 1 sffc primarygroup  104287 Dec 11 17:57 make_var.wasm
-rwxr-xr-x 1 sffc primarygroup    8052 Dec 11 17:57 make.wasm
-rwxr-xr-x 1 sffc primarygroup    8994 Dec 11 17:57 permyriad.wasm
-rwxr-xr-x 1 sffc primarygroup   31682 Dec 11 17:57 syntatically_canonicalize_locales.wasm
-rwxr-xr-x 1 sffc primarygroup 2712024 Dec 11 17:57 timezone_picker.wasm
-rwxr-xr-x 1 sffc primarygroup  341108 Dec 11 17:57 tui.wasm
-rwxr-xr-x 1 sffc primarygroup    9156 Dec 11 17:57 unicode_bmp_blocks_selector.wasm
-rwxr-xr-x 1 sffc primarygroup   23988 Dec 11 17:57 unread_emails.wasm
-rwxr-xr-x 1 sffc primarygroup  306727 Dec 11 17:57 work_log.wasm
-rwxr-xr-x 1 sffc primarygroup   19494 Dec 11 17:57 writeable_message.wasm
-rwxr-xr-x 1 sffc primarygroup    3420 Dec 11 17:57 yoke_derive.wasm
-rwxr-xr-x 1 sffc primarygroup    3415 Dec 11 17:57 zf_derive.wasm
-rwxr-xr-x 1 sffc primarygroup     559 Dec 11 17:57 zv_serde.wasm

After:

Details

-rwxr-xr-x 1 sffc primarygroup   52750 Dec 11 17:58 and_list.wasm
-rwxr-xr-x 1 sffc primarygroup   41593 Dec 11 17:58 casemapping.wasm
-rwxr-xr-x 1 sffc primarygroup  276873 Dec 11 17:58 chrome_transliterators.wasm
-rwxr-xr-x 1 sffc primarygroup 3737310 Dec 11 17:58 chrono_jiff.wasm
-rwxr-xr-x 1 sffc primarygroup   36041 Dec 11 17:58 code_line_diff.wasm
-rwxr-xr-x 1 sffc primarygroup   47052 Dec 11 17:58 date_try_from_fields.wasm
-rwxr-xr-x 1 sffc primarygroup    9306 Dec 11 17:58 derives.wasm
-rwxr-xr-x 1 sffc primarygroup   22222 Dec 11 17:58 elevator_floors.wasm
-rwxr-xr-x 1 sffc primarygroup     415 Dec 11 17:58 experimental_segmenter.wasm
-rwxr-xr-x 1 sffc primarygroup   14863 Dec 11 17:58 filter_langids.wasm
-rwxr-xr-x 1 sffc primarygroup   12318 Dec 11 17:58 first_weekday_for_region.wasm
-rwxr-xr-x 1 sffc primarygroup    1492 Dec 11 17:58 iso_date_manipulations.wasm
-rwxr-xr-x 1 sffc primarygroup   10836 Dec 11 17:58 language_names_hash_map.wasm
-rwxr-xr-x 1 sffc primarygroup   16187 Dec 11 17:58 language_names_lite_map.wasm
-rwxr-xr-x 1 sffc primarygroup   28387 Dec 11 17:58 litemap_bincode.wasm
-rwxr-xr-x 1 sffc primarygroup   16921 Dec 11 17:58 litemap_postcard.wasm
-rwxr-xr-x 1 sffc primarygroup  104287 Dec 11 17:58 make_var.wasm
-rwxr-xr-x 1 sffc primarygroup    8052 Dec 11 17:58 make.wasm
-rwxr-xr-x 1 sffc primarygroup    8994 Dec 11 17:58 permyriad.wasm
-rwxr-xr-x 1 sffc primarygroup   31682 Dec 11 17:58 syntatically_canonicalize_locales.wasm
-rwxr-xr-x 1 sffc primarygroup 2712024 Dec 11 17:58 timezone_picker.wasm
-rwxr-xr-x 1 sffc primarygroup  335574 Dec 11 17:58 tui.wasm
-rwxr-xr-x 1 sffc primarygroup    9156 Dec 11 17:58 unicode_bmp_blocks_selector.wasm
-rwxr-xr-x 1 sffc primarygroup   23988 Dec 11 17:58 unread_emails.wasm
-rwxr-xr-x 1 sffc primarygroup  301186 Dec 11 17:58 work_log.wasm
-rwxr-xr-x 1 sffc primarygroup   19494 Dec 11 17:58 writeable_message.wasm
-rwxr-xr-x 1 sffc primarygroup    3420 Dec 11 17:58 yoke_derive.wasm
-rwxr-xr-x 1 sffc primarygroup    3415 Dec 11 17:58 zf_derive.wasm
-rwxr-xr-x 1 sffc primarygroup     559 Dec 11 17:58 zv_serde.wasm

Differences:

Example	Before	After
chrome_transliterators.wasm	276517	276873
chrono_jiff.wasm	3409535	3737310
tui.wasm	341108	335574
work_log.wasm	306727	301186

tui and work_log get smaller, but chrono_jiff, which is already big, gets bigger. 🤔

[sffc]

Things to highlight:

1. The build time reduction is fairly substantial (about 20%), but only on the icu_datetime crate
2. There might be a small (<5%) performance regression on some microbenchmarks
3. Code size measurements give mixed signals
4. The trait unfortunately wasn't implemented with EncodeAsVarULE in mind, so this is a breaking change in icu_provider in a trait that isn't used by anything except hello_world (which lives in that crate) and is only implemented by macros that are versioned along with the crate. Do we care? Please speak up.

[dpulls]

🎉 All dependencies have been resolved !

[sffc]

Please review the whole PR; the commits got messy due to merges, and there isn't that much code being changed.

[sffc]

@Manishearth Do you think the metrics posted above justify the change?

[Manishearth]

Yes! I think the performance slowdown is basically noise and the build time improvement is important for one of our larger baked data crates.

Edit: hadn't really realized the size of the codesize loss in the chrono test. ignore!

[robertbastian]

The numbers don't look very conclusive to me, I don't think build time is that important given that runtime performance will suffer.

[robertbastian]

should this change?

[Manishearth]

Yeah, we looked closer at this and determined we shouldn't land this, but that this was a good proof of concept. It was important to learn that depending on the data architecture, data deduplication may suffer with this type of optimization (which explains the large chrono_jiff codesize)