Add C++20 likely/unlikely hints on quick-tricks hot paths

[tameware]

Summary

• Mark rare void / no-opponents suit cases in QuickTricks as [[unlikely]] and the common else arms as [[likely]].
• No behavioral change; attributes only guide code layout and branch prediction.
• Roughly sub-1% but consistent performance improvement on batches; instrumentation confirms the quick-tricks hints match observed hit rates.

Test plan

• bazel test //library/tests:unit_tests //library/tests/system:context_equivalence_test
• CI green on this PR
• Optional: compare benchmark.sh before/after on a small hand list to spot-check timing

[Copilot]

Pull request overview

This PR adds C++20 [[likely]] / [[unlikely]] branch prediction hints to hot-path conditionals in the alpha-beta search and quick-tricks logic, aiming to improve generated code layout and branch prediction without changing behavior.

Changes:

• Annotates early cutoff / leaf checks in ab_search.cpp with [[likely]] / [[unlikely]].
• Marks rare suit/void/trump-edge cases in QuickTricks as [[unlikely]] and common paths as [[likely]].

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
library/src/ab_search.cpp	Adds [[likely]]/[[unlikely]] hints to cutoff and QT/LT result branches in AB search.
library/src/quick_tricks.cpp	Adds [[likely]]/[[unlikely]] hints around rare void/partner-only/opponent-less suit cases in quick-tricks.

[tameware]

Zero difference over a large sample when inverted. I'll also benchmark removing from this if entirely - presumably I'll find the same.

[zzcgumn]

How was the data collected to determine how likely a path is?

[wopdevries]

The likelihood assumptions follow from alpha-beta search theory: most nodes hit an early cutoff (target reached or tricks exhausted) before reaching move generation. The "fall-through to move search" path is therefore the rare case, making cutoff branches [[unlikely]]. This is standard practice in chess/bridge engine optimization. The benchmark results across 11 runs on different hand set sizes support the annotation being correct in practice.

[zzcgumn]

Happy to merge this as we do see a performance improvement.

[zzcgumn]

@wopdevries, call my old-fashioned but I prefer data supporting claims like "most nodes hit an early cutoff" over this is expected based on what usually happens in alpha-beta searches with pruning. We should merge this as it does improve the performance, but I am still wondering if we can do even better by collecting search statistics.

[tameware]

Per my comments on the related Issue, not all of these help. I think I’ve isolated the ones that do. I’ll confirm via further benchmarking before I merge them.

…

On Tue, Jul 21, 2026 at 6:25 PM Martin Nygren ***@***.***> wrote: *zzcgumn* left a comment (dds-bridge/dds#252) <#252 (comment)> @wopdevries <https://github.com/wopdevries>, call my old-fashioned but I prefer data supporting claims like "most nodes hit an early cutoff" over this is expected based on what usually happens in alpha-beta searches with pruning. We should merge this as it does improve the performance, but I am still wondering if we can do even better by collecting search statistics. — Reply to this email directly, view it on GitHub <#252?email_source=notifications&email_token=ABC4PYCYKUG6MJLTPRJ3ZID5F6KOXA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBTGY2DKMBZGE32M4TFMFZW63VGMFZXG2LHN2SWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5036450917>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABC4PYG4HWVFVJUZCMKYJNL5F6KOXAVCNFSNUABEKJSXA33TNF2G64TZHMZDMOJYHAZDANR3JFZXG5LFHM2DSMRVG42TQOJUG2QXMAQ> . You are receiving this because you were assigned.Message ID: ***@***.***>

[tameware]

Performance improvement is below 1%, but consistent.

Instrumentation shows that the hints in quick_tricks.cpp are correct:

$ ../dtest_instrumented  --file hands/list100.txt -s calc
DDS DLL
-------
System                      Apple
Word size                 64 bits
Compiler                    clang
Constructor                      
Version                     3.0.0
Memory max (MB)             25463
Threads (MB)       20-30 / 95-160
Number of cores                18
Number of threads               1
Thread sizes             0 S, 1 L
Threading                 None(*)

dtest worker threads: auto
test program
-------------
System                      Apple
Word size                 64 bits
Compiler                    clang

Hand no.                     Time
      40 (  40.0%)           1365
      80 (  80.0%)           1907
     100 ( 100.0%)            854

Timer name             Hand stats
Number of hands               100
User time (ms)               4126
Avg user time (ms)          41.26
Sys time (ms)               66233
Avg sys time (ms)          662.33
Ratio                       16.05
QuickTricks branch hint stats
branch                      hint        hits          pair%     note
own_only                    unlikely    11139887        9.3%     
not_own_only                likely      108206095      90.7%     
own_empty                   unlikely    7148955       -         
own_long_no_otrump          unlikely    314630         14.1%     
own_long_has_otrump         likely      1920904        85.9%     
own_long_nt_or_trump        likely      1755398       -         
part_trump_no_opps          unlikely    1632538       -         
part_side_no_opps           unlikely    6537206       -         
comm_part_only              unlikely    922663          2.3%     
comm_not_part_only          likely      38735095       97.7%     
comm_part_no_otrump         unlikely    67468          13.3%     
comm_part_has_otrump        likely      440590         86.7%     
comm_part_nt_or_trump       likely      414605        -         
comm_part_trump_no_opps     unlikely    366186        -         
comm_part_side_no_opps      unlikely    1081507       -         
winner_rank_zero            unlikely    0             -        

[wopdevries]

The instrumentation data beautifully answers @zzcgumn's question. All [[likely]]/[[unlikely]] annotations in quick_tricks.cpp are confirmed correct by actual hit counts. The dominant paths (not_own_only 90.7%, comm_not_part_only 97.7%) match the hints exactly.

I also measured on Linux (Intel i7-4770, Z87) with perf stat, 1000 hands single-threaded:

Branch miss rate: 4.32% (1.6B misses / 37.2B branches)

Ready to merge — the data supports it.

[tameware]

Tomorrow I'll revert ab_search.cpp to remove it from the PR and submit with just the quick_tricks.cpp hints.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 2 changed files in this pull request and generated no new comments.